Add [1.8.1] changelog compare link and update [Unreleased] base

Remove incorrect agent-framework-foundry changelog entry for #6259
Python: bump ag-ui and azurefunctions for 1.8.1 release
2026-06-16 21:04:09 +08:00 · 2026-06-09 20:05:32 +00:00 · 2026-06-09 18:37:20 +00:00 · 2026-06-09 18:14:41 +00:00 · 2026-06-09 18:12:45 +00:00 · 2026-06-09 18:02:27 +00:00
159 changed files with 13911 additions and 965 deletions
@@ -0,0 +1,55 @@
+<svg width="256" height="256" viewBox="0 0 256 256" fill="none" xmlns="http://www.w3.org/2000/svg">
+<rect width="256" height="256" transform="matrix(-1 0 0 1 256 0)" fill="white"/>
+<path d="M163.006 36.0262C155.129 31.4785 145.424 31.4759 137.533 36.0284L104.002 55.3877C107.838 53.2763 112.07 52.2274 116.296 52.2248C120.633 52.2231 124.976 53.3247 128.871 55.5411L174.091 81.6479C175.11 82.2362 175.738 83.3236 175.738 84.5004V151.263C175.738 152.716 177.313 153.621 178.568 152.895L196.967 142.28C204.845 137.732 209.704 129.326 209.704 120.221V77.7299C209.704 68.6342 204.855 60.227 196.967 55.6697L190.983 52.2141C190.65 52.0008 190.313 51.7921 189.969 51.5933L163.006 36.0262Z" fill="url(#paint0_linear_481_4810)"/>
+<path d="M163.006 36.0262C155.129 31.4785 145.424 31.4759 137.533 36.0284L104.002 55.3877C107.838 53.2763 112.07 52.2274 116.296 52.2248C120.633 52.2231 124.976 53.3247 128.871 55.5411L174.091 81.6479C175.11 82.2362 175.738 83.3236 175.738 84.5004V151.263C175.738 152.716 177.313 153.621 178.568 152.895L196.967 142.28C204.845 137.732 209.704 129.326 209.704 120.221V77.7299C209.704 68.6342 204.855 60.227 196.967 55.6697L190.983 52.2141C190.65 52.0008 190.313 51.7921 189.969 51.5933L163.006 36.0262Z" fill="url(#paint1_linear_481_4810)"/>
+<path d="M103.548 55.6397L103.557 55.6451L104.002 55.3877C103.851 55.471 103.698 55.5531 103.548 55.6397Z" fill="url(#paint2_linear_481_4810)"/>
+<path d="M103.548 55.6397L103.557 55.6451L104.002 55.3877C103.851 55.471 103.698 55.5531 103.548 55.6397Z" fill="url(#paint3_linear_481_4810)"/>
+<path d="M116.308 52.2209C111.903 52.2271 107.507 53.3498 103.561 55.6366C95.6702 60.1891 90.8239 68.6019 90.8231 77.6986L90.8223 167.846C90.8222 169.786 92.871 171.041 94.599 170.16L103.523 165.611C116.573 158.959 124.788 145.549 124.788 130.902V62.9894C124.778 58.6476 129.49 55.9209 133.25 58.0698L128.879 55.5453C124.976 53.3242 120.645 52.2192 116.308 52.2209Z" fill="url(#paint4_linear_481_4810)"/>
+<path d="M95.3068 221.682C103.184 226.23 112.889 226.232 120.78 221.68L154.311 202.32C150.475 204.432 146.243 205.481 142.018 205.483C137.68 205.485 133.337 204.383 129.442 202.167L84.2226 176.06C83.2035 175.472 82.5757 174.384 82.5757 173.208L82.5757 106.445C82.5757 104.992 81 104.087 79.7451 104.813L61.3465 115.428C53.4682 119.976 48.6091 128.382 48.6089 137.487L48.6089 179.978C48.6089 189.074 53.4585 197.481 61.3465 202.038L67.3303 205.494C67.6629 205.707 68.0002 205.916 68.3446 206.115L95.3068 221.682Z" fill="url(#paint5_linear_481_4810)"/>
+<path d="M95.3068 221.682C103.184 226.23 112.889 226.232 120.78 221.68L154.311 202.32C150.475 204.432 146.243 205.481 142.018 205.483C137.68 205.485 133.337 204.383 129.442 202.167L84.2226 176.06C83.2035 175.472 82.5757 174.384 82.5757 173.208L82.5757 106.445C82.5757 104.992 81 104.087 79.7451 104.813L61.3465 115.428C53.4682 119.976 48.6091 128.382 48.6089 137.487L48.6089 179.978C48.6089 189.074 53.4585 197.481 61.3465 202.038L67.3303 205.494C67.6629 205.707 68.0002 205.916 68.3446 206.115L95.3068 221.682Z" fill="url(#paint6_linear_481_4810)"/>
+<path d="M154.765 202.068L154.756 202.063L154.311 202.32C154.463 202.237 154.615 202.155 154.765 202.068Z" fill="url(#paint7_linear_481_4810)"/>
+<path d="M154.765 202.068L154.756 202.063L154.311 202.32C154.463 202.237 154.615 202.155 154.765 202.068Z" fill="url(#paint8_linear_481_4810)"/>
+<path d="M142.003 205.487C146.408 205.481 150.805 204.358 154.751 202.071C162.641 197.519 167.488 189.106 167.488 180.009L167.489 89.8618C167.489 87.9222 165.44 86.667 163.712 87.5479L154.788 92.0972C141.739 98.7494 133.523 112.159 133.523 126.806L133.523 194.719C133.533 199.06 128.821 201.787 125.061 199.638L129.432 202.163C133.336 204.384 137.666 205.489 142.003 205.487Z" fill="url(#paint9_linear_481_4810)"/>
+<defs>
+<linearGradient id="paint0_linear_481_4810" x1="207.413" y1="61.882" x2="148.999" y2="163.067" gradientUnits="userSpaceOnUse">
+<stop stop-color="#9189F7"/>
+<stop offset="1" stop-color="#4135E9"/>
+</linearGradient>
+<linearGradient id="paint1_linear_481_4810" x1="188.235" y1="204.189" x2="264.21" y2="152.592" gradientUnits="userSpaceOnUse">
+<stop stop-color="#4F42FD"/>
+<stop offset="1" stop-color="#7274FF"/>
+</linearGradient>
+<linearGradient id="paint2_linear_481_4810" x1="207.413" y1="61.882" x2="148.999" y2="163.067" gradientUnits="userSpaceOnUse">
+<stop stop-color="#9189F7"/>
+<stop offset="1" stop-color="#4135E9"/>
+</linearGradient>
+<linearGradient id="paint3_linear_481_4810" x1="188.235" y1="204.189" x2="264.21" y2="152.592" gradientUnits="userSpaceOnUse">
+<stop stop-color="#4F42FD"/>
+<stop offset="1" stop-color="#7274FF"/>
+</linearGradient>
+<linearGradient id="paint4_linear_481_4810" x1="93.1761" y1="128.826" x2="66.2399" y2="104.746" gradientUnits="userSpaceOnUse">
+<stop offset="0.25" stop-color="#4F42FD"/>
+<stop offset="1" stop-color="#2C08AC"/>
+</linearGradient>
+<linearGradient id="paint5_linear_481_4810" x1="50.9004" y1="195.826" x2="109.315" y2="94.6412" gradientUnits="userSpaceOnUse">
+<stop stop-color="#9189F7"/>
+<stop offset="1" stop-color="#4135E9"/>
+</linearGradient>
+<linearGradient id="paint6_linear_481_4810" x1="70.0787" y1="53.5188" x2="-5.89657" y2="105.116" gradientUnits="userSpaceOnUse">
+<stop stop-color="#4F42FD"/>
+<stop offset="1" stop-color="#7274FF"/>
+</linearGradient>
+<linearGradient id="paint7_linear_481_4810" x1="50.9004" y1="195.826" x2="109.315" y2="94.6412" gradientUnits="userSpaceOnUse">
+<stop stop-color="#9189F7"/>
+<stop offset="1" stop-color="#4135E9"/>
+</linearGradient>
+<linearGradient id="paint8_linear_481_4810" x1="70.0787" y1="53.5188" x2="-5.89657" y2="105.116" gradientUnits="userSpaceOnUse">
+<stop stop-color="#4F42FD"/>
+<stop offset="1" stop-color="#7274FF"/>
+</linearGradient>
+<linearGradient id="paint9_linear_481_4810" x1="165.135" y1="128.882" x2="192.072" y2="152.962" gradientUnits="userSpaceOnUse">
+<stop offset="0.25" stop-color="#4F42FD"/>
+<stop offset="1" stop-color="#2C08AC"/>
+</linearGradient>
+</defs>
+</svg>
@@ -0,0 +1,5 @@
+<svg width="256" height="256" viewBox="0 0 256 256" fill="none" xmlns="http://www.w3.org/2000/svg">
+<rect width="256" height="256" transform="matrix(-1 0 0 1 256 0)" fill="white"/>
+<path d="M167.489 89.8618C167.489 87.9224 165.441 86.667 163.713 87.5473L154.788 92.0971C141.739 98.7493 133.523 112.159 133.523 126.806V194.718C133.533 199.053 128.837 201.777 125.08 199.647L84.2227 176.06C83.2036 175.472 82.5762 174.384 82.5762 173.207V106.445C82.5759 104.992 80.9999 104.087 79.7451 104.813L61.3467 115.428C53.4685 119.976 48.6096 128.382 48.6094 137.487V179.978C48.6094 189.074 53.4588 197.481 61.3467 202.039L67.3301 205.494C67.6627 205.707 68.0004 205.916 68.3447 206.115L95.3066 221.682C103.184 226.23 112.889 226.232 120.779 221.679L154.312 202.32C154.271 202.342 154.23 202.362 154.189 202.384C154.283 202.333 154.375 202.281 154.468 202.229L154.312 202.32C154.463 202.237 154.615 202.154 154.765 202.068L154.762 202.065C162.646 197.511 167.487 189.102 167.488 180.009L167.489 89.8618Z" fill="black"/>
+<path d="M163.007 36.0259C155.13 31.4781 145.424 31.4763 137.533 36.0288L104.002 55.3882C104.029 55.3732 104.057 55.359 104.084 55.3442C104.02 55.3794 103.956 55.4159 103.893 55.4516L104.002 55.3882C103.851 55.4714 103.699 55.5536 103.549 55.6401L103.552 55.6411C95.6664 60.1948 90.8241 68.6053 90.8232 77.6987L90.8223 167.846C90.8223 169.786 92.8707 171.04 94.5986 170.16L103.523 165.611C116.573 158.959 124.788 145.549 124.788 130.902V62.9897C124.778 58.6479 129.49 55.9211 133.25 58.0698L174.091 81.6479C175.11 82.2363 175.737 83.3238 175.737 84.5005V151.263C175.738 152.716 177.314 153.621 178.568 152.895L196.967 142.28C204.845 137.732 209.704 129.326 209.704 120.221V77.73C209.704 68.6343 204.855 60.2267 196.967 55.6694L190.982 52.2143C190.65 52.0011 190.313 51.792 189.969 51.5932L163.007 36.0259Z" fill="black"/>
+</svg>
@@ -0,0 +1,5 @@
+<svg width="256" height="256" viewBox="0 0 256 256" fill="none" xmlns="http://www.w3.org/2000/svg">
+<rect width="256" height="256" transform="matrix(-1 0 0 1 256 0)" fill="black"/>
+<path d="M167.489 89.8618C167.489 87.9224 165.441 86.667 163.713 87.5473L154.788 92.0971C141.739 98.7493 133.523 112.159 133.523 126.806V194.718C133.533 199.053 128.837 201.777 125.08 199.647L84.2227 176.06C83.2036 175.472 82.5762 174.384 82.5762 173.207V106.445C82.5759 104.992 80.9999 104.087 79.7451 104.813L61.3467 115.428C53.4685 119.976 48.6096 128.382 48.6094 137.487V179.978C48.6094 189.074 53.4588 197.481 61.3467 202.039L67.3301 205.494C67.6627 205.707 68.0004 205.916 68.3447 206.115L95.3066 221.682C103.184 226.23 112.889 226.232 120.779 221.679L154.312 202.32C154.271 202.342 154.23 202.362 154.189 202.384C154.283 202.333 154.375 202.281 154.468 202.229L154.312 202.32C154.463 202.237 154.615 202.154 154.765 202.068L154.762 202.065C162.646 197.511 167.487 189.102 167.488 180.009L167.489 89.8618Z" fill="white"/>
+<path d="M163.007 36.0259C155.13 31.4781 145.424 31.4763 137.533 36.0288L104.002 55.3882C104.029 55.3732 104.057 55.359 104.084 55.3442C104.02 55.3794 103.956 55.4159 103.893 55.4516L104.002 55.3882C103.851 55.4714 103.699 55.5536 103.549 55.6401L103.552 55.6411C95.6664 60.1948 90.8241 68.6053 90.8232 77.6987L90.8223 167.846C90.8223 169.786 92.8707 171.04 94.5986 170.16L103.523 165.611C116.573 158.959 124.788 145.549 124.788 130.902V62.9897C124.778 58.6479 129.49 55.9211 133.25 58.0698L174.091 81.6479C175.11 82.2363 175.737 83.3238 175.737 84.5005V151.263C175.738 152.716 177.314 153.621 178.568 152.895L196.967 142.28C204.845 137.732 209.704 129.326 209.704 120.221V77.73C209.704 68.6343 204.855 60.2267 196.967 55.6694L190.982 52.2143C190.65 52.0011 190.313 51.792 189.969 51.5932L163.007 36.0259Z" fill="white"/>
+</svg>
@@ -0,0 +1,4 @@
+<svg width="256" height="256" viewBox="0 0 256 256" fill="none" xmlns="http://www.w3.org/2000/svg">
+<rect width="256" height="256" transform="matrix(-1 0 0 1 256 0)" fill="white"/>
+<path d="M48.6094 179.978V137.487C48.6096 128.382 53.4685 119.976 61.3467 115.428L79.7451 104.813C80.9999 104.087 82.5759 104.992 82.5762 106.445V173.207C82.5762 174.384 83.2036 175.472 84.2227 176.06L125.08 199.647C128.837 201.777 133.533 199.053 133.523 194.718V126.806C133.523 112.388 141.484 99.1684 154.18 92.4135L154.788 92.0971L163.713 87.5473C165.441 86.667 167.489 87.9224 167.489 89.8618L167.488 180.009L167.474 180.86C167.181 189.624 162.399 197.653 154.762 202.065L154.765 202.068C154.615 202.154 154.463 202.237 154.312 202.32L154.468 202.229C154.375 202.281 154.283 202.333 154.189 202.384C154.23 202.362 154.271 202.342 154.312 202.32L120.779 221.679L120.034 222.092C112.534 226.093 103.539 226.092 96.0508 222.095L95.3066 221.682L68.3447 206.115C68.0004 205.916 67.6627 205.707 67.3301 205.494L61.3467 202.039C53.7053 197.624 48.9149 189.595 48.623 180.829L48.6094 179.978ZM175.737 84.5005C175.737 83.3974 175.186 82.3728 174.277 81.7641L174.091 81.6479L133.25 58.0698C129.49 55.9211 124.778 58.6479 124.788 62.9897V130.902L124.782 131.587C124.53 145.966 116.369 159.063 103.523 165.611L94.5986 170.16L94.4355 170.236C92.799 170.938 90.9459 169.803 90.8281 168.026L90.8223 167.846L90.8232 77.6987C90.8241 68.6053 95.6664 60.1948 103.552 55.6411L103.549 55.6401C103.699 55.5536 103.851 55.4714 104.002 55.3882L103.893 55.4516C103.956 55.4159 104.02 55.3794 104.084 55.3442C104.057 55.359 104.029 55.3732 104.002 55.3882L137.533 36.0288C145.424 31.4763 155.13 31.4781 163.007 36.0259L189.969 51.5932C190.313 51.792 190.65 52.0011 190.982 52.2143L196.967 55.6694C204.855 60.2267 209.704 68.6343 209.704 77.73V120.221L209.689 121.073C209.397 129.848 204.599 137.874 196.967 142.28L178.568 152.895C177.314 153.621 175.738 152.716 175.737 151.263V84.5005ZM176.814 149.866L176.819 149.864L176.832 149.856C176.826 149.859 176.82 149.862 176.814 149.866ZM137.023 194.71L137.019 195.038C136.802 201.878 129.341 206.087 123.354 202.692L123.33 202.678L82.4727 179.091C80.3698 177.877 79.0762 175.634 79.0762 173.207V109.239L63.0957 118.458C56.5124 122.259 52.3745 129.183 52.1221 136.752L52.1094 137.487V179.978C52.1094 187.823 56.2909 195.075 63.0957 199.007H63.0967L69.0801 202.462L69.1504 202.503L69.2188 202.547C69.5212 202.741 69.8111 202.92 70.0947 203.083L97.0566 218.651L97.6982 219.007C104.372 222.569 112.434 222.453 119.029 218.648L152.514 199.316L152.512 199.312C152.581 199.274 152.65 199.236 152.752 199.178L152.753 199.181C152.775 199.169 152.796 199.158 152.815 199.147L153.011 199.035C159.81 195.107 163.987 187.854 163.988 180.009V91.3344L156.378 95.2153C144.501 101.27 137.023 113.475 137.023 126.806V194.71ZM94.3223 166.372L101.934 162.493C113.811 156.438 121.288 144.233 121.288 130.902V62.9897C121.278 55.9521 128.901 51.5532 134.986 55.0307L135 55.0385L175.841 78.6167L176.036 78.7339C178.023 79.9714 179.237 82.15 179.237 84.5005V148.468L195.217 139.249C202.013 135.326 206.204 128.075 206.204 120.221V77.73C206.204 69.8848 202.022 62.6318 195.217 58.6997L189.232 55.2456L189.162 55.2046L189.094 55.1606C188.788 54.9649 188.5 54.787 188.219 54.6245L161.257 39.0571C154.463 35.135 146.091 35.1311 139.282 39.0591L139.283 39.06L105.765 58.4106L105.767 58.4135C105.713 58.443 105.723 58.4383 105.602 58.5063L105.6 58.5034C105.535 58.5391 105.481 58.5691 105.433 58.5962L105.302 58.6723C98.5016 62.5995 94.324 69.8532 94.3232 77.6987L94.3223 166.372Z" fill="black"/>
+</svg>
@@ -0,0 +1,4 @@
+<svg width="256" height="256" viewBox="0 0 256 256" fill="none" xmlns="http://www.w3.org/2000/svg">
+<rect width="256" height="256" transform="matrix(-1 0 0 1 256 0)" fill="black"/>
+<path d="M48.6094 179.978V137.487C48.6096 128.382 53.4685 119.976 61.3467 115.428L79.7451 104.813C80.9999 104.087 82.5759 104.992 82.5762 106.445V173.207C82.5762 174.384 83.2036 175.472 84.2227 176.06L125.08 199.647C128.837 201.777 133.533 199.053 133.523 194.718V126.806C133.523 112.388 141.484 99.1684 154.18 92.4135L154.788 92.0971L163.713 87.5473C165.441 86.667 167.489 87.9224 167.489 89.8618L167.488 180.009L167.474 180.86C167.181 189.624 162.399 197.653 154.762 202.065L154.765 202.068C154.615 202.154 154.463 202.237 154.312 202.32L154.468 202.229C154.375 202.281 154.283 202.333 154.189 202.384C154.23 202.362 154.271 202.342 154.312 202.32L120.779 221.679L120.034 222.092C112.534 226.093 103.539 226.092 96.0508 222.095L95.3066 221.682L68.3447 206.115C68.0004 205.916 67.6627 205.707 67.3301 205.494L61.3467 202.039C53.7053 197.624 48.9149 189.595 48.623 180.829L48.6094 179.978ZM175.737 84.5005C175.737 83.3974 175.186 82.3728 174.277 81.7641L174.091 81.6479L133.25 58.0698C129.49 55.9211 124.778 58.6479 124.788 62.9897V130.902L124.782 131.587C124.53 145.966 116.369 159.063 103.523 165.611L94.5986 170.16L94.4355 170.236C92.799 170.938 90.9459 169.803 90.8281 168.026L90.8223 167.846L90.8232 77.6987C90.8241 68.6053 95.6664 60.1948 103.552 55.6411L103.549 55.6401C103.699 55.5536 103.851 55.4714 104.002 55.3882L103.893 55.4516C103.956 55.4159 104.02 55.3794 104.084 55.3442C104.057 55.359 104.029 55.3732 104.002 55.3882L137.533 36.0288C145.424 31.4763 155.13 31.4781 163.007 36.0259L189.969 51.5932C190.313 51.792 190.65 52.0011 190.982 52.2143L196.967 55.6694C204.855 60.2267 209.704 68.6343 209.704 77.73V120.221L209.689 121.073C209.397 129.848 204.599 137.874 196.967 142.28L178.568 152.895C177.314 153.621 175.738 152.716 175.737 151.263V84.5005ZM176.814 149.866L176.819 149.864L176.832 149.856C176.826 149.859 176.82 149.862 176.814 149.866ZM137.023 194.71L137.019 195.038C136.802 201.878 129.341 206.087 123.354 202.692L123.33 202.678L82.4727 179.091C80.3698 177.877 79.0762 175.634 79.0762 173.207V109.239L63.0957 118.458C56.5124 122.259 52.3745 129.183 52.1221 136.752L52.1094 137.487V179.978C52.1094 187.823 56.2909 195.075 63.0957 199.007H63.0967L69.0801 202.462L69.1504 202.503L69.2188 202.547C69.5212 202.741 69.8111 202.92 70.0947 203.083L97.0566 218.651L97.6982 219.007C104.372 222.569 112.434 222.453 119.029 218.648L152.514 199.316L152.512 199.312C152.581 199.274 152.65 199.236 152.752 199.178L152.753 199.181C152.775 199.169 152.796 199.158 152.815 199.147L153.011 199.035C159.81 195.107 163.987 187.854 163.988 180.009V91.3344L156.378 95.2153C144.501 101.27 137.023 113.475 137.023 126.806V194.71ZM94.3223 166.372L101.934 162.493C113.811 156.438 121.288 144.233 121.288 130.902V62.9897C121.278 55.9521 128.901 51.5532 134.986 55.0307L135 55.0385L175.841 78.6167L176.036 78.7339C178.023 79.9714 179.237 82.15 179.237 84.5005V148.468L195.217 139.249C202.013 135.326 206.204 128.075 206.204 120.221V77.73C206.204 69.8848 202.022 62.6318 195.217 58.6997L189.232 55.2456L189.162 55.2046L189.094 55.1606C188.788 54.9649 188.5 54.787 188.219 54.6245L161.257 39.0571C154.463 35.135 146.091 35.1311 139.282 39.0591L139.283 39.06L105.765 58.4106L105.767 58.4135C105.713 58.443 105.723 58.4383 105.602 58.5063L105.6 58.5034C105.535 58.5391 105.481 58.5691 105.433 58.5962L105.302 58.6723C98.5016 62.5995 94.324 69.8532 94.3232 77.6987L94.3223 166.372Z" fill="white"/>
+</svg>
@@ -99,7 +99,7 @@
    <PackageVersion Include="Microsoft.SemanticKernel.Connectors.InMemory" Version="1.67.0-preview" />
    <PackageVersion Include="Microsoft.SemanticKernel.Connectors.Qdrant" Version="1.67.0-preview" />
    <!-- Agent SDKs -->
-    <PackageVersion Include="GitHub.Copilot.SDK" Version="1.0.0-beta.2" />
+    <PackageVersion Include="GitHub.Copilot.SDK" Version="1.0.0" />
    <PackageVersion Include="Microsoft.Agents.CopilotStudio.Client" Version="1.3.171-beta" />
    <!-- M365 Agents SDK -->
    <PackageVersion Include="AdaptiveCards" Version="3.1.0" />
@@ -6,6 +6,7 @@

    <Nullable>enable</Nullable>
    <ImplicitUsings>enable</ImplicitUsings>
+    <NoWarn>$(NoWarn);GHCP001</NoWarn>
  </PropertyGroup>

  <ItemGroup>
@@ -2,21 +2,22 @@

 // This sample shows how to create a GitHub Copilot agent with shell command permissions.

-using GitHub.Copilot.SDK;
+using GitHub.Copilot;
+using GitHub.Copilot.Rpc;
 using Microsoft.Agents.AI;

 // Permission handler that prompts the user for approval
-static Task<PermissionRequestResult> PromptPermission(PermissionRequest request, PermissionInvocation invocation)
+static Task<PermissionDecision> PromptPermission(PermissionRequest request, PermissionInvocation invocation)
 {
    Console.WriteLine($"\n[Permission Request: {request.Kind}]");
    Console.Write("Approve? (y/n): ");

    string? input = Console.ReadLine()?.Trim().ToUpperInvariant();
-    PermissionRequestResultKind kind = input is "Y" or "YES"
-                                     ? PermissionRequestResultKind.Approved
-                                     : PermissionRequestResultKind.Rejected;
+    PermissionDecision decision = input is "Y" or "YES"
+                                     ? PermissionDecision.ApproveOnce()
+                                     : PermissionDecision.Reject();

-    return Task.FromResult(new PermissionRequestResult { Kind = kind });
+    return Task.FromResult(decision);
 }

 // Create and start a Copilot client
@@ -36,7 +36,7 @@ dotnet run
 You can customize the agent by providing additional configuration:

 ```csharp
-using GitHub.Copilot.SDK;
+using GitHub.Copilot;
 using Microsoft.Agents.AI;

 // Create and start a Copilot client
@@ -79,8 +79,10 @@ AIAgent agent =
    .GetProjectOpenAIClient()
    .GetResponsesClient()
    .AsIChatClient(deploymentName)
-    .AsHarnessAgent(MaxContextWindowTokens, MaxOutputTokens, new HarnessAgentOptions
+    .AsHarnessAgent(new HarnessAgentOptions
    {
+        MaxContextWindowTokens = MaxContextWindowTokens,
+        MaxOutputTokens = MaxOutputTokens,
        Name = "ResearchAgent",
        Description = "A research assistant that plans and executes research tasks.",
        DisableFileAccess = true,                           // If enabled, this would allow the agent to read/write files in a working directory
@@ -44,8 +44,10 @@ AIAgent webSearchAgent =
    .GetProjectOpenAIClient()
    .GetResponsesClient()
    .AsIChatClient(deploymentName)
-    .AsHarnessAgent(MaxContextWindowTokens, MaxOutputTokens, new HarnessAgentOptions
+    .AsHarnessAgent(new HarnessAgentOptions
    {
+        MaxContextWindowTokens = MaxContextWindowTokens,
+        MaxOutputTokens = MaxOutputTokens,
        Name = "WebSearchAgent",
        Description = "An agent that can search the web to find information.",
        OpenTelemetrySourceName = TracingSourceName,
@@ -92,8 +94,10 @@ AIAgent parentAgent =
    .GetProjectOpenAIClient()
    .GetResponsesClient()
    .AsIChatClient(deploymentName)
-    .AsHarnessAgent(MaxContextWindowTokens, MaxOutputTokens, new HarnessAgentOptions
+    .AsHarnessAgent(new HarnessAgentOptions
    {
+        MaxContextWindowTokens = MaxContextWindowTokens,
+        MaxOutputTokens = MaxOutputTokens,
        Name = "StockPriceResearcher",
        Description = "An agent that researches stock prices using background agents.",
        OpenTelemetrySourceName = TracingSourceName,
@@ -68,8 +68,10 @@ AIAgent agent =
    .GetProjectOpenAIClient()
    .GetResponsesClient()
    .AsIChatClient(deploymentName)
-    .AsHarnessAgent(MaxContextWindowTokens, MaxOutputTokens, new HarnessAgentOptions
+    .AsHarnessAgent(new HarnessAgentOptions
    {
+        MaxContextWindowTokens = MaxContextWindowTokens,
+        MaxOutputTokens = MaxOutputTokens,
        Name = "DataAnalyst",
        Description = "A data analyst assistant that reads, analyzes, and processes data files.",
        OpenTelemetrySourceName = TracingSourceName,
@@ -89,8 +89,10 @@ AIAgent agent =
    .GetProjectOpenAIClient()
    .GetResponsesClient()
    .AsIChatClient(deploymentName)
-    .AsHarnessAgent(MaxContextWindowTokens, MaxOutputTokens, new HarnessAgentOptions
+    .AsHarnessAgent(new HarnessAgentOptions
    {
+        MaxContextWindowTokens = MaxContextWindowTokens,
+        MaxOutputTokens = MaxOutputTokens,
        Name = "CodeExecutionAgent",
        Description = "A technical assistant with sandboxed code execution and skill-based workflows.",
        OpenTelemetrySourceName = TracingSourceName,
@@ -44,18 +44,33 @@ public static class HostedFoundryMemoryProviderScopes
        session => new FoundryMemoryProvider.State(new FoundryMemoryProviderScope(GetRequiredHostedContext(session).ChatId));

    /// <summary>
-    /// Returns a <c>stateInitializer</c> that scopes memories per (user, chat) pair, using
-    /// <c>"{UserId}:{ChatId}"</c> as the partition key. Use this when memories should be visible
-    /// only to the same user within the same conversation.
+    /// Returns a <c>stateInitializer</c> that scopes memories per (user, chat) pair, composing
+    /// <see cref="HostedSessionContext.UserId"/> and <see cref="HostedSessionContext.ChatId"/> into a
+    /// single delimiter-safe partition key. Use this when memories should be visible only to the same
+    /// user within the same conversation.
    /// </summary>
+    /// <remarks>
+    /// Both identity values are opaque strings that may contain any characters, including the <c>:</c>
+    /// delimiter. To keep the composite key injective (so two distinct (user, chat) pairs can never
+    /// collide), each part is escaped (<c>\</c> becomes <c>\\</c>, then <c>:</c> becomes <c>\:</c>) before
+    /// being joined with a <c>::</c> separator.
+    /// </remarks>
    /// <returns>A delegate suitable for the <c>stateInitializer</c> argument of <see cref="FoundryMemoryProvider"/>.</returns>
    public static Func<AgentSession?, FoundryMemoryProvider.State> PerUserAndChat() =>
        session =>
        {
            var ctx = GetRequiredHostedContext(session);
-            return new FoundryMemoryProvider.State(new FoundryMemoryProviderScope($"{ctx.UserId}:{ctx.ChatId}"));
+            return new FoundryMemoryProvider.State(
+                new FoundryMemoryProviderScope($"{EscapeScopePart(ctx.UserId)}::{EscapeScopePart(ctx.ChatId)}"));
        };

+    /// <summary>
+    /// Escapes special characters in a scope part so that distinct (user, chat) pairs produce distinct
+    /// composite scope keys. Backslashes are escaped first (<c>\</c> becomes <c>\\</c>), then colons
+    /// (<c>:</c> becomes <c>\:</c>), ensuring the <c>{user}::{chat}</c> format is unambiguous.
+    /// </summary>
+    private static string EscapeScopePart(string part) => part.Replace("\\", "\\\\").Replace(":", "\\:");
+
    private static HostedSessionContext GetRequiredHostedContext(AgentSession? session) =>
        session?.GetHostedContext()
            ?? throw new InvalidOperationException(
@@ -6,7 +6,7 @@ using Microsoft.Agents.AI.GitHub.Copilot;
 using Microsoft.Extensions.AI;
 using Microsoft.Shared.Diagnostics;

-namespace GitHub.Copilot.SDK;
+namespace GitHub.Copilot;

 /// <summary>
 /// Provides extension methods for <see cref="CopilotClient"/>
@@ -9,7 +9,7 @@ using System.Text.Json;
 using System.Threading;
 using System.Threading.Channels;
 using System.Threading.Tasks;
-using GitHub.Copilot.SDK;
+using GitHub.Copilot;
 using Microsoft.Extensions.AI;
 using Microsoft.Shared.Diagnostics;

@@ -169,7 +169,7 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable
            Channel<AgentResponseUpdate> channel = Channel.CreateUnbounded<AgentResponseUpdate>();

            // Subscribe to session events
-            using IDisposable subscription = copilotSession.On(evt =>
+            using IDisposable subscription = copilotSession.On<SessionEvent>(evt =>
            {
                switch (evt)
                {
@@ -210,7 +210,7 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable
                string prompt = string.Join("\n", messages.Select(m => m.Text));

                // Handle DataContent as attachments
-                (List<UserMessageAttachmentFile>? attachments, tempDir) = await ProcessDataContentAttachmentsAsync(
+                (List<AttachmentFile>? attachments, tempDir) = await ProcessDataContentAttachmentsAsync(
                    messages,
                    cancellationToken).ConfigureAwait(false);

@@ -262,10 +262,7 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable

    private async Task EnsureClientStartedAsync(CancellationToken cancellationToken)
    {
-        if (this._copilotClient.State != ConnectionState.Connected)
-        {
-            await this._copilotClient.StartAsync(cancellationToken).ConfigureAwait(false);
-        }
+        await this._copilotClient.StartAsync(cancellationToken).ConfigureAwait(false);
    }

    private ResumeSessionConfig CreateResumeConfig()
@@ -275,36 +272,18 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable

    /// <summary>
    /// Copies all supported properties from a source <see cref="SessionConfig"/> into a new instance
-    /// with <see cref="SessionConfig.Streaming"/> set to <c>true</c>.
+    /// with <see cref="SessionConfigBase.Streaming"/> set to <c>true</c>.
    /// </summary>
    internal static SessionConfig CopySessionConfig(SessionConfig source)
    {
-        return new SessionConfig
-        {
-            Model = source.Model,
-            ReasoningEffort = source.ReasoningEffort,
-            Tools = source.Tools,
-            SystemMessage = source.SystemMessage,
-            AvailableTools = source.AvailableTools,
-            ExcludedTools = source.ExcludedTools,
-            Provider = source.Provider,
-            OnPermissionRequest = source.OnPermissionRequest,
-            OnUserInputRequest = source.OnUserInputRequest,
-            Hooks = source.Hooks,
-            WorkingDirectory = source.WorkingDirectory,
-            ConfigDir = source.ConfigDir,
-            McpServers = source.McpServers,
-            CustomAgents = source.CustomAgents,
-            SkillDirectories = source.SkillDirectories,
-            DisabledSkills = source.DisabledSkills,
-            InfiniteSessions = source.InfiniteSessions,
-            Streaming = true
-        };
+        SessionConfig copy = source.Clone();
+        copy.Streaming = true;
+        return copy;
    }

    /// <summary>
    /// Copies all supported properties from a source <see cref="SessionConfig"/> into a new
-    /// <see cref="ResumeSessionConfig"/> with <see cref="ResumeSessionConfig.Streaming"/> set to <c>true</c>.
+    /// <see cref="ResumeSessionConfig"/> with <see cref="SessionConfigBase.Streaming"/> set to <c>true</c>.
    /// </summary>
    internal static ResumeSessionConfig CopyResumeSessionConfig(SessionConfig? source)
    {
@@ -321,7 +300,7 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable
            OnUserInputRequest = source?.OnUserInputRequest,
            Hooks = source?.Hooks,
            WorkingDirectory = source?.WorkingDirectory,
-            ConfigDir = source?.ConfigDir,
+            ConfigDirectory = source?.ConfigDirectory,
            McpServers = source?.McpServers,
            CustomAgents = source?.CustomAgents,
            SkillDirectories = source?.SkillDirectories,
@@ -394,10 +373,10 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable

        AdditionalPropertiesDictionary<long>? additionalCounts = null;

-        if (usageEvent.Data.CacheWriteTokens is double cacheWriteTokens)
+        if (usageEvent.Data.CacheWriteTokens is long cacheWriteTokens)
        {
            additionalCounts ??= [];
-            additionalCounts[nameof(AssistantUsageData.CacheWriteTokens)] = (long)cacheWriteTokens;
+            additionalCounts[nameof(AssistantUsageData.CacheWriteTokens)] = cacheWriteTokens;
        }

        if (usageEvent.Data.Cost is double cost)
@@ -406,10 +385,10 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable
            additionalCounts[nameof(AssistantUsageData.Cost)] = (long)cost;
        }

-        if (usageEvent.Data.Duration is double duration)
+        if (usageEvent.Data.Duration is TimeSpan duration)
        {
            additionalCounts ??= [];
-            additionalCounts[nameof(AssistantUsageData.Duration)] = (long)duration;
+            additionalCounts[nameof(AssistantUsageData.Duration)] = (long)duration.TotalMilliseconds;
        }

        return additionalCounts;
@@ -432,7 +411,7 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable

    private static SessionConfig? GetSessionConfig(IList<AITool>? tools, string? instructions)
    {
-        List<AIFunction>? mappedTools = tools is { Count: > 0 } ? tools.OfType<AIFunction>().ToList() : null;
+        List<AIFunctionDeclaration>? mappedTools = tools is { Count: > 0 } ? tools.OfType<AIFunctionDeclaration>().ToList() : null;
        SystemMessageConfig? systemMessage = instructions is not null ? new SystemMessageConfig { Mode = SystemMessageMode.Append, Content = instructions } : null;

        if (mappedTools is null && systemMessage is null)
@@ -443,11 +422,11 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable
        return new SessionConfig { Tools = mappedTools, SystemMessage = systemMessage };
    }

-    private static async Task<(List<UserMessageAttachmentFile>? Attachments, string? TempDir)> ProcessDataContentAttachmentsAsync(
+    private static async Task<(List<AttachmentFile>? Attachments, string? TempDir)> ProcessDataContentAttachmentsAsync(
        IEnumerable<ChatMessage> messages,
        CancellationToken cancellationToken)
    {
-        List<UserMessageAttachmentFile>? attachments = null;
+        List<AttachmentFile>? attachments = null;
        string? tempDir = null;
        foreach (ChatMessage message in messages)
        {
@@ -461,7 +440,7 @@ public sealed class GitHubCopilotAgent : AIAgent, IAsyncDisposable
                    string tempFilePath = await dataContent.SaveToAsync(tempDir, cancellationToken).ConfigureAwait(false);

                    attachments ??= [];
-                    attachments.Add(new UserMessageAttachmentFile
+                    attachments.Add(new AttachmentFile
                    {
                        Path = tempFilePath,
                        DisplayName = Path.GetFileName(tempFilePath)
@@ -4,6 +4,7 @@
    <VersionSuffix>preview</VersionSuffix>
    <!-- GitHub.Copilot.SDK only supports .NET 8.0+ -->
    <TargetFrameworks>$(TargetFrameworksCore)</TargetFrameworks>
+    <NoWarn>$(NoWarn);GHCP001</NoWarn>
  </PropertyGroup>

  <PropertyGroup>
@@ -16,23 +16,16 @@ public static class ChatClientHarnessExtensions
 {
    /// <summary>
    /// Creates a new <see cref="HarnessAgent"/> that wraps this <see cref="IChatClient"/> with a pre-configured
-    /// pipeline including function invocation, per-service-call chat history persistence, and in-loop compaction.
+    /// pipeline including function invocation, per-service-call chat history persistence, optional in-loop compaction, and a rich set
+    /// of default context providers and agent decorators.
    /// </summary>
    /// <param name="chatClient">
    /// The <see cref="IChatClient"/> that provides access to the underlying AI model.
    /// </param>
-    /// <param name="maxContextWindowTokens">
-    /// The maximum number of tokens the model's context window supports (e.g., 1,050,000 for gpt-5.4).
-    /// Used to configure the compaction strategy.
-    /// </param>
-    /// <param name="maxOutputTokens">
-    /// The maximum number of output tokens the model can generate per response (e.g., 128,000 for gpt-5.4).
-    /// Used to configure the compaction strategy.
-    /// </param>
    /// <param name="options">
    /// Optional configuration options for the agent, including instructions override, tools,
-    /// additional context providers, and chat history provider.
-    /// When <see langword="null"/>, the agent uses built-in default settings.
+    /// additional context providers, chat history provider, and compaction settings.
+    /// When <see langword="null"/>, the agent uses built-in default settings with compaction disabled.
    /// </param>
    /// <param name="loggerFactory">
    /// Optional logger factory for creating loggers used by the agent and its components.
@@ -43,10 +36,8 @@ public static class ChatClientHarnessExtensions
    /// <returns>A new <see cref="HarnessAgent"/> instance.</returns>
    public static HarnessAgent AsHarnessAgent(
        this IChatClient chatClient,
-        int maxContextWindowTokens,
-        int maxOutputTokens,
        HarnessAgentOptions? options = null,
        ILoggerFactory? loggerFactory = null,
        IServiceProvider? services = null) =>
-        new(chatClient, maxContextWindowTokens, maxOutputTokens, options, loggerFactory, services);
+        new(chatClient, options, loggerFactory, services);
 }
@@ -18,50 +18,65 @@ namespace Microsoft.Agents.AI;

 /// <summary>
 /// A pre-configured <see cref="DelegatingAIAgent"/> that wraps a <see cref="ChatClientAgent"/> with
-/// function invocation, per-service-call chat history persistence, in-loop compaction, and a rich set
+/// function invocation, per-service-call chat history persistence, optional in-loop compaction, and a rich set
 /// of default context providers and agent decorators.
 /// </summary>
 /// <remarks>
 /// <para>
-/// <see cref="HarnessAgent"/> assembles the following pipeline from a caller-supplied <see cref="IChatClient"/>:
+/// <see cref="HarnessAgent"/> provides an opinionated, batteries-included agent suitable for
+/// interactive agentic scenarios such as research, coding, data analysis, and general task automation.
+/// It assembles a full pipeline from a caller-supplied <see cref="IChatClient"/> so that callers
+/// only need to configure the parts they want to customize.
+/// </para>
+/// <para>
+/// <strong>Chat client pipeline (inner to outer):</strong>
 /// <list type="number">
-/// <item><description><see cref="FunctionInvokingChatClient"/> — automatic function/tool invocation.</description></item>
-/// <item><description><see cref="MessageInjectingChatClient"/> — allows external code to inject messages into the conversation mid-stream.</description></item>
-/// <item><description><see cref="PerServiceCallChatHistoryPersistingChatClient"/> — persists chat history after every individual service call within a function-invocation loop.</description></item>
-/// <item><description><see cref="AIContextProviderChatClient"/> with a <see cref="CompactionProvider"/> — applies context-window compaction before each call so long function-invocation loops do not overflow the context window.</description></item>
+/// <item><description><see cref="FunctionInvokingChatClient"/> — automatic function/tool invocation with configurable iteration limits.</description></item>
+/// <item><description><see cref="MessageInjectingChatClient"/> — allows external code to inject messages into the conversation mid-stream (e.g., for user interrupts).</description></item>
+/// <item><description><see cref="PerServiceCallChatHistoryPersistingChatClient"/> — persists chat history after every individual service call within a function-invocation loop, enabling crash recovery and history inspection.</description></item>
+/// <item><description><see cref="AIContextProviderChatClient"/> with a <see cref="CompactionProvider"/> — applies context-window compaction before each call so long function-invocation loops do not overflow the context window. Only included when <see cref="HarnessAgentOptions.MaxContextWindowTokens"/> and <see cref="HarnessAgentOptions.MaxOutputTokens"/> are both provided.</description></item>
 /// </list>
 /// </para>
 /// <para>
-/// By default, the following context providers are included (each can be disabled via <see cref="HarnessAgentOptions"/>):
+/// <strong>Context providers (each enabled by default, individually disableable via <see cref="HarnessAgentOptions"/>):</strong>
 /// <list type="bullet">
-/// <item><description><see cref="TodoProvider"/> — todo list management.</description></item>
-/// <item><description><see cref="AgentModeProvider"/> — agent mode tracking (plan/execute).</description></item>
-/// <item><description><see cref="FileMemoryProvider"/> — file-based session memory.</description></item>
-/// <item><description><see cref="FileAccessProvider"/> — shared file access.</description></item>
-/// <item><description><see cref="AgentSkillsProvider"/> — skill discovery and loading.</description></item>
+/// <item><description><see cref="TodoProvider"/> — persistent todo list that the agent uses to track multi-step plans. Disable with <see cref="HarnessAgentOptions.DisableTodoProvider"/>.</description></item>
+/// <item><description><see cref="AgentModeProvider"/> — mode tracking (e.g., "plan" vs "execute") that the agent uses to structure its work. Disable with <see cref="HarnessAgentOptions.DisableAgentModeProvider"/>.</description></item>
+/// <item><description><see cref="FileMemoryProvider"/> — file-based session memory allowing the agent to persist notes and artifacts across turns. Disable with <see cref="HarnessAgentOptions.DisableFileMemory"/>.</description></item>
+/// <item><description><see cref="FileAccessProvider"/> — shared file access providing read/write tools for a working directory. Disable with <see cref="HarnessAgentOptions.DisableFileAccess"/>.</description></item>
+/// <item><description><see cref="AgentSkillsProvider"/> — discovers and loads skill definitions from the file system, enabling dynamic tool sets. Disable with <see cref="HarnessAgentOptions.DisableAgentSkillsProvider"/>.</description></item>
 /// </list>
 /// </para>
 /// <para>
-/// The agent is also wrapped with the following decorators by default (each can be disabled):
+/// <strong>Optional context providers (enabled via <see cref="HarnessAgentOptions"/>):</strong>
 /// <list type="bullet">
-/// <item><description><see cref="ToolApprovalAgent"/> — "don't ask again" tool approval rules.</description></item>
-/// <item><description><see cref="OpenTelemetryAgent"/> — OpenTelemetry instrumentation.</description></item>
+/// <item><description><see cref="BackgroundAgentsProvider"/> — enables delegation to background agents for parallel work. Enable by setting <see cref="HarnessAgentOptions.BackgroundAgents"/>.</description></item>
+/// <item><description><c>ShellEnvironmentProvider</c> — injects OS/shell/CWD information and a shell execution tool. Enable by setting <c>HarnessAgentOptions.ShellExecutor</c> (.NET only).</description></item>
 /// </list>
 /// </para>
 /// <para>
-/// A <see cref="HostedWebSearchTool"/> is added to the chat options by default (can be disabled via
-/// <see cref="HarnessAgentOptions.DisableWebSearch"/>).
+/// <strong>Agent decorators (each enabled by default, individually disableable):</strong>
+/// <list type="bullet">
+/// <item><description><see cref="ToolApprovalAgent"/> — "don't ask again" tool approval rules enabling safe unattended execution. Disable with <see cref="HarnessAgentOptions.DisableToolApproval"/>.</description></item>
+/// <item><description><see cref="OpenTelemetryAgent"/> — OpenTelemetry instrumentation following semantic conventions for generative AI. Disable with <see cref="HarnessAgentOptions.DisableOpenTelemetry"/>.</description></item>
+/// </list>
 /// </para>
 /// <para>
-/// The underlying <see cref="ChatClientAgent"/> is configured with
-/// <see cref="ChatClientAgentOptions.UseProvidedChatClientAsIs"/> and
-/// <see cref="ChatClientAgentOptions.RequirePerServiceCallChatHistoryPersistence"/> set to <see langword="true"/>
-/// to match the manually-assembled pipeline.
+/// <strong>Default tools:</strong>
+/// <list type="bullet">
+/// <item><description><see cref="HostedWebSearchTool"/> — a hosted web search tool added to chat options by default. Disable with <see cref="HarnessAgentOptions.DisableWebSearch"/>.</description></item>
+/// </list>
 /// </para>
 /// <para>
-/// When no <see cref="HarnessAgentOptions.ChatHistoryProvider"/> is supplied, the agent defaults to an
-/// <see cref="InMemoryChatHistoryProvider"/> whose chat reducer applies the same compaction strategy,
-/// keeping in-memory history from growing unboundedly across sessions.
+/// <strong>Chat history:</strong> When no <see cref="HarnessAgentOptions.ChatHistoryProvider"/> is supplied,
+/// the agent defaults to an <see cref="InMemoryChatHistoryProvider"/>. If compaction is enabled, the provider
+/// is configured with a compaction-based chat reducer to keep in-memory history bounded. Otherwise, no reducer
+/// is applied.
+/// </para>
+/// <para>
+/// <strong>Default instructions:</strong> The agent includes built-in system instructions (<see cref="DefaultInstructions"/>)
+/// that guide general tool usage and reasoning patterns. These can be overridden via <see cref="HarnessAgentOptions.HarnessInstructions"/>
+/// and combined with agent-specific instructions via <see cref="ChatOptions.Instructions"/>.
 /// </para>
 /// </remarks>
 [Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
@@ -90,21 +105,13 @@ public sealed class HarnessAgent : DelegatingAIAgent
    /// </summary>
    /// <param name="chatClient">
    /// The <see cref="IChatClient"/> that provides access to the underlying AI model.
-    /// The agent wraps this client in a function-invocation, per-service-call persistence,
-    /// and compaction pipeline automatically.
-    /// </param>
-    /// <param name="maxContextWindowTokens">
-    /// The maximum number of tokens the model's context window supports (e.g., 1,050,000 for gpt-5.4).
-    /// Used to configure the compaction strategy.
-    /// </param>
-    /// <param name="maxOutputTokens">
-    /// The maximum number of output tokens the model can generate per response (e.g., 128,000 for gpt-5.4).
-    /// Used to configure the compaction strategy and to limit the model's output.
+    /// The agent wraps this client in a function-invocation and per-service-call persistence pipeline.
+    /// When compaction is enabled via <paramref name="options"/>, a compaction decorator is also added.
    /// </param>
    /// <param name="options">
    /// Optional configuration options for the agent, including instructions override, tools,
-    /// additional context providers, and chat history provider.
-    /// When <see langword="null"/>, the agent uses built-in default settings.
+    /// additional context providers, chat history provider, and compaction settings.
+    /// When <see langword="null"/>, the agent uses built-in default settings with compaction disabled.
    /// </param>
    /// <param name="loggerFactory">
    /// Optional logger factory for creating loggers used by the agent and its components.
@@ -116,29 +123,28 @@ public sealed class HarnessAgent : DelegatingAIAgent
    /// <paramref name="chatClient"/> is <see langword="null"/>.
    /// </exception>
    /// <exception cref="ArgumentOutOfRangeException">
-    /// <paramref name="maxContextWindowTokens"/> is not positive, or
-    /// <paramref name="maxOutputTokens"/> is negative or greater than or equal to <paramref name="maxContextWindowTokens"/>.
+    /// <see cref="HarnessAgentOptions.MaxContextWindowTokens"/> is not positive, or
+    /// <see cref="HarnessAgentOptions.MaxOutputTokens"/> is negative or greater than or equal to
+    /// <see cref="HarnessAgentOptions.MaxContextWindowTokens"/> (when both are provided).
    /// </exception>
-    public HarnessAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options = null, ILoggerFactory? loggerFactory = null, IServiceProvider? services = null)
+    public HarnessAgent(IChatClient chatClient, HarnessAgentOptions? options = null, ILoggerFactory? loggerFactory = null, IServiceProvider? services = null)
        : base(BuildAgent(
            Throw.IfNull(chatClient),
-            maxContextWindowTokens,
-            maxOutputTokens,
            options,
            loggerFactory,
            services))
    {
    }

-    private static AIAgent BuildAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options, ILoggerFactory? loggerFactory, IServiceProvider? services)
+    private static AIAgent BuildAgent(IChatClient chatClient, HarnessAgentOptions? options, ILoggerFactory? loggerFactory, IServiceProvider? services)
    {
-        ChatClientAgent innerAgent = BuildInnerAgent(chatClient, maxContextWindowTokens, maxOutputTokens, options, loggerFactory, services);
+        ChatClientAgent innerAgent = BuildInnerAgent(chatClient, options, loggerFactory, services);

        AIAgentBuilder builder = innerAgent.AsBuilder();

        if (options?.DisableToolApproval is not true)
        {
-            builder.UseToolApproval();
+            builder.UseToolApproval(options?.ToolApprovalAgentOptions);
        }

        if (options?.DisableOpenTelemetry is not true)
@@ -149,17 +155,35 @@ public sealed class HarnessAgent : DelegatingAIAgent
        return builder.Build(services);
    }

-    private static ChatClientAgent BuildInnerAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options, ILoggerFactory? loggerFactory, IServiceProvider? services)
+    private static ChatClientAgent BuildInnerAgent(IChatClient chatClient, HarnessAgentOptions? options, ILoggerFactory? loggerFactory, IServiceProvider? services)
    {
-        var compactionStrategy = new ContextWindowCompactionStrategy(
-            maxContextWindowTokens: maxContextWindowTokens,
-            maxOutputTokens: maxOutputTokens);
+        // Determine compaction strategy:
+        // 1. DisableCompaction = true → no compaction
+        // 2. Custom CompactionStrategy provided → use it (ignore token params)
+        // 3. Both token params provided → build default ContextWindowCompactionStrategy
+        // 4. Otherwise → no compaction
+        CompactionStrategy? compactionStrategy = null;
+        if (options?.DisableCompaction is not true)
+        {
+            if (options?.CompactionStrategy is CompactionStrategy customStrategy)
+            {
+                compactionStrategy = customStrategy;
+            }
+            else if (options?.MaxContextWindowTokens is int maxCtx && options?.MaxOutputTokens is int maxOut)
+            {
+                compactionStrategy = new ContextWindowCompactionStrategy(
+                    maxContextWindowTokens: maxCtx,
+                    maxOutputTokens: maxOut);
+            }
+        }

        ChatHistoryProvider chatHistoryProvider = options?.ChatHistoryProvider
-            ?? new InMemoryChatHistoryProvider(new InMemoryChatHistoryProviderOptions
-            {
-                ChatReducer = compactionStrategy.AsChatReducer(),
-            });
+            ?? (compactionStrategy is not null
+                ? new InMemoryChatHistoryProvider(new InMemoryChatHistoryProviderOptions
+                {
+                    ChatReducer = compactionStrategy.AsChatReducer(),
+                })
+                : new InMemoryChatHistoryProvider());

        string harnessInstructions = options?.HarnessInstructions ?? DefaultInstructions;
        string? agentInstructions = options?.ChatOptions?.Instructions;
@@ -172,20 +196,34 @@ public sealed class HarnessAgent : DelegatingAIAgent
            (false, false) => $"{harnessInstructions}\n\n{agentInstructions}",
        };

-        ChatOptions chatOptions = BuildChatOptions(options, instructions, maxOutputTokens);
+        ChatOptions chatOptions = BuildChatOptions(options, instructions, options?.MaxOutputTokens);

-        var compactionProvider = new CompactionProvider(compactionStrategy, loggerFactory: loggerFactory);
+        CompactionProvider? compactionProvider = compactionStrategy is not null
+            ? new CompactionProvider(compactionStrategy, loggerFactory: loggerFactory)
+            : null;

        IEnumerable<AIContextProvider> contextProviders = BuildContextProviders(options, loggerFactory);

-        return chatClient
-            .AsBuilder()
+        ChatClientBuilder chatClientBuilder = chatClient.AsBuilder();
+
+        if (options?.DisableNonApprovalRequiredFunctionBypassing is not true)
+        {
+            chatClientBuilder.UseNonApprovalRequiredFunctionBypassing();
+        }
+
+        ChatClientBuilder pipeline = chatClientBuilder
            .UseFunctionInvocation(loggerFactory, configure: options?.MaximumIterationsPerRequest is int maxIterations
                ? ficc => ficc.MaximumIterationsPerRequest = maxIterations
                : null)
            .UseMessageInjection()
-            .UsePerServiceCallChatHistoryPersistence()
-            .UseAIContextProviders(compactionProvider)
+            .UsePerServiceCallChatHistoryPersistence();
+
+        if (compactionProvider is not null)
+        {
+            pipeline = pipeline.UseAIContextProviders(compactionProvider);
+        }
+
+        return pipeline
            .BuildAIAgent(new ChatClientAgentOptions
            {
                Id = options?.Id,
@@ -203,11 +241,15 @@ public sealed class HarnessAgent : DelegatingAIAgent
            services);
    }

-    private static ChatOptions BuildChatOptions(HarnessAgentOptions? options, string instructions, int maxOutputTokens)
+    private static ChatOptions BuildChatOptions(HarnessAgentOptions? options, string instructions, int? maxOutputTokens)
    {
        ChatOptions result = options?.ChatOptions?.Clone() ?? new ChatOptions();
        result.Instructions = instructions;
-        result.MaxOutputTokens ??= maxOutputTokens;
+
+        if (maxOutputTokens.HasValue)
+        {
+            result.MaxOutputTokens ??= maxOutputTokens.Value;
+        }

        if (options?.DisableWebSearch is not true)
        {
@@ -2,6 +2,7 @@

 using System.Collections.Generic;
 using System.Diagnostics.CodeAnalysis;
+using Microsoft.Agents.AI.Compaction;
 #if NET
 using Microsoft.Agents.AI.Tools.Shell;
 #endif
@@ -31,6 +32,68 @@ public sealed class HarnessAgentOptions
    /// </summary>
    public string? Description { get; set; }

+    /// <summary>
+    /// Gets or sets the maximum number of tokens the model's context window supports (e.g., 1,050,000 for gpt-5.4).
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// When both <see cref="MaxContextWindowTokens"/> and <see cref="MaxOutputTokens"/> are provided (and no
+    /// custom <see cref="CompactionStrategy"/> is set), a default <see cref="ContextWindowCompactionStrategy"/>
+    /// is constructed from these values to prevent function-invocation loops from overflowing the context window.
+    /// </para>
+    /// <para>
+    /// Ignored when <see cref="CompactionStrategy"/> is provided or when <see cref="DisableCompaction"/> is
+    /// <see langword="true"/>.
+    /// </para>
+    /// </remarks>
+    public int? MaxContextWindowTokens { get; set; }
+
+    /// <summary>
+    /// Gets or sets the maximum number of output tokens the model can generate per response (e.g., 128,000 for gpt-5.4).
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// When set, this value is used as the default for <see cref="ChatOptions"/>.<see cref="ChatOptions.MaxOutputTokens"/>
+    /// when not explicitly configured.
+    /// </para>
+    /// <para>
+    /// For compaction purposes, this value is used together with <see cref="MaxContextWindowTokens"/> to construct a
+    /// default <see cref="ContextWindowCompactionStrategy"/> — but only when no custom <see cref="CompactionStrategy"/>
+    /// is provided and <see cref="DisableCompaction"/> is <see langword="false"/>.
+    /// </para>
+    /// </remarks>
+    public int? MaxOutputTokens { get; set; }
+
+    /// <summary>
+    /// Gets or sets a custom <see cref="Compaction.CompactionStrategy"/> to use for in-loop context-window compaction.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// When provided, this strategy is used directly and <see cref="MaxContextWindowTokens"/> and
+    /// <see cref="MaxOutputTokens"/> are ignored for compaction purposes (<see cref="MaxOutputTokens"/> is still
+    /// used as the default for <see cref="ChatOptions"/>.<see cref="ChatOptions.MaxOutputTokens"/> if set).
+    /// </para>
+    /// <para>
+    /// When <see langword="null"/> and both <see cref="MaxContextWindowTokens"/> and <see cref="MaxOutputTokens"/>
+    /// are provided, a default <see cref="ContextWindowCompactionStrategy"/> is constructed from those values.
+    /// </para>
+    /// <para>
+    /// This property is ignored when <see cref="DisableCompaction"/> is <see langword="true"/>.
+    /// </para>
+    /// </remarks>
+    public CompactionStrategy? CompactionStrategy { get; set; }
+
+    /// <summary>
+    /// Gets or sets a value indicating whether in-loop compaction is disabled.
+    /// </summary>
+    /// <remarks>
+    /// When <see langword="true"/>, compaction is disabled regardless of <see cref="CompactionStrategy"/>,
+    /// <see cref="MaxContextWindowTokens"/>, or <see cref="MaxOutputTokens"/> settings. No
+    /// <see cref="CompactionProvider"/> is added to the chat client pipeline, and the default
+    /// <see cref="InMemoryChatHistoryProvider"/> is configured without a chat reducer.
+    /// </remarks>
+    public bool DisableCompaction { get; set; }
+
    /// <summary>
    /// Gets or sets additional chat options such as tools for the agent to use.
    /// </summary>
@@ -68,9 +131,9 @@ public sealed class HarnessAgentOptions
    /// Gets or sets the <see cref="ChatHistoryProvider"/> to use for storing chat history.
    /// </summary>
    /// <remarks>
-    /// When <see langword="null"/>, the agent defaults to an <see cref="InMemoryChatHistoryProvider"/>
-    /// configured with a compaction-based chat reducer derived from the <c>maxContextWindowTokens</c>
-    /// and <c>maxOutputTokens</c> constructor parameters of <see cref="HarnessAgent"/>.
+    /// When <see langword="null"/>, the agent defaults to an <see cref="InMemoryChatHistoryProvider"/>.
+    /// If <see cref="MaxContextWindowTokens"/> and <see cref="MaxOutputTokens"/> are both provided,
+    /// the default provider is configured with a compaction-based chat reducer; otherwise, no reducer is applied.
    /// </remarks>
    public ChatHistoryProvider? ChatHistoryProvider { get; set; }

@@ -101,6 +164,29 @@ public sealed class HarnessAgentOptions
    /// </remarks>
    public bool DisableToolApproval { get; set; }

+    /// <summary>
+    /// Gets or sets the options for the <see cref="ToolApprovalAgent"/> middleware.
+    /// </summary>
+    /// <remarks>
+    /// When <see langword="null"/>, the <see cref="ToolApprovalAgent"/> uses default settings.
+    /// This property has no effect when <see cref="DisableToolApproval"/> is <see langword="true"/>.
+    /// </remarks>
+    public ToolApprovalAgentOptions? ToolApprovalAgentOptions { get; set; }
+
+    /// <summary>
+    /// Gets or sets a value indicating whether bypassing of approval requests for tools that do not
+    /// require approval is disabled.
+    /// </summary>
+    /// <remarks>
+    /// When <see langword="false"/> (the default), the underlying chat client pipeline includes the decorator
+    /// added by <see cref="ChatClientBuilderExtensions.UseNonApprovalRequiredFunctionBypassing"/> above the
+    /// function invocation middleware.
+    /// This stores automatically approved function calls for tools that do not require approval in the session
+    /// state when they are returned alongside tools that do, so that only tools that truly require human
+    /// approval are surfaced to the caller.
+    /// </remarks>
+    public bool DisableNonApprovalRequiredFunctionBypassing { get; set; }
+
    /// <summary>
    /// Gets or sets a value indicating whether the <see cref="FileMemoryProvider"/> is disabled.
    /// </summary>
@@ -21,6 +21,18 @@ namespace Microsoft.Agents.AI.Hosting;
 /// from the ambient <see cref="HttpContext"/>.
 /// </para>
 /// <para>
+/// <strong>Security warning:</strong> The configured <see cref="ClaimsIdentitySessionIsolationKeyProviderOptions.ClaimType"/>
+/// must uniquely identify the principal within the served population. Display names, usernames, email
+/// aliases, and other mutable or non-unique claims are <strong>unsafe</strong> isolation keys unless the
+/// host can prove their uniqueness across all callers: two distinct principals that share the same value
+/// would receive the same isolation key and could read or overwrite one another's persisted sessions.
+/// The default claim type is <see cref="ClaimTypes.NameIdentifier"/>, a stable unique subject identifier
+/// that is typically populated from the OpenID Connect <c>sub</c> claim via the default JWT inbound claim
+/// mapping (note that this differs from Entra's object identifier <c>oid</c> claim; override
+/// <see cref="ClaimsIdentitySessionIsolationKeyProviderOptions.ClaimType"/> if you need <c>oid</c> or your
+/// provider maps a different claim).
+/// </para>
+/// <para>
 /// If the <see cref="HttpContext"/> is unavailable, the user is not authenticated, or the specified claim
 /// is missing, the provider returns <see langword="null"/>. The consuming <see cref="IsolationKeyScopedAgentSessionStore"/>
 /// will then enforce strict or pass-through behavior based on its configuration.
@@ -60,18 +72,24 @@ public class ClaimsIdentitySessionIsolationKeyProvider : SessionIsolationKeyProv
    /// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests.</param>
    /// <returns>
    /// A task that represents the asynchronous operation. The task result contains the value of the
-    /// configured claim type from the current user's identity, or <see langword="null"/> if the claim
-    /// is not present or the HTTP context is unavailable.
+    /// configured claim type from the current user's identity, or <see langword="null"/> if the HTTP
+    /// context is unavailable, the user is not authenticated, or the claim is not present.
    /// </returns>
    /// <remarks>
-    /// This method retrieves the claim value from <c>HttpContext.User.Claims</c>. If multiple claims
-    /// of the specified type exist, the first match is returned.
+    /// This method only reads claims from an authenticated principal: if the current request has no
+    /// authenticated user, it returns <see langword="null"/> rather than trusting claims on an
+    /// unauthenticated identity. The claim value is retrieved from <c>HttpContext.User.Claims</c>; if
+    /// multiple claims of the specified type exist, the first match is returned.
    /// </remarks>
    public override ValueTask<string?> GetSessionIsolationKeyAsync(CancellationToken cancellationToken = default)
    {
-        Claim? claim = this._httpContextAccessor?
-                           .HttpContext?
-                           .User?.Claims.FirstOrDefault(c => c.Type == this._claimType);
+        ClaimsPrincipal? user = this._httpContextAccessor?.HttpContext?.User;
+        if (user?.Identity?.IsAuthenticated != true)
+        {
+            return new ValueTask<string?>((string?)null);
+        }
+
+        Claim? claim = user?.Claims.FirstOrDefault(c => c.Type == this._claimType);

        return new ValueTask<string?>(claim?.Value);
    }
@@ -14,17 +14,30 @@ public class ClaimsIdentitySessionIsolationKeyProviderOptions
    /// </summary>
    /// <remarks>
    /// <para>
-    /// Defaults to <see cref="ClaimsIdentity.DefaultNameClaimType"/>, which typically corresponds to
-    /// the user's name or unique identifier claim.
+    /// Defaults to <see cref="ClaimTypes.NameIdentifier"/>, which corresponds to a stable, unique
+    /// subject identifier for the authenticated principal. For OpenID Connect tokens (including those
+    /// issued by Microsoft Entra ID), this is typically populated from the <c>sub</c> claim via the
+    /// default JWT inbound claim mapping. Note that <c>sub</c> is distinct from Entra's object
+    /// identifier (<c>oid</c>) claim; if you require the <c>oid</c> claim, or your provider does not map
+    /// a unique identifier onto <see cref="ClaimTypes.NameIdentifier"/>, override <see cref="ClaimType"/>
+    /// with the appropriate claim type.
+    /// </para>
+    /// <para>
+    /// <strong>Security warning:</strong> The configured claim must uniquely identify the principal
+    /// within the served population. Display names (<see cref="ClaimsIdentity.DefaultNameClaimType"/>
+    /// / <see cref="ClaimTypes.Name"/>), usernames, email aliases, and other mutable or non-unique
+    /// claims are <strong>unsafe</strong> isolation keys unless the host can prove their uniqueness
+    /// across all callers. Two distinct principals that share the same value for a non-unique claim
+    /// would receive the same session-isolation key and could read or overwrite one another's
+    /// persisted sessions. Only override this value with a claim that is guaranteed unique and stable.
    /// </para>
    /// <para>
    /// Common alternatives include:
    /// <list type="bullet">
-    /// <item><description><c>ClaimTypes.NameIdentifier</c> — Stable user identifier</description></item>
-    /// <item><description><c>ClaimTypes.Email</c> — Email address</description></item>
-    /// <item><description>Custom claim types specific to your authentication provider</description></item>
+    /// <item><description>A composite of tenant and subject identifiers — required for multi-tenant hosts where the subject is only unique per tenant</description></item>
+    /// <item><description>Custom claim types specific to your authentication provider, provided they are unique and stable</description></item>
    /// </list>
    /// </para>
    /// </remarks>
-    public string ClaimType { get; set; } = ClaimsIdentity.DefaultNameClaimType;
+    public string ClaimType { get; set; } = ClaimTypes.NameIdentifier;
 }
@@ -1,6 +1,7 @@
 // Copyright (c) Microsoft. All rights reserved.

 using System;
+using System.Security.Claims;
 using Microsoft.AspNetCore.Http;
 using Microsoft.Extensions.DependencyInjection;

@@ -19,8 +20,28 @@ public static class ServiceCollectionExtensions
    /// <param name="options"> Optional configuration for the claims-based session isolation key provider.</param>
    /// <returns>The <see cref="IServiceCollection"/> so that additional calls can be chained.</returns>
    /// <remarks>
+    /// <para>
    /// This method requires <see cref="IHttpContextAccessor"/> to be registered in the service collection.
    /// Ensure that <c>services.AddHttpContextAccessor()</c> has been called before using this method.
+    /// </para>
+    /// <para>
+    /// When <paramref name="options"/> is not supplied, the isolation key is derived from the
+    /// <see cref="ClaimTypes.NameIdentifier"/> claim, a stable unique subject identifier. For OpenID
+    /// Connect tokens (including Microsoft Entra ID), this is typically mapped from the <c>sub</c> claim
+    /// by the default JWT inbound claim mapping. Authentication schemes that do not project a unique
+    /// identifier onto <see cref="ClaimTypes.NameIdentifier"/> (or hosts that require a different claim
+    /// such as Entra's <c>oid</c>) should override
+    /// <see cref="ClaimsIdentitySessionIsolationKeyProviderOptions.ClaimType"/>; otherwise the key may be
+    /// absent, which causes strict-mode session stores to fail.
+    /// </para>
+    /// <para>
+    /// <strong>Security warning:</strong> If you override
+    /// <see cref="ClaimsIdentitySessionIsolationKeyProviderOptions.ClaimType"/>, the chosen claim must
+    /// uniquely identify the principal within the served population. Display names, usernames, email
+    /// aliases, and other mutable or non-unique claims are <strong>unsafe</strong> isolation keys unless
+    /// the host can prove their uniqueness across all callers, because distinct principals that share the
+    /// same claim value would receive the same isolation key and could access one another's sessions.
+    /// </para>
    /// </remarks>
    public static IServiceCollection UseClaimsBasedSessionIsolation(
        this IServiceCollection services,
@@ -49,7 +49,7 @@ internal sealed class ForeachExecutor : DeclarativeActionExecutor<Foreach>
        EvaluationResult<DataValue> expressionResult = this.Evaluator.GetValue(this.Model.Items);
        if (expressionResult.Value is TableDataValue tableValue)
        {
-            this._values = [.. tableValue.Values.Select(value => value.Properties.Values.First().ToFormula())];
+            this._values = [.. tableValue.Values.Select(ToLoopValue)];
        }
        else
        {
@@ -99,6 +99,15 @@ internal sealed class ForeachExecutor : DeclarativeActionExecutor<Foreach>
        }
    }

+    // Power Fx wraps scalar array literals (`=[1, 2, 3]`) as `Table({Value: 1}, ...)`. Unwrap that single-column
+    // `Value`-record shape so `Local.LoopValue` is the scalar; multi-field and other shapes pass through unchanged.
+    private static FormulaValue ToLoopValue(DataValue value) =>
+        value is RecordDataValue record
+            && record.Properties.Count == 1
+            && record.Properties.TryGetValue("Value", out DataValue? singleColumn)
+                ? singleColumn.ToFormula()
+                : value.ToFormula();
+
    /// <inheritdoc/>
    /// <remarks>
    /// Persists the iteration cursor (<see cref="_index"/>), the materialized item snapshot
@@ -5,4 +5,5 @@ namespace Microsoft.Agents.AI.Workflows.Specialized.Magentic;
 internal static class MagenticConstants
 {
    public const string MagenticTaskContextKey = nameof(MagenticTaskContextKey);
+    public const string CurrentSpeakerStateKey = nameof(CurrentSpeakerStateKey);
 }
@@ -90,6 +90,7 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta

    private MagenticTaskContext? _taskContext;
    private PortBinding? _planReviewPort;
+    private string? _currentSpeakerExecutorId;

    protected override ProtocolBuilder ConfigureProtocol(ProtocolBuilder protocolBuilder)
    {
@@ -196,15 +197,46 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta
        else
        {
            // Subsequent turns: agent returned control, go directly to coordination (progress ledger only, no replan).
-            // Capture the participant's reply into the manager-visible chat history so the progress ledger can see it.
            if (messages is { Count: > 0 })
            {
+                // Capture the participant's reply into the manager-visible chat history so the progress ledger can see it.
                this._taskContext.ChatHistory.AddRange(messages);
+
+                // Share the reply with the other participants except the replier
+                await this.BroadcastReplyToOtherParticipantsAsync(messages, context, cancellationToken).ConfigureAwait(false);
            }
            await this.RunCoordinationRoundAsync(this._taskContext, context, cancellationToken).ConfigureAwait(false);
        }
    }

+    /// <summary>
+    /// Forwards a participant's reply to every other participant so they share the running conversation.
+    /// The messages are buffered (no <see cref="TurnToken"/> is sent) - they only become context for the participant's next turn.
+    /// </summary>
+    private ValueTask BroadcastReplyToOtherParticipantsAsync(
+        List<ChatMessage> messages, IWorkflowContext context, CancellationToken cancellationToken)
+    {
+        // Without a known current speaker we cannot exclude the reply's author, so skip the broadcast
+        // rather than risk echoing the reply back to its own author. This covers the window after a
+        // checkpoint restore but before any delegation has set the current speaker.
+        if (string.IsNullOrEmpty(this._currentSpeakerExecutorId))
+        {
+            return default;
+        }
+
+        List<Task>? sendTasks = null;
+        foreach (AIAgent agent in team)
+        {
+            string executorId = AIAgentHostExecutor.IdFor(agent);
+            if (string.Equals(executorId, this._currentSpeakerExecutorId, StringComparison.Ordinal))
+            {
+                continue;
+            }
+            (sendTasks ??= []).Add(context.SendMessageAsync(messages, executorId, cancellationToken).AsTask());
+        }
+        return sendTasks is null ? default : new ValueTask(Task.WhenAll(sendTasks));
+    }
+
    private ChatMessage? _fullTaskLedgerMessage;
    private ValueTask DelegateToTeamAsync(MagenticTaskContext taskContext, IWorkflowContext context, CancellationToken cancellationToken)
    {
@@ -287,15 +319,18 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta
            return;
        }

+        string nextExecutorId = AIAgentHostExecutor.IdFor(nextAgent);
+
        if (!string.IsNullOrWhiteSpace(taskContext.ProgressLedger.InstructionOrQuestion))
        {
            ChatMessage instruction = new(ChatRole.Assistant, taskContext.ProgressLedger.InstructionOrQuestion);
            taskContext.ChatHistory.Add(instruction);

-            await context.SendMessageAsync(instruction, cancellationToken).ConfigureAwait(false);
+            // Target the instruction at the chosen speaker only.
+            await context.SendMessageAsync(instruction, nextExecutorId, cancellationToken).ConfigureAwait(false);
        }

-        string nextExecutorId = AIAgentHostExecutor.IdFor(nextAgent);
+        this._currentSpeakerExecutorId = nextExecutorId;
        await context.SendMessageAsync(new TurnToken(taskContext.EmitUpdateEvents), nextExecutorId, cancellationToken).ConfigureAwait(false);
    }

@@ -303,6 +338,7 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta
    {
        bool wasStalled = taskContext.IsStalled;
        taskContext.Reset();
+        this._currentSpeakerExecutorId = null;
        await context.SendMessageAsync(new ResetChatSignal(), cancellationToken: cancellationToken).ConfigureAwait(false);

        await this.UpdatePlanAndDelegateAsync(taskContext, context, cancellationToken, replanAfterStall: wasStalled).ConfigureAwait(false);
@@ -313,9 +349,9 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta
        List<ChatMessage> messages = [await this._manager.PrepareFinalAnswerAsync(taskContext, context, cancellationToken).ConfigureAwait(false)];
        await context.YieldOutputAsync(messages, cancellationToken).ConfigureAwait(false);
        taskContext.IsTerminated = true;
+        this._currentSpeakerExecutorId = null;
    }

-    private const string CurrentTurnEmitUpdateEventsKey = nameof(CurrentTurnEmitUpdateEventsKey);
    protected internal override async ValueTask OnCheckpointingAsync(IWorkflowContext context, CancellationToken cancellationToken = default)
    {
        Task contextStateTask = this._taskContext == null
@@ -325,14 +361,21 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta
                                                              cancellationToken: cancellationToken)
                                       .AsTask();

+        Task currentSpeakerTask = context.QueueStateUpdateAsync(MagenticConstants.CurrentSpeakerStateKey,
+                                                            this._currentSpeakerExecutorId,
+                                                            cancellationToken: cancellationToken)
+                                        .AsTask();
+
        await Task.WhenAll(base.OnCheckpointingAsync(context, cancellationToken).AsTask(),
-                           contextStateTask).ConfigureAwait(false);
+                           contextStateTask,
+                           currentSpeakerTask).ConfigureAwait(false);
    }

    protected internal override async ValueTask OnCheckpointRestoredAsync(IWorkflowContext context, CancellationToken cancellationToken = default)
    {
-        await Task.WhenAll(base.OnCheckpointRestoredAsync(context, cancellationToken).AsTask(), LoadContextStateAsync())
-                  .ConfigureAwait(false);
+        await Task.WhenAll(base.OnCheckpointRestoredAsync(context, cancellationToken).AsTask(),
+                            LoadContextStateAsync(),
+                            LoadCurrentSpeakerAsync()).ConfigureAwait(false);

        async Task LoadContextStateAsync()
        {
@@ -344,5 +387,11 @@ internal class MagenticOrchestrator(AIAgent managerAgent, List<AIAgent> team, Ta
                this._taskContext = new MagenticTaskContext(state, team, limits, []);
            }
        }
+
+        async Task LoadCurrentSpeakerAsync()
+        {
+            this._currentSpeakerExecutorId = await context.ReadStateAsync<string?>(MagenticConstants.CurrentSpeakerStateKey, cancellationToken: cancellationToken)
+                                                        .ConfigureAwait(false);
+        }
    }
 }
@@ -38,6 +38,8 @@ namespace Microsoft.Agents.AI;
 /// </remarks>
 public sealed partial class ChatClientAgent : AIAgent
 {
+    private const string AGUIProviderName = "ag-ui";
+
    private readonly ChatClientAgentOptions? _agentOptions;
    private readonly HashSet<string> _aiContextProviderStateKeys;
    private readonly AIAgentMetadata _agentMetadata;
@@ -562,6 +564,7 @@ public sealed partial class ChatClientAgent : AIAgent
        requestChatOptions.ModelId ??= this._agentOptions.ChatOptions.ModelId;
        requestChatOptions.PresencePenalty ??= this._agentOptions.ChatOptions.PresencePenalty;
        requestChatOptions.ResponseFormat ??= this._agentOptions.ChatOptions.ResponseFormat;
+        requestChatOptions.Reasoning ??= this._agentOptions.ChatOptions.Reasoning;
        requestChatOptions.Seed ??= this._agentOptions.ChatOptions.Seed;
        requestChatOptions.Temperature ??= this._agentOptions.ChatOptions.Temperature;
        requestChatOptions.TopP ??= this._agentOptions.ChatOptions.TopP;
@@ -815,7 +818,7 @@ public sealed partial class ChatClientAgent : AIAgent

        if (!string.IsNullOrWhiteSpace(responseConversationId))
        {
-            if (this._agentOptions?.ChatHistoryProvider is not null)
+            if (!IsAGUIProviderName(this._agentMetadata.ProviderName) && this._agentOptions?.ChatHistoryProvider is not null)
            {
                // The agent has a ChatHistoryProvider configured, but the service returned a conversation id,
                // meaning the service manages chat history server-side. Both cannot be used simultaneously.
@@ -929,6 +932,9 @@ public sealed partial class ChatClientAgent : AIAgent
        }
    }

+    private static bool IsAGUIProviderName(string? providerName) =>
+        string.Equals(providerName, AGUIProviderName, StringComparison.Ordinal);
+
    /// <summary>
    /// Ensures that <see cref="AIAgent.CurrentRunContext"/> contains the resolved session.
    /// </summary>
@@ -976,12 +982,17 @@ public sealed partial class ChatClientAgent : AIAgent

    private ChatHistoryProvider? ResolveChatHistoryProvider(ChatOptions? chatOptions)
    {
-        ChatHistoryProvider? provider = chatOptions?.ConversationId is null ? this.ChatHistoryProvider : null;
+        ChatHistoryProvider? provider =
+            chatOptions?.ConversationId is null || IsAGUIProviderName(this._agentMetadata.ProviderName)
+            ? this.ChatHistoryProvider
+            : null;

        // If someone provided an override ChatHistoryProvider via AdditionalProperties, we should use that instead.
        if (chatOptions?.AdditionalProperties?.TryGetValue(out ChatHistoryProvider? overrideProvider) is true)
        {
-            if (this._agentOptions?.ThrowOnChatHistoryProviderConflict is true && string.IsNullOrWhiteSpace(chatOptions?.ConversationId) is false)
+            if (!IsAGUIProviderName(this._agentMetadata.ProviderName) &&
+                this._agentOptions?.ThrowOnChatHistoryProviderConflict is true &&
+                string.IsNullOrWhiteSpace(chatOptions?.ConversationId) is false)
            {
                throw new InvalidOperationException(
                    $"Only {nameof(ChatClientAgentSession.ConversationId)} or {nameof(this.ChatHistoryProvider)} may be used, but not both. The current {nameof(ChatClientAgentSession)} has a {nameof(ChatClientAgentSession.ConversationId)} indicating server-side chat history management, but an override {nameof(this.ChatHistoryProvider)} was provided via {nameof(AgentRunOptions.AdditionalProperties)}.");
@@ -181,6 +181,36 @@ public sealed class ChatClientAgentOptions
    [Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
    public bool EnableMessageInjection { get; set; }

+    /// <summary>
+    /// Gets or sets a value indicating whether to store automatically approved function calls in the session state
+    /// for tools that do not require approval when they are returned alongside tools that do.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <see cref="FunctionInvokingChatClient"/> has an all-or-nothing behavior for approvals: when any tool
+    /// in a response is an <see cref="ApprovalRequiredAIFunction"/>, it converts all <see cref="FunctionCallContent"/>
+    /// items to <see cref="ToolApprovalRequestContent"/>, even for tools that do not require approval.
+    /// </para>
+    /// <para>
+    /// Setting this property to <see langword="true"/> injects an <see cref="NonApprovalRequiredFunctionBypassingChatClient"/>
+    /// decorator above <see cref="FunctionInvokingChatClient"/> in the pipeline. This decorator identifies approval
+    /// requests for non-approval-required tools, removes them from the response, and stores them in the session.
+    /// On the next request, the stored items are automatically re-injected as approved, so the caller only needs
+    /// to handle approval requests for tools that truly require human approval.
+    /// </para>
+    /// <para>
+    /// This option has no effect when <see cref="UseProvidedChatClientAsIs"/> is <see langword="true"/>.
+    /// When using a custom chat client stack, you can add an <see cref="NonApprovalRequiredFunctionBypassingChatClient"/>
+    /// manually via the <see cref="ChatClientBuilderExtensions.UseNonApprovalRequiredFunctionBypassing"/>
+    /// extension method.
+    /// </para>
+    /// </remarks>
+    /// <value>
+    /// Default is <see langword="false"/>.
+    /// </value>
+    [Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
+    public bool EnableNonApprovalRequiredFunctionBypassing { get; set; }
+
    /// <summary>
    /// Creates a new instance of <see cref="ChatClientAgentOptions"/> with the same values as this instance.
    /// </summary>
@@ -199,5 +229,6 @@ public sealed class ChatClientAgentOptions
            ThrowOnChatHistoryProviderConflict = this.ThrowOnChatHistoryProviderConflict,
            RequirePerServiceCallChatHistoryPersistence = this.RequirePerServiceCallChatHistoryPersistence,
            EnableMessageInjection = this.EnableMessageInjection,
+            EnableNonApprovalRequiredFunctionBypassing = this.EnableNonApprovalRequiredFunctionBypassing,
        };
 }
@@ -148,4 +148,35 @@ public static class ChatClientBuilderExtensions
    {
        return builder.Use(innerClient => new MessageInjectingChatClient(innerClient));
    }
+
+    /// <summary>
+    /// Adds an <see cref="NonApprovalRequiredFunctionBypassingChatClient"/> to the chat client pipeline.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// This decorator should be positioned above the <see cref="FunctionInvokingChatClient"/> in the pipeline
+    /// so that it can intercept approval requests for tools that do not require approval. When
+    /// <see cref="FunctionInvokingChatClient"/> converts all function calls to approval requests (because at
+    /// least one tool requires approval), this decorator removes the requests for non-approval-required tools,
+    /// stores them in the session, and automatically re-injects them as approved on the next request.
+    /// </para>
+    /// <para>
+    /// This extension method is intended for use with custom chat client stacks when
+    /// <see cref="ChatClientAgentOptions.UseProvidedChatClientAsIs"/> is <see langword="true"/>.
+    /// When <see cref="ChatClientAgentOptions.UseProvidedChatClientAsIs"/> is <see langword="false"/> (the default),
+    /// the <see cref="ChatClientAgent"/> automatically injects this decorator when
+    /// <see cref="ChatClientAgentOptions.EnableNonApprovalRequiredFunctionBypassing"/> is <see langword="true"/>.
+    /// </para>
+    /// <para>
+    /// This decorator only works within the context of a running <see cref="ChatClientAgent"/> with
+    /// an active session, and will throw an exception if used in any other stack.
+    /// </para>
+    /// </remarks>
+    /// <param name="builder">The <see cref="ChatClientBuilder"/> to add the decorator to.</param>
+    /// <returns>The <paramref name="builder"/> for chaining.</returns>
+    [Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
+    public static ChatClientBuilder UseNonApprovalRequiredFunctionBypassing(this ChatClientBuilder builder)
+    {
+        return builder.Use(innerClient => new NonApprovalRequiredFunctionBypassingChatClient(innerClient));
+    }
 }
@@ -53,6 +53,17 @@ public static class ChatClientExtensions
    {
        var chatBuilder = chatClient.AsBuilder();

+        // NonApprovalRequiredFunctionBypassingChatClient is registered before FunctionInvokingChatClient so that
+        // it sits above FICC in the pipeline. ChatClientBuilder.Build applies factories in reverse order,
+        // making the first Use() call outermost. By adding this decorator first, the resulting pipeline is:
+        //   NonApprovalRequiredFunctionBypassingChatClient → FunctionInvokingChatClient → ChatHistoryPersistingChatClient → leaf IChatClient
+        // This allows the decorator to intercept FICC's responses and remove approval requests for tools
+        // that don't actually require approval, storing them for automatic re-injection on the next request.
+        if (options?.EnableNonApprovalRequiredFunctionBypassing is true)
+        {
+            chatBuilder.Use(innerClient => new NonApprovalRequiredFunctionBypassingChatClient(innerClient));
+        }
+
        if (chatClient.GetService<FunctionInvokingChatClient>() is null)
        {
            chatBuilder.Use((innerClient, services) =>
@@ -0,0 +1,285 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Runtime.CompilerServices;
+using System.Threading;
+using System.Threading.Tasks;
+using Microsoft.Extensions.AI;
+
+namespace Microsoft.Agents.AI;
+
+/// <summary>
+/// A delegating chat client that automatically removes <see cref="ToolApprovalRequestContent"/> for tools
+/// that do not actually require approval, storing auto-approved results in the session for transparent
+/// re-injection on the next request.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <see cref="FunctionInvokingChatClient"/> has an all-or-nothing behavior for approvals: when any tool
+/// in a response is an <see cref="ApprovalRequiredAIFunction"/>, it converts all <see cref="FunctionCallContent"/>
+/// items to <see cref="ToolApprovalRequestContent"/> — even for tools that do not require approval. This
+/// decorator sits above <see cref="FunctionInvokingChatClient"/> in the pipeline and transparently handles
+/// the non-approval-required items so callers only see approval requests for tools that truly need them.
+/// </para>
+/// <para>
+/// On outbound responses, the decorator identifies <see cref="ToolApprovalRequestContent"/> items for tools
+/// that are not wrapped in <see cref="ApprovalRequiredAIFunction"/>, removes them from the response, and
+/// stores them in the session's <see cref="AgentSessionStateBag"/>. On the next inbound request, the stored
+/// items are re-injected as pre-approved <see cref="ToolApprovalResponseContent"/> so that
+/// <see cref="FunctionInvokingChatClient"/> can process them alongside the caller's human-approved responses.
+/// </para>
+/// <para>
+/// This decorator requires an active <see cref="AIAgent.CurrentRunContext"/> with a non-null
+/// <see cref="AgentRunContext.Session"/>. An <see cref="InvalidOperationException"/> is thrown if no
+/// run context or session is available.
+/// </para>
+/// </remarks>
+internal sealed class NonApprovalRequiredFunctionBypassingChatClient : DelegatingChatClient
+{
+    /// <summary>
+    /// The key used in <see cref="AgentSessionStateBag"/> to store pending auto-approved function calls
+    /// between agent runs.
+    /// </summary>
+    internal const string StateBagKey = "_autoApprovedFunctionCalls";
+
+    /// <summary>
+    /// Initializes a new instance of the <see cref="NonApprovalRequiredFunctionBypassingChatClient"/> class.
+    /// </summary>
+    /// <param name="innerClient">The underlying chat client (typically a <see cref="FunctionInvokingChatClient"/>).</param>
+    public NonApprovalRequiredFunctionBypassingChatClient(IChatClient innerClient)
+        : base(innerClient)
+    {
+    }
+
+    /// <inheritdoc/>
+    public override async Task<ChatResponse> GetResponseAsync(
+        IEnumerable<ChatMessage> messages,
+        ChatOptions? options = null,
+        CancellationToken cancellationToken = default)
+    {
+        var session = GetRequiredSession();
+        var autoApprovableNames = this.GetAutoApprovableToolNames(options);
+
+        messages = InjectPendingAutoApprovals(messages, session);
+
+        var response = await base.GetResponseAsync(messages, options, cancellationToken).ConfigureAwait(false);
+
+        RemoveAutoApprovedFromMessages(response.Messages, autoApprovableNames, session);
+
+        return response;
+    }
+
+    /// <inheritdoc/>
+    public override async IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
+        IEnumerable<ChatMessage> messages,
+        ChatOptions? options = null,
+        [EnumeratorCancellation] CancellationToken cancellationToken = default)
+    {
+        var session = GetRequiredSession();
+        var autoApprovableNames = this.GetAutoApprovableToolNames(options);
+
+        messages = InjectPendingAutoApprovals(messages, session);
+        List<ToolApprovalRequestContent>? autoApproved = null;
+
+        try
+        {
+            await foreach (var update in base.GetStreamingResponseAsync(messages, options, cancellationToken).ConfigureAwait(false))
+            {
+                if (FilterUpdateContents(update, autoApprovableNames, ref autoApproved))
+                {
+                    yield return update;
+                }
+            }
+        }
+        finally
+        {
+            if (autoApproved is { Count: > 0 })
+            {
+                session.StateBag.SetValue(StateBagKey, autoApproved, AgentJsonUtilities.DefaultOptions);
+            }
+        }
+    }
+
+    /// <summary>
+    /// Gets the current <see cref="AgentSession"/> from the ambient run context.
+    /// </summary>
+    /// <exception cref="InvalidOperationException">No run context or session is available.</exception>
+    private static AgentSession GetRequiredSession()
+    {
+        var runContext = AIAgent.CurrentRunContext
+            ?? throw new InvalidOperationException(
+                $"{nameof(NonApprovalRequiredFunctionBypassingChatClient)} can only be used within the context of a running AIAgent. " +
+                "Ensure that the chat client is being invoked as part of an AIAgent.RunAsync or AIAgent.RunStreamingAsync call.");
+
+        return runContext.Session
+            ?? throw new InvalidOperationException(
+                $"{nameof(NonApprovalRequiredFunctionBypassingChatClient)} requires a session. " +
+                "Ensure the agent has a resolved session before invoking the chat client.");
+    }
+
+    /// <summary>
+    /// Checks the session for stored auto-approvals from a previous turn and injects them as
+    /// a user message containing <see cref="ToolApprovalResponseContent"/> items appended to the input messages.
+    /// </summary>
+    /// <remarks>
+    /// All stored requests are unconditionally injected as approved responses regardless of whether the
+    /// tool set has changed, because the LLM requires a complete set of tool call responses for a prior turn.
+    /// </remarks>
+    private static IEnumerable<ChatMessage> InjectPendingAutoApprovals(
+        IEnumerable<ChatMessage> messages,
+        AgentSession session)
+    {
+        if (!session.StateBag.TryGetValue<List<ToolApprovalRequestContent>>(
+            StateBagKey,
+            out var pendingRequests,
+            AgentJsonUtilities.DefaultOptions)
+            || pendingRequests is not { Count: > 0 })
+        {
+            return messages;
+        }
+
+        session.StateBag.TryRemoveValue(StateBagKey);
+
+        List<AIContent> approvalResponses = [];
+        foreach (var request in pendingRequests)
+        {
+            approvalResponses.Add(request.CreateResponse(approved: true));
+        }
+
+        var userMessage = new ChatMessage(ChatRole.User, approvalResponses);
+        return messages.Concat([userMessage]);
+    }
+
+    /// <summary>
+    /// Builds a set of tool names that do not require approval and can be auto-approved,
+    /// by checking all available tools from <see cref="ChatOptions.Tools"/> and
+    /// <see cref="FunctionInvokingChatClient.AdditionalTools"/>.
+    /// </summary>
+    private HashSet<string> GetAutoApprovableToolNames(ChatOptions? options)
+    {
+        var ficc = this.GetService<FunctionInvokingChatClient>();
+
+        var allTools = (options?.Tools ?? Enumerable.Empty<AITool>())
+            .Concat(ficc?.AdditionalTools ?? Enumerable.Empty<AITool>());
+
+        return new HashSet<string>(
+            allTools
+                .OfType<AIFunction>()
+                .Where(static f => f.GetService<ApprovalRequiredAIFunction>() is null)
+                .Select(static f => f.Name),
+            StringComparer.Ordinal);
+    }
+
+    /// <summary>
+    /// Determines whether a <see cref="ToolApprovalRequestContent"/> can be auto-approved because
+    /// the underlying tool is not an <see cref="ApprovalRequiredAIFunction"/>.
+    /// </summary>
+    /// <returns>
+    /// <see langword="true"/> if the approval request is for a known tool that does not require approval
+    /// and can be auto-approved; <see langword="false"/> otherwise.
+    /// </returns>
+    private static bool IsAutoApprovable(ToolApprovalRequestContent approval, HashSet<string> autoApprovableNames)
+    {
+        if (approval.ToolCall is not FunctionCallContent fcc)
+        {
+            // Non-function tool calls cannot be auto-approved.
+            return false;
+        }
+
+        // Auto-approve only if the tool is known and explicitly does NOT require approval.
+        // Unknown tools are not in the set and are treated as approval-required (safe default).
+        return autoApprovableNames.Contains(fcc.Name);
+    }
+
+    /// <summary>
+    /// Scans response messages for auto-approvable <see cref="ToolApprovalRequestContent"/> items,
+    /// removes them from the messages, and stores them in the session for the next request.
+    /// </summary>
+    private static void RemoveAutoApprovedFromMessages(
+        IList<ChatMessage> messages,
+        HashSet<string> autoApprovableNames,
+        AgentSession session)
+    {
+        List<ToolApprovalRequestContent>? autoApproved = null;
+
+        foreach (var message in messages)
+        {
+            for (int i = message.Contents.Count - 1; i >= 0; i--)
+            {
+                if (message.Contents[i] is ToolApprovalRequestContent approval
+                    && IsAutoApprovable(approval, autoApprovableNames))
+                {
+                    (autoApproved ??= []).Add(approval);
+                    message.Contents.RemoveAt(i);
+                }
+            }
+        }
+
+        // Remove messages that are now empty after filtering.
+        for (int i = messages.Count - 1; i >= 0; i--)
+        {
+            if (messages[i].Contents.Count == 0)
+            {
+                messages.RemoveAt(i);
+            }
+        }
+
+        if (autoApproved is { Count: > 0 })
+        {
+            session.StateBag.SetValue(StateBagKey, autoApproved, AgentJsonUtilities.DefaultOptions);
+        }
+    }
+
+    /// <summary>
+    /// Filters auto-approvable <see cref="ToolApprovalRequestContent"/> items from a streaming update's
+    /// contents, collecting them for later storage.
+    /// </summary>
+    /// <returns>
+    /// <see langword="true"/> if the update should be yielded (has remaining content or had no
+    /// approval content to begin with); <see langword="false"/> if the update is now empty and
+    /// should be skipped.
+    /// </returns>
+    private static bool FilterUpdateContents(
+        ChatResponseUpdate update,
+        HashSet<string> autoApprovableNames,
+        ref List<ToolApprovalRequestContent>? autoApproved)
+    {
+        bool hasApprovalContent = false;
+        List<AIContent> filteredContents = [];
+        bool removedAny = false;
+
+        for (int i = 0; i < update.Contents.Count; i++)
+        {
+            var content = update.Contents[i];
+
+            if (content is ToolApprovalRequestContent approval)
+            {
+                hasApprovalContent = true;
+
+                if (IsAutoApprovable(approval, autoApprovableNames))
+                {
+                    (autoApproved ??= []).Add(approval);
+                    removedAny = true;
+                }
+                else
+                {
+                    filteredContents.Add(content);
+                }
+            }
+            else
+            {
+                filteredContents.Add(content);
+            }
+        }
+
+        if (removedAny)
+        {
+            update.Contents = filteredContents;
+        }
+
+        // Yield the update unless it was purely auto-approvable approval content (now empty).
+        return update.Contents.Count > 0 || !hasApprovalContent;
+    }
+}
@@ -51,20 +51,22 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
 {
    private readonly ProviderSessionState<ToolApprovalState> _sessionState;
    private readonly JsonSerializerOptions _jsonSerializerOptions;
+    private readonly Func<FunctionCallContent, ValueTask<bool>>[]? _autoApprovalRules;

    /// <summary>
    /// Initializes a new instance of the <see cref="ToolApprovalAgent"/> class.
    /// </summary>
    /// <param name="innerAgent">The underlying agent to delegate to.</param>
-    /// <param name="jsonSerializerOptions">
-    /// Optional <see cref="JsonSerializerOptions"/> used for serializing argument values when storing rules
-    /// and for persisting state. When <see langword="null"/>, <see cref="AgentJsonUtilities.DefaultOptions"/> is used.
+    /// <param name="options">
+    /// Optional <see cref="ToolApprovalAgentOptions"/> for configuring serialization and auto-approval rules.
+    /// When <see langword="null"/>, default settings are used.
    /// </param>
    /// <exception cref="ArgumentNullException"><paramref name="innerAgent"/> is <see langword="null"/>.</exception>
-    public ToolApprovalAgent(AIAgent innerAgent, JsonSerializerOptions? jsonSerializerOptions = null)
+    public ToolApprovalAgent(AIAgent innerAgent, ToolApprovalAgentOptions? options = null)
        : base(innerAgent)
    {
-        this._jsonSerializerOptions = jsonSerializerOptions ?? AgentJsonUtilities.DefaultOptions;
+        this._jsonSerializerOptions = options?.JsonSerializerOptions ?? AgentJsonUtilities.DefaultOptions;
+        this._autoApprovalRules = options?.AutoApprovalRules?.ToArray();
        this._sessionState = new ProviderSessionState<ToolApprovalState>(
            _ => new ToolApprovalState(),
            "toolApprovalState",
@@ -79,7 +81,7 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
        CancellationToken cancellationToken = default)
    {
        // Steps 1–2: Unwrap AlwaysApprove wrappers, process any queued approval requests.
-        var (state, callerMessages, nextQueuedItem) = this.PrepareInboundMessages(messages, session);
+        var (state, callerMessages, nextQueuedItem) = await this.PrepareInboundMessagesAsync(messages, session).ConfigureAwait(false);

        if (nextQueuedItem is not null)
        {
@@ -98,7 +100,7 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
            var response = await this.InnerAgent.RunAsync(processedMessages, session, options, cancellationToken).ConfigureAwait(false);

            // Classify approval requests: auto-approve matching, queue excess, keep first unapproved.
-            bool allAutoApproved = this.ProcessAndQueueOutboundApprovalRequests(response.Messages, state, session);
+            bool allAutoApproved = await this.ProcessAndQueueOutboundApprovalRequestsAsync(response.Messages, state, session).ConfigureAwait(false);

            if (!allAutoApproved)
            {
@@ -119,7 +121,7 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
        [EnumeratorCancellation] CancellationToken cancellationToken = default)
    {
        // Steps 1–2: Unwrap AlwaysApprove wrappers, process any queued approval requests.
-        var (state, callerMessages, nextQueuedItem) = this.PrepareInboundMessages(messages, session);
+        var (state, callerMessages, nextQueuedItem) = await this.PrepareInboundMessagesAsync(messages, session).ConfigureAwait(false);

        if (nextQueuedItem is not null)
        {
@@ -197,7 +199,7 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
                yield break;
            }

-            // 4. Classify the collected approval requests against standing rules.
+            // 4. Classify the collected approval requests against standing rules and auto-approval rules.
            List<ToolApprovalRequestContent> unapproved = [];
            foreach (var tarc in streamedApprovalRequests)
            {
@@ -206,6 +208,11 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
                    state.CollectedApprovalResponses.Add(
                        tarc.CreateResponse(approved: true, reason: "Auto-approved by standing rule"));
                }
+                else if (await this.MatchesAutoApprovalRuleAsync(tarc).ConfigureAwait(false))
+                {
+                    state.CollectedApprovalResponses.Add(
+                        tarc.CreateResponse(approved: true, reason: "Auto-approved by auto-approval rule"));
+                }
                else
                {
                    unapproved.Add(tarc);
@@ -291,9 +298,9 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
    }

    /// <summary>
-    /// Re-evaluates queued approval requests against current rules and auto-approves any that now match.
+    /// Re-evaluates queued approval requests against current rules and auto-approval rules, and auto-approves any that now match.
    /// </summary>
-    private void DrainAutoApprovableFromQueue(ToolApprovalState state)
+    private async ValueTask DrainAutoApprovableFromQueueAsync(ToolApprovalState state)
    {
        for (int i = state.QueuedApprovalRequests.Count - 1; i >= 0; i--)
        {
@@ -303,6 +310,12 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
                    state.QueuedApprovalRequests[i].CreateResponse(approved: true, reason: "Auto-approved by standing rule"));
                state.QueuedApprovalRequests.RemoveAt(i);
            }
+            else if (await this.MatchesAutoApprovalRuleAsync(state.QueuedApprovalRequests[i]).ConfigureAwait(false))
+            {
+                state.CollectedApprovalResponses.Add(
+                    state.QueuedApprovalRequests[i].CreateResponse(approved: true, reason: "Auto-approved by auto-approval rule"));
+                state.QueuedApprovalRequests.RemoveAt(i);
+            }
        }
    }

@@ -318,8 +331,8 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
    /// A tuple of (state, processed caller messages, next queued item or <see langword="null"/> if the queue is resolved).
    /// When the returned item is non-null, the caller should return/yield it without calling the inner agent.
    /// </returns>
-    private (ToolApprovalState State, List<ChatMessage> CallerMessages, ToolApprovalRequestContent? NextQueuedItem)
-        PrepareInboundMessages(IEnumerable<ChatMessage> messages, AgentSession? session)
+    private async ValueTask<(ToolApprovalState State, List<ChatMessage> CallerMessages, ToolApprovalRequestContent? NextQueuedItem)>
+        PrepareInboundMessagesAsync(IEnumerable<ChatMessage> messages, AgentSession? session)
    {
        var state = this._sessionState.GetOrInitializeState(session);

@@ -337,7 +350,7 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent

            // Re-evaluate remaining queued items — the caller may have added new rules
            // (e.g., "always approve this tool") that resolve additional items.
-            this.DrainAutoApprovableFromQueue(state);
+            await this.DrainAutoApprovableFromQueueAsync(state).ConfigureAwait(false);

            if (state.QueuedApprovalRequests.Count > 0)
            {
@@ -386,15 +399,18 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
    /// <see langword="true"/> if all TARc items were auto-approved (caller should re-invoke the inner agent);
    /// <see langword="false"/> otherwise.
    /// </returns>
-    private bool ProcessAndQueueOutboundApprovalRequests(
+    private async ValueTask<bool> ProcessAndQueueOutboundApprovalRequestsAsync(
        IList<ChatMessage> responseMessages,
        ToolApprovalState state,
        AgentSession? session)
    {
-        // Pass 1: Scan all response messages and classify each approval request as
-        //         auto-approved (matches a standing rule) or unapproved (needs caller decision).
-        var autoApproved = new List<ToolApprovalRequestContent>();
+        // Pass 1: Scan all response messages and classify each approval request.
+        //         Auto-approved requests (matching a standing rule or auto-approval rule) have their
+        //         responses collected immediately, preserving the original request order, and are
+        //         marked for removal. Unapproved requests are collected for the caller to decide.
+        var toRemove = new HashSet<ToolApprovalRequestContent>();
        var unapproved = new List<ToolApprovalRequestContent>();
+        int autoApprovedCount = 0;

        foreach (var message in responseMessages)
        {
@@ -404,7 +420,17 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
                {
                    if (MatchesRule(tarc, state.Rules, this._jsonSerializerOptions))
                    {
-                        autoApproved.Add(tarc);
+                        state.CollectedApprovalResponses.Add(
+                            tarc.CreateResponse(approved: true, reason: "Auto-approved by standing rule"));
+                        toRemove.Add(tarc);
+                        autoApprovedCount++;
+                    }
+                    else if (await this.MatchesAutoApprovalRuleAsync(tarc).ConfigureAwait(false))
+                    {
+                        state.CollectedApprovalResponses.Add(
+                            tarc.CreateResponse(approved: true, reason: "Auto-approved by auto-approval rule"));
+                        toRemove.Add(tarc);
+                        autoApprovedCount++;
                    }
                    else
                    {
@@ -415,18 +441,12 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
        }

        // Nothing to process: no auto-approved items and at most one unapproved (no queueing needed).
-        if (autoApproved.Count == 0 && unapproved.Count <= 1)
+        // No responses were collected above in this case, so state is unmodified and safe to leave.
+        if (autoApprovedCount == 0 && unapproved.Count <= 1)
        {
            return false;
        }

-        // Store auto-approved responses for later injection into the inner agent.
-        foreach (var tarc in autoApproved)
-        {
-            state.CollectedApprovalResponses.Add(
-                tarc.CreateResponse(approved: true, reason: "Auto-approved by standing rule"));
-        }
-
        // If every approval request was auto-approved, strip them all and signal the caller
        // to re-invoke the inner agent immediately with the collected responses.
        if (unapproved.Count == 0)
@@ -439,14 +459,10 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
        // Pass 2: Keep only the first unapproved request in the response (for the caller to decide).
        //         Queue the remaining unapproved requests for subsequent one-at-a-time delivery.
        //         Remove all auto-approved and queued items from the response messages.
-        var toRemove = new HashSet<ToolApprovalRequestContent>(autoApproved);
-        if (unapproved.Count > 1)
+        for (int i = 1; i < unapproved.Count; i++)
        {
-            for (int i = 1; i < unapproved.Count; i++)
-            {
-                toRemove.Add(unapproved[i]);
-                state.QueuedApprovalRequests.Add(unapproved[i]);
-            }
+            toRemove.Add(unapproved[i]);
+            state.QueuedApprovalRequests.Add(unapproved[i]);
        }

        // Walk messages in reverse and strip marked items.
@@ -663,8 +679,36 @@ public sealed class ToolApprovalAgent : DelegatingAIAgent
    }

    /// <summary>
-    /// Compares stored rule arguments against actual function call arguments for an exact match.
+    /// Checks whether a <see cref="ToolApprovalRequestContent"/> is approved by any of the configured
+    /// auto-approval rules (heuristic functions).
    /// </summary>
+    /// <returns>
+    /// <see langword="true"/> if any auto-approval rule returns <see langword="true"/> for the function call;
+    /// <see langword="false"/> if no rules are configured, the request is not a function call, or no rule approves it.
+    /// </returns>
+    private async ValueTask<bool> MatchesAutoApprovalRuleAsync(ToolApprovalRequestContent request)
+    {
+        if (this._autoApprovalRules is not { Length: > 0 })
+        {
+            return false;
+        }
+
+        if (request.ToolCall is not FunctionCallContent functionCall)
+        {
+            return false;
+        }
+
+        foreach (var rule in this._autoApprovalRules)
+        {
+            if (await rule(functionCall).ConfigureAwait(false))
+            {
+                return true;
+            }
+        }
+
+        return false;
+    }
+
    private static bool ArgumentsMatch(IDictionary<string, string> ruleArguments, IDictionary<string, object?>? callArguments, JsonSerializerOptions jsonSerializerOptions)
    {
        if (callArguments is null)
@@ -1,7 +1,6 @@
 // Copyright (c) Microsoft. All rights reserved.

 using System.Diagnostics.CodeAnalysis;
-using System.Text.Json;
 using Microsoft.Shared.DiagnosticIds;
 using Microsoft.Shared.Diagnostics;

@@ -17,9 +16,9 @@ public static class ToolApprovalAgentBuilderExtensions
    /// Adds tool approval middleware to the agent pipeline, enabling "don't ask again" approval behavior.
    /// </summary>
    /// <param name="builder">The <see cref="AIAgentBuilder"/> to which tool approval support will be added.</param>
-    /// <param name="jsonSerializerOptions">
-    /// Optional <see cref="JsonSerializerOptions"/> used for serializing argument values when storing rules
-    /// and for persisting state. When <see langword="null"/>, <see cref="AgentJsonUtilities.DefaultOptions"/> is used.
+    /// <param name="options">
+    /// Optional <see cref="ToolApprovalAgentOptions"/> for configuring serialization and auto-approval rules.
+    /// When <see langword="null"/>, default settings are used.
    /// </param>
    /// <returns>The <see cref="AIAgentBuilder"/> with tool approval middleware added, enabling method chaining.</returns>
    /// <exception cref="System.ArgumentNullException"><paramref name="builder"/> is <see langword="null"/>.</exception>
@@ -32,6 +31,6 @@ public static class ToolApprovalAgentBuilderExtensions
    /// </remarks>
    public static AIAgentBuilder UseToolApproval(
        this AIAgentBuilder builder,
-        JsonSerializerOptions? jsonSerializerOptions = null)
-        => Throw.IfNull(builder).Use(innerAgent => new ToolApprovalAgent(innerAgent, jsonSerializerOptions));
+        ToolApprovalAgentOptions? options = null)
+        => Throw.IfNull(builder).Use(innerAgent => new ToolApprovalAgent(innerAgent, options));
 }
@@ -0,0 +1,45 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+using System;
+using System.Collections.Generic;
+using System.Diagnostics.CodeAnalysis;
+using System.Text.Json;
+using System.Threading.Tasks;
+using Microsoft.Extensions.AI;
+using Microsoft.Shared.DiagnosticIds;
+
+namespace Microsoft.Agents.AI;
+
+/// <summary>
+/// Options for configuring the <see cref="ToolApprovalAgent"/> middleware.
+/// </summary>
+[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
+public class ToolApprovalAgentOptions
+{
+    /// <summary>
+    /// Gets or sets the <see cref="System.Text.Json.JsonSerializerOptions"/> used for serializing argument values
+    /// when storing rules and for persisting state.
+    /// </summary>
+    /// <remarks>
+    /// When <see langword="null"/>, <see cref="AgentJsonUtilities.DefaultOptions"/> is used.
+    /// </remarks>
+    public JsonSerializerOptions? JsonSerializerOptions { get; set; }
+
+    /// <summary>
+    /// Gets or sets a collection of heuristic functions that can automatically approve function calls
+    /// that would otherwise require user approval.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Each function receives a <see cref="FunctionCallContent"/> representing the tool call that requires approval
+    /// and returns a <see cref="ValueTask{Boolean}"/> that resolves to <see langword="true"/> to auto-approve
+    /// the call, or <see langword="false"/> to continue evaluating the next rule.
+    /// </para>
+    /// <para>
+    /// Auto-approval rules are evaluated after standing rules (derived from prior user approvals) but before
+    /// prompting the user. Rules are evaluated in order; the first rule returning <see langword="true"/>
+    /// causes the function call to be auto-approved.
+    /// </para>
+    /// </remarks>
+    public IEnumerable<Func<FunctionCallContent, ValueTask<bool>>>? AutoApprovalRules { get; set; }
+}
@@ -243,6 +243,46 @@ public sealed class AGUIAgentTests
        Assert.Contains(updates, u => u.Text == "Hello");
    }

+    [Fact]
+    public async Task RunStreamingAsync_WithSession_SendsFullHistoryAfterThreadIdIsSetAsync()
+    {
+        // Arrange
+        var captureHandler = new StateCapturingTestDelegatingHandler();
+        captureHandler.AddResponse(
+        [
+            new RunStartedEvent { ThreadId = "thread1", RunId = "run1" },
+            new TextMessageStartEvent { MessageId = "msg1", Role = AGUIRoles.Assistant },
+            new TextMessageContentEvent { MessageId = "msg1", Delta = "First response" },
+            new TextMessageEndEvent { MessageId = "msg1" },
+            new RunFinishedEvent { ThreadId = "thread1", RunId = "run1" }
+        ]);
+        captureHandler.AddResponse(
+        [
+            new RunStartedEvent { ThreadId = "thread1", RunId = "run2" },
+            new TextMessageStartEvent { MessageId = "msg2", Role = AGUIRoles.Assistant },
+            new TextMessageContentEvent { MessageId = "msg2", Delta = "Second response" },
+            new TextMessageEndEvent { MessageId = "msg2" },
+            new RunFinishedEvent { ThreadId = "thread1", RunId = "run2" }
+        ]);
+        using HttpClient httpClient = new(captureHandler);
+
+        var chatClient = new AGUIChatClient(httpClient, "http://localhost/agent", null, AGUIJsonSerializerContext.Default.Options);
+        AIAgent agent = chatClient.AsAIAgent(instructions: null, name: "agent1", description: "Test agent", tools: []);
+        AgentSession session = await agent.CreateSessionAsync();
+
+        // Act
+        await foreach (var _ in agent.RunStreamingAsync([new ChatMessage(ChatRole.User, "First")], session))
+        {
+        }
+
+        await foreach (var _ in agent.RunStreamingAsync([new ChatMessage(ChatRole.User, "Second")], session))
+        {
+        }
+
+        // Assert
+        Assert.Equal([1, 3], captureHandler.CapturedMessageCounts);
+    }
+
    [Fact]
    public async Task DeserializeSession_WithValidState_ReturnsChatClientAgentSessionAsync()
    {
@@ -1686,10 +1726,12 @@ internal sealed class CapturingTestDelegatingHandler : DelegatingHandler
 internal sealed class StateCapturingTestDelegatingHandler : DelegatingHandler
 {
    private readonly Queue<Func<HttpRequestMessage, Task<HttpResponseMessage>>> _responseFactories = new();
+    private readonly List<int> _capturedMessageCounts = [];

    public bool RequestWasMade { get; private set; }
    public JsonElement? CapturedState { get; private set; }
    public int CapturedMessageCount { get; private set; }
+    public IReadOnlyList<int> CapturedMessageCounts => this._capturedMessageCounts;

    public void AddResponse(BaseEvent[] events)
    {
@@ -1714,6 +1756,7 @@ internal sealed class StateCapturingTestDelegatingHandler : DelegatingHandler
                this.CapturedState = input.State;
            }
            this.CapturedMessageCount = input.Messages.Count();
+            this._capturedMessageCounts.Add(this.CapturedMessageCount);
        }

        if (this._responseFactories.Count == 0)
@@ -182,7 +182,7 @@ public sealed class WorkflowConsoleAppSamplesValidation(ITestOutputHelper output
        }
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "KeyNotFoundException in workflow execution. See https://github.com/microsoft/agent-framework/issues/6404")]
    public async Task WorkflowEventsSampleValidationAsync()
    {
        using CancellationTokenSource testTimeoutCts = this.CreateTestTimeoutCts(s_testTimeout);
@@ -278,7 +278,7 @@ public sealed class WorkflowConsoleAppSamplesValidation(ITestOutputHelper output
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "KeyNotFoundException in workflow execution. See https://github.com/microsoft/agent-framework/issues/6404")]
    public async Task WorkflowSharedStateSampleValidationAsync()
    {
        using CancellationTokenSource testTimeoutCts = this.CreateTestTimeoutCts(s_testTimeout);
@@ -376,7 +376,7 @@ public sealed class WorkflowConsoleAppSamplesValidation(ITestOutputHelper output
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "KeyNotFoundException in workflow execution. See https://github.com/microsoft/agent-framework/issues/6404")]
    public async Task SubWorkflowsSampleValidationAsync()
    {
        using CancellationTokenSource testTimeoutCts = this.CreateTestTimeoutCts(s_testTimeout);
@@ -452,7 +452,7 @@ public sealed class WorkflowConsoleAppSamplesValidation(ITestOutputHelper output
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "KeyNotFoundException in workflow execution. See https://github.com/microsoft/agent-framework/issues/6404")]
    public async Task WorkflowHITLSampleValidationAsync()
    {
        using CancellationTokenSource testTimeoutCts = this.CreateTestTimeoutCts(s_testTimeout);
@@ -43,7 +43,7 @@ public class HostedFoundryMemoryProviderScopesTests
    }

    [Fact]
-    public void PerUserAndChat_ComposesUserAndChatWithColon()
+    public void PerUserAndChat_ComposesUserAndChatWithEscapedSeparator()
    {
        // Arrange
        var session = CreateTaggedSession(TestUserId, TestChatId);
@@ -54,7 +54,51 @@ public class HostedFoundryMemoryProviderScopesTests

        // Assert
        Assert.NotNull(state);
-        Assert.Equal($"{TestUserId}:{TestChatId}", state.Scope.Scope);
+        Assert.Equal($"{TestUserId}::{TestChatId}", state.Scope.Scope);
+    }
+
+    [Fact]
+    public void PerUserAndChat_EscapesColonsInUserAndChat()
+    {
+        // Arrange
+        var session = CreateTaggedSession("alice:finance", "q2:final");
+        var initializer = HostedFoundryMemoryProviderScopes.PerUserAndChat();
+
+        // Act
+        var state = initializer(session);
+
+        // Assert - colons inside each part are escaped as \: , parts joined with ::
+        Assert.Equal(@"alice\:finance::q2\:final", state.Scope.Scope);
+    }
+
+    [Fact]
+    public void PerUserAndChat_EscapesBackslashesInUserAndChat()
+    {
+        // Arrange
+        var session = CreateTaggedSession(@"alice\corp", @"chat\1");
+        var initializer = HostedFoundryMemoryProviderScopes.PerUserAndChat();
+
+        // Act
+        var state = initializer(session);
+
+        // Assert - backslashes escaped first as \\ , parts joined with ::
+        Assert.Equal(@"alice\\corp::chat\\1", state.Scope.Scope);
+    }
+
+    [Fact]
+    public void PerUserAndChat_DistinctContextsDoNotCollide()
+    {
+        // Arrange - two distinct (UserId, ChatId) pairs that collide under raw-colon composition.
+        var sessionA = CreateTaggedSession("alice:finance", "q2");
+        var sessionB = CreateTaggedSession("alice", "finance:q2");
+        var initializer = HostedFoundryMemoryProviderScopes.PerUserAndChat();
+
+        // Act
+        var scopeA = initializer(sessionA).Scope.Scope;
+        var scopeB = initializer(sessionB).Scope.Scope;
+
+        // Assert
+        Assert.NotEqual(scopeA, scopeB);
    }

    [Fact]
@@ -4,7 +4,8 @@ using System;
 using System.Collections.Generic;
 using System.Linq;
 using System.Threading.Tasks;
-using GitHub.Copilot.SDK;
+using GitHub.Copilot;
+using GitHub.Copilot.Rpc;
 using Microsoft.Extensions.AI;

 namespace Microsoft.Agents.AI.GitHub.Copilot.IntegrationTests;
@@ -13,8 +14,8 @@ public class GitHubCopilotAgentTests
 {
    private const string SkipReason = "Integration tests require GitHub Copilot CLI installed. For local execution only.";

-    private static Task<PermissionRequestResult> OnPermissionRequestAsync(PermissionRequest request, PermissionInvocation invocation)
-        => Task.FromResult(new PermissionRequestResult { Kind = PermissionRequestResultKind.Approved });
+    private static Task<PermissionDecision> OnPermissionRequestAsync(PermissionRequest request, PermissionInvocation invocation)
+        => Task.FromResult(PermissionDecision.ApproveOnce());

    [Fact(Skip = SkipReason)]
    public async Task RunAsync_WithSimplePrompt_ReturnsResponseAsync()
@@ -3,6 +3,7 @@
  <PropertyGroup>
    <!-- GitHub.Copilot.SDK only supports .NET 8.0+ -->
    <TargetFrameworks>$(TargetFrameworksCore)</TargetFrameworks>
+    <NoWarn>$(NoWarn);GHCP001</NoWarn>
  </PropertyGroup>

  <ItemGroup>
@@ -2,7 +2,7 @@

 using System;
 using System.Collections.Generic;
-using GitHub.Copilot.SDK;
+using GitHub.Copilot;
 using Microsoft.Extensions.AI;

 namespace Microsoft.Agents.AI.GitHub.Copilot.UnitTests;
@@ -16,7 +16,7 @@ public sealed class CopilotClientExtensionsTests
    public void AsAIAgent_WithAllParameters_ReturnsGitHubCopilotAgentWithSpecifiedProperties()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());

        const string TestId = "test-agent-id";
        const string TestName = "Test Agent";
@@ -37,7 +37,7 @@ public sealed class CopilotClientExtensionsTests
    public void AsAIAgent_WithMinimalParameters_ReturnsGitHubCopilotAgent()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());

        // Act
        var agent = copilotClient.AsAIAgent(ownsClient: false, tools: null);
@@ -61,7 +61,7 @@ public sealed class CopilotClientExtensionsTests
    public void AsAIAgent_WithOwnsClient_ReturnsAgentThatOwnsClient()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());

        // Act
        var agent = copilotClient.AsAIAgent(ownsClient: true, tools: null);
@@ -75,7 +75,7 @@ public sealed class CopilotClientExtensionsTests
    public void AsAIAgent_WithTools_ReturnsAgentWithTools()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());
        List<AITool> tools = [AIFunctionFactory.Create(() => "test", "TestFunc", "Test function")];

        // Act
@@ -3,7 +3,8 @@
 using System;
 using System.Collections.Generic;
 using System.Threading.Tasks;
-using GitHub.Copilot.SDK;
+using GitHub.Copilot;
+using GitHub.Copilot.Rpc;
 using Microsoft.Extensions.AI;

 namespace Microsoft.Agents.AI.GitHub.Copilot.UnitTests;
@@ -17,7 +18,7 @@ public sealed class GitHubCopilotAgentTests
    public void Constructor_WithCopilotClient_InitializesPropertiesCorrectly()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());
        const string TestId = "test-id";
        const string TestName = "test-name";
        const string TestDescription = "test-description";
@@ -42,7 +43,7 @@ public sealed class GitHubCopilotAgentTests
    public void Constructor_WithDefaultParameters_UsesBaseProperties()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());

        // Act
        var agent = new GitHubCopilotAgent(copilotClient, ownsClient: false, tools: null);
@@ -58,7 +59,7 @@ public sealed class GitHubCopilotAgentTests
    public async Task CreateSessionAsync_ReturnsGitHubCopilotAgentSessionAsync()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());
        var agent = new GitHubCopilotAgent(copilotClient, ownsClient: false, tools: null);

        // Act
@@ -73,7 +74,7 @@ public sealed class GitHubCopilotAgentTests
    public async Task CreateSessionAsync_WithSessionId_ReturnsSessionWithSessionIdAsync()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());
        var agent = new GitHubCopilotAgent(copilotClient, ownsClient: false, tools: null);
        const string TestSessionId = "test-session-id";

@@ -90,7 +91,7 @@ public sealed class GitHubCopilotAgentTests
    public void Constructor_WithTools_InitializesCorrectly()
    {
        // Arrange
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());
        List<AITool> tools = [AIFunctionFactory.Create(() => "test", "TestFunc", "Test function")];

        // Act
@@ -105,12 +106,12 @@ public sealed class GitHubCopilotAgentTests
    public void CopySessionConfig_CopiesAllProperties()
    {
        // Arrange
-        List<AIFunction> tools = [AIFunctionFactory.Create(() => "test", "TestFunc", "Test function")];
+        List<AIFunctionDeclaration> tools = [AIFunctionFactory.Create(() => "test", "TestFunc", "Test function")];
        var hooks = new SessionHooks();
        var infiniteSessions = new InfiniteSessionConfig();
        var systemMessage = new SystemMessageConfig { Mode = SystemMessageMode.Append, Content = "Be helpful" };
-        PermissionRequestHandler permissionHandler = (_, _) => Task.FromResult(new PermissionRequestResult());
-        UserInputHandler userInputHandler = (_, _) => Task.FromResult(new UserInputResponse { Answer = "input" });
+        Func<PermissionRequest, PermissionInvocation, Task<PermissionDecision>> permissionHandler = (_, _) => Task.FromResult(PermissionDecision.ApproveOnce());
+        Func<UserInputRequest, UserInputInvocation, Task<UserInputResponse>> userInputHandler = (_, _) => Task.FromResult(new UserInputResponse { Answer = "input" });
        var mcpServers = new Dictionary<string, McpServerConfig> { ["server1"] = new McpStdioServerConfig() };

        var source = new SessionConfig
@@ -122,7 +123,7 @@ public sealed class GitHubCopilotAgentTests
            AvailableTools = ["tool1", "tool2"],
            ExcludedTools = ["tool3"],
            WorkingDirectory = "/workspace",
-            ConfigDir = "/config",
+            ConfigDirectory = "/config",
            Hooks = hooks,
            InfiniteSessions = infiniteSessions,
            OnPermissionRequest = permissionHandler,
@@ -137,17 +138,15 @@ public sealed class GitHubCopilotAgentTests
        // Assert
        Assert.Equal("gpt-4o", result.Model);
        Assert.Equal("high", result.ReasoningEffort);
-        Assert.Same(tools, result.Tools);
-        Assert.Same(systemMessage, result.SystemMessage);
+        Assert.Equal(systemMessage, result.SystemMessage);
        Assert.Equal(new List<string> { "tool1", "tool2" }, result.AvailableTools);
        Assert.Equal(new List<string> { "tool3" }, result.ExcludedTools);
        Assert.Equal("/workspace", result.WorkingDirectory);
-        Assert.Equal("/config", result.ConfigDir);
+        Assert.Equal("/config", result.ConfigDirectory);
        Assert.Same(hooks, result.Hooks);
        Assert.Same(infiniteSessions, result.InfiniteSessions);
        Assert.Same(permissionHandler, result.OnPermissionRequest);
        Assert.Same(userInputHandler, result.OnUserInputRequest);
-        Assert.Same(mcpServers, result.McpServers);
        Assert.Equal(new List<string> { "skill1" }, result.DisabledSkills);
        Assert.True(result.Streaming);
    }
@@ -156,12 +155,12 @@ public sealed class GitHubCopilotAgentTests
    public void CopyResumeSessionConfig_CopiesAllProperties()
    {
        // Arrange
-        List<AIFunction> tools = [AIFunctionFactory.Create(() => "test", "TestFunc", "Test function")];
+        List<AIFunctionDeclaration> tools = [AIFunctionFactory.Create(() => "test", "TestFunc", "Test function")];
        var hooks = new SessionHooks();
        var infiniteSessions = new InfiniteSessionConfig();
        var systemMessage = new SystemMessageConfig { Mode = SystemMessageMode.Append, Content = "Be helpful" };
-        PermissionRequestHandler permissionHandler = (_, _) => Task.FromResult(new PermissionRequestResult());
-        UserInputHandler userInputHandler = (_, _) => Task.FromResult(new UserInputResponse { Answer = "input" });
+        Func<PermissionRequest, PermissionInvocation, Task<PermissionDecision>> permissionHandler = (_, _) => Task.FromResult(PermissionDecision.ApproveOnce());
+        Func<UserInputRequest, UserInputInvocation, Task<UserInputResponse>> userInputHandler = (_, _) => Task.FromResult(new UserInputResponse { Answer = "input" });
        var mcpServers = new Dictionary<string, McpServerConfig> { ["server1"] = new McpStdioServerConfig() };

        var source = new SessionConfig
@@ -173,7 +172,7 @@ public sealed class GitHubCopilotAgentTests
            AvailableTools = ["tool1", "tool2"],
            ExcludedTools = ["tool3"],
            WorkingDirectory = "/workspace",
-            ConfigDir = "/config",
+            ConfigDirectory = "/config",
            Hooks = hooks,
            InfiniteSessions = infiniteSessions,
            OnPermissionRequest = permissionHandler,
@@ -193,7 +192,7 @@ public sealed class GitHubCopilotAgentTests
        Assert.Equal(new List<string> { "tool1", "tool2" }, result.AvailableTools);
        Assert.Equal(new List<string> { "tool3" }, result.ExcludedTools);
        Assert.Equal("/workspace", result.WorkingDirectory);
-        Assert.Equal("/config", result.ConfigDir);
+        Assert.Equal("/config", result.ConfigDirectory);
        Assert.Same(hooks, result.Hooks);
        Assert.Same(infiniteSessions, result.InfiniteSessions);
        Assert.Same(permissionHandler, result.OnPermissionRequest);
@@ -218,7 +217,7 @@ public sealed class GitHubCopilotAgentTests
        Assert.Null(result.OnUserInputRequest);
        Assert.Null(result.Hooks);
        Assert.Null(result.WorkingDirectory);
-        Assert.Null(result.ConfigDir);
+        Assert.Null(result.ConfigDirectory);
        Assert.True(result.Streaming);
    }

@@ -233,7 +232,7 @@ public sealed class GitHubCopilotAgentTests
                Content = "Some streamed content that was already delivered via delta events"
            }
        };
-        CopilotClient copilotClient = new(new CopilotClientOptions { AutoStart = false });
+        CopilotClient copilotClient = new(new CopilotClientOptions());
        const string TestId = "agent-id";
        var agent = new GitHubCopilotAgent(copilotClient, ownsClient: false, id: TestId, tools: null);
        AgentResponseUpdate result = agent.ConvertToAgentResponseUpdate(assistantMessage);
@@ -3,6 +3,7 @@
  <PropertyGroup>
    <!-- GitHub.Copilot.SDK only supports .NET 8.0+ -->
    <TargetFrameworks>$(TargetFrameworksCore)</TargetFrameworks>
+    <NoWarn>$(NoWarn);GHCP001</NoWarn>
  </PropertyGroup>

  <ItemGroup>
@@ -27,6 +27,7 @@ public class HarnessAgentOptionsTests
        Assert.Null(options.ChatHistoryProvider);
        Assert.Null(options.AIContextProviders);
        Assert.False(options.DisableToolApproval);
+        Assert.False(options.DisableNonApprovalRequiredFunctionBypassing);
        Assert.False(options.DisableFileMemory);
        Assert.False(options.DisableFileAccess);
        Assert.False(options.DisableWebSearch);
@@ -80,6 +81,7 @@ public class HarnessAgentOptionsTests
            AIContextProviders = contextProviders,
            MaximumIterationsPerRequest = 42,
            DisableToolApproval = true,
+            DisableNonApprovalRequiredFunctionBypassing = true,
            DisableFileMemory = true,
            FileMemoryStore = fileMemoryStore,
            DisableFileAccess = true,
@@ -112,6 +114,7 @@ public class HarnessAgentOptionsTests
        Assert.Same(contextProviders, options.AIContextProviders);
        Assert.Equal(42, options.MaximumIterationsPerRequest);
        Assert.True(options.DisableToolApproval);
+        Assert.True(options.DisableNonApprovalRequiredFunctionBypassing);
        Assert.True(options.DisableFileMemory);
        Assert.Same(fileMemoryStore, options.FileMemoryStore);
        Assert.True(options.DisableFileAccess);
@@ -21,9 +21,12 @@ public class HarnessAgentTests

    /// <summary>
    /// Creates a HarnessAgent with all default features disabled to isolate tests for specific behaviors.
+    /// Compaction is enabled by default for backward compatibility with existing tests.
    /// </summary>
    private static HarnessAgentOptions CreateAllDisabledOptions() => new()
    {
+        MaxContextWindowTokens = TestMaxContextWindowTokens,
+        MaxOutputTokens = TestMaxOutputTokens,
        DisableToolApproval = true,
        DisableOpenTelemetry = true,
        DisableFileMemory = true,
@@ -43,7 +46,7 @@ public class HarnessAgentTests
    public void Constructor_ThrowsWhenChatClientIsNull()
    {
        // Act & Assert
-        Assert.Throws<ArgumentNullException>(() => new HarnessAgent(null!, TestMaxContextWindowTokens, TestMaxOutputTokens));
+        Assert.Throws<ArgumentNullException>(() => new HarnessAgent(null!));
    }

    /// <summary>
@@ -54,9 +57,10 @@ public class HarnessAgentTests
    {
        // Arrange
        var chatClient = new Mock<IChatClient>().Object;
+        var options = new HarnessAgentOptions { MaxContextWindowTokens = 0, MaxOutputTokens = TestMaxOutputTokens };

        // Act & Assert
-        Assert.Throws<ArgumentOutOfRangeException>(() => new HarnessAgent(chatClient, 0, TestMaxOutputTokens));
+        Assert.Throws<ArgumentOutOfRangeException>(() => new HarnessAgent(chatClient, options));
    }

    /// <summary>
@@ -67,9 +71,10 @@ public class HarnessAgentTests
    {
        // Arrange
        var chatClient = new Mock<IChatClient>().Object;
+        var options = new HarnessAgentOptions { MaxContextWindowTokens = 100_000, MaxOutputTokens = 100_000 };

        // Act & Assert
-        Assert.Throws<ArgumentOutOfRangeException>(() => new HarnessAgent(chatClient, 100_000, 100_000));
+        Assert.Throws<ArgumentOutOfRangeException>(() => new HarnessAgent(chatClient, options));
    }

    /// <summary>
@@ -82,7 +87,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens);
+        var agent = new HarnessAgent(chatClient);

        // Assert
        Assert.NotNull(agent);
@@ -105,7 +110,7 @@ public class HarnessAgentTests
        options.Description = "A test agent";

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);

        // Assert
        Assert.Equal("TestAgent", agent.Name);
@@ -124,7 +129,7 @@ public class HarnessAgentTests
        options.Id = "my-agent-id";

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);

        // Assert
        Assert.Equal("my-agent-id", agent.Id);
@@ -144,7 +149,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -164,7 +169,7 @@ public class HarnessAgentTests
        options.ChatOptions = new ChatOptions { Temperature = 0.5f };

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -184,7 +189,7 @@ public class HarnessAgentTests
        options.ChatOptions = new ChatOptions { Instructions = "You are a custom assistant." };

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -205,7 +210,7 @@ public class HarnessAgentTests
        options.HarnessInstructions = "Custom harness rules.";

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -226,7 +231,7 @@ public class HarnessAgentTests
        options.ChatOptions = new ChatOptions { Instructions = "You are a research agent." };

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -247,7 +252,7 @@ public class HarnessAgentTests
        options.ChatOptions = new ChatOptions { Instructions = "Agent only instructions." };

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -267,7 +272,7 @@ public class HarnessAgentTests
        options.HarnessInstructions = string.Empty;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -289,7 +294,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -310,7 +315,7 @@ public class HarnessAgentTests
        options.ChatHistoryProvider = customProvider;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -332,7 +337,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -353,7 +358,7 @@ public class HarnessAgentTests
        var rawClient = mockClient.Object;

        // Act
-        var agent = new HarnessAgent(rawClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(rawClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — the pipeline wraps the raw client, so the outer client is not the same object.
@@ -378,7 +383,7 @@ public class HarnessAgentTests
        options.AIContextProviders = [customProvider];

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — the custom provider should appear in the inner agent's AIContextProviders.
@@ -398,7 +403,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -432,7 +437,7 @@ public class HarnessAgentTests
        var options = CreateAllDisabledOptions();
        options.ChatOptions = new ChatOptions { Tools = [tool] };

-        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(mockClient.Object, options);
        var session = await agent.CreateSessionAsync();

        // Act
@@ -459,8 +464,10 @@ public class HarnessAgentTests
        };

        // Act
-        _ = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, new HarnessAgentOptions
+        _ = new HarnessAgent(chatClient, new HarnessAgentOptions
        {
+            MaxContextWindowTokens = TestMaxContextWindowTokens,
+            MaxOutputTokens = TestMaxOutputTokens,
            ChatOptions = sourceChatOptions,
        });

@@ -483,7 +490,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());

        // Assert
        Assert.Same(agent, agent.GetService<HarnessAgent>());
@@ -499,7 +506,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());

        // Assert
        Assert.NotNull(agent.GetService<ChatClientAgent>());
@@ -524,7 +531,7 @@ public class HarnessAgentTests
                It.IsAny<CancellationToken>()))
            .ReturnsAsync(new ChatResponse(new ChatMessage(ChatRole.Assistant, "Hello!")));

-        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(mockClient.Object, CreateAllDisabledOptions());
        var session = await agent.CreateSessionAsync();

        // Act
@@ -565,7 +572,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = chatClient.AsHarnessAgent(TestMaxContextWindowTokens, TestMaxOutputTokens);
+        var agent = chatClient.AsHarnessAgent();

        // Assert
        Assert.NotNull(agent);
@@ -586,7 +593,7 @@ public class HarnessAgentTests
        options.ChatOptions = new ChatOptions { Instructions = "Custom instructions" };

        // Act
-        var agent = chatClient.AsHarnessAgent(TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = chatClient.AsHarnessAgent(options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -603,7 +610,7 @@ public class HarnessAgentTests
    public void AsHarnessAgent_ThrowsWhenChatClientIsNull()
    {
        // Act & Assert
-        Assert.Throws<ArgumentNullException>(() => ((IChatClient)null!).AsHarnessAgent(TestMaxContextWindowTokens, TestMaxOutputTokens));
+        Assert.Throws<ArgumentNullException>(() => ((IChatClient)null!).AsHarnessAgent());
    }

    #endregion
@@ -622,7 +629,7 @@ public class HarnessAgentTests
        options.DisableToolApproval = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);

        // Assert
        Assert.NotNull(agent.GetService<ToolApprovalAgent>());
@@ -638,12 +645,148 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());

        // Assert
        Assert.Null(agent.GetService<ToolApprovalAgent>());
    }

+    /// <summary>
+    /// Verify that ToolApprovalAgentOptions auto-approval rules are passed through and actually used.
+    /// </summary>
+    [Fact]
+    public async Task ToolApproval_AutoApprovalRulesAreAppliedAsync()
+    {
+        // Arrange — inner client returns an approval request on first call, then final response on second.
+        var callCount = 0;
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "ReadTool"));
+
+        var mockClient = new Mock<IChatClient>();
+        mockClient
+            .Setup(c => c.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return new ChatResponse(new ChatMessage(ChatRole.Assistant, [approvalRequest]));
+                }
+
+                return new ChatResponse(new ChatMessage(ChatRole.Assistant, "Done"));
+            });
+
+        var options = CreateAllDisabledOptions();
+        options.DisableToolApproval = false;
+        options.ToolApprovalAgentOptions = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]
+        };
+
+        var agent = new HarnessAgent(mockClient.Object, options);
+        var session = await agent.CreateSessionAsync();
+
+        // Act
+        var response = await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+
+        // Assert — the auto-approval rule approved the request, so we get "Done" (not an approval request)
+        Assert.Equal(2, callCount);
+        Assert.Equal("Done", response.Text);
+    }
+
+    #endregion
+
+    #region Feature: NonApprovalRequiredFunctionBypassing
+
+    /// <summary>
+    /// Verify that by default, when a response contains a mix of tools that require approval and tools that do not,
+    /// only the approval-required tool is surfaced to the caller. The non-approval-required tool is bypassed
+    /// (stored as auto-approved) by the <c>NonApprovalRequiredFunctionBypassingChatClient</c> decorator.
+    /// </summary>
+    [Fact]
+    public async Task NonApprovalRequiredFunctionBypassing_BypassesNonApprovalToolsByDefaultAsync()
+    {
+        // Arrange — the model requests both a normal tool and an approval-required tool in the same turn.
+        var normalTool = AIFunctionFactory.Create(() => "result", "NormalTool");
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "ApprovalTool"));
+
+        var mockClient = new Mock<IChatClient>();
+        mockClient
+            .Setup(c => c.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .ReturnsAsync(() => new ChatResponse(new ChatMessage(ChatRole.Assistant,
+            [
+                new FunctionCallContent("call1", "NormalTool"),
+                new FunctionCallContent("call2", "ApprovalTool"),
+            ])));
+
+        // Disable ToolApproval so the approval requests surface in the response instead of being handled.
+        var options = CreateAllDisabledOptions();
+        options.ChatOptions = new ChatOptions { Tools = [normalTool, approvalTool] };
+
+        var agent = new HarnessAgent(mockClient.Object, options);
+        var session = await agent.CreateSessionAsync();
+
+        // Act
+        var response = await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+
+        // Assert — only the approval-required tool surfaces as an approval request; the normal tool is bypassed.
+        var approvalRequests = response.Messages
+            .SelectMany(m => m.Contents)
+            .OfType<ToolApprovalRequestContent>()
+            .ToList();
+        var approvalRequest = Assert.Single(approvalRequests);
+        Assert.Equal("ApprovalTool", Assert.IsType<FunctionCallContent>(approvalRequest.ToolCall).Name);
+    }
+
+    /// <summary>
+    /// Verify that when bypassing is disabled, all tools (including those that do not require approval) are surfaced
+    /// as approval requests, reflecting the all-or-nothing behavior of <see cref="FunctionInvokingChatClient"/>.
+    /// </summary>
+    [Fact]
+    public async Task NonApprovalRequiredFunctionBypassing_SurfacesAllApprovalsWhenDisabledAsync()
+    {
+        // Arrange — the model requests both a normal tool and an approval-required tool in the same turn.
+        var normalTool = AIFunctionFactory.Create(() => "result", "NormalTool");
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "ApprovalTool"));
+
+        var mockClient = new Mock<IChatClient>();
+        mockClient
+            .Setup(c => c.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .ReturnsAsync(() => new ChatResponse(new ChatMessage(ChatRole.Assistant,
+            [
+                new FunctionCallContent("call1", "NormalTool"),
+                new FunctionCallContent("call2", "ApprovalTool"),
+            ])));
+
+        var options = CreateAllDisabledOptions();
+        options.DisableNonApprovalRequiredFunctionBypassing = true;
+        options.ChatOptions = new ChatOptions { Tools = [normalTool, approvalTool] };
+
+        var agent = new HarnessAgent(mockClient.Object, options);
+        var session = await agent.CreateSessionAsync();
+
+        // Act
+        var response = await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+
+        // Assert — both tools surface as approval requests because bypassing is disabled.
+        var approvalRequests = response.Messages
+            .SelectMany(m => m.Contents)
+            .OfType<ToolApprovalRequestContent>()
+            .Select(r => ((FunctionCallContent)r.ToolCall).Name)
+            .ToList();
+        Assert.Equal(2, approvalRequests.Count);
+        Assert.Contains("NormalTool", approvalRequests);
+        Assert.Contains("ApprovalTool", approvalRequests);
+    }
+
    #endregion

    #region Feature: OpenTelemetry
@@ -660,7 +803,7 @@ public class HarnessAgentTests
        options.DisableOpenTelemetry = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);

        // Assert
        Assert.NotNull(agent.GetService<OpenTelemetryAgent>());
@@ -676,7 +819,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());

        // Assert
        Assert.Null(agent.GetService<OpenTelemetryAgent>());
@@ -695,7 +838,7 @@ public class HarnessAgentTests
        options.OpenTelemetrySourceName = "MyApp.AgentTracing";

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);

        // Assert
        Assert.NotNull(agent.GetService<OpenTelemetryAgent>());
@@ -722,7 +865,7 @@ public class HarnessAgentTests
        var options = CreateAllDisabledOptions();
        options.DisableWebSearch = false;

-        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(mockClient.Object, options);
        var session = await agent.CreateSessionAsync();

        // Act
@@ -747,7 +890,7 @@ public class HarnessAgentTests
            .Callback<IEnumerable<ChatMessage>, ChatOptions?, CancellationToken>((_, opts, _) => capturedOptions = opts)
            .ReturnsAsync(new ChatResponse(new ChatMessage(ChatRole.Assistant, "Done")));

-        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(mockClient.Object, CreateAllDisabledOptions());
        var session = await agent.CreateSessionAsync();

        // Act
@@ -780,7 +923,7 @@ public class HarnessAgentTests
        options.DisableWebSearch = false;
        options.ChatOptions = new ChatOptions { Tools = [userTool] };

-        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(mockClient.Object, options);
        var session = await agent.CreateSessionAsync();

        // Act
@@ -808,7 +951,7 @@ public class HarnessAgentTests
        options.DisableTodoProvider = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -826,7 +969,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -853,7 +996,7 @@ public class HarnessAgentTests
        options.DisableAgentModeProvider = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -871,7 +1014,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -902,7 +1045,7 @@ public class HarnessAgentTests
        };

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — AgentModeProvider should be present (we can't easily inspect its internal options,
@@ -927,7 +1070,7 @@ public class HarnessAgentTests
        options.DisableFileMemory = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -945,7 +1088,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -970,7 +1113,7 @@ public class HarnessAgentTests
        options.FileMemoryStore = customStore;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — FileMemoryProvider should be present with the custom store.
@@ -994,7 +1137,7 @@ public class HarnessAgentTests
        options.DisableFileAccess = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1012,7 +1155,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1037,7 +1180,7 @@ public class HarnessAgentTests
        options.FileAccessStore = customStore;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — FileAccessProvider should be present with the custom store.
@@ -1061,7 +1204,7 @@ public class HarnessAgentTests
        options.DisableAgentSkillsProvider = false;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1079,7 +1222,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1104,7 +1247,7 @@ public class HarnessAgentTests
        options.AgentSkillsSource = customSource;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — AgentSkillsProvider should be present.
@@ -1128,7 +1271,7 @@ public class HarnessAgentTests
        options.MaximumIterationsPerRequest = 42;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();
        var ficc = innerAgent!.ChatClient.GetService<FunctionInvokingChatClient>();

@@ -1147,7 +1290,7 @@ public class HarnessAgentTests
        var chatClient = new Mock<IChatClient>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions());
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
        var innerAgent = agent.GetService<ChatClientAgent>();
        var ficc = innerAgent!.ChatClient.GetService<FunctionInvokingChatClient>();

@@ -1175,7 +1318,7 @@ public class HarnessAgentTests
            .ReturnsAsync(new ChatResponse(new ChatMessage(ChatRole.Assistant, "Done")));

        // Act
-        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens);
+        var agent = new HarnessAgent(mockClient.Object);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — agent wrappers
@@ -1218,7 +1361,7 @@ public class HarnessAgentTests
        options.BackgroundAgents = [bgAgentMock.Object];

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1238,7 +1381,7 @@ public class HarnessAgentTests
        options.BackgroundAgents = null;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1261,7 +1404,7 @@ public class HarnessAgentTests
        options.BackgroundAgents = Array.Empty<AIAgent>();

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1292,7 +1435,7 @@ public class HarnessAgentTests
        options.BackgroundAgentsProviderOptions = providerOptions;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();
        var bgProvider = innerAgent!.AIContextProviders!.OfType<BackgroundAgentsProvider>().Single();

@@ -1329,7 +1472,7 @@ public class HarnessAgentTests
        options.BackgroundAgents = [agent1Mock.Object, agent2Mock.Object];

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();
        var bgProvider = innerAgent!.AIContextProviders!.OfType<BackgroundAgentsProvider>().Single();

@@ -1370,7 +1513,7 @@ public class HarnessAgentTests
        options.ShellExecutor = executorMock.Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1390,7 +1533,7 @@ public class HarnessAgentTests
        options.ShellExecutor = null;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert
@@ -1422,7 +1565,7 @@ public class HarnessAgentTests
        options.ShellExecutor = executorMock.Object;

        // Act
-        var agent = new HarnessAgent(chatClientMock.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClientMock.Object, options);
        var session = await agent.CreateSessionAsync();
        await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);

@@ -1451,7 +1594,7 @@ public class HarnessAgentTests
        options.ShellEnvironmentProviderOptions = envOptions;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var agent = new HarnessAgent(chatClient, options);
        var innerAgent = agent.GetService<ChatClientAgent>();

        // Assert — provider should exist (options wiring is validated by the provider's behavior)
@@ -1475,7 +1618,7 @@ public class HarnessAgentTests
        var loggerFactory = new Mock<ILoggerFactory>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), loggerFactory);
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions(), loggerFactory);

        // Assert
        Assert.NotNull(agent);
@@ -1492,7 +1635,7 @@ public class HarnessAgentTests
        var services = new Mock<IServiceProvider>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), services: services);
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions(), services: services);

        // Assert
        Assert.NotNull(agent);
@@ -1510,7 +1653,7 @@ public class HarnessAgentTests
        var services = new Mock<IServiceProvider>().Object;

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), loggerFactory, services);
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions(), loggerFactory, services);

        // Assert
        Assert.NotNull(agent);
@@ -1528,7 +1671,7 @@ public class HarnessAgentTests
        var services = new Mock<IServiceProvider>().Object;

        // Act
-        var agent = chatClient.AsHarnessAgent(TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), loggerFactory, services);
+        var agent = chatClient.AsHarnessAgent(CreateAllDisabledOptions(), loggerFactory, services);

        // Assert
        Assert.NotNull(agent);
@@ -1550,6 +1693,8 @@ public class HarnessAgentTests
        // Act — use options that leave CompactionProvider and AgentSkillsProvider enabled
        var options = new HarnessAgentOptions
        {
+            MaxContextWindowTokens = TestMaxContextWindowTokens,
+            MaxOutputTokens = TestMaxOutputTokens,
            DisableToolApproval = true,
            DisableOpenTelemetry = true,
            DisableFileMemory = true,
@@ -1558,7 +1703,7 @@ public class HarnessAgentTests
            DisableTodoProvider = true,
            DisableAgentModeProvider = true,
        };
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options, mockLoggerFactory.Object);
+        var agent = new HarnessAgent(chatClient, options, mockLoggerFactory.Object);

        // Assert — CreateLogger should have been called by one or more downstream components
        Assert.NotNull(agent);
@@ -1580,7 +1725,7 @@ public class HarnessAgentTests
            .Returns(null!);

        // Act
-        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), services: mockServices.Object);
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions(), services: mockServices.Object);

        // Assert — the service provider should have been queried during pipeline construction
        Assert.NotNull(agent);
@@ -1588,4 +1733,91 @@ public class HarnessAgentTests
    }

    #endregion
+
+    #region Compaction Opt-in
+
+    /// <summary>
+    /// Verify that constructing without token values succeeds (compaction disabled).
+    /// </summary>
+    [Fact]
+    public void Constructor_SucceedsWithoutTokenValues()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var options = new HarnessAgentOptions
+        {
+            DisableToolApproval = true,
+            DisableOpenTelemetry = true,
+            DisableFileMemory = true,
+            DisableFileAccess = true,
+            DisableWebSearch = true,
+            DisableTodoProvider = true,
+            DisableAgentModeProvider = true,
+            DisableAgentSkillsProvider = true,
+        };
+
+        // Act
+        var agent = new HarnessAgent(chatClient, options);
+
+        // Assert — compaction should be disabled (no chat reducer)
+        var innerAgent = agent.GetService<ChatClientAgent>();
+        Assert.NotNull(innerAgent);
+        var historyProvider = innerAgent!.ChatHistoryProvider as InMemoryChatHistoryProvider;
+        Assert.NotNull(historyProvider);
+        Assert.Null(historyProvider!.ChatReducer);
+    }
+
+    /// <summary>
+    /// Verify that when only MaxContextWindowTokens is provided (no MaxOutputTokens), compaction is disabled.
+    /// </summary>
+    [Fact]
+    public void Constructor_SucceedsWithOnlyMaxContextWindowTokens()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var options = new HarnessAgentOptions
+        {
+            MaxContextWindowTokens = TestMaxContextWindowTokens,
+            DisableToolApproval = true,
+            DisableOpenTelemetry = true,
+            DisableFileMemory = true,
+            DisableFileAccess = true,
+            DisableWebSearch = true,
+            DisableTodoProvider = true,
+            DisableAgentModeProvider = true,
+            DisableAgentSkillsProvider = true,
+        };
+
+        // Act
+        var agent = new HarnessAgent(chatClient, options);
+
+        // Assert — compaction should be disabled (only one token value provided)
+        var innerAgent = agent.GetService<ChatClientAgent>();
+        Assert.NotNull(innerAgent);
+        var historyProvider = innerAgent!.ChatHistoryProvider as InMemoryChatHistoryProvider;
+        Assert.NotNull(historyProvider);
+        Assert.Null(historyProvider!.ChatReducer);
+    }
+
+    /// <summary>
+    /// Verify that when both token values are provided, the agent is constructed successfully with compaction.
+    /// </summary>
+    [Fact]
+    public void Constructor_SucceedsWithBothTokenValues()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+
+        // Act
+        var agent = new HarnessAgent(chatClient, CreateAllDisabledOptions());
+
+        // Assert — compaction should be enabled (chat reducer configured)
+        var innerAgent = agent.GetService<ChatClientAgent>();
+        Assert.NotNull(innerAgent);
+        var historyProvider = innerAgent!.ChatHistoryProvider as InMemoryChatHistoryProvider;
+        Assert.NotNull(historyProvider);
+        Assert.NotNull(historyProvider!.ChatReducer);
+    }
+
+    #endregion
 }
@@ -60,7 +60,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        await Task.CompletedTask;
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task SingleAgentSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "01_SingleAgent");
@@ -148,7 +148,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task MultiAgentOrchestrationConcurrentSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "03_AgentOrchestration_Concurrency");
@@ -198,7 +198,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task MultiAgentOrchestrationConditionalsSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "04_AgentOrchestration_Conditionals");
@@ -216,7 +216,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task SingleAgentOrchestrationHITLSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "05_AgentOrchestration_HITL");
@@ -272,7 +272,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task LongRunningToolsSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "06_LongRunningTools");
@@ -362,7 +362,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task AgentAsMcpToolAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "07_AgentAsMcpTool");
@@ -402,7 +402,7 @@ public sealed class SamplesValidation(ITestOutputHelper outputHelper) : IAsyncLi
        });
    }

-    [RetryFact(2, 5000)]
+    [RetryFact(2, 5000, Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task ReliableStreamingSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "08_ReliableStreaming");
@@ -62,7 +62,7 @@ public sealed class WorkflowSamplesValidation(ITestOutputHelper outputHelper) :
        return default;
    }

-    [Fact]
+    [Fact(Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task SequentialWorkflowSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "01_SequentialWorkflow");
@@ -168,7 +168,7 @@ public sealed class WorkflowSamplesValidation(ITestOutputHelper outputHelper) :
        });
    }

-    [Fact]
+    [Fact(Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task HITLWorkflowSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "03_WorkflowHITL");
@@ -277,7 +277,7 @@ public sealed class WorkflowSamplesValidation(ITestOutputHelper outputHelper) :
        });
    }

-    [Fact]
+    [Fact(Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task WorkflowMcpToolSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "04_WorkflowMcpTool");
@@ -333,7 +333,7 @@ public sealed class WorkflowSamplesValidation(ITestOutputHelper outputHelper) :
        });
    }

-    [Fact]
+    [Fact(Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task WorkflowAndAgentsSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "05_WorkflowAndAgents");
@@ -385,7 +385,7 @@ public sealed class WorkflowSamplesValidation(ITestOutputHelper outputHelper) :
        });
    }

-    [Fact]
+    [Fact(Skip = "Azure Functions Core Tools v4 cannot auto-detect worker runtime in CI. See https://github.com/microsoft/agent-framework/issues/6402")]
    public async Task ConcurrentWorkflowSampleValidationAsync()
    {
        string samplePath = Path.Combine(s_samplesPath, "02_ConcurrentWorkflow");
@@ -16,6 +16,7 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
    private const string TestUserId = "test-user-id";
    private const string CustomClaimType = "custom-claim-type";
    private const string CustomClaimValue = "custom-claim-value";
+    private const string TestAuthenticationType = "TestAuth";

    private readonly Mock<IHttpContextAccessor> _httpContextAccessorMock;

@@ -101,7 +102,7 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
    public async Task GetSessionIsolationKeyAsyncExtractsDefaultClaimTypeAsync()
    {
        // Arrange
-        this.SetupHttpContextWithClaim(ClaimsIdentity.DefaultNameClaimType, TestUserId);
+        this.SetupHttpContextWithClaim(ClaimTypes.NameIdentifier, TestUserId);
        var provider = new ClaimsIdentitySessionIsolationKeyProvider(this._httpContextAccessorMock.Object);

        // Act
@@ -111,6 +112,25 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
        Assert.Equal(TestUserId, result);
    }

+    /// <summary>
+    /// Verify that the default claim type is the stable, unique NameIdentifier claim rather than the
+    /// non-unique display name claim. This guards against the session-isolation collision described in
+    /// the security report where two principals sharing the same name claim received the same key.
+    /// </summary>
+    [Fact]
+    public async Task GetSessionIsolationKeyAsyncIgnoresNameClaimByDefaultAsync()
+    {
+        // Arrange - only a display-name claim is present; the default provider must not use it.
+        this.SetupHttpContextWithClaim(ClaimsIdentity.DefaultNameClaimType, TestUserId);
+        var provider = new ClaimsIdentitySessionIsolationKeyProvider(this._httpContextAccessorMock.Object);
+
+        // Act
+        string? result = await provider.GetSessionIsolationKeyAsync();
+
+        // Assert
+        Assert.Null(result);
+    }
+
    /// <summary>
    /// Verify that GetSessionIsolationKeyAsync uses custom claim type when specified.
    /// </summary>
@@ -191,10 +211,10 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
        const string SecondValue = "second-value";
        var claims = new[]
        {
-            new Claim(ClaimsIdentity.DefaultNameClaimType, FirstValue),
-            new Claim(ClaimsIdentity.DefaultNameClaimType, SecondValue),
+            new Claim(ClaimTypes.NameIdentifier, FirstValue),
+            new Claim(ClaimTypes.NameIdentifier, SecondValue),
        };
-        var identity = new ClaimsIdentity(claims);
+        var identity = new ClaimsIdentity(claims, TestAuthenticationType);
        var principal = new ClaimsPrincipal(identity);

        var httpContext = new DefaultHttpContext
@@ -219,7 +239,7 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
    public async Task GetSessionIsolationKeyAsyncHandlesEmptyClaimValueAsync()
    {
        // Arrange
-        this.SetupHttpContextWithClaim(ClaimsIdentity.DefaultNameClaimType, string.Empty);
+        this.SetupHttpContextWithClaim(ClaimTypes.NameIdentifier, string.Empty);
        var provider = new ClaimsIdentitySessionIsolationKeyProvider(this._httpContextAccessorMock.Object);

        // Act
@@ -229,6 +249,66 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
        Assert.Equal(string.Empty, result);
    }

+    /// <summary>
+    /// Regression test for the session-isolation collision security report: two distinct authenticated
+    /// principals that share the same display-name claim but have different stable identifiers and tenants
+    /// must produce distinct isolation keys under the default options.
+    /// </summary>
+    [Fact]
+    public async Task GetSessionIsolationKeyAsyncDistinctForPrincipalsSharingNameClaimAsync()
+    {
+        // Arrange - both principals share the same name claim but differ by NameIdentifier and tenant.
+        const string CommonName = "John Doe";
+
+        var principalA = CreatePrincipal(
+            new Claim(ClaimsIdentity.DefaultNameClaimType, CommonName),
+            new Claim(ClaimTypes.NameIdentifier, "oid-user-a"),
+            new Claim("http://schemas.microsoft.com/identity/claims/tenantid", "tenant-a"));
+
+        var principalB = CreatePrincipal(
+            new Claim(ClaimsIdentity.DefaultNameClaimType, CommonName),
+            new Claim(ClaimTypes.NameIdentifier, "oid-user-b"),
+            new Claim("http://schemas.microsoft.com/identity/claims/tenantid", "tenant-b"));
+
+        var provider = new ClaimsIdentitySessionIsolationKeyProvider(this._httpContextAccessorMock.Object);
+
+        // Act
+        this._httpContextAccessorMock.Setup(x => x.HttpContext).Returns(new DefaultHttpContext { User = principalA });
+        string? principalAKey = await provider.GetSessionIsolationKeyAsync();
+
+        this._httpContextAccessorMock.Setup(x => x.HttpContext).Returns(new DefaultHttpContext { User = principalB });
+        string? principalBKey = await provider.GetSessionIsolationKeyAsync();
+
+        // Assert
+        Assert.Equal("oid-user-a", principalAKey);
+        Assert.Equal("oid-user-b", principalBKey);
+        Assert.NotEqual(principalAKey, principalBKey);
+    }
+
+    /// <summary>
+    /// Verify that GetSessionIsolationKeyAsync returns null when the request's user is not authenticated,
+    /// even if a claim of the configured type is present. The provider must not derive an isolation key
+    /// from claims on an unauthenticated identity.
+    /// </summary>
+    [Fact]
+    public async Task GetSessionIsolationKeyAsyncReturnsNullWhenUserNotAuthenticatedAsync()
+    {
+        // Arrange - identity has the claim but no authentication type, so IsAuthenticated is false.
+        var claims = new[] { new Claim(ClaimTypes.NameIdentifier, TestUserId) };
+        var unauthenticatedIdentity = new ClaimsIdentity(claims);
+        var principal = new ClaimsPrincipal(unauthenticatedIdentity);
+        var httpContext = new DefaultHttpContext { User = principal };
+        this._httpContextAccessorMock.Setup(x => x.HttpContext).Returns(httpContext);
+        var provider = new ClaimsIdentitySessionIsolationKeyProvider(this._httpContextAccessorMock.Object);
+
+        // Act
+        string? result = await provider.GetSessionIsolationKeyAsync();
+
+        // Assert
+        Assert.False(unauthenticatedIdentity.IsAuthenticated);
+        Assert.Null(result);
+    }
+
    #endregion

    #region Helper Methods
@@ -236,7 +316,7 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
    private void SetupHttpContextWithClaim(string claimType, string claimValue)
    {
        var claims = new[] { new Claim(claimType, claimValue) };
-        var identity = new ClaimsIdentity(claims);
+        var identity = new ClaimsIdentity(claims, TestAuthenticationType);
        var principal = new ClaimsPrincipal(identity);

        var httpContext = new DefaultHttpContext
@@ -247,5 +327,8 @@ public class ClaimsIdentitySessionIsolationKeyProviderTests
        this._httpContextAccessorMock.Setup(x => x.HttpContext).Returns(httpContext);
    }

+    private static ClaimsPrincipal CreatePrincipal(params Claim[] claims)
+        => new(new ClaimsIdentity(claims, TestAuthenticationType));
+
    #endregion
 }
@@ -134,6 +134,7 @@ public class ChatClientAgentOptionsTests
            ClearOnChatHistoryProviderConflict = false,
            WarnOnChatHistoryProviderConflict = false,
            ThrowOnChatHistoryProviderConflict = false,
+            EnableNonApprovalRequiredFunctionBypassing = true,
        };

        // Act
@@ -150,6 +151,7 @@ public class ChatClientAgentOptionsTests
        Assert.Equal(original.ClearOnChatHistoryProviderConflict, clone.ClearOnChatHistoryProviderConflict);
        Assert.Equal(original.WarnOnChatHistoryProviderConflict, clone.WarnOnChatHistoryProviderConflict);
        Assert.Equal(original.ThrowOnChatHistoryProviderConflict, clone.ThrowOnChatHistoryProviderConflict);
+        Assert.Equal(original.EnableNonApprovalRequiredFunctionBypassing, clone.EnableNonApprovalRequiredFunctionBypassing);

        // ChatOptions should be cloned, not the same reference
        Assert.NotSame(original.ChatOptions, clone.ChatOptions);
@@ -347,6 +347,115 @@ public class ChatClientAgent_ChatOptionsMergingTests
        Assert.Equal(expectedSetting, capturedChatOptions.RawRepresentationFactory(null!));
    }

+    /// <summary>
+    /// Verify that <see cref="ChatOptions.Reasoning"/> from the request takes priority over the agent's.
+    /// </summary>
+    [Fact]
+    public async Task ChatOptionsMergingUsesRequestReasoningOverAgentReasoningAsync()
+    {
+        // Arrange
+        var agentReasoning = new ReasoningOptions { Effort = ReasoningEffort.Low, Output = ReasoningOutput.Full };
+        var requestReasoning = new ReasoningOptions { Effort = ReasoningEffort.High, Output = ReasoningOutput.Full };
+
+        Mock<IChatClient> mockService = new();
+        ChatOptions? capturedChatOptions = null;
+        mockService.Setup(
+            s => s.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<IEnumerable<ChatMessage>, ChatOptions, CancellationToken>((msgs, opts, ct) =>
+                capturedChatOptions = opts)
+            .ReturnsAsync(new ChatResponse([new(ChatRole.Assistant, "response")]));
+
+        ChatClientAgent agent = new(mockService.Object, options: new()
+        {
+            ChatOptions = new ChatOptions { Reasoning = agentReasoning }
+        });
+        var messages = new List<ChatMessage> { new(ChatRole.User, "test") };
+
+        // Act
+        await agent.RunAsync(messages, options: new ChatClientAgentRunOptions(new ChatOptions { Reasoning = requestReasoning }));
+
+        // Assert
+        Assert.NotNull(capturedChatOptions);
+        Assert.NotNull(capturedChatOptions.Reasoning);
+        Assert.Equal(requestReasoning.Effort, capturedChatOptions.Reasoning.Effort);
+        Assert.Equal(requestReasoning.Output, capturedChatOptions.Reasoning.Output);
+    }
+
+    /// <summary>
+    /// Verify that <see cref="ChatOptions.Reasoning"/> falls back to the agent's when the request has none.
+    /// </summary>
+    [Fact]
+    public async Task ChatOptionsMergingFallsBackToAgentReasoningWhenRequestHasNoneAsync()
+    {
+        // Arrange
+        var agentReasoning = new ReasoningOptions { Effort = ReasoningEffort.Low, Output = ReasoningOutput.Full };
+
+        Mock<IChatClient> mockService = new();
+        ChatOptions? capturedChatOptions = null;
+        mockService.Setup(
+            s => s.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<IEnumerable<ChatMessage>, ChatOptions, CancellationToken>((msgs, opts, ct) =>
+                capturedChatOptions = opts)
+            .ReturnsAsync(new ChatResponse([new(ChatRole.Assistant, "response")]));
+
+        ChatClientAgent agent = new(mockService.Object, options: new()
+        {
+            ChatOptions = new ChatOptions { Reasoning = agentReasoning }
+        });
+        var messages = new List<ChatMessage> { new(ChatRole.User, "test") };
+
+        // Act
+        await agent.RunAsync(messages, options: new ChatClientAgentRunOptions(new ChatOptions()));
+
+        // Assert
+        Assert.NotNull(capturedChatOptions);
+        Assert.NotNull(capturedChatOptions.Reasoning);
+        Assert.Equal(agentReasoning.Effort, capturedChatOptions.Reasoning.Effort);
+        Assert.Equal(agentReasoning.Output, capturedChatOptions.Reasoning.Output);
+    }
+
+    /// <summary>
+    /// Verify that <see cref="ChatOptions.Reasoning"/> from the request is used when the agent has none.
+    /// </summary>
+    [Fact]
+    public async Task ChatOptionsMergingUsesRequestReasoningWhenAgentHasNoneAsync()
+    {
+        // Arrange
+        var requestReasoning = new ReasoningOptions { Effort = ReasoningEffort.High, Output = ReasoningOutput.Full };
+
+        Mock<IChatClient> mockService = new();
+        ChatOptions? capturedChatOptions = null;
+        mockService.Setup(
+            s => s.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<IEnumerable<ChatMessage>, ChatOptions, CancellationToken>((msgs, opts, ct) =>
+                capturedChatOptions = opts)
+            .ReturnsAsync(new ChatResponse([new(ChatRole.Assistant, "response")]));
+
+        ChatClientAgent agent = new(mockService.Object, options: new()
+        {
+            ChatOptions = new ChatOptions()
+        });
+        var messages = new List<ChatMessage> { new(ChatRole.User, "test") };
+
+        // Act
+        await agent.RunAsync(messages, options: new ChatClientAgentRunOptions(new ChatOptions { Reasoning = requestReasoning }));
+
+        // Assert
+        Assert.NotNull(capturedChatOptions);
+        Assert.NotNull(capturedChatOptions.Reasoning);
+        Assert.Equal(requestReasoning.Effort, capturedChatOptions.Reasoning.Effort);
+        Assert.Equal(requestReasoning.Output, capturedChatOptions.Reasoning.Output);
+    }
+
    /// <summary>
    /// Verify that ChatOptions merging handles all scalar properties correctly.
    /// </summary>
@@ -0,0 +1,574 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Threading;
+using System.Threading.Tasks;
+using Microsoft.Extensions.AI;
+using Moq;
+
+namespace Microsoft.Agents.AI.UnitTests;
+
+public class NonApprovalRequiredFunctionBypassingChatClientTests
+{
+    #region GetResponseAsync Tests
+
+    [Fact]
+    public async Task GetResponseAsync_NoApprovalContent_PassesThroughUnchangedAsync()
+    {
+        // Arrange
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, "Hello")])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+
+        // Act
+        var response = await RunWithAgentContextAsync(decorator, session);
+
+        // Assert
+        Assert.Single(response.Messages);
+        Assert.Equal("Hello", response.Messages[0].Text);
+        Assert.Equal(0, session.StateBag.Count);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_AllToolsRequireApproval_PassesThroughUnchangedAsync()
+    {
+        // Arrange
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "approvalTool"));
+        var fcc = new FunctionCallContent("call1", "approvalTool");
+        var approval = new ToolApprovalRequestContent("req1", fcc);
+
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, [approval])])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [approvalTool] };
+
+        // Act
+        var response = await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — approval request should remain
+        Assert.Single(response.Messages);
+        var contents = response.Messages[0].Contents;
+        Assert.Single(contents);
+        Assert.IsType<ToolApprovalRequestContent>(contents[0]);
+        Assert.Equal(0, session.StateBag.Count);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_MixedApproval_RemovesNonApprovalItemsAsync()
+    {
+        // Arrange
+        var normalTool = AIFunctionFactory.Create(() => "result", "normalTool");
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "approvalTool"));
+
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var fccApproval = new FunctionCallContent("call2", "approvalTool");
+        var approvalNormal = new ToolApprovalRequestContent("req1", fccNormal);
+        var approvalRequired = new ToolApprovalRequestContent("req2", fccApproval);
+
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([
+                new ChatMessage(ChatRole.Assistant, [approvalNormal, approvalRequired])
+            ])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [normalTool, approvalTool] };
+
+        // Act
+        var response = await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — only the approval-required item remains in the response
+        Assert.Single(response.Messages);
+        var contents = response.Messages[0].Contents;
+        Assert.Single(contents);
+        var remainingApproval = Assert.IsType<ToolApprovalRequestContent>(contents[0]);
+        Assert.Equal("req2", remainingApproval.RequestId);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_MixedApproval_StoresAutoApprovedInSessionAsync()
+    {
+        // Arrange
+        var normalTool = AIFunctionFactory.Create(() => "result", "normalTool");
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "approvalTool"));
+
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var fccApproval = new FunctionCallContent("call2", "approvalTool");
+        var approvalNormal = new ToolApprovalRequestContent("req1", fccNormal);
+        var approvalRequired = new ToolApprovalRequestContent("req2", fccApproval);
+
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([
+                new ChatMessage(ChatRole.Assistant, [approvalNormal, approvalRequired])
+            ])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [normalTool, approvalTool] };
+
+        // Act
+        await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — the auto-approved item should be stored in the session
+        Assert.True(session.StateBag.TryGetValue<List<ToolApprovalRequestContent>>(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey, out var stored, AgentJsonUtilities.DefaultOptions));
+        Assert.NotNull(stored);
+        Assert.Single(stored!);
+        Assert.Equal("req1", stored![0].RequestId);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_AllNonApproval_RemovesAllApprovalsAndRemovesEmptyMessageAsync()
+    {
+        // Arrange
+        var normalTool = AIFunctionFactory.Create(() => "result", "normalTool");
+
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var approvalNormal = new ToolApprovalRequestContent("req1", fccNormal);
+
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([
+                new ChatMessage(ChatRole.Assistant, [approvalNormal])
+            ])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [normalTool] };
+
+        // Act
+        var response = await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — the message should be removed since it's now empty
+        Assert.Empty(response.Messages);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_NextRequest_InjectsStoredAutoApprovalsAsync()
+    {
+        // Arrange
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var storedApproval = new ToolApprovalRequestContent("req1", fccNormal);
+
+        var session = new ChatClientAgentSession();
+        session.StateBag.SetValue(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey,
+            new List<ToolApprovalRequestContent> { storedApproval },
+            AgentJsonUtilities.DefaultOptions);
+
+        IEnumerable<ChatMessage>? capturedMessages = null;
+        var innerClient = CreateMockChatClient((messages, _, _) =>
+        {
+            capturedMessages = messages.ToList();
+            return Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, "Done")]));
+        });
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var options = new ChatOptions { Tools = [AIFunctionFactory.Create(() => "result", "normalTool")] };
+
+        // Act
+        await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — the inner client should receive injected messages
+        Assert.NotNull(capturedMessages);
+        var messagesList = capturedMessages!.ToList();
+
+        // Original user message + user message with approved responses.
+        Assert.Equal(2, messagesList.Count);
+        Assert.Equal(ChatRole.User, messagesList[0].Role);
+
+        // User message with the auto-approved ToolApprovalResponseContent
+        Assert.Equal(ChatRole.User, messagesList[1].Role);
+        var userContent = messagesList[1].Contents.OfType<ToolApprovalResponseContent>().ToList();
+        Assert.Single(userContent);
+        Assert.Equal("req1", userContent[0].RequestId);
+        Assert.True(userContent[0].Approved);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_NextRequest_ClearsStoredAfterInjectionAsync()
+    {
+        // Arrange
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var storedApproval = new ToolApprovalRequestContent("req1", fccNormal);
+
+        var session = new ChatClientAgentSession();
+        session.StateBag.SetValue(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey,
+            new List<ToolApprovalRequestContent> { storedApproval },
+            AgentJsonUtilities.DefaultOptions);
+
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, "Done")])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var options = new ChatOptions { Tools = [AIFunctionFactory.Create(() => "result", "normalTool")] };
+
+        // Act
+        await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — the stored data should be cleared after successful injection
+        Assert.False(session.StateBag.TryGetValue<List<ToolApprovalRequestContent>>(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey, out _, AgentJsonUtilities.DefaultOptions));
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_UnknownTool_TreatedAsApprovalRequiredAsync()
+    {
+        // Arrange — tool is not in ChatOptions.Tools
+        var fccUnknown = new FunctionCallContent("call1", "unknownTool");
+        var approvalUnknown = new ToolApprovalRequestContent("req1", fccUnknown);
+
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([
+                new ChatMessage(ChatRole.Assistant, [approvalUnknown])
+            ])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [] };
+
+        // Act
+        var response = await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — unknown tool should NOT be auto-approved
+        Assert.Single(response.Messages);
+        Assert.Single(response.Messages[0].Contents);
+        Assert.IsType<ToolApprovalRequestContent>(response.Messages[0].Contents[0]);
+        Assert.Equal(0, session.StateBag.Count);
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_StoredRequestToolSetChanged_StillInjectsAsApprovedAsync()
+    {
+        // Arrange — tool was previously non-approval-required but is now wrapped in ApprovalRequiredAIFunction.
+        // The LLM still requires a complete set of responses, so we inject unconditionally.
+        var fccTool = new FunctionCallContent("call1", "changingTool");
+        var storedApproval = new ToolApprovalRequestContent("req1", fccTool);
+
+        var session = new ChatClientAgentSession();
+        session.StateBag.SetValue(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey,
+            new List<ToolApprovalRequestContent> { storedApproval },
+            AgentJsonUtilities.DefaultOptions);
+
+        IEnumerable<ChatMessage>? capturedMessages = null;
+        var innerClient = CreateMockChatClient((messages, _, _) =>
+        {
+            capturedMessages = messages.ToList();
+            return Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, "Done")]));
+        });
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+
+        // The tool is now wrapped in ApprovalRequiredAIFunction — but we still inject unconditionally
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "changingTool"));
+        var options = new ChatOptions { Tools = [approvalTool] };
+
+        // Act
+        await RunWithAgentContextAsync(decorator, session, options);
+
+        // Assert — the stored request should still be injected as approved
+        Assert.NotNull(capturedMessages);
+        var messagesList = capturedMessages!.ToList();
+        Assert.Equal(2, messagesList.Count);
+        var userContent = messagesList[1].Contents.OfType<ToolApprovalResponseContent>().ToList();
+        Assert.Single(userContent);
+        Assert.Equal("req1", userContent[0].RequestId);
+        Assert.True(userContent[0].Approved);
+
+        // Session should be cleared
+        Assert.False(session.StateBag.TryGetValue<List<ToolApprovalRequestContent>>(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey, out _, AgentJsonUtilities.DefaultOptions));
+    }
+
+    #endregion
+
+    #region GetStreamingResponseAsync Tests
+
+    [Fact]
+    public async Task GetStreamingResponseAsync_NoApprovalContent_PassesThroughUnchangedAsync()
+    {
+        // Arrange
+        var innerClient = CreateMockStreamingChatClient((_, _, _) =>
+            ToAsyncEnumerableAsync(
+                new ChatResponseUpdate(ChatRole.Assistant, "Hello")));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+
+        // Act
+        var updates = new List<ChatResponseUpdate>();
+        await RunStreamingWithAgentContextAsync(decorator, session, updates);
+
+        // Assert
+        Assert.Single(updates);
+        Assert.Equal("Hello", updates[0].Text);
+        Assert.Equal(0, session.StateBag.Count);
+    }
+
+    [Fact]
+    public async Task GetStreamingResponseAsync_MixedApproval_FiltersNonApprovalItemsAsync()
+    {
+        // Arrange
+        var normalTool = AIFunctionFactory.Create(() => "result", "normalTool");
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "approvalTool"));
+
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var fccApproval = new FunctionCallContent("call2", "approvalTool");
+        var approvalNormal = new ToolApprovalRequestContent("req1", fccNormal);
+        var approvalRequired = new ToolApprovalRequestContent("req2", fccApproval);
+
+        var innerClient = CreateMockStreamingChatClient((_, _, _) =>
+            ToAsyncEnumerableAsync(
+                new ChatResponseUpdate(ChatRole.Assistant, "text"),
+                new ChatResponseUpdate { Contents = [approvalNormal, approvalRequired] }));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [normalTool, approvalTool] };
+
+        // Act
+        var updates = new List<ChatResponseUpdate>();
+        await RunStreamingWithAgentContextAsync(decorator, session, updates, options);
+
+        // Assert — text update + filtered approval update
+        Assert.Equal(2, updates.Count);
+        Assert.Equal("text", updates[0].Text);
+
+        // Second update should only have the approval-required item
+        var approvalContents = updates[1].Contents.OfType<ToolApprovalRequestContent>().ToList();
+        Assert.Single(approvalContents);
+        Assert.Equal("req2", approvalContents[0].RequestId);
+    }
+
+    [Fact]
+    public async Task GetStreamingResponseAsync_MixedApproval_StoresAutoApprovedInSessionAsync()
+    {
+        // Arrange
+        var normalTool = AIFunctionFactory.Create(() => "result", "normalTool");
+        var approvalTool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(() => "result", "approvalTool"));
+
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var fccApproval = new FunctionCallContent("call2", "approvalTool");
+        var approvalNormal = new ToolApprovalRequestContent("req1", fccNormal);
+        var approvalRequired = new ToolApprovalRequestContent("req2", fccApproval);
+
+        var innerClient = CreateMockStreamingChatClient((_, _, _) =>
+            ToAsyncEnumerableAsync(
+                new ChatResponseUpdate { Contents = [approvalNormal, approvalRequired] }));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [normalTool, approvalTool] };
+
+        // Act
+        var updates = new List<ChatResponseUpdate>();
+        await RunStreamingWithAgentContextAsync(decorator, session, updates, options);
+
+        // Assert — the auto-approved item should be stored in the session
+        Assert.True(session.StateBag.TryGetValue<List<ToolApprovalRequestContent>>(
+            NonApprovalRequiredFunctionBypassingChatClient.StateBagKey, out var stored, AgentJsonUtilities.DefaultOptions));
+        Assert.NotNull(stored);
+        Assert.Single(stored!);
+        Assert.Equal("req1", stored![0].RequestId);
+    }
+
+    [Fact]
+    public async Task GetStreamingResponseAsync_AllNonApproval_SkipsEmptyUpdateAsync()
+    {
+        // Arrange
+        var normalTool = AIFunctionFactory.Create(() => "result", "normalTool");
+
+        var fccNormal = new FunctionCallContent("call1", "normalTool");
+        var approvalNormal = new ToolApprovalRequestContent("req1", fccNormal);
+
+        var innerClient = CreateMockStreamingChatClient((_, _, _) =>
+            ToAsyncEnumerableAsync(
+                new ChatResponseUpdate(ChatRole.Assistant, "text"),
+                new ChatResponseUpdate { Contents = [approvalNormal] }));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+        var session = new ChatClientAgentSession();
+        var options = new ChatOptions { Tools = [normalTool] };
+
+        // Act
+        var updates = new List<ChatResponseUpdate>();
+        await RunStreamingWithAgentContextAsync(decorator, session, updates, options);
+
+        // Assert — the approval update should be skipped entirely
+        Assert.Single(updates);
+        Assert.Equal("text", updates[0].Text);
+    }
+
+    #endregion
+
+    #region Error Handling Tests
+
+    [Fact]
+    public async Task GetResponseAsync_NoRunContext_ThrowsInvalidOperationExceptionAsync()
+    {
+        // Arrange
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, "response")])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+
+        // Act & Assert — calling directly without agent context
+        await Assert.ThrowsAsync<InvalidOperationException>(
+            () => decorator.GetResponseAsync([new ChatMessage(ChatRole.User, "test")]));
+    }
+
+    [Fact]
+    public async Task GetResponseAsync_NoSession_ThrowsInvalidOperationExceptionAsync()
+    {
+        // Arrange
+        var innerClient = CreateMockChatClient((_, _, _) =>
+            Task.FromResult(new ChatResponse([new ChatMessage(ChatRole.Assistant, "response")])));
+
+        var decorator = new NonApprovalRequiredFunctionBypassingChatClient(innerClient);
+
+        // Act & Assert — run with null session
+        await Assert.ThrowsAsync<InvalidOperationException>(
+            () => RunWithAgentContextAsync(decorator, session: null!));
+    }
+
+    #endregion
+
+    #region Builder Extension Tests
+
+    [Fact]
+    public void UseNonApprovalRequiredFunctionBypassing_AddsDecoratorToPipeline()
+    {
+        // Arrange
+        var innerClient = new Mock<IChatClient>().Object;
+
+        // Act
+        var pipeline = innerClient.AsBuilder()
+            .UseNonApprovalRequiredFunctionBypassing()
+            .Build();
+
+        // Assert
+        Assert.NotNull(pipeline.GetService<NonApprovalRequiredFunctionBypassingChatClient>());
+    }
+
+    [Fact]
+    public void WithDefaultAgentMiddleware_EnableNonApprovalRequiredFunctionBypassing_InjectsDecorator()
+    {
+        // Arrange
+        var innerClient = new Mock<IChatClient>().Object;
+        var options = new ChatClientAgentOptions { EnableNonApprovalRequiredFunctionBypassing = true };
+
+        // Act
+        var pipeline = innerClient.WithDefaultAgentMiddleware(options);
+
+        // Assert
+        Assert.NotNull(pipeline.GetService<NonApprovalRequiredFunctionBypassingChatClient>());
+    }
+
+    [Fact]
+    public void WithDefaultAgentMiddleware_EnableNonApprovalRequiredFunctionBypassingFalse_DoesNotInjectDecorator()
+    {
+        // Arrange
+        var innerClient = new Mock<IChatClient>().Object;
+        var options = new ChatClientAgentOptions { EnableNonApprovalRequiredFunctionBypassing = false };
+
+        // Act
+        var pipeline = innerClient.WithDefaultAgentMiddleware(options);
+
+        // Assert
+        Assert.Null(pipeline.GetService<NonApprovalRequiredFunctionBypassingChatClient>());
+    }
+
+    #endregion
+
+    #region Helpers
+
+    private static async Task<ChatResponse> RunWithAgentContextAsync(
+        NonApprovalRequiredFunctionBypassingChatClient decorator,
+        AgentSession? session,
+        ChatOptions? options = null)
+    {
+        ChatResponse? capturedResponse = null;
+
+        var agent = new TestAIAgent
+        {
+            RunAsyncFunc = async (messages, agentSession, agentOptions, ct) =>
+            {
+                capturedResponse = await decorator.GetResponseAsync(messages, options, ct);
+                return new AgentResponse(capturedResponse);
+            }
+        };
+
+        await agent.RunAsync([new ChatMessage(ChatRole.User, "Hello")], session);
+        return capturedResponse!;
+    }
+
+    private static Task<ChatResponse> RunWithAgentContextAsync(
+        NonApprovalRequiredFunctionBypassingChatClient decorator,
+        AgentSession session)
+        => RunWithAgentContextAsync(decorator, session, options: null);
+
+    private static async Task RunStreamingWithAgentContextAsync(
+        NonApprovalRequiredFunctionBypassingChatClient decorator,
+        AgentSession session,
+        List<ChatResponseUpdate> updates,
+        ChatOptions? options = null)
+    {
+        var agent = new TestAIAgent
+        {
+            RunAsyncFunc = async (messages, agentSession, agentOptions, ct) =>
+            {
+                await foreach (var update in decorator.GetStreamingResponseAsync(messages, options, ct))
+                {
+                    updates.Add(update);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "done")]);
+            }
+        };
+
+        await agent.RunAsync([new ChatMessage(ChatRole.User, "Hello")], session);
+    }
+
+    private static IChatClient CreateMockChatClient(
+        Func<IEnumerable<ChatMessage>, ChatOptions?, CancellationToken, Task<ChatResponse>> onGetResponse)
+    {
+        var mock = new Mock<IChatClient>();
+        mock.Setup(c => c.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions?>(),
+                It.IsAny<CancellationToken>()))
+            .Returns((IEnumerable<ChatMessage> m, ChatOptions? o, CancellationToken ct) => onGetResponse(m, o, ct));
+        return mock.Object;
+    }
+
+    private static IChatClient CreateMockStreamingChatClient(
+        Func<IEnumerable<ChatMessage>, ChatOptions?, CancellationToken, IAsyncEnumerable<ChatResponseUpdate>> onGetStreamingResponse)
+    {
+        var mock = new Mock<IChatClient>();
+        mock.Setup(c => c.GetStreamingResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions?>(),
+                It.IsAny<CancellationToken>()))
+            .Returns((IEnumerable<ChatMessage> m, ChatOptions? o, CancellationToken ct) => onGetStreamingResponse(m, o, ct));
+        return mock.Object;
+    }
+
+    private static async IAsyncEnumerable<ChatResponseUpdate> ToAsyncEnumerableAsync(params ChatResponseUpdate[] updates)
+    {
+        foreach (var update in updates)
+        {
+            yield return update;
+        }
+
+        await Task.CompletedTask;
+    }
+
+    #endregion
+}
@@ -59,15 +59,15 @@ public class ToolApprovalAgentBuilderExtensionsTests
    /// Verify that UseToolApproval with custom JsonSerializerOptions works correctly.
    /// </summary>
    [Fact]
-    public void UseToolApproval_WithCustomJsonSerializerOptions_ReturnsToolApprovalAgent()
+    public void UseToolApproval_WithCustomOptions_ReturnsToolApprovalAgent()
    {
        // Arrange
        var mockAgent = new Mock<AIAgent>();
        var builder = new AIAgentBuilder(mockAgent.Object);
-        var options = new JsonSerializerOptions();
+        var options = new ToolApprovalAgentOptions { JsonSerializerOptions = new JsonSerializerOptions() };

        // Act
-        var result = builder.UseToolApproval(jsonSerializerOptions: options).Build();
+        var result = builder.UseToolApproval(options: options).Build();

        // Assert
        Assert.IsType<ToolApprovalAgent>(result);
@@ -47,14 +47,14 @@ public class ToolApprovalAgentTests
    }

    /// <summary>
-    /// Verify that constructor accepts custom JsonSerializerOptions.
+    /// Verify that constructor accepts custom options.
    /// </summary>
    [Fact]
-    public void Constructor_CustomJsonSerializerOptions_CreatesInstanceAsync()
+    public void Constructor_CustomOptions_CreatesInstance()
    {
        // Arrange
        var innerAgent = new Mock<AIAgent>().Object;
-        var options = new JsonSerializerOptions();
+        var options = new ToolApprovalAgentOptions { JsonSerializerOptions = new JsonSerializerOptions() };

        // Act
        var agent = new ToolApprovalAgent(innerAgent, options);
@@ -1535,4 +1535,311 @@ public class ToolApprovalAgentTests
    }

    #endregion
+
+    #region Auto-Approval Rules (Heuristics)
+
+    /// <summary>
+    /// Verify that an auto-approval rule can approve a function call that would otherwise need user approval.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_AutoApprovalRule_ApprovesMatchingToolAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "ReadTool"));
+
+        // Inner agent: first call returns approval request, second returns final response.
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+            });
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        var response = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, "Hi")],
+            session);
+
+        // Assert — the approval request was auto-approved, inner agent called twice
+        Assert.Equal(2, callCount);
+        Assert.Equal("Done", response.Text);
+    }
+
+    /// <summary>
+    /// Verify that when auto-approval rule does not match, request is surfaced to the caller.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_AutoApprovalRule_DoesNotMatchSurfacesToCallerAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "DangerousTool"));
+
+        var innerAgent = CreateMockAgent(new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]));
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]  // Only approves ReadTool
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        var response = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, "Hi")],
+            session);
+
+        // Assert — request surfaced to caller since heuristic doesn't match
+        var requests = response.Messages.SelectMany(m => m.Contents).OfType<ToolApprovalRequestContent>().ToList();
+        Assert.Single(requests);
+        Assert.Equal("DangerousTool", ((FunctionCallContent)requests[0].ToolCall).Name);
+    }
+
+    /// <summary>
+    /// Verify that multiple auto-approval rules are evaluated in order; first match wins.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_MultipleAutoApprovalRules_FirstMatchWinsAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "SpecialTool"));
+
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+            });
+
+        var rule1Called = false;
+        var rule2Called = false;
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules =
+            [
+                fcc => { rule1Called = true; return new ValueTask<bool>(fcc.Name == "SpecialTool"); },
+                fcc => { rule2Called = true; return new ValueTask<bool>(true); }  // Should not be reached
+            ]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+
+        // Assert — first rule matched, second was never called
+        Assert.True(rule1Called);
+        Assert.False(rule2Called);
+    }
+
+    /// <summary>
+    /// Verify that standing rules are evaluated before auto-approval rules.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_StandingRuleTakesPrecedenceOverAutoApprovalRuleAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "MyTool"));
+
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount <= 2)
+                {
+                    return new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+            });
+
+        var heuristicCalled = false;
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => { heuristicCalled = true; return new ValueTask<bool>(true); }]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Call 1: heuristic should be called (no standing rule yet)
+        var response1 = await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+        Assert.True(heuristicCalled);
+        Assert.Equal("Done", response1.Text);
+
+        // Now establish a standing rule by sending AlwaysApprove
+        heuristicCalled = false;
+        callCount = 0;
+        var alwaysApprove = new AlwaysApproveToolApprovalResponseContent(
+            approvalRequest.CreateResponse(approved: true),
+            alwaysApproveTool: true,
+            alwaysApproveToolWithArguments: false);
+
+        // Call 2: standing rule should match first, heuristic should NOT be called
+        var response2 = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, [alwaysApprove])],
+            session);
+        Assert.False(heuristicCalled);
+        Assert.Equal("Done", response2.Text);
+    }
+
+    /// <summary>
+    /// Verify that when a batch contains a mix of heuristic-approved and standing-rule-approved
+    /// requests, the collected approval responses preserve the original request order rather than
+    /// being grouped by approval kind.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_MixedAutoApprovals_PreserveOriginalOrderAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+
+        // Batch ordering: first request is approved by a heuristic, second by a standing rule.
+        var heuristicRequest = new ToolApprovalRequestContent("reqA", new FunctionCallContent("callA", "HeuristicTool"));
+        var standingRequest = new ToolApprovalRequestContent("reqB", new FunctionCallContent("callB", "StandingTool"));
+
+        var batchResponse = new AgentResponse([new ChatMessage(ChatRole.Assistant, [heuristicRequest, standingRequest])]);
+        var finalResponse = new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+
+        var callCount = 0;
+        List<ChatMessage>? secondCallMessages = null;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .Callback<IEnumerable<ChatMessage>, AgentSession?, AgentRunOptions?, CancellationToken>((msgs, _, _, _) =>
+            {
+                callCount++;
+                if (callCount == 2)
+                {
+                    secondCallMessages = msgs.ToList();
+                }
+            })
+            .ReturnsAsync(() => callCount == 1 ? batchResponse : finalResponse);
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "HeuristicTool")]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Establish a standing rule for "StandingTool" via an AlwaysApprove response in the same call.
+        var alwaysApprove = standingRequest.CreateAlwaysApproveToolResponse("User said always");
+
+        // Act — both requests auto-approve (heuristic + standing rule), so the inner agent is re-invoked.
+        var response = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, [alwaysApprove])],
+            session);
+
+        // Assert — inner agent re-called and final response returned.
+        Assert.Equal(2, callCount);
+        Assert.Equal("Done", response.Text);
+
+        // The injected approval responses must preserve the original request order: reqA before reqB,
+        // even though reqA was approved by a heuristic and reqB by a standing rule.
+        Assert.NotNull(secondCallMessages);
+        var injected = secondCallMessages!
+            .SelectMany(m => m.Contents)
+            .OfType<ToolApprovalResponseContent>()
+            .Where(r => r.RequestId is "reqA" or "reqB")
+            .ToList();
+        Assert.Equal(2, injected.Count);
+        Assert.Equal("reqA", injected[0].RequestId);
+        Assert.Equal("reqB", injected[1].RequestId);
+        Assert.All(injected, r => Assert.True(r.Approved));
+    }
+
+    /// <summary>
+    /// Verify that auto-approval rules work in the streaming path.
+    /// </summary>
+    [Fact]
+    public async Task RunStreamingAsync_AutoApprovalRule_ApprovesMatchingToolAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "ReadTool"));
+
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<IAsyncEnumerable<AgentResponseUpdate>>("RunCoreStreamingAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .Returns(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return ToAsyncEnumerableAsync([new AgentResponseUpdate(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return ToAsyncEnumerableAsync([new AgentResponseUpdate(ChatRole.Assistant, "Done")]);
+            });
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        var updates = new List<AgentResponseUpdate>();
+        await foreach (var update in agent.RunStreamingAsync([new ChatMessage(ChatRole.User, "Hi")], session))
+        {
+            updates.Add(update);
+        }
+
+        // Assert — the approval request was auto-approved, inner agent streamed twice
+        Assert.Equal(2, callCount);
+        Assert.Single(updates);
+        Assert.Equal("Done", updates[0].Text);
+    }
+
+    #endregion
 }
@@ -142,6 +142,95 @@ public sealed class ForeachExecutorTest(ITestOutputHelper output) : WorkflowActi
            indexName: "CurrentIndex");
    }

+    [Fact]
+    public async Task ForeachTakeNextWithMultiFieldRecordAsync()
+    {
+        // Arrange
+        const string CurrentValueName = "CurrentValue";
+        this.SetVariableState(CurrentValueName);
+
+        TableDataValue tableValue = DataValue.TableFromRecords(
+            DataValue.RecordFromFields(
+                new KeyValuePair<string, DataValue>("name", new StringDataValue("Alice")),
+                new KeyValuePair<string, DataValue>("role", new StringDataValue("Engineer"))));
+
+        Foreach model = this.CreateModel(
+            displayName: nameof(ForeachTakeNextWithMultiFieldRecordAsync),
+            items: ValueExpression.Literal(tableValue),
+            valueName: CurrentValueName,
+            indexName: null);
+        ForeachExecutor action = new(model, this.State);
+
+        // Act
+        await this.ExecuteAsync(action, ForeachExecutor.Steps.Next(action.Id), action.TakeNextAsync);
+
+        // Assert
+        RecordValue currentValue = Assert.IsType<RecordValue>(this.State.Get(CurrentValueName), exactMatch: false);
+        Assert.Equal("Alice", currentValue.GetField("name").ToObject());
+        Assert.Equal("Engineer", currentValue.GetField("role").ToObject());
+    }
+
+    /// <summary>
+    /// Power Fx wraps scalar array literals such as <c>=[1, 2, 3]</c> as <c>Table({Value: 1}, ...)</c>;
+    /// the loop value must expose the bare scalar, not the single-column wrapper record.
+    /// </summary>
+    [Fact]
+    public async Task ForeachTakeNextWithSingleColumnValueRecordAsync()
+    {
+        // Arrange
+        const string CurrentValueName = "CurrentValue";
+        this.SetVariableState(CurrentValueName);
+
+        TableDataValue tableValue = DataValue.TableFromRecords(
+            DataValue.RecordFromFields(new KeyValuePair<string, DataValue>("Value", new NumberDataValue(1))),
+            DataValue.RecordFromFields(new KeyValuePair<string, DataValue>("Value", new NumberDataValue(2))),
+            DataValue.RecordFromFields(new KeyValuePair<string, DataValue>("Value", new NumberDataValue(3))));
+
+        Foreach model = this.CreateModel(
+            displayName: nameof(ForeachTakeNextWithSingleColumnValueRecordAsync),
+            items: ValueExpression.Literal(tableValue),
+            valueName: CurrentValueName,
+            indexName: null);
+        ForeachExecutor action = new(model, this.State);
+
+        // Act
+        await this.ExecuteAsync(action, ForeachExecutor.Steps.Next(action.Id), action.TakeNextAsync);
+
+        // Assert
+        FormulaValue currentValue = this.State.Get(CurrentValueName);
+        Assert.IsNotType<RecordValue>(currentValue, exactMatch: false);
+        Assert.Equal(1m, currentValue.ToObject());
+    }
+
+    /// <summary>
+    /// Single-field records whose only field is NOT named <c>Value</c> are not Power Fx auto-wraps;
+    /// they are preserved as records so the field name remains accessible inside the loop body.
+    /// </summary>
+    [Fact]
+    public async Task ForeachTakeNextWithSingleFieldNonValueRecordAsync()
+    {
+        // Arrange
+        const string CurrentValueName = "CurrentValue";
+        this.SetVariableState(CurrentValueName);
+
+        TableDataValue tableValue = DataValue.TableFromRecords(
+            DataValue.RecordFromFields(new KeyValuePair<string, DataValue>("name", new StringDataValue("Alice"))));
+
+        Foreach model = this.CreateModel(
+            displayName: nameof(ForeachTakeNextWithSingleFieldNonValueRecordAsync),
+            items: ValueExpression.Literal(tableValue),
+            valueName: CurrentValueName,
+            indexName: null);
+        ForeachExecutor action = new(model, this.State);
+
+        // Act
+        await this.ExecuteAsync(action, ForeachExecutor.Steps.Next(action.Id), action.TakeNextAsync);
+
+        // Assert
+        RecordValue currentValue = Assert.IsType<RecordValue>(this.State.Get(CurrentValueName), exactMatch: false);
+        Assert.Equal("Alice", currentValue.GetField("name").ToObject());
+    }
+
    [Fact]
    public async Task ForeachTakeLastAsync()
    {
@@ -419,6 +419,82 @@ public class MagenticOrchestrationTests
            "final-answer synthesis must see what participants actually said");
    }

+    [Fact]
+    public async Task Participant_Receives_Prior_Participant_Response_Not_InstructionAsync()
+    {
+        // Regression: each participant must see prior participants' *responses* (the running conversation),
+        // not their *instructions*. Previously the orchestrator broadcast the per-round instruction to every
+        // participant (untargeted fan-out) and never broadcast replies, so a later speaker received the earlier
+        // speaker's instruction and never its answer.
+        const string HealthInstruction = "HEALTH_CHECKER_INSTRUCTION_check_framework";
+        const string DatabaseInstruction = "DATABASE_CHECKER_INSTRUCTION_check_database";
+        const string HealthEchoPrefix = "HC_RESPONSE::";
+        const string DatabaseEchoPrefix = "DB_RESPONSE::";
+
+        List<ChatMessage> facts = CreatePlanResponse("Facts");
+        List<ChatMessage> plan = CreatePlanResponse("Plan");
+        List<ChatMessage> round1Ledger = CreateProgressLedgerResponse(
+            isRequestSatisfied: false,
+            isInLoop: false,
+            isProgressBeingMade: true,
+            nextSpeaker: "HealthChecker",
+            instructionOrQuestion: HealthInstruction);
+        List<ChatMessage> round2Ledger = CreateProgressLedgerResponse(
+            isRequestSatisfied: false,
+            isInLoop: false,
+            isProgressBeingMade: true,
+            nextSpeaker: "DatabaseChecker",
+            instructionOrQuestion: DatabaseInstruction);
+        List<ChatMessage> round3Ledger = CreateProgressLedgerResponse(
+            isRequestSatisfied: true,
+            isInLoop: false,
+            isProgressBeingMade: true,
+            nextSpeaker: "DatabaseChecker",
+            instructionOrQuestion: "Done");
+        List<ChatMessage> finalAnswer = CreateFinalAnswerResponse("All systems checked");
+
+        TestReplayAgent manager = new(
+            [facts, plan, round1Ledger, round2Ledger, round3Ledger, finalAnswer],
+            name: "Manager");
+        RecordingEchoAgent healthChecker = new(name: "HealthChecker", prefix: HealthEchoPrefix);
+        RecordingEchoAgent databaseChecker = new(name: "DatabaseChecker", prefix: DatabaseEchoPrefix);
+
+        Workflow workflow = new MagenticWorkflowBuilder(manager)
+            .AddParticipants(healthChecker, databaseChecker)
+            .RequirePlanSignoff(false)
+            .Build();
+
+        WorkflowRunResult runResult = await RunMagenticWorkflowAsync(
+            workflow,
+            [new ChatMessage(ChatRole.User, "Check system health")]);
+
+        runResult.Result.Should().NotBeNull();
+        runResult.Result![0].Text.Should().Contain("All systems checked");
+
+        // Each participant takes exactly one turn.
+        healthChecker.RecordedInputs.Should().ContainSingle();
+        databaseChecker.RecordedInputs.Should().ContainSingle();
+
+        // The first speaker receives its own instruction.
+        List<ChatMessage> healthInput = healthChecker.RecordedInputs[0];
+        healthInput.Should().Contain(m => m.Text.Contains(HealthInstruction), "the first speaker receives its own instruction");
+
+        // The second speaker must see the first speaker's RESPONSE (authored by HealthChecker, carrying the echo
+        // prefix that only the response — not the raw instruction — has), plus its own instruction.
+        List<ChatMessage> databaseInput = databaseChecker.RecordedInputs[0];
+        databaseInput.Should().Contain(
+            m => m.AuthorName == "HealthChecker" && m.Text.Contains(HealthEchoPrefix),
+            "the next speaker must receive the prior participant's response (the running conversation)");
+        databaseInput.Should().Contain(m => m.Text.Contains(DatabaseInstruction),
+            "the next speaker must receive its own instruction");
+
+        // The leaked-instruction bug: the second speaker must not receive HealthChecker's instruction as a
+        // bare message (it should only appear, if at all, embedded in HealthChecker's prefixed response).
+        databaseInput.Should().NotContain(
+            m => m.AuthorName != "HealthChecker" && m.Text.Trim() == HealthInstruction,
+            "the prior speaker's instruction must not leak into the next speaker's context as a standalone message");
+    }
+
    [Fact]
    public async Task PlanReview_Revised_Triggers_ReplanAsync()
    {
@@ -0,0 +1,37 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+using System.Collections.Generic;
+using System.Linq;
+using System.Runtime.CompilerServices;
+using System.Threading;
+using Microsoft.Extensions.AI;
+
+namespace Microsoft.Agents.AI.Workflows.UnitTests;
+
+/// <summary>
+/// A <see cref="TestEchoAgent"/> that records the input messages it receives on each call.
+/// Used by tests that need to assert what context a participant was actually handed - for example,
+/// that a later speaker sees prior participants' <em>responses</em> (the running conversation) rather
+/// than their <em>instructions</em>.
+/// </summary>
+internal sealed class RecordingEchoAgent(string? id = null, string? name = null, string? prefix = null)
+    : TestEchoAgent(id, name, prefix)
+{
+    public List<List<ChatMessage>> RecordedInputs { get; } = [];
+
+    protected override async IAsyncEnumerable<AgentResponseUpdate> RunCoreStreamingAsync(
+        IEnumerable<ChatMessage> messages,
+        AgentSession? session = null,
+        AgentRunOptions? options = null,
+        [EnumeratorCancellation] CancellationToken cancellationToken = default)
+    {
+        // Materialize once so the deferred input is recorded and replayed identically.
+        List<ChatMessage> recorded = messages.ToList();
+        this.RecordedInputs.Add(recorded);
+
+        await foreach (AgentResponseUpdate update in base.RunCoreStreamingAsync(recorded, session, options, cancellationToken))
+        {
+            yield return update;
+        }
+    }
+}
@@ -7,6 +7,26 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0

 ## [Unreleased]

+## [1.8.1] - 2026-06-09
+
+### Added
+- **agent-framework-core**: Add MCP client OTel spans per GenAI semantic conventions ([#6349](https://github.com/microsoft/agent-framework/pull/6349))
+- **agent-framework-core**: Add MCP long-running task support ([#6319](https://github.com/microsoft/agent-framework/pull/6319))
+
+### Changed
+- **agent-framework-claude**: Bump `claude-agent-sdk` to 0.2.87 ([#6248](https://github.com/microsoft/agent-framework/pull/6248))
+- **agent-framework-core**: Document checkpoint storage security model and deserialization trust boundaries ([#6295](https://github.com/microsoft/agent-framework/pull/6295))
+- **agent-framework-azurefunctions**: Document checkpoint storage security model and deserialization trust boundaries ([#6295](https://github.com/microsoft/agent-framework/pull/6295))
+
+### Fixed
+- **agent-framework-core**: Filter MCP tool kwargs to declared params via allowlist ([#6399](https://github.com/microsoft/agent-framework/pull/6399))
+- **agent-framework-core**: Fix per-service-call history persistence with server-storing clients ([#6310](https://github.com/microsoft/agent-framework/pull/6310))
+- **agent-framework-openai**: Use `getattr` for non-OpenAI provider response compatibility ([#6270](https://github.com/microsoft/agent-framework/pull/6270))
+- **agent-framework-foundry-hosting**: Refactor workflow-as-agent pending request handling ([#6259](https://github.com/microsoft/agent-framework/pull/6259))
+- **agent-framework-gemini**: Make Gemini honor declarative `outputSchema`, not just JSON mode ([#5893](https://github.com/microsoft/agent-framework/pull/5893))
+- **agent-framework-mem0**: Isolate entity retrieval and correct `app_id` payload ([#6242](https://github.com/microsoft/agent-framework/pull/6242))
+- **agent-framework-ag-ui**: Match AG-UI approval responses to requested arguments ([#6376](https://github.com/microsoft/agent-framework/pull/6376))
+
 ## [1.8.0] - 2026-06-04

 ### Added
@@ -1169,7 +1189,8 @@ Release candidate for **agent-framework-core** and **agent-framework-azure-ai**

 For more information, see the [announcement blog post](https://devblogs.microsoft.com/foundry/introducing-microsoft-agent-framework-the-open-source-engine-for-agentic-ai-apps/).

-[Unreleased]: https://github.com/microsoft/agent-framework/compare/python-1.8.0...HEAD
+[Unreleased]: https://github.com/microsoft/agent-framework/compare/python-1.8.1...HEAD
+[1.8.1]: https://github.com/microsoft/agent-framework/compare/python-1.8.0...python-1.8.1
 [1.8.0]: https://github.com/microsoft/agent-framework/compare/python-1.7.0...python-1.8.0
 [1.7.0]: https://github.com/microsoft/agent-framework/compare/python-1.6.0...python-1.7.0
 [1.6.0]: https://github.com/microsoft/agent-framework/compare/python-1.5.0...python-1.6.0
@@ -9,7 +9,7 @@ from typing import Any, cast
 from ag_ui.core import BaseEvent
 from agent_framework import SupportsAgentRun

-from ._agent_run import run_agent_stream
+from ._agent_run import PendingApprovalEntry, run_agent_stream


 class AgentConfig:
@@ -107,7 +107,7 @@ class AgentFrameworkAgent:
        # Populated when approval requests are emitted; consumed when responses arrive.
        # Prevents bypass, function name spoofing, and replay attacks.
        # Bounded to prevent unbounded growth from abandoned approval requests.
-        self._pending_approvals: OrderedDict[str, str] = OrderedDict()
+        self._pending_approvals: OrderedDict[str, PendingApprovalEntry] = OrderedDict()
        self._pending_approvals_max_size: int = 10_000

    async def run(
@@ -8,7 +8,7 @@ import json
 import logging
 import uuid
 from collections.abc import AsyncIterable, Awaitable
-from typing import TYPE_CHECKING, Any, cast
+from typing import TYPE_CHECKING, Any, TypedDict, cast

 from ag_ui.core import (
    BaseEvent,
@@ -56,6 +56,7 @@ from ._run_common import (
    _stringify_tool_result,  # type: ignore
 )
 from ._utils import (
+    canonical_function_arguments,
    convert_agui_tools_to_agent_framework,
    generate_event_id,
    get_conversation_id_from_update,
@@ -407,7 +408,33 @@ def _make_approval_tool_result_events(resolved_approval_results: list[Content])
    return events


-def _evict_oldest_approvals(registry: dict[str, str], max_size: int = 10_000) -> None:
+class _PendingApproval(TypedDict):
+    """Pending approval details for a requested function call."""
+
+    name: str
+    arguments: str | None
+
+
+PendingApprovalEntry = _PendingApproval | str
+
+
+def _make_pending_approval_entry(name: str, arguments: str | None) -> _PendingApproval:
+    return {"name": name, "arguments": arguments}
+
+
+def _pending_approval_name(entry: PendingApprovalEntry) -> str | None:
+    if isinstance(entry, str):
+        return entry
+    return entry["name"]
+
+
+def _pending_approval_arguments(entry: PendingApprovalEntry) -> str | None:
+    if isinstance(entry, str):
+        return None
+    return entry["arguments"]
+
+
+def _evict_oldest_approvals(registry: dict[str, PendingApprovalEntry], max_size: int = 10_000) -> None:
    """Evict the oldest entries from the pending-approvals registry (LRU).

    Only effective when *registry* is an ``OrderedDict``;  plain dicts are
@@ -427,7 +454,7 @@ async def _resolve_approval_responses(
    tools: list[Any],
    agent: SupportsAgentRun,
    run_kwargs: dict[str, Any],
-    pending_approvals: dict[str, str] | None = None,
+    pending_approvals: dict[str, PendingApprovalEntry] | None = None,
    thread_id: str = "",
 ) -> list[Content]:
    """Execute approved function calls and replace approval content with results.
@@ -480,7 +507,8 @@ async def _resolve_approval_responses(
                invalid_ids.add(resp_id)
                continue

-            pending_name = pending_approvals[registry_key]
+            pending_entry = pending_approvals[registry_key]
+            pending_name = _pending_approval_name(pending_entry)
            if resp_name != pending_name:
                logger.warning(
                    "Rejected approval response id=%s: function name mismatch (response=%s, pending=%s)",
@@ -491,6 +519,16 @@ async def _resolve_approval_responses(
                invalid_ids.add(resp_id)
                continue

+            pending_arguments = _pending_approval_arguments(pending_entry)
+            response_arguments = canonical_function_arguments(resp.function_call)
+            if pending_arguments is not None and response_arguments != pending_arguments:
+                logger.warning(
+                    "Rejected approval response id=%s: function arguments mismatch",
+                    resp_id,
+                )
+                invalid_ids.add(resp_id)
+                continue
+
            # Valid — consume entry to prevent replay
            del pending_approvals[registry_key]
            if resp.approved:
@@ -714,7 +752,7 @@ async def run_agent_stream(
    input_data: dict[str, Any],
    agent: SupportsAgentRun,
    config: AgentConfig,
-    pending_approvals: dict[str, str] | None = None,
+    pending_approvals: dict[str, PendingApprovalEntry] | None = None,
 ) -> AsyncGenerator[BaseEvent]:
    """Run agent and yield AG-UI events.

@@ -917,7 +955,10 @@ async def run_agent_stream(
            # Register pending approval requests so we can validate responses later
            if content_type == "function_approval_request" and pending_approvals is not None:
                if content.id and content.function_call and content.function_call.name:
-                    pending_approvals[f"{thread_id}:{content.id}"] = content.function_call.name
+                    pending_approvals[f"{thread_id}:{content.id}"] = _make_pending_approval_entry(
+                        content.function_call.name,
+                        canonical_function_arguments(content.function_call),
+                    )
                    # Evict oldest entries if the registry exceeds a safe bound (LRU)
                    _evict_oldest_approvals(pending_approvals, max_size=10_000)
                else:
@@ -56,6 +56,22 @@ def safe_json_parse(value: Any) -> dict[str, Any] | None:
    return None


+def canonical_function_arguments(function_call: Any) -> str | None:
+    """Return a stable representation of function-call arguments."""
+    if function_call is None:
+        return None
+
+    try:
+        parsed_arguments = function_call.parse_arguments()
+    except Exception:
+        parsed_arguments = getattr(function_call, "arguments", None)
+
+    if parsed_arguments is None:
+        parsed_arguments = {}
+
+    return json.dumps(make_json_safe(parsed_arguments), sort_keys=True, separators=(",", ":"))
+
+
 def get_role_value(message: Any) -> str:
    """Extract role string from a message object.

@@ -35,7 +35,7 @@ from ._run_common import (
    _extract_resume_payload,
    _normalize_resume_interrupts,
 )
-from ._utils import generate_event_id, make_json_safe
+from ._utils import canonical_function_arguments, generate_event_id, make_json_safe

 logger = logging.getLogger(__name__)

@@ -324,6 +324,29 @@ def _coerce_response_for_request(request_event: Any, value: Any) -> Any | None:
    return candidate


+def _approval_response_matches_request(request_id: str, request_event: Any, response: Any) -> bool:
+    """Check whether an approval response matches the pending approval request."""
+    request_data = getattr(request_event, "data", None)
+    if not isinstance(request_data, Content) or request_data.type != "function_approval_request":
+        return True
+
+    if not isinstance(response, Content) or response.type != "function_approval_response":
+        return False
+
+    if str(getattr(response, "id", "")) != request_id:
+        return False
+
+    request_call = getattr(request_data, "function_call", None)
+    response_call = getattr(response, "function_call", None)
+    if request_call is None or response_call is None:
+        return False
+
+    if getattr(response_call, "name", None) != getattr(request_call, "name", None):
+        return False
+
+    return canonical_function_arguments(response_call) == canonical_function_arguments(request_call)
+
+
 def _single_pending_response_from_value(pending_events: dict[str, Any], value: Any) -> dict[str, Any]:
    """Map a scalar resume payload to the single pending request (if unambiguous)."""
    if value is None or len(pending_events) != 1:
@@ -343,6 +366,13 @@ def _single_pending_response_from_value(pending_events: dict[str, Any], value: A
        )
        return {}

+    if not _approval_response_matches_request(str(request_id), request_event, coerced_value):
+        logger.info(
+            "Ignoring pending request response for request_id=%s: approval response does not match pending request",
+            request_id,
+        )
+        return {}
+
    return {str(request_id): coerced_value}


@@ -372,6 +402,12 @@ def _coerce_responses_for_pending_requests(
                _response_type_name(request_event),
            )
            continue
+        if not _approval_response_matches_request(request_key, request_event, coerced_value):
+            logger.info(
+                "Ignoring resume response for request_id=%s: approval response does not match pending request",
+                request_key,
+            )
+            continue
        normalized[request_key] = coerced_value
    return normalized

@@ -1,6 +1,6 @@
 [project]
 name = "agent-framework-ag-ui"
-version = "1.0.0rc3"
+version = "1.0.0rc4"
 description = "AG-UI protocol integration for Agent Framework"
 readme = "README.md"
 license-files = ["LICENSE"]
@@ -22,7 +22,7 @@ classifiers = [
  "Typing :: Typed",
 ]
 dependencies = [
-    "agent-framework-core>=1.6.0,<2",
+    "agent-framework-core>=1.8.1,<2",
    "ag-ui-protocol>=0.1.16,<0.2",
    "fastapi>=0.115.0,<0.133.1",
    "uvicorn[standard]>=0.30.0,<1"
@@ -1407,6 +1407,92 @@ async def test_fabricated_rejection_without_pending_approval_is_blocked(streamin
                assert False, "Fabricated rejection response leaked as function_result into LLM messages"


+async def test_approval_argument_mismatch_is_blocked(streaming_chat_client_stub):
+    """An approval response must not execute changed arguments for the pending call."""
+    from agent_framework import tool
+    from agent_framework.ag_ui import AgentFrameworkAgent
+
+    executed_args: list[dict[str, Any]] = []
+
+    @tool(
+        name="update_record",
+        description="Update a record",
+        approval_mode="always_require",
+    )
+    def update_record(record_id: str, value: str) -> str:
+        executed_args.append({"record_id": record_id, "value": value})
+        return f"updated {record_id} to {value}"
+
+    async def stream_fn_approval(
+        messages: MutableSequence[Message], options: ChatOptions, **kwargs: Any
+    ) -> AsyncIterator[ChatResponseUpdate]:
+        yield ChatResponseUpdate(
+            contents=[
+                Content.from_function_call(
+                    name="update_record",
+                    call_id="call_update_001",
+                    arguments={"record_id": "alpha", "value": "approved"},
+                )
+            ]
+        )
+
+    wrapper = AgentFrameworkAgent(
+        agent=Agent(
+            client=streaming_chat_client_stub(stream_fn_approval),
+            name="test_agent",
+            instructions="Test",
+            tools=[update_record],
+        )
+    )
+    thread_id = "thread-argument-mismatch-test"
+
+    events1: list[Any] = []
+    async for event in wrapper.run({"thread_id": thread_id, "messages": [{"role": "user", "content": "update"}]}):
+        events1.append(event)
+
+    assert any("call_update_001" in k for k in wrapper._pending_approvals)
+
+    async def stream_fn_post(
+        messages: MutableSequence[Message], options: ChatOptions, **kwargs: Any
+    ) -> AsyncIterator[ChatResponseUpdate]:
+        yield ChatResponseUpdate(contents=[Content.from_text(text="Done")])
+
+    wrapper.agent = Agent(
+        client=streaming_chat_client_stub(stream_fn_post),
+        name="test_agent",
+        instructions="Test",
+        tools=[update_record],
+    )
+
+    turn2_input: dict[str, Any] = {
+        "thread_id": thread_id,
+        "messages": [
+            {
+                "role": "user",
+                "content": "approve",
+                "function_approvals": [
+                    {
+                        "id": "call_update_001",
+                        "call_id": "call_update_001",
+                        "name": "update_record",
+                        "approved": True,
+                        "arguments": {"record_id": "beta", "value": "changed"},
+                    }
+                ],
+            },
+        ],
+    }
+
+    events2: list[Any] = []
+    async for event in wrapper.run(turn2_input):
+        events2.append(event)
+
+    assert executed_args == []
+    assert any("call_update_001" in k for k in wrapper._pending_approvals), (
+        "Pending approval should be preserved after argument mismatch for legitimate retry"
+    )
+
+
 async def test_state_update_end_to_end_via_real_tool_invocation(streaming_chat_client_stub):
    """End-to-end coverage for issue #3167: a real ``@tool`` returning ``state_update`` must
    emit a deterministic STATE_SNAPSHOT through the full pipeline.
@@ -1352,6 +1352,70 @@ async def test_workflow_run_approval_via_messages_approved() -> None:
    assert not resumed_finished.get("interrupt")


+async def test_workflow_run_approval_argument_mismatch_keeps_interrupt_pending() -> None:
+    """Workflow approval responses must not resume with changed function arguments."""
+
+    handled_responses: list[dict[str, Any]] = []
+
+    class ApprovalExecutor(Executor):
+        def __init__(self) -> None:
+            super().__init__(id="approval_executor")
+
+        @handler
+        async def start(self, message: Any, ctx: WorkflowContext) -> None:
+            del message
+            function_call = Content.from_function_call(
+                call_id="refund-call",
+                name="submit_refund",
+                arguments={"order_id": "12345", "amount": "$89.99"},
+            )
+            approval_request = Content.from_function_approval_request(id="approval-1", function_call=function_call)
+            await ctx.request_info(approval_request, Content, request_id="approval-1")
+
+        @response_handler
+        async def handle_approval(self, original_request: Content, response: Content, ctx: WorkflowContext) -> None:
+            del original_request
+            if response.function_call is not None:
+                handled_responses.append(response.function_call.parse_arguments() or {})
+            await ctx.yield_output("handled")
+
+    workflow = WorkflowBuilder(start_executor=ApprovalExecutor()).build()
+    first_events = [
+        event async for event in run_workflow_stream({"messages": [{"role": "user", "content": "go"}]}, workflow)
+    ]
+    first_finished = [event for event in first_events if event.type == "RUN_FINISHED"][0].model_dump()
+    interrupt_payload = cast(list[dict[str, Any]], first_finished.get("interrupt"))
+    assert isinstance(interrupt_payload, list) and len(interrupt_payload) == 1
+
+    resumed_events = [
+        event
+        async for event in run_workflow_stream(
+            {
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": "",
+                        "function_approvals": [
+                            {
+                                "approved": True,
+                                "id": "approval-1",
+                                "call_id": "refund-call",
+                                "name": "submit_refund",
+                                "arguments": {"order_id": "99999", "amount": "$1000.00"},
+                            }
+                        ],
+                    }
+                ],
+            },
+            workflow,
+        )
+    ]
+
+    assert handled_responses == []
+    resumed_finished = [event for event in resumed_events if event.type == "RUN_FINISHED"][0].model_dump()
+    assert resumed_finished.get("interrupt")
+
+
 async def test_workflow_run_approval_via_messages_denied() -> None:
    """Denied approval response sent via messages (function_approvals) should satisfy the pending request."""

@@ -14,6 +14,24 @@ This module adds:
 - reconstruct_to_type: for HITL responses where external data (without type markers)
  needs to be reconstructed to a known type
 - resolve_type: resolves 'module:class' type keys to Python types
+
+Security Model
+--------------
+The underlying Azure Durable Functions storage (Azure Storage account) is the
+trusted persistence layer for serialized checkpoint data.  The
+``RestrictedUnpickler`` in the core encoding module provides defense-in-depth
+type filtering, but checkpoint storage itself must be properly access-controlled:
+
+- Ensure the Azure Storage account used by Durable Functions is not publicly
+  writable and uses appropriate RBAC / shared-access policies.
+- Never route untrusted user input directly into ``deserialize_value`` without
+  first calling :func:`strip_pickle_markers` to neutralize injection of
+  pickle markers into the data path.
+- Configure your checkpoint storage with ``allowed_checkpoint_types`` (or call
+  ``decode_checkpoint_value(..., allowed_types=...)`` directly) to restrict the set of types that can be deserialized.
+
+See :mod:`agent_framework._workflows._checkpoint_encoding` for the full
+security model documentation.
 """

 from __future__ import annotations
@@ -4,7 +4,7 @@ description = "Azure Functions integration for Microsoft Agent Framework."
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.0.0b260604"
+version = "1.0.0b260609"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -22,7 +22,7 @@ classifiers = [
  "Typing :: Typed",
 ]
 dependencies = [
-    "agent-framework-core>=1.8.0,<2",
+    "agent-framework-core>=1.8.1,<2",
    "agent-framework-durabletask>=1.0.0b260604,<2",
    "azure-functions>=1.24.0,<2",
    "azure-functions-durable>=1.3.1,<2",
@@ -221,9 +221,31 @@ class ClaudeAgentOptions(TypedDict, total=False):
    thinking: ThinkingConfig
    """Extended thinking configuration (adaptive, enabled, or disabled)."""

-    effort: Literal["low", "medium", "high", "max"]
+    effort: Literal["low", "medium", "high", "xhigh", "max"]
    """Effort level for thinking depth."""

+    skills: list[str] | Literal["all"]
+    """Skills to enable for the main session. Use ``"all"`` for every discovered skill,
+    a list of named skills, or ``[]`` to suppress all skills."""
+
+    session_id: str
+    """Use a specific session ID (must be a valid UUID) instead of auto-generated."""
+
+    task_budget: dict[str, int]
+    """API-side task budget in tokens for pacing tool use."""
+
+    include_hook_events: bool
+    """When True, hook lifecycle events are emitted in the message stream."""
+
+    strict_mcp_config: bool
+    """When True, only use MCP servers passed via ``mcp_servers``, ignoring all others."""
+
+    continue_conversation: bool
+    """Continue the most recent conversation instead of starting a new one."""
+
+    fork_session: bool
+    """When True, resumed sessions fork to a new session ID."""
+
    on_function_approval: FunctionApprovalCallback
    """Approval callback for ``FunctionTool`` instances declared with
    ``approval_mode="always_require"``. The callback is awaited (sync or async)
@@ -4,7 +4,7 @@ description = "Claude Agent SDK integration for Microsoft Agent Framework."
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.0.0b260521"
+version = "1.0.0b260609"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -23,8 +23,8 @@ classifiers = [
  "Typing :: Typed",
 ]
 dependencies = [
-    "agent-framework-core>=1.6.0,<2",
-    "claude-agent-sdk>=0.1.36,<0.1.49",
+    "agent-framework-core>=1.8.1,<2",
+    "claude-agent-sdk>=0.1.36,<0.3",
 ]

 [tool.uv]
@@ -76,6 +76,21 @@ agent_framework/
 - **`SkillScriptRunner`** - Protocol for file-based script execution. Any callable matching `(skill, script, args) -> Any` satisfies it. Code-defined scripts do not use a runner.
 - **`SkillsProvider`** - Context provider (extends `ContextProvider`) that discovers file-based skills from `SKILL.md` files and/or accepts code-defined `Skill` instances. Follows progressive disclosure: advertise → load → read resources / run scripts.

+### Model Context Protocol (`_mcp.py`)
+
+- **`MCPTool`** - Base wrapper that owns the MCP `ClientSession` and exposes the remote server's tools as `FunctionTool`s.
+- **`MCPStdioTool`** / **`MCPStreamableHTTPTool`** / **`MCPWebsocketTool`** - Transport-specific subclasses.
+- **Argument allowlist (`_prepare_call_kwargs`)** - Before each `tools/call`, kwargs are filtered to an **allowlist** built from the tool's declared parameters (`inputSchema.properties`) plus any user-configured extras. Framework runtime kwargs injected through the function-invocation pipeline (e.g. `thread`, `conversation_id`, `chat_options`, `options`, `response_format`) are stripped by default rather than forwarded. A tool that declares no usable `properties` (including schemas with `additionalProperties: true`) forwards only the configured extras. The `_MCP_FRAMEWORK_DENYLIST` is a safety net for framework-named params a server *declares* in its schema (those are dropped); names explicitly opted in via `additional_tool_argument_names` always win. The reserved `_meta` key is extracted as MCP request metadata, never forwarded as an argument.
+- **`additional_tool_argument_names`** (constructor arg on all `MCPTool` subclasses) - Opt extra argument names back into the allowlist. Accepts a `Sequence[str]` (applied to every tool) or a `Mapping[str, Sequence[str]]` keyed by **remote tool name**, where the reserved key `"*"` denotes global extras. It is configured only in user code at construction; there is **no per-call/runtime override**, so a model-issued tool call cannot change which names pass through. To use a server that accepts `additionalProperties: true`, list the extra names here and then either (1) manually extend that tool's `inputSchema` (via the `.functions` list after connecting) so the model is prompted to supply them, or (2) supply the values yourself via `function_invocation_kwargs`. If a name is supplied by both the model and `function_invocation_kwargs`, the model-supplied value wins.
+- **`MCPTaskOptions`** (experimental, `MCP_LONG_RUNNING_TASKS` feature, **frozen**) - Per-tool-instance options controlling the SEP-2663 long-running task lifecycle. When the server advertises a tool with `execution.taskSupport == "required"`, `MCPTool.call_tool` transparently routes through `call_tool_as_task`, which sends an augmented `tools/call`, polls `tasks/get` until terminal, and reinterprets `tasks/result` as a normal `CallToolResult`. Instances are immutable; replace via `MCPTool.task_options = MCPTaskOptions(...)`. Fields:
+  - `default_ttl: timedelta | None` — forwarded to the server as `params.task.ttl` (milliseconds). When `None`, the server's default applies.
+  - `cancel_remote_task_on_local_cancellation: bool = True` — only gates the `CancelledError` path. Abandonment paths (see below) always cancel.
+  - `max_task_wait: timedelta | None` — client-side deadline for the whole post-create lifecycle (poll + result fetch). When exceeded, raises `ToolExecutionException` and fires a best-effort `tasks/cancel`. `None` (default) means no client-side bound. Bounds sleeps, sends, AND reconnects via `asyncio.wait_for`.
+- **Permissive fallback**: servers that ignore the augmentation (return `CallToolResult` directly) or reject the unknown `task` field with `METHOD_NOT_FOUND` / `INVALID_PARAMS` fall back to the plain `session.call_tool(...)` path so legacy servers keep working. An unparseable success response (server accepted the augmented call but returned a payload that is neither `CreateTaskResult` nor `CallToolResult`) **does not** fall back — it raises `ToolExecutionException` to avoid double-executing a side-effecting tool.
+- **Submit-vs-track reconnect policy**: a dropped connection before a `task_id` is known raises `ToolExecutionException("connection lost; task state unknown")` without re-issuing the augmented `tools/call`, so a server that accepted the request but lost the response cannot be made to start the same operation twice; once a `task_id` exists, `tasks/get` / `tasks/result` reconnect once and retry against the same id (a shared `_send_with_one_reconnect` helper).
+- **Cancel-on-abandonment vs terminal failure**: any path where the remote task may still be running (max-wait exceeded, hard `McpError` in poll, malformed `tasks/get`, second connection loss in poll/fetch, reconnect failure) fires best-effort `tasks/cancel` before raising. Terminal failures (`failed`/`cancelled`/`input_required` server-side, `completed+isError`, malformed `tasks/result` after server completed) do **not** cancel — the server is already done. `_MCPTaskAbandoned` is the private marker distinguishing the two.
+- **Transient poll retry**: a slow `tasks/get` that surfaces as `McpError(code=408 REQUEST_TIMEOUT)` is retried (bounded by `max_task_wait`). All other non-connection `McpError`s during poll are treated as abandonment. `tasks/result` does not get transient retry — the server has already completed, so a slow payload fetch is anomalous.
+
 ### File Access Harness (`_harness/_file_access.py`)

 - **`AgentFileStore`** - Abstract async store backing the file-access harness. Implementations expose `write_file`, `read_file`, `delete_file`, `list_files`, `file_exists`, `search_files`, and `create_directory` over forward-slash relative paths.
@@ -124,7 +124,7 @@ from ._harness._todo import (
    TodoSessionStore,
    TodoStore,
 )
-from ._mcp import MCPStdioTool, MCPStreamableHTTPTool, MCPWebsocketTool
+from ._mcp import MCPStdioTool, MCPStreamableHTTPTool, MCPTaskOptions, MCPWebsocketTool
 from ._middleware import (
    AgentContext,
    AgentMiddleware,
@@ -444,12 +444,13 @@ __all__ = [
    "InlineSkillResource",
    "InlineSkillScript",
    "LocalEvaluator",
-    "MCPStdioTool",
-    "MCPStreamableHTTPTool",
-    "MCPWebsocketTool",
    "MCPSkill",
    "MCPSkillResource",
    "MCPSkillsSource",
+    "MCPStdioTool",
+    "MCPStreamableHTTPTool",
+    "MCPTaskOptions",
+    "MCPWebsocketTool",
    "MemoryContextProvider",
    "MemoryFileStore",
    "MemoryIndexEntry",
@@ -92,12 +92,16 @@ OptionsCoT = TypeVar(
 def _merge_options(base: dict[str, Any], override: dict[str, Any]) -> dict[str, Any]:
    """Merge two options dicts, with override values taking precedence.

+    ``None`` is treated as "unset": ``None`` overrides are skipped so they don't clobber a base
+    value, and the merged result is stripped of any remaining ``None`` values in a final pass so
+    unset options are never forwarded (e.g. an unset ``store`` is left for the service to default).
+
    Args:
        base: The base options dict.
        override: The override options dict (values take precedence).

    Returns:
-        A new merged options dict.
+        A new merged options dict containing no ``None`` values.
    """
    result = dict(base)

@@ -123,7 +127,7 @@ def _merge_options(base: dict[str, Any], override: dict[str, Any]) -> dict[str,
            result["instructions"] = f"{result['instructions']}\n{value}"
        else:
            result[key] = value
-    return result
+    return {key: value for key, value in result.items() if value is not None}


 def _sanitize_agent_name(agent_name: str | None) -> str | None:
@@ -460,6 +464,9 @@ class BaseAgent(SerializationMixin):
        if provider_session is None and self.context_providers:
            provider_session = AgentSession()

+        # When per-service-call persistence is enabled, the per-service-call middleware owns
+        # HistoryProvider persistence (in both the local and service-managed cases), so skip
+        # them on the once-per-run path to avoid double persistence.
        per_service_call_history_required = self.require_per_service_call_history_persistence and any(
            isinstance(provider, HistoryProvider) for provider in self.context_providers
        )
@@ -686,11 +693,16 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]
            description: A brief description of the agent's purpose.
            context_providers: Context providers to include during agent invocation.
            middleware: List of middleware to intercept agent and function invocations.
-            require_per_service_call_history_persistence: When True, history providers are invoked
-                around each model call instead of once per ``run()`` when the service
-                is not already storing history. If service-side storage is active for
-                the run, the agent skips local history providers and relies on the
-                service-managed conversation instead.
+            require_per_service_call_history_persistence: When True (and a HistoryProvider is
+                present), the provider always persists history via per-service-call middleware,
+                regardless of whether the client stores history server-side. If the client does
+                not store history, the middleware also loads providers around each model call and
+                drives the function loop with a local conversation; if it does, loading is skipped
+                (the service-managed conversation is the source of truth) and the middleware only
+                persists. A warning is logged for providers with ``load_messages=True`` when
+                loading is skipped because service-side storage is active. When no HistoryProvider
+                is present, this flag has no effect (no middleware is installed and nothing is
+                persisted).
            default_options: A TypedDict containing chat options. When using a typed agent like
                ``Agent[OpenAIChatOptions]``, this enables IDE autocomplete for
                provider-specific options including temperature, max_tokens, model,
@@ -791,22 +803,20 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]
        self,
        *,
        session: AgentSession | None,
-        options: Mapping[str, Any] | None,
+        conversation_id: str | None,
        service_stores_history: bool,
    ) -> list[HistoryProvider]:
        history_providers = self._get_history_providers()
        if not self.require_per_service_call_history_persistence or not history_providers:
            return []

-        conversation_id = (
-            session.service_session_id
-            if session and session.service_session_id
-            else cast(str | None, (options or {}).get("conversation_id") or self.default_options.get("conversation_id"))
-        )
-        if service_stores_history:
-            return []
-
-        if conversation_id is not None:
+        # A live service-managed session id takes precedence over the resolved conversation id.
+        if session and session.service_session_id:
+            conversation_id = session.service_session_id
+        # Without service-side storage the middleware persists locally and drives the function
+        # loop with a local sentinel, which cannot be reconciled with an existing service-managed
+        # conversation. When the service stores history, an existing conversation id is expected.
+        if conversation_id is not None and not service_stores_history:
            raise AgentInvalidRequestException(
                "require_per_service_call_history_persistence cannot be used "
                "with an existing service-managed conversation."
@@ -1167,18 +1177,34 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]

        input_messages = normalize_messages(messages)

-        # `store` in runtime or agent options takes precedence over client-level storage
-        # indicators. An explicit `store=False` forces local (in-memory) history injection,
-        # even if the client is configured to use service-side storage by default.
-        store_ = opts.get("store", self.default_options.get("store", getattr(self.client, "STORES_BY_DEFAULT", False)))
+        # Combine agent-level defaults with runtime options up front so the decisions below read
+        # `store` from a single place rather than introspecting both dicts. _merge_options applies
+        # the same precedence used for the actual client call (runtime wins; unset/None falls back
+        # to the agent default).
+        effective_options = _merge_options(self.default_options, opts)
+
+        # `store` in runtime or agent options takes precedence over the client's default
+        # storage behavior. An explicit `store=False` forces local (in-memory) history
+        # injection even when the client stores server-side by default; an explicit
+        # `store=True` forces service-side storage. A `store=None`/unset value means the
+        # service falls back to its own default.
+        explicit_store = effective_options.get("store")
+        # Internal behavior hint: will the service own history for this run? Only when the
+        # user left `store` unset do we fall back to the client's STORES_BY_DEFAULT.
+        service_stores_history = (
+            explicit_store if explicit_store is not None else getattr(self.client, "STORES_BY_DEFAULT", False)
+        )
+        # Resolve conversation_id from the same combined view so an agent-level default is honored
+        # when the runtime omits it (a live session id still takes precedence below).
+        effective_conversation_id = effective_options.get("conversation_id")
        # Auto-inject InMemoryHistoryProvider when session is provided, no context providers
        # registered, and no service-side storage indicators
        if (
            session is not None
            and not self.context_providers
            and not session.service_session_id
-            and not opts.get("conversation_id")
-            and not store_
+            and not effective_conversation_id
+            and not service_stores_history
        ):
            self.context_providers.append(InMemoryHistoryProvider())

@@ -1188,10 +1214,30 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]

        per_service_call_history_providers = self._resolve_per_service_call_history_providers(
            session=active_session,
-            options=opts,
-            service_stores_history=bool(store_),
+            conversation_id=effective_conversation_id,
+            service_stores_history=service_stores_history,
        )

+        # When require_per_service_call_history_persistence is set together with a
+        # HistoryProvider, the per-service-call middleware (installed below) always persists
+        # the provider. ``service_stores_history`` only selects how the middleware behaves:
+        # - service does not store: the middleware also loads providers and drives the function
+        #   loop with a local sentinel conversation id, or
+        # - service stores: the middleware skips loading (the service owns history) and simply
+        #   persists each service call while the real conversation id flows through.
+        # In the service-managed case loading is skipped, so warn for providers that expect to load.
+        history_providers = self._get_history_providers()
+        if self.require_per_service_call_history_persistence and history_providers and service_stores_history:
+            for provider in history_providers:
+                if provider.load_messages:
+                    logger.warning(
+                        "HistoryProvider '%s' has load_messages=True but the chat client stores history "
+                        "server-side; skipping local history load and relying on the service-managed "
+                        "conversation. Set store=False to load from the provider, or load_messages=False "
+                        "to silence this warning.",
+                        provider.source_id,
+                    )
+
        session_context, chat_options = await self._prepare_session_and_messages(
            session=active_session,
            input_messages=input_messages,
@@ -1265,8 +1311,8 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]
        }
        if model is not None:
            run_opts["model"] = model
-        # Remove None values and merge with chat_options
-        run_opts = {k: v for k, v in run_opts.items() if v is not None}
+        # _merge_options strips unset (None) options, so e.g. an unset `store` is not forwarded
+        # and the service decides its own default.
        co = _merge_options(chat_options, run_opts)

        # Build session_messages from session context: context messages + input messages
@@ -1280,6 +1326,7 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]
                agent=self,
                session=active_session,
                providers=per_service_call_history_providers,
+                service_stores_history=service_stores_history,
            )
            existing_middleware = effective_client_kwargs.get("middleware")
            if isinstance(existing_middleware, Sequence) and not isinstance(existing_middleware, (str, bytes)):
@@ -1319,7 +1366,7 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]
            "input_messages": input_messages,
            "session_messages": session_messages,
            "agent_name": agent_name,
-            "suppress_response_id": bool(per_service_call_history_providers),
+            "suppress_response_id": bool(per_service_call_history_providers) and not service_stores_history,
            "chat_options": co,
            "compaction_strategy": compaction_strategy or self.compaction_strategy,
            "tokenizer": tokenizer or self.tokenizer,
@@ -1413,11 +1460,15 @@ class RawAgent(BaseAgent, Generic[OptionsCoT]):  # type: ignore[misc]
            options=options or {},
        )

+        # When per-service-call persistence is enabled, the per-service-call middleware owns
+        # HistoryProvider loading (it loads locally when the service does not store history, or
+        # relies on the service when it does), so skip them on the once-per-run before_run path.
        per_service_call_history_required = self.require_per_service_call_history_persistence and bool(
            self._get_history_providers()
        )

-        # Run before_run providers (forward order, skip HistoryProvider when per-service-call persistence owns history)
+        # Run before_run providers (forward order, skip HistoryProvider when per-service-call
+        # persistence owns loading)
        for provider in self.context_providers:
            if per_service_call_history_required and isinstance(provider, HistoryProvider):
                continue
@@ -604,10 +604,13 @@ class BaseChatClient(SerializationMixin, ABC, Generic[OptionsCoT]):
                and dict literals are accepted without specialized option typing.
            context_providers: Context providers to include during agent invocation.
            middleware: List of middleware to intercept agent and function invocations.
-            require_per_service_call_history_persistence: Whether to require per-service-call
-                chat history persistence. When enabled, history providers are invoked around
-                each model call instead of once per ``run()`` when the service is not already
-                storing history.
+            require_per_service_call_history_persistence: When enabled (and a HistoryProvider is
+                present), the provider always persists history after each model call. If the
+                client does not store history server-side, history providers are also loaded and
+                injected around each model call; if it does, provider loading is skipped and the
+                service-managed conversation is the source of truth (persistence still happens
+                after each model call). When no HistoryProvider is present, this flag has no
+                effect (no middleware is installed and nothing is persisted).
            function_invocation_configuration: Optional function invocation configuration override.
            compaction_strategy: Optional agent-level compaction override. When omitted,
                client-level compaction defaults remain in effect for each call.
@@ -58,6 +58,7 @@ class ExperimentalFeature(str, Enum):
    FOUNDRY_PREVIEW_TOOLS = "FOUNDRY_PREVIEW_TOOLS"
    FUNCTIONAL_WORKFLOWS = "FUNCTIONAL_WORKFLOWS"
    HARNESS = "HARNESS"
+    MCP_LONG_RUNNING_TASKS = "MCP_LONG_RUNNING_TASKS"
    MCP_SKILLS = "MCP_SKILLS"
    PROGRESSIVE_TOOLS = "PROGRESSIVE_TOOLS"
    SKILLS = "SKILLS"
@@ -16,6 +16,7 @@ from __future__ import annotations
 import asyncio
 import copy
 import json
+import logging
 import threading
 import uuid
 import weakref
@@ -36,6 +37,8 @@ if TYPE_CHECKING:
    from ._middleware import MiddlewareTypes


+logger = logging.getLogger("agent_framework")
+
 # Registry of known types for state deserialization
 _STATE_TYPE_REGISTRY: dict[str, type] = {}

@@ -580,6 +583,7 @@ class PerServiceCallHistoryPersistingMiddleware(ChatMiddleware):
        agent: SupportsAgentRun,
        session: AgentSession,
        providers: Sequence[HistoryProvider],
+        service_stores_history: bool = False,
    ) -> None:
        """Initialize the middleware.

@@ -587,10 +591,16 @@ class PerServiceCallHistoryPersistingMiddleware(ChatMiddleware):
            agent: The agent that owns the history providers.
            session: The active session for the current run.
            providers: The history providers participating in per-service-call persistence.
+            service_stores_history: When True, the chat client stores history server-side. The
+                middleware then skips loading providers and leaves the real conversation id
+                untouched, persisting each service call without driving the function loop with a
+                local sentinel. When False, the middleware loads providers and uses a local
+                sentinel conversation id so the function loop runs without service-side storage.
        """
        self._agent = agent
        self._session = session
        self._providers = list(providers)
+        self._service_stores_history = service_stores_history

    async def _prepare_service_call_context(self, messages: Sequence[Message]) -> SessionContext:
        """Create a per-call SessionContext and load history providers into it."""
@@ -602,6 +612,9 @@ class PerServiceCallHistoryPersistingMiddleware(ChatMiddleware):
        )
        for source_id, source_messages in context_messages.items():
            service_call_context.extend_messages(source_id, source_messages)
+        # When the service stores history, it owns loading; the providers are write-only sinks.
+        if self._service_stores_history:
+            return service_call_context
        for provider in self._providers:
            if not provider.load_messages:
                continue
@@ -652,17 +665,35 @@ class PerServiceCallHistoryPersistingMiddleware(ChatMiddleware):
        response: ChatResponse,
    ) -> ChatResponse:
        """Persist a model response and apply the local follow-up sentinel when needed."""
-        if response.conversation_id is not None and not is_local_history_conversation_id(response.conversation_id):
+        if (
+            not self._service_stores_history
+            and response.conversation_id is not None
+            and not is_local_history_conversation_id(response.conversation_id)
+        ):
            raise ChatClientInvalidResponseException(
                "require_per_service_call_history_persistence cannot be used "
                "when the chat client returns a real conversation_id."
            )

+        # In storing mode the service is expected to echo a conversation id that the next run
+        # resumes from. If it comes back empty, the provider still captures this turn but there is
+        # no service id to load from next time, so cross-turn history can be lost silently. Warn
+        # every time so this uncommon, easy-to-miss failure mode cannot fail quietly.
+        if self._service_stores_history and response.conversation_id is None:
+            logger.warning(
+                "require_per_service_call_history_persistence is enabled with a chat client that "
+                "stores history server-side, but the client returned no conversation_id; cross-turn "
+                "history may not resume. Set store=False to load and resume from the HistoryProvider "
+                "instead."
+            )
+
        await self._persist_service_call_response(
            service_call_context=service_call_context,
            response=response,
        )
-        if _response_contains_follow_up_request(response):
+        # The local sentinel only applies when the service does not store history; when it does,
+        # the real conversation id already drives function-loop continuation.
+        if not self._service_stores_history and _response_contains_follow_up_request(response):
            response.mark_internal_conversation_id()
            response.conversation_id = LOCAL_HISTORY_CONVERSATION_ID
        return response
@@ -681,8 +712,12 @@ class PerServiceCallHistoryPersistingMiddleware(ChatMiddleware):
                result type for streaming or non-streaming execution.
        """
        service_call_context = await self._prepare_service_call_context(context.messages)
-        context.messages = service_call_context.get_messages(include_input=True)
-        self._strip_local_conversation_id(context)
+        # When the service stores history, leave the outgoing messages and the real conversation
+        # id untouched (pass-through); the middleware only persists. Otherwise reconstruct the
+        # outgoing messages from the loaded local history and strip the local sentinel.
+        if not self._service_stores_history:
+            context.messages = service_call_context.get_messages(include_input=True)
+            self._strip_local_conversation_id(context)

        await call_next()

@@ -2,7 +2,6 @@

 from __future__ import annotations

-import json
 import logging
 import sys
 import uuid
@@ -12,7 +11,6 @@ from datetime import datetime, timezone
 from typing import TYPE_CHECKING, Any, ClassVar, Literal, cast, overload

 from .._agents import BaseAgent
-from .._serialization import make_json_safe
 from .._sessions import (
    AgentSession,
    ContextProvider,
@@ -30,11 +28,12 @@ from .._types import (
    UsageDetails,
    add_usage_details,
 )
-from ..exceptions import AgentInvalidRequestException, AgentInvalidResponseException
+from ..exceptions import AgentException, AgentInvalidRequestException, AgentInvalidResponseException
 from ._checkpoint import CheckpointStorage
 from ._events import (
    AGENT_FORWARDED_EVENT_TYPES,
    WorkflowEvent,
+    WorkflowRunState,
 )
 from ._message_utils import normalize_messages_input
 from ._typing_utils import is_instance_of, is_type_compatible
@@ -59,27 +58,24 @@ class WorkflowAgent(BaseAgent):
    @dataclass
    class RequestInfoFunctionArgs:
        request_id: str
-        data: Any
+        request_event: WorkflowEvent

        def to_dict(self) -> dict[str, Any]:
-            return {"request_id": self.request_id, "data": make_json_safe(self.data)}
-
-        def to_json(self) -> str:
-            return json.dumps(self.to_dict())
+            return {"request_id": self.request_id, "request_event": self.request_event.to_dict()}

        @classmethod
        def from_dict(cls, payload: dict[str, Any]) -> WorkflowAgent.RequestInfoFunctionArgs:
-            return cls(request_id=payload.get("request_id", ""), data=payload.get("data"))
+            if "request_id" not in payload or "request_event" not in payload:
+                raise ValueError(
+                    "Invalid payload for RequestInfoFunctionArgs. 'request_id' and 'request_event' are required."
+                )
+            if not payload["request_id"]:
+                raise ValueError("request_id cannot be empty.")

-        @classmethod
-        def from_json(cls, raw: str) -> WorkflowAgent.RequestInfoFunctionArgs:
-            try:
-                parsed: Any = json.loads(raw)
-            except json.JSONDecodeError as exc:
-                raise ValueError(f"RequestInfoFunctionArgs JSON payload is malformed: {exc}") from exc
-            if not isinstance(parsed, dict):
-                raise ValueError("RequestInfoFunctionArgs JSON payload must decode to a mapping")
-            return cls.from_dict(cast(dict[str, Any], parsed))
+            return cls(
+                request_id=payload.get("request_id", ""),
+                request_event=WorkflowEvent.from_dict(payload.get("request_event", {})),
+            )

    def __init__(
        self,
@@ -129,16 +125,11 @@ class WorkflowAgent(BaseAgent):
            **kwargs,
        )
        self._workflow: Workflow = workflow
-        self._pending_requests: dict[str, WorkflowEvent[Any]] = {}

    @property
    def workflow(self) -> Workflow:
        return self._workflow

-    @property
-    def pending_requests(self) -> dict[str, WorkflowEvent[Any]]:
-        return self._pending_requests
-
    # region Run Methods

    @overload
@@ -182,7 +173,7 @@ class WorkflowAgent(BaseAgent):

        Args:
            messages: The message(s) to send to the workflow. Required for new runs,
-                should be None when resuming from checkpoint.
+                could be None if only restoring the underlying workflow from a checkpoint.

        Keyword Args:
            stream: If True, returns an async iterable of updates. If False (default),
@@ -416,101 +407,79 @@ class WorkflowAgent(BaseAgent):
        Yields:
            WorkflowEvent objects from the workflow execution.
        """
-        # Determine the execution mode based on state.
-        # The streaming flag controls the workflow's internal streaming mode,
-        # which affects executor behavior (e.g. AgentExecutor emits different event
-        # types in streaming vs non-streaming mode).
-        if bool(self.pending_requests):
-            function_responses = self._process_pending_requests(input_messages)
+        # Restore the workflow state if a checkpoint is provided
+        if checkpoint_id is not None:
+            if checkpoint_storage is None:
+                raise AgentInvalidRequestException("checkpoint_storage must be provided when checkpoint_id is provided")
+            logger.debug(f"Restoring workflow from checkpoint {checkpoint_id}")
+            # Restore the workflow from checkpoint
            if streaming:
-                async for event in self.workflow.run(
-                    responses=function_responses,
-                    stream=True,
-                    function_invocation_kwargs=function_invocation_kwargs,
-                    client_kwargs=client_kwargs,
-                ):
-                    yield event
-            else:
-                for event in await self.workflow.run(
-                    responses=function_responses,
-                    function_invocation_kwargs=function_invocation_kwargs,
-                    client_kwargs=client_kwargs,
-                ):
-                    yield event
-
-        elif checkpoint_id is not None:
-            # Restore the prior workflow state from the checkpoint. Shared
-            # state (e.g. accumulated conversation history maintained by the
-            # workflow's executors) survives across turns because Workflow.run
-            # no longer wipes state per call. Callers who want to deliver a
-            # new user message after restore should make a second
-            # `workflow.run(message=...)` call - they are NOT mutually
-            # exclusive on the same instance, but each must be its own call.
-            if streaming:
-                async for event in self.workflow.run(
+                async for _ in self.workflow.run(
                    stream=True,
                    checkpoint_id=checkpoint_id,
                    checkpoint_storage=checkpoint_storage,
+                ):
+                    pass
+            else:
+                _ = await self.workflow.run(
+                    checkpoint_id=checkpoint_id,
+                    checkpoint_storage=checkpoint_storage,
+                )
+            if not input_messages:
+                logger.info("No input messages provided; the workflow has been restored to the checkpoint state.")
+                return
+
+        final_state = self._workflow.status
+        logger.debug(f"Workflow state: {final_state}")
+
+        if final_state == WorkflowRunState.IDLE_WITH_PENDING_REQUESTS:
+            # Extract function responses from input messages, and ensure that
+            # only function responses are present in messages if there is any
+            # pending request.
+            # NOTE: It is possible that some pending requests are not fulfilled,
+            # and we will let the workflow to handle this -- the agent does not
+            # have an opinion on this.
+            function_responses = self._extract_function_responses(input_messages)
+            if streaming:
+                async for event in self.workflow.run(
+                    responses=function_responses,
+                    stream=True,
+                    checkpoint_storage=checkpoint_storage,
                    function_invocation_kwargs=function_invocation_kwargs,
                    client_kwargs=client_kwargs,
                ):
                    yield event
            else:
                for event in await self.workflow.run(
-                    checkpoint_id=checkpoint_id,
+                    responses=function_responses,
+                    checkpoint_storage=checkpoint_storage,
+                    function_invocation_kwargs=function_invocation_kwargs,
+                    client_kwargs=client_kwargs,
+                ):
+                    yield event
+        elif final_state == WorkflowRunState.IDLE:
+            if streaming:
+                async for event in self.workflow.run(
+                    message=input_messages,
+                    stream=True,
+                    checkpoint_storage=checkpoint_storage,
+                    function_invocation_kwargs=function_invocation_kwargs,
+                    client_kwargs=client_kwargs,
+                ):
+                    yield event
+            else:
+                for event in await self.workflow.run(
+                    message=input_messages,
                    checkpoint_storage=checkpoint_storage,
                    function_invocation_kwargs=function_invocation_kwargs,
                    client_kwargs=client_kwargs,
                ):
                    yield event
-
        else:
-            if streaming:
-                async for event in self.workflow.run(
-                    message=input_messages,
-                    stream=True,
-                    checkpoint_storage=checkpoint_storage,
-                    function_invocation_kwargs=function_invocation_kwargs,
-                    client_kwargs=client_kwargs,
-                ):
-                    yield event
-            else:
-                for event in await self.workflow.run(
-                    message=input_messages,
-                    checkpoint_storage=checkpoint_storage,
-                    function_invocation_kwargs=function_invocation_kwargs,
-                    client_kwargs=client_kwargs,
-                ):
-                    yield event
+            raise AgentException(f"The underlying workflow is in an invalid state to restart: {final_state}.")

    # endregion Run Methods

-    def _process_pending_requests(self, input_messages: Sequence[Message]) -> dict[str, Any]:
-        """Process pending requests by extracting function responses and updating state.
-
-        Args:
-            input_messages: Input messages that may contain function responses.
-
-        Returns:
-            A dictionary mapping request IDs to their response data.
-        """
-        logger.info(f"Continuing workflow to address {len(self.pending_requests)} requests")
-
-        # Extract function responses from input messages, and ensure that
-        # only function responses are present in messages if there is any
-        # pending request.
-        function_responses = self._extract_function_responses(input_messages)
-
-        # Pop pending requests if fulfilled.
-        for request_id in list(self.pending_requests.keys()):
-            if request_id in function_responses:
-                self.pending_requests.pop(request_id)
-
-        # NOTE: It is possible that some pending requests are not fulfilled,
-        # and we will let the workflow to handle this -- the agent does not
-        # have an opinion on this.
-        return function_responses
-
    def _convert_workflow_events_to_agent_response(
        self,
        response_id: str,
@@ -528,10 +497,10 @@ class WorkflowAgent(BaseAgent):

        for output_event in output_events:
            if output_event.type == "request_info":
-                function_call, approval_request = self._process_request_info_event(output_event)
+                request_content = self._process_request_info_event(output_event)
                messages.append(
                    Message(
-                        contents=[function_call, approval_request],
+                        contents=[request_content],
                        role="assistant",
                        author_name=output_event.source_executor_id,
                        message_id=str(uuid.uuid4()),
@@ -598,38 +567,6 @@ class WorkflowAgent(BaseAgent):
            raw_representation=raw_representations,
        )

-    def _process_request_info_event(
-        self,
-        event: WorkflowEvent[Any],
-    ) -> tuple[Content, Content]:
-        """Convert a request_info event to FunctionCallContent and FunctionApprovalRequestContent.
-
-        Args:
-            event: A WorkflowEvent with type='request_info'.
-
-        Returns:
-            A tuple of (FunctionCallContent, FunctionApprovalRequestContent).
-        """
-        request_id = event.request_id
-        if not request_id:
-            raise ValueError("request_info event must have a request_id")
-
-        self.pending_requests[request_id] = event
-
-        args = self.RequestInfoFunctionArgs(request_id=request_id, data=event.data).to_dict()
-
-        function_call = Content.from_function_call(
-            call_id=request_id,
-            name=self.REQUEST_INFO_FUNCTION_NAME,
-            arguments=args,
-        )
-        approval_request = Content.from_function_approval_request(
-            id=request_id,
-            function_call=function_call,
-            additional_properties={"request_id": request_id},
-        )
-        return function_call, approval_request
-
    def _convert_workflow_event_to_agent_response_updates(
        self,
        response_id: str,
@@ -731,85 +668,72 @@ class WorkflowAgent(BaseAgent):
            ]

        if event.type == "request_info":
-            # Store the pending request for later correlation
-            request_id = event.request_id
-            if not request_id:
-                raise ValueError("request_info event must have a request_id")
-
-            self.pending_requests[request_id] = event
-
-            args = self.RequestInfoFunctionArgs(request_id=request_id, data=event.data).to_dict()
-
-            function_call = Content.from_function_call(
-                call_id=request_id,
-                name=self.REQUEST_INFO_FUNCTION_NAME,
-                arguments=args,
-            )
-            approval_request = Content.from_function_approval_request(
-                id=request_id,
-                function_call=function_call,
-                additional_properties={"request_id": request_id},
-            )
+            request_content = self._process_request_info_event(event)
            return [
                AgentResponseUpdate(
-                    contents=[function_call, approval_request],
+                    contents=[request_content],
                    role="assistant",
                    author_name=self.name,
                    response_id=response_id,
                    message_id=str(uuid.uuid4()),
                    created_at=datetime.now(tz=timezone.utc).strftime("%Y-%m-%dT%H:%M:%S.%fZ"),
+                    raw_representation=event,
                )
            ]

        # Ignore workflow-internal events
        return []

+    def _process_request_info_event(
+        self,
+        event: WorkflowEvent[Any],
+    ) -> Content:
+        """Convert a request_info event to FunctionApprovalRequestContent.
+
+        Args:
+            event: A WorkflowEvent with type='request_info'.
+
+        Returns:
+            A content object representing the request info. The content can be a `function_approval_request`
+            or a `function_call` depending on the structure of the event data.
+
+        Note:
+            If the event data is already a FunctionApprovalRequestContent, it will be returned as-is.
+        """
+        if isinstance(event.data, Content) and event.data.user_input_request:
+            # Return the event data as-is if it's already a properly formed FunctionApprovalRequestContent
+            return event.data
+
+        request_id = event.request_id
+        args = self.RequestInfoFunctionArgs(request_id=request_id, request_event=event).to_dict()
+
+        return Content.from_function_call(
+            call_id=request_id,
+            name=self.REQUEST_INFO_FUNCTION_NAME,
+            arguments=args,
+        )
+
    def _extract_function_responses(self, input_messages: Sequence[Message]) -> dict[str, Any]:
-        """Extract function responses from input messages."""
+        """Extract function responses from input messages.
+
+        The responses are for pending requests that the workflow is waiting on, and
+        will be passed to the workflow. The pending requests are processed to either
+        `function_approval_request` or `function_call` content by `_process_request_info_event`.
+        """
        function_responses: dict[str, Any] = {}
        for message in input_messages:
            for content in message.contents:
                if content.type == "function_approval_response":
-                    # Parse the function arguments to recover request payload
-                    arguments_payload = content.function_call.arguments  # type: ignore[attr-defined, union-attr]
-                    if isinstance(arguments_payload, str):
-                        try:
-                            parsed_args = self.RequestInfoFunctionArgs.from_json(arguments_payload)
-                        except ValueError as exc:
-                            raise AgentInvalidResponseException(
-                                "FunctionApprovalResponseContent arguments must decode to a mapping."
-                            ) from exc
-                    elif isinstance(arguments_payload, dict):
-                        parsed_args = self.RequestInfoFunctionArgs.from_dict(arguments_payload)
-                    else:
-                        raise AgentInvalidResponseException(
-                            "FunctionApprovalResponseContent arguments must be a mapping or JSON string."
-                        )
-
-                    request_id = parsed_args.request_id or content.id  # type: ignore[attr-defined]
-                    if not content.approved:  # type: ignore[attr-defined]
-                        raise AgentInvalidResponseException(f"Request '{request_id}' was not approved by the caller.")
-
-                    if request_id in self.pending_requests:
-                        function_responses[request_id] = parsed_args.data
-                    elif bool(self.pending_requests):
-                        raise AgentInvalidRequestException(
-                            "Only responses for pending requests are allowed when there are outstanding approvals."
-                        )
+                    request_id: str = content.id  # type: ignore[assignment]
+                    function_responses[request_id] = content
                elif content.type == "function_result":
-                    request_id = content.call_id  # type: ignore[attr-defined]
-                    if request_id in self.pending_requests:
-                        response_data = content.result if hasattr(content, "result") else str(content)  # type: ignore[attr-defined]
-                        function_responses[request_id] = response_data
-                    elif bool(self.pending_requests):
-                        raise AgentInvalidRequestException(
-                            "Only function responses for pending requests are allowed while requests are outstanding."
-                        )
+                    response_data = content.result if hasattr(content, "result") else str(content)  # type: ignore[attr-defined]
+                    function_responses[content.call_id] = response_data  # type: ignore
                else:
-                    if bool(self.pending_requests):
-                        raise AgentInvalidResponseException(
-                            "Unexpected content type while awaiting request info responses."
-                        )
+                    raise AgentInvalidResponseException(
+                        "Unexpected content type while awaiting request info responses."
+                    )
+
        return function_responses

    def _extract_contents(self, data: Any) -> list[Content]:
@@ -429,15 +429,30 @@ class AgentExecutor(Executor):
            function_invocation_kwargs=function_invocation_kwargs,
            client_kwargs=client_kwargs,
        )
-        await ctx.yield_output(response)

        # Handle any user input requests
        if response.user_input_requests:
+            user_input_request_count = len(response.user_input_requests)
+            total_message_content_count = sum(len(msg.contents) for msg in response.messages)
+            if user_input_request_count != total_message_content_count:
+                logger.warning(
+                    "Response %s contains %d user input requests but total message contents are %d. "
+                    "This indicates the response contains both user input requests and message contents. "
+                    "Double check if this is the intended behavior, as non user input request contents in "
+                    "this response will not be emitted.",
+                    response.response_id,
+                    user_input_request_count,
+                    total_message_content_count,
+                )
            for user_input_request in response.user_input_requests:
                self._pending_agent_requests[user_input_request.id] = user_input_request  # type: ignore[index]
-                await ctx.request_info(user_input_request, Content)
+                await ctx.request_info(user_input_request, Content, request_id=user_input_request.id)
            return None

+        # Only yield output if the response is complete and not waiting for user input.
+        # This is to avoid emitting two events of different types ('output' and 'request_info')
+        # that carry the same payload.
+        await ctx.yield_output(response)
        return response

    async def _run_agent_streaming(self, ctx: WorkflowContext[Never, AgentResponseUpdate]) -> AgentResponse | None:
@@ -472,9 +487,25 @@ class AgentExecutor(Executor):
        )
        async for update in stream:
            updates.append(update)
-            await ctx.yield_output(update)
            if update.user_input_requests:
+                user_input_request_count = len(update.user_input_requests)
+                total_message_content_count = len(update.contents)
+                if user_input_request_count != total_message_content_count:
+                    logger.warning(
+                        "Response update %s contains %d user input requests but total message contents are %d. "
+                        "This indicates the response update contains both user input requests and message contents. "
+                        "Double check if this is the intended behavior, as non user input request contents will "
+                        "not be emitted.",
+                        update.response_id,
+                        user_input_request_count,
+                        total_message_content_count,
+                    )
                streamed_user_input_requests.extend(update.user_input_requests)
+            else:
+                # Only yield output events for updates that do not contain user input requests.
+                # This is to avoid emitting two events of different types ('output' and 'request_info')
+                # that carry the same payload.
+                await ctx.yield_output(update)

        # Prefer stream finalization when available so result hooks run
        # (e.g., thread conversation updates). Fall back to reconstructing from updates
@@ -509,7 +540,7 @@ class AgentExecutor(Executor):
        if user_input_requests:
            for user_input_request in user_input_requests:
                self._pending_agent_requests[user_input_request.id] = user_input_request  # type: ignore[index]
-                await ctx.request_info(user_input_request, Content)
+                await ctx.request_info(user_input_request, Content, request_id=user_input_request.id)
            return None

        return response
@@ -13,6 +13,35 @@ during deserialization.  The default built-in safe set covers common Python
 value types (primitives, datetime, uuid, ...), all ``agent_framework`` internal
 types, and all ``openai.types`` types.  Callers can extend the set by passing
 additional ``"module:qualname"`` strings.
+
+Security Model
+--------------
+Checkpoint storage is treated as a **trusted data source**.  The serialization
+format uses Python's ``pickle`` module which can execute arbitrary code during
+deserialization.  The ``RestrictedUnpickler`` provides a defense-in-depth
+allowlist that limits instantiable classes, but it is **not** a security
+boundary — certain allowlisted builtins (e.g. ``getattr``) are required for
+legitimate object reconstruction (enums, named tuples) and cannot be removed
+without breaking compatibility.
+
+Developers **must** ensure that:
+
+1. The checkpoint storage backend (file system, Cosmos DB, Azure Blob, Durable
+   Functions storage) is access-controlled and not writable by untrusted
+   parties.
+2. Data flowing into ``decode_checkpoint_value`` originates exclusively from
+   the application's own checkpoint storage — never from user-supplied HTTP
+   requests, message payloads, or other untrusted sources.
+3. The ``allowed_types`` parameter is specified whenever possible to restrict
+   the set of reconstructible types to the minimum required by the application.
+4. Never pass untrusted external input to ``decode_checkpoint_value``. If you
+   must accept external JSON that might contain checkpoint markers, sanitize it
+   first (for example, :func:`agent_framework_azurefunctions._serialization.strip_pickle_markers`).
+
+The allowlist is a mitigation that reduces attack surface but does not
+eliminate the inherent risks of deserializing untrusted pickle data.  Treat
+your checkpoint storage with the same access controls you would apply to
+application secrets or database credentials.
 """

 from __future__ import annotations
@@ -360,6 +360,22 @@ class Workflow(DictConvertible):
        # Flag to prevent concurrent workflow executions
        self._is_running = False

+        # Current run-level status of this workflow instance. Updated in lockstep with
+        # the status events emitted from `_run_workflow_with_tracing`. Defaults to IDLE
+        # for a freshly built workflow that has not yet been run.
+        self._status: WorkflowRunState = WorkflowRunState.IDLE
+
+    @property
+    def status(self) -> WorkflowRunState:
+        """Return the current run-level status of this workflow instance.
+
+        Mirrors the most recent status event emitted by the workflow. Safe to read at
+        any time: workflows run on a single asyncio event loop, and the underlying
+        attribute is a single enum reference whose assignment is atomic under the
+        CPython GIL, so no locking is required.
+        """
+        return self._status
+
    def _ensure_not_running(self) -> None:
        """Ensure the workflow is not already running."""
        if self._is_running:
@@ -513,8 +529,9 @@ class Workflow(DictConvertible):
                with _framework_event_origin():
                    started = WorkflowEvent.started()
                yield started  # noqa: RUF070
+                self._status = WorkflowRunState.IN_PROGRESS
                with _framework_event_origin():
-                    in_progress = WorkflowEvent.status(WorkflowRunState.IN_PROGRESS)
+                    in_progress = WorkflowEvent.status(self._status)
                yield in_progress  # noqa: RUF070

                # Per-run reset for fresh-message runs only. We deliberately
@@ -569,17 +586,20 @@ class Workflow(DictConvertible):

                    if event.type == "request_info" and not emitted_in_progress_pending:
                        emitted_in_progress_pending = True
+                        self._status = WorkflowRunState.IN_PROGRESS_PENDING_REQUESTS
                        with _framework_event_origin():
-                            pending_status = WorkflowEvent.status(WorkflowRunState.IN_PROGRESS_PENDING_REQUESTS)
+                            pending_status = WorkflowEvent.status(self._status)
                        yield pending_status  # noqa: RUF070
                # Workflow runs until idle - emit final status based on whether requests are pending
                if saw_request:
+                    self._status = WorkflowRunState.IDLE_WITH_PENDING_REQUESTS
                    with _framework_event_origin():
-                        terminal_status = WorkflowEvent.status(WorkflowRunState.IDLE_WITH_PENDING_REQUESTS)
+                        terminal_status = WorkflowEvent.status(self._status)
                    yield terminal_status
                else:
+                    self._status = WorkflowRunState.IDLE
                    with _framework_event_origin():
-                        terminal_status = WorkflowEvent.status(WorkflowRunState.IDLE)
+                        terminal_status = WorkflowEvent.status(self._status)
                    yield terminal_status

                span.add_event(OtelAttr.WORKFLOW_COMPLETED)
@@ -593,6 +613,7 @@ class Workflow(DictConvertible):
                with _framework_event_origin():
                    failed_event = WorkflowEvent.failed(details)
                yield failed_event  # noqa: RUF070
+                self._status = WorkflowRunState.FAILED
                with _framework_event_origin():
                    failed_status = WorkflowEvent.status(WorkflowRunState.FAILED)
                yield failed_status  # noqa: RUF070
@@ -80,6 +80,7 @@ __all__ = [
    "EmbeddingTelemetryLayer",
    "OtelAttr",
    "configure_otel_providers",
+    "create_mcp_client_span",
    "create_metric_views",
    "create_resource",
    "disable_instrumentation",
@@ -87,6 +88,7 @@ __all__ = [
    "enable_sensitive_telemetry",
    "get_meter",
    "get_tracer",
+    "set_mcp_span_error",
 ]


@@ -110,7 +112,6 @@ INNER_ACCUMULATED_USAGE: Final[contextvars.ContextVar[UsageDetails | None]] = co
    "inner_accumulated_usage", default=None
 )

-
 OTEL_METRICS: Final[str] = "__otel_metrics__"
 TOKEN_USAGE_BUCKET_BOUNDARIES: Final[tuple[float, ...]] = (
    1,
@@ -292,6 +293,14 @@ class OtelAttr(str, Enum):
    AGENT_CREATE_OPERATION = "create_agent"
    AGENT_INVOKE_OPERATION = "invoke_agent"

+    # MCP attributes (https://opentelemetry.io/docs/specs/semconv/gen-ai/mcp/)
+    MCP_METHOD_NAME = "mcp.method.name"
+    MCP_PROTOCOL_VERSION = "mcp.protocol.version"
+    MCP_SESSION_ID = "mcp.session.id"
+    PROMPT_NAME = "gen_ai.prompt.name"
+    NETWORK_TRANSPORT = "network.transport"
+    NETWORK_PROTOCOL_NAME = "network.protocol.name"
+
    # Agent Framework specific attributes
    MEASUREMENT_FUNCTION_TAG_NAME = "agent_framework.function.name"
    MEASUREMENT_FUNCTION_INVOCATION_DURATION = "agent_framework.function.invocation.duration"
@@ -2013,6 +2022,61 @@ def get_function_span(
    )


+# region MCP span helpers
+
+
+@contextlib.contextmanager
+def create_mcp_client_span(
+    method_name: str,
+    target: str | None = None,
+    attributes: dict[str, Any] | None = None,
+) -> Generator[trace.Span, Any, Any]:
+    """Create an MCP client span per OTel MCP semantic conventions.
+
+    Span name follows the format ``{mcp.method.name} {target}`` when a target
+    is available, otherwise just ``{mcp.method.name}``.
+
+    See: https://opentelemetry.io/docs/specs/semconv/gen-ai/mcp/#client
+
+    Args:
+        method_name: The MCP method name (e.g. ``initialize``, ``tools/call``).
+        target: Optional low-cardinality target (tool name, prompt name).
+        attributes: Additional span attributes.
+    """
+    span_name = f"{method_name} {target}" if target else method_name
+    attrs: dict[str, Any] = {OtelAttr.MCP_METHOD_NAME: method_name}
+    if attributes:
+        attrs.update(attributes)
+    tracer = get_tracer() if OBSERVABILITY_SETTINGS.ENABLED else trace.NoOpTracer()
+    span = tracer.start_span(span_name, kind=trace.SpanKind.CLIENT, attributes=attrs)
+    with trace.use_span(
+        span=span,
+        end_on_exit=True,
+        record_exception=True,
+        set_status_on_exception=True,
+    ) as current_span:
+        yield current_span
+
+
+def set_mcp_span_error(
+    span: trace.Span,
+    error_type: str,
+    description: str | None = None,
+) -> None:
+    """Set error status and ``error.type`` on an MCP span.
+
+    Args:
+        span: The span to mark as errored.
+        error_type: The error type string (e.g. ``tool_error``, exception class name).
+        description: Optional description (e.g. JSON-RPC error message).
+    """
+    span.set_attribute(OtelAttr.ERROR_TYPE, error_type)
+    span.set_status(trace.StatusCode.ERROR, description=description)
+
+
+# endregion
+
+
@contextlib.contextmanager
 def _activate_span(span: trace.Span) -> Generator[None]:
    """Attach ``span`` as the current span in the OpenTelemetry context.
@@ -4,7 +4,7 @@ description = "Microsoft Agent Framework for building AI Agents with Python. Thi
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.8.0"
+version = "1.8.1"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -3,6 +3,7 @@
 import contextlib
 import inspect
 import json
+import logging
 from collections.abc import AsyncIterable, Awaitable, Callable, MutableSequence, Sequence
 from typing import Any, cast
 from unittest.mock import AsyncMock, MagicMock, patch
@@ -42,6 +43,8 @@ from agent_framework._mcp import MCPTool, _build_prefixed_mcp_name, _normalize_m
 from agent_framework._middleware import FunctionInvocationContext
 from agent_framework.exceptions import AgentInvalidRequestException, ChatClientInvalidResponseException

+from .conftest import MockBaseChatClient
+

 class _FixedTokenizer:
    def __init__(self, token_count: int) -> None:
@@ -609,6 +612,7 @@ async def test_streaming_per_service_call_persistence_hides_response_id_from_aft

 async def test_per_service_call_persistence_uses_real_service_storage_when_client_stores_by_default(
    chat_client_base: SupportsChatGetResponse,
+    caplog: pytest.LogCaptureFixture,
 ) -> None:
    provider = _RecordingHistoryProvider()

@@ -649,15 +653,22 @@ async def test_per_service_call_persistence_uses_real_service_storage_when_clien
        require_per_service_call_history_persistence=True,
    )

-    result = await agent.run("What's the weather in Seattle?", session=session)
+    with caplog.at_level(logging.WARNING, logger="agent_framework"):
+        result = await agent.run("What's the weather in Seattle?", session=session)

    provider_state = session.state[provider.source_id]

    assert result.text == "It is sunny in Seattle."
    assert result.response_id == "resp_call_2"
    assert chat_client_base.call_count == 2
+    # The service owns the conversation, so the provider never loads (issue #5798).
    assert "get_call_count" not in provider_state
-    assert "save_call_count" not in provider_state
+    # Persistence is owned by the per-service-call middleware: it persists once per service call
+    # (issue #5798: the provider must never be silently bypassed when the service stores history).
+    # This run makes two service calls (function call + final answer), so it persists twice.
+    assert provider_state["save_call_count"] == 2
+    # load_messages=True while the service stores history surfaces a warning.
+    assert any("load_messages" in record.message for record in caplog.records)
    assert session.service_session_id == "resp_service_managed"


@@ -1996,6 +2007,19 @@ def test_merge_options_none_values_ignored():
    assert result["key2"] == "value2"


+def test_merge_options_drops_none_base_values():
+    """Test _merge_options strips None values so unset options are never forwarded."""
+    base = {"store": None, "temperature": 0.5}
+    override = {"top_p": 0.9}
+
+    result = _merge_options(base, override)
+
+    # An unset base value (e.g. store=None from default_options) must not survive the merge.
+    assert "store" not in result
+    assert result["temperature"] == 0.5
+    assert result["top_p"] == 0.9
+
+
 def test_merge_options_runtime_model_overrides_default_model() -> None:
    """Test _merge_options lets a runtime model override a default model."""
    result = _merge_options({"model": "default-model"}, {"model": "runtime-model"})
@@ -2658,3 +2682,449 @@ async def test_as_tool_raises_on_user_input_request(client: SupportsChatGetRespo
    assert len(exc_info.value.contents) == 1
    assert exc_info.value.contents[0].type == "oauth_consent_request"
    assert exc_info.value.contents[0].consent_link == "https://login.microsoftonline.com/consent"
+
+
+# region Per-service-call history persistence scenario matrix
+#
+# The driving field is ``require_per_service_call_history_persistence``. Every scenario runs a
+# single agent run that makes **two service calls** -- a function call followed by a final
+# completion -- so the *timing* of persistence is observable:
+#
+# * When the flag is ``True``, the per-service-call middleware persists the provider **after each
+#   service call**. So the function-call turn is already saved by the time the second (final)
+#   service call starts. This holds regardless of whether the chat client stores history
+#   server-side (the bug in issue #5798 was that a storing client silently bypassed persistence).
+# * When the flag is ``False``, the provider persists **once, at the end of the run** -- nothing is
+#   saved between the two service calls.
+#
+# ``SpyChatClient.saves_before_call`` records ``provider.save_calls`` at the start of every service
+# call, so ``[0, 1]`` means "the function-call turn was persisted before the final call" and
+# ``[0, 0]`` means "no persistence happened mid-run". The client's ``store`` / ``STORES_BY_DEFAULT``
+# only selects *how* the middleware behaves -- never *whether* the provider persists.
+
+_PSC_SERVICE_CONVERSATION_ID = "svc-conversation"
+
+_psc_stream_params = pytest.mark.parametrize("stream", [False, True], ids=["sync", "stream"])
+
+
+@tool(name="lookup_weather", approval_mode="never_require")
+def _psc_lookup_weather(location: str) -> str:
+    return f"Weather in {location}: sunny"
+
+
+def _psc_function_call_script() -> list[tuple[str, ...]]:
+    """A fresh function-call-then-final-completion script (the client mutates it)."""
+    return [
+        ("call", "call_1", "lookup_weather", '{"location": "Seattle"}'),
+        ("text", "It is sunny in Seattle."),
+    ]
+
+
+class _PscSpyHistoryProvider(HistoryProvider):
+    """In-memory history provider that records load/save calls for assertions."""
+
+    def __init__(self, source_id: str = "spy_history", **kwargs: Any) -> None:
+        super().__init__(source_id, **kwargs)
+        self._messages: list[Message] = []
+        self.get_calls: int = 0
+        self.save_calls: int = 0
+        self.saved_batches: list[list[Message]] = []
+
+    async def get_messages(
+        self, session_id: str | None, *, state: dict[str, Any] | None = None, **kwargs: Any
+    ) -> list[Message]:
+        self.get_calls += 1
+        return list(self._messages)
+
+    async def save_messages(
+        self,
+        session_id: str | None,
+        messages: Sequence[Message],
+        *,
+        state: dict[str, Any] | None = None,
+        **kwargs: Any,
+    ) -> None:
+        self.save_calls += 1
+        self.saved_batches.append(list(messages))
+        self._messages.extend(messages)
+
+    @property
+    def stored_messages(self) -> list[Message]:
+        return list(self._messages)
+
+
+class _PscSpyChatClient(MockBaseChatClient):
+    """Chat client that scripts a function-call/final-completion sequence.
+
+    It records, at the start of each service call, how many provider saves have already happened
+    (``saves_before_call``), what messages it received, and what options it saw. When the effective
+    ``store`` is truthy it returns a stable ``conversation_id`` to mimic a server-managed
+    conversation, so the framework propagates ``session.service_session_id``.
+    """
+
+    def __init__(
+        self,
+        *,
+        provider: _PscSpyHistoryProvider,
+        stores_by_default: bool = False,
+        script: list[tuple[str, ...]] | None = None,
+        echo_conversation_id: bool = True,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.STORES_BY_DEFAULT = stores_by_default  # type: ignore[attr-defined]
+        self._provider = provider
+        self._script = list(script) if script is not None else [("text", "ok")]
+        self._echo_conversation_id = echo_conversation_id
+        self.received_messages: list[list[Message]] = []
+        self.received_options: list[dict[str, Any]] = []
+        self.saves_before_call: list[int] = []
+
+    def _effective_store(self, options: dict[str, Any]) -> bool:
+        store = options.get("store")
+        if store is None:
+            return bool(self.STORES_BY_DEFAULT)
+        return bool(store)
+
+    def _next_contents(self) -> list[Content]:
+        turn = self._script.pop(0) if self._script else ("text", "ok")
+        if turn[0] == "call":
+            _, call_id, name, args = turn
+            return [Content.from_function_call(call_id=call_id, name=name, arguments=args)]
+        return [Content.from_text(turn[1])]
+
+    def _inner_get_response(  # type: ignore[override]
+        self,
+        *,
+        messages: MutableSequence[Message],
+        stream: bool,
+        options: dict[str, Any],
+        **kwargs: Any,
+    ) -> Awaitable[ChatResponse] | ResponseStream[ChatResponseUpdate, ChatResponse]:
+        self.received_messages.append(list(messages))
+        self.received_options.append(dict(options))
+        self.saves_before_call.append(self._provider.save_calls)
+        store_and_echo = self._effective_store(options) and self._echo_conversation_id
+        conv_id = _PSC_SERVICE_CONVERSATION_ID if store_and_echo else None
+        contents = self._next_contents()
+
+        if stream:
+
+            async def _stream() -> AsyncIterable[ChatResponseUpdate]:
+                self.call_count += 1
+                yield ChatResponseUpdate(
+                    contents=contents,
+                    role="assistant",
+                    finish_reason="stop",
+                    conversation_id=conv_id,
+                )
+
+            def _finalize(updates: Sequence[ChatResponseUpdate]) -> ChatResponse:
+                response = ChatResponse.from_updates(updates, output_format_type=options.get("response_format"))
+                if conv_id:
+                    response.conversation_id = conv_id
+                return response
+
+            return ResponseStream(_stream(), finalizer=_finalize)
+
+        async def _get() -> ChatResponse:
+            self.call_count += 1
+            return ChatResponse(
+                messages=Message(role="assistant", contents=contents),
+                conversation_id=conv_id,
+            )
+
+        return _get()
+
+
+def _psc_build_agent(
+    client: _PscSpyChatClient,
+    provider: _PscSpyHistoryProvider,
+    *,
+    require_per_service_call_history_persistence: bool,
+    default_options: dict[str, Any] | None = None,
+) -> Agent:
+    kwargs: dict[str, Any] = {}
+    if default_options is not None:
+        kwargs["default_options"] = default_options
+    return Agent(
+        client=client,
+        tools=[_psc_lookup_weather],
+        context_providers=[provider],
+        require_per_service_call_history_persistence=require_per_service_call_history_persistence,
+        **kwargs,
+    )
+
+
+async def _psc_run(agent: Agent, text: str, session: AgentSession, *, stream: bool) -> str:
+    if stream:
+        chunks: list[str] = []
+        async for update in agent.run(text, session=session, stream=True):
+            chunks.append(update.text or "")
+        return "".join(chunks)
+    result = await agent.run(text, session=session)
+    return result.text
+
+
+# driver=True (the contract under test): persistence happens per service call
+
+
+@_psc_stream_params
+async def test_psc_flag_on_store_false_persists_after_each_service_call(stream: bool) -> None:
+    """Mode A (flag on, service does not store): function-call turn is persisted before the final call."""
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=False, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    text = await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert text == "It is sunny in Seattle."
+    # Two service calls: function call, then final completion.
+    assert client.call_count == 2
+    # The contract: the function-call turn was persisted *before* the second service call started.
+    assert client.saves_before_call == [0, 1]
+    assert provider.save_calls == 2
+    # Mode A loads local history (the middleware injects it before each service call).
+    assert provider.get_calls >= 1
+    # No service-side storage, so no conversation id is propagated.
+    assert session.service_session_id is None
+
+
+@_psc_stream_params
+async def test_psc_flag_on_stores_by_default_persists_after_each_service_call(
+    stream: bool, caplog: pytest.LogCaptureFixture
+) -> None:
+    """Mode B (flag on, service stores by default): still persists per service call, but skips load (issue #5798)."""
+    provider = _PscSpyHistoryProvider()  # load_messages=True by default
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    with caplog.at_level(logging.WARNING, logger="agent_framework"):
+        text = await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert text == "It is sunny in Seattle."
+    assert client.call_count == 2
+    # The invariant the bug violated: persistence still happens per service call when the service stores.
+    assert client.saves_before_call == [0, 1]
+    assert provider.save_calls == 2
+    # The service owns loading, so the provider is never asked to load.
+    assert provider.get_calls == 0
+    # A warning surfaces the bypassed load (load_messages=True).
+    assert any("load_messages" in record.message for record in caplog.records)
+    # The real service conversation id propagates to the session.
+    assert session.service_session_id == _PSC_SERVICE_CONVERSATION_ID
+
+
+@_psc_stream_params
+async def test_psc_flag_on_store_only_provider_no_load_no_warning(
+    stream: bool, caplog: pytest.LogCaptureFixture
+) -> None:
+    """Mode B with a store-only provider (load_messages=False): persists per call, no load, no warning."""
+    provider = _PscSpyHistoryProvider(load_messages=False)
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    with caplog.at_level(logging.WARNING, logger="agent_framework"):
+        await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert client.saves_before_call == [0, 1]
+    assert provider.save_calls == 2
+    assert provider.get_calls == 0
+    assert not any("load_messages" in record.message for record in caplog.records)
+
+
+@_psc_stream_params
+async def test_psc_flag_on_store_false_override_behaves_as_mode_a(stream: bool) -> None:
+    """Flag on + storing client but store=False override: falls back to Mode A (local, per call)."""
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(
+        client, provider, require_per_service_call_history_persistence=True, default_options={"store": False}
+    )
+    session = agent.create_session()
+
+    await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert client.saves_before_call == [0, 1]
+    assert provider.save_calls == 2
+    assert provider.get_calls >= 1
+    # store=False forces local handling, so no real service conversation id.
+    assert session.service_session_id is None
+
+
+@_psc_stream_params
+async def test_psc_flag_on_store_none_treated_as_absent(stream: bool, caplog: pytest.LogCaptureFixture) -> None:
+    """Flag on + storing client + explicit store=None: None is "unset", so the storing default applies (Mode B)."""
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(
+        client, provider, require_per_service_call_history_persistence=True, default_options={"store": None}
+    )
+    session = agent.create_session()
+
+    with caplog.at_level(logging.WARNING, logger="agent_framework"):
+        await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert client.saves_before_call == [0, 1]
+    assert provider.save_calls == 2
+    assert provider.get_calls == 0
+    assert session.service_session_id == _PSC_SERVICE_CONVERSATION_ID
+    assert any("load_messages" in record.message for record in caplog.records)
+    # store=None must not be forwarded to the client; the service decides its own default.
+    assert all("store" not in options for options in client.received_options)
+
+
+@_psc_stream_params
+async def test_psc_flag_on_respects_store_outputs_flag(stream: bool) -> None:
+    """Flag on: the provider's store_inputs/store_outputs flags still apply per service call."""
+    provider = _PscSpyHistoryProvider(store_inputs=True, store_outputs=False)
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert provider.save_calls == 2
+    # Outputs disabled, so no assistant/tool-call messages were stored, only user/tool inputs.
+    assert provider.stored_messages
+    assert all(message.role != "assistant" for message in provider.stored_messages)
+
+
+# driver=False (control): persistence happens once, at the end of the run
+
+
+@_psc_stream_params
+async def test_psc_flag_off_store_false_persists_once_at_end(stream: bool) -> None:
+    """Flag off + non-storing client: nothing is persisted mid-run; one save at the end."""
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=False, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=False)
+    session = agent.create_session()
+
+    text = await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert text == "It is sunny in Seattle."
+    assert client.call_count == 2
+    # The control contract: no save happened between the function call and the final completion.
+    assert client.saves_before_call == [0, 0]
+    assert provider.save_calls == 1
+
+
+@_psc_stream_params
+async def test_psc_flag_off_stores_by_default_persists_once_at_end(stream: bool) -> None:
+    """Flag off + storing client: once-per-run persistence, and the service conversation id propagates."""
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=False)
+    session = agent.create_session()
+
+    await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert client.saves_before_call == [0, 0]
+    assert provider.save_calls == 1
+    assert session.service_session_id == _PSC_SERVICE_CONVERSATION_ID
+
+
+@_psc_stream_params
+async def test_psc_flag_on_storing_with_existing_conversation_id_does_not_raise(stream: bool) -> None:
+    """Allow side of the guard: flag on + storing client + an existing conversation_id resumes (no raise).
+
+    The non-storing path raises on an existing service-managed conversation id, but with a storing
+    client the run must proceed and the service conversation id must propagate to the session.
+    """
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    if stream:
+        chunks: list[str] = []
+        async for update in agent.run(
+            "What's the weather in Seattle?",
+            session=session,
+            stream=True,
+            options={"conversation_id": "existing_conversation"},
+        ):
+            chunks.append(update.text or "")
+        text = "".join(chunks)
+    else:
+        result = await agent.run(
+            "What's the weather in Seattle?",
+            session=session,
+            options={"conversation_id": "existing_conversation"},
+        )
+        text = result.text
+
+    assert text == "It is sunny in Seattle."
+    # Persistence still happens per service call, and the real service id propagates to the session.
+    assert provider.save_calls == 2
+    assert provider.get_calls == 0
+    assert session.service_session_id == _PSC_SERVICE_CONVERSATION_ID
+
+
+@_psc_stream_params
+async def test_psc_flag_on_storing_two_runs_same_session(stream: bool) -> None:
+    """Storing mode across two runs on one session: persistence keeps happening, id is stable, no load.
+
+    The second run exercises the precedence branch where the session already carries a
+    service_session_id, which must continue to skip provider loading and keep persisting.
+    """
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(provider=provider, stores_by_default=True, script=_psc_function_call_script())
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    assert provider.save_calls == 2
+    assert provider.get_calls == 0
+    first_run_service_id = session.service_session_id
+    assert first_run_service_id == _PSC_SERVICE_CONVERSATION_ID
+
+    # Reset the scripted client for a second run on the same session.
+    client._script = _psc_function_call_script()
+    client.call_count = 0
+    client.saves_before_call = []
+
+    await _psc_run(agent, "And in Portland?", session, stream=stream)
+
+    # Persistence keeps happening on the second run (two more saves), still per service call.
+    assert client.saves_before_call == [2, 3]
+    assert provider.save_calls == 4
+    # Loading stays skipped and the service conversation id stays stable across runs.
+    assert provider.get_calls == 0
+    assert session.service_session_id == first_run_service_id
+
+
+@_psc_stream_params
+async def test_psc_flag_on_storing_without_conversation_id_warns_every_call(
+    stream: bool, caplog: pytest.LogCaptureFixture
+) -> None:
+    """Storing mode but the client returns no conversation_id: warn on every service call.
+
+    Without an echoed conversation id the next run has nothing to resume from, so cross-turn
+    history can be lost silently. The warning fires per service call (no dedup) so the uncommon
+    failure mode cannot pass unnoticed.
+    """
+    provider = _PscSpyHistoryProvider()
+    client = _PscSpyChatClient(
+        provider=provider,
+        stores_by_default=True,
+        script=_psc_function_call_script(),
+        echo_conversation_id=False,
+    )
+    agent = _psc_build_agent(client, provider, require_per_service_call_history_persistence=True)
+    session = agent.create_session()
+
+    with caplog.at_level(logging.WARNING, logger="agent_framework"):
+        await _psc_run(agent, "What's the weather in Seattle?", session, stream=stream)
+
+    # Persistence still happens, but no service id is captured to resume from.
+    assert provider.save_calls == 2
+    assert session.service_session_id is None
+    # Two service calls -> the warning is emitted twice (one per call, not deduped).
+    missing_id_warnings = [r for r in caplog.records if "returned no conversation_id" in r.message]
+    assert len(missing_id_warnings) == 2
@@ -0,0 +1,376 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Tests for MCP client span instrumentation per OTel GenAI Semantic Conventions.
+
+See: https://opentelemetry.io/docs/specs/semconv/gen-ai/mcp/#client
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import AsyncMock, Mock
+
+import pytest
+from mcp import types
+from mcp.shared.exceptions import McpError
+from mcp.types import ErrorData
+from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
+from opentelemetry.trace import SpanKind, StatusCode
+
+from agent_framework import MCPStdioTool, MCPStreamableHTTPTool, MCPWebsocketTool
+from agent_framework._mcp import MCPTool
+from agent_framework.exceptions import ToolExecutionException
+from agent_framework.observability import OtelAttr
+
+# region helpers
+
+
+def _make_connected_mcp_tool(
+    name: str = "test-mcp",
+    *,
+    supports_tools: bool = True,
+    supports_prompts: bool = True,
+) -> MCPTool:
+    """Create an MCPTool with a mocked session, ready for testing."""
+    tool = MCPTool(name=name)
+    tool.session = AsyncMock()
+    tool.is_connected = True
+    tool._supports_tools = supports_tools
+    tool._supports_prompts = supports_prompts
+    tool.load_tools_flag = True
+    tool.load_prompts_flag = True
+    return tool
+
+
+def _make_tool_list_result(
+    tools: list[dict[str, Any]] | None = None,
+) -> Mock:
+    """Create a mock ListToolsResult."""
+    if tools is None:
+        tools = [{"name": "get-weather", "description": "Get weather", "inputSchema": {"type": "object"}}]
+    result = Mock()
+    result.tools = [
+        types.Tool(name=t["name"], description=t.get("description", ""), inputSchema=t.get("inputSchema", {}))
+        for t in tools
+    ]
+    result.nextCursor = None
+    return result
+
+
+def _make_prompt_list_result(
+    prompts: list[dict[str, Any]] | None = None,
+) -> Mock:
+    """Create a mock ListPromptsResult."""
+    if prompts is None:
+        prompts = [{"name": "analyze-code", "description": "Analyze code"}]
+    result = Mock()
+    result.prompts = [
+        types.Prompt(name=p["name"], description=p.get("description", ""), arguments=None) for p in prompts
+    ]
+    result.nextCursor = None
+    return result
+
+
+def _make_call_tool_result(text: str = "result", is_error: bool = False) -> Mock:
+    """Create a mock CallToolResult."""
+    result = Mock()
+    result.isError = is_error
+    result.content = [types.TextContent(type="text", text=text)]
+    return result
+
+
+def _make_get_prompt_result(text: str = "prompt result") -> types.GetPromptResult:
+    """Create a mock GetPromptResult."""
+    return types.GetPromptResult(
+        description="test prompt",
+        messages=[
+            types.PromptMessage(
+                role="user",
+                content=types.TextContent(type="text", text=text),
+            )
+        ],
+    )
+
+
+# endregion
+
+
+# region initialize span
+
+
+async def test_mcp_initialize_span(span_exporter: InMemorySpanExporter):
+    """session.initialize() should produce an MCP CLIENT span named 'initialize'."""
+    tool = MCPTool(name="test-server")
+
+    mock_session_cls = AsyncMock()
+    init_result = Mock()
+    init_result.capabilities = None
+    init_result.protocolVersion = "2025-06-18"
+    mock_session_cls.initialize = AsyncMock(return_value=init_result)
+
+    # Create a mock transport context manager
+    mock_transport = AsyncMock()
+    mock_transport.__aenter__ = AsyncMock(return_value=(Mock(), Mock()))
+    mock_transport.__aexit__ = AsyncMock(return_value=False)
+
+    # Mock get_mcp_client and the session creation
+    tool.session = None
+    tool.load_tools_flag = False
+    tool.load_prompts_flag = False
+
+    span_exporter.clear()
+
+    with pytest.MonkeyPatch.context() as m:
+        m.setattr(tool, "get_mcp_client", lambda: mock_transport)
+
+        async def patched_connect(self_: Any, *, reset: bool = False, load_configured: bool = True) -> None:
+            # Simulate _connect_on_owner: create initialize span and call session.initialize()
+            from agent_framework._mcp import create_mcp_client_span
+            from agent_framework.observability import OtelAttr
+
+            with create_mcp_client_span("initialize", attributes=self_._mcp_base_span_attributes()) as init_span:
+                result = await mock_session_cls.initialize()
+                protocol_version = getattr(result, "protocolVersion", None)
+                if protocol_version:
+                    init_span.set_attribute(OtelAttr.MCP_PROTOCOL_VERSION, protocol_version)
+
+            self_.session = mock_session_cls
+            self_.is_connected = True
+
+        m.setattr(MCPTool, "_connect_on_owner", patched_connect)
+        await tool.connect()
+
+    mock_session_cls.initialize.assert_awaited_once()
+    spans = span_exporter.get_finished_spans()
+    init_spans = [s for s in spans if s.name == "initialize"]
+    assert len(init_spans) == 1
+    span = init_spans[0]
+    assert span.kind == SpanKind.CLIENT
+    assert span.attributes[OtelAttr.MCP_METHOD_NAME] == "initialize"
+    assert span.attributes.get(OtelAttr.MCP_PROTOCOL_VERSION) == "2025-06-18"
+
+
+# endregion
+
+
+# region tools/list span
+
+
+async def test_mcp_tools_list_span(span_exporter: InMemorySpanExporter):
+    """session.list_tools() should produce an MCP CLIENT span named 'tools/list'."""
+    tool = _make_connected_mcp_tool()
+    tool.session.list_tools = AsyncMock(return_value=_make_tool_list_result())
+
+    span_exporter.clear()
+    await tool.load_tools()
+
+    spans = span_exporter.get_finished_spans()
+    list_spans = [s for s in spans if s.name == "tools/list"]
+    assert len(list_spans) == 1
+    span = list_spans[0]
+    assert span.kind == SpanKind.CLIENT
+    assert span.attributes[OtelAttr.MCP_METHOD_NAME] == "tools/list"
+
+
+# endregion
+
+
+# region prompts/list span
+
+
+async def test_mcp_prompts_list_span(span_exporter: InMemorySpanExporter):
+    """session.list_prompts() should produce an MCP CLIENT span named 'prompts/list'."""
+    tool = _make_connected_mcp_tool()
+    tool.session.list_prompts = AsyncMock(return_value=_make_prompt_list_result())
+
+    span_exporter.clear()
+    await tool.load_prompts()
+
+    spans = span_exporter.get_finished_spans()
+    list_spans = [s for s in spans if s.name == "prompts/list"]
+    assert len(list_spans) == 1
+    span = list_spans[0]
+    assert span.kind == SpanKind.CLIENT
+    assert span.attributes[OtelAttr.MCP_METHOD_NAME] == "prompts/list"
+
+
+# endregion
+
+
+# region tools/call span
+
+
+async def test_mcp_tools_call_creates_client_span_when_no_parent(span_exporter: InMemorySpanExporter):
+    """Direct call_tool() without FunctionTool wrapper creates new MCP CLIENT span."""
+    tool = _make_connected_mcp_tool()
+    tool.session.call_tool = AsyncMock(return_value=_make_call_tool_result("hello"))
+
+    span_exporter.clear()
+    result = await tool.call_tool("get-weather", city="Seattle")
+
+    assert result is not None
+    spans = span_exporter.get_finished_spans()
+    call_spans = [s for s in spans if "tools/call" in s.name]
+    assert len(call_spans) == 1
+    span = call_spans[0]
+    assert span.kind == SpanKind.CLIENT
+    assert span.name == "tools/call get-weather"
+    assert span.attributes[OtelAttr.MCP_METHOD_NAME] == "tools/call"
+    assert span.attributes[OtelAttr.TOOL_NAME] == "get-weather"
+
+
+async def test_mcp_tools_call_tool_error_sets_error_type(span_exporter: InMemorySpanExporter):
+    """When CallToolResult.isError is true, error.type should be 'tool_error' per MCP spec."""
+    tool = _make_connected_mcp_tool()
+    tool.session.call_tool = AsyncMock(return_value=_make_call_tool_result("bad input", is_error=True))
+
+    span_exporter.clear()
+    with pytest.raises(ToolExecutionException):
+        await tool.call_tool("get-weather", city="invalid")
+
+    spans = span_exporter.get_finished_spans()
+    call_spans = [s for s in spans if "tools/call" in s.name]
+    assert len(call_spans) == 1
+    span = call_spans[0]
+    assert span.attributes.get(OtelAttr.ERROR_TYPE) == "tool_error"
+    assert span.status.status_code == StatusCode.ERROR
+
+
+async def test_mcp_tools_call_mcp_error_sets_error_type(span_exporter: InMemorySpanExporter):
+    """When session.call_tool() raises McpError, error.type should be the exception class name."""
+    tool = _make_connected_mcp_tool()
+    tool.session.call_tool = AsyncMock(side_effect=McpError(ErrorData(code=-32600, message="invalid request")))
+
+    span_exporter.clear()
+    with pytest.raises(ToolExecutionException):
+        await tool.call_tool("get-weather")
+
+    spans = span_exporter.get_finished_spans()
+    call_spans = [s for s in spans if "tools/call" in s.name]
+    assert len(call_spans) == 1
+    span = call_spans[0]
+    assert span.attributes.get(OtelAttr.ERROR_TYPE) == "McpError"
+    assert span.status.status_code == StatusCode.ERROR
+
+
+# endregion
+
+
+# region prompts/get span
+
+
+async def test_mcp_prompts_get_creates_client_span(span_exporter: InMemorySpanExporter):
+    """get_prompt() should always create a new MCP CLIENT span (not enrich execute_tool)."""
+    tool = _make_connected_mcp_tool()
+    tool.session.get_prompt = AsyncMock(return_value=_make_get_prompt_result("code analysis"))
+
+    span_exporter.clear()
+    result = await tool.get_prompt("analyze-code", language="python")
+
+    assert "code analysis" in result
+    spans = span_exporter.get_finished_spans()
+    prompt_spans = [s for s in spans if "prompts/get" in s.name]
+    assert len(prompt_spans) == 1
+    span = prompt_spans[0]
+    assert span.kind == SpanKind.CLIENT
+    assert span.name == "prompts/get analyze-code"
+    assert span.attributes[OtelAttr.MCP_METHOD_NAME] == "prompts/get"
+    assert span.attributes[OtelAttr.PROMPT_NAME] == "analyze-code"
+
+
+async def test_mcp_prompts_get_mcp_error_sets_error_type(span_exporter: InMemorySpanExporter):
+    """When session.get_prompt() raises McpError, the span should have error.type and ERROR status."""
+    tool = _make_connected_mcp_tool()
+    tool.session.get_prompt = AsyncMock(
+        side_effect=McpError(ErrorData(code=-32602, message="prompt not found"))
+    )
+
+    span_exporter.clear()
+    with pytest.raises(ToolExecutionException):
+        await tool.get_prompt("missing-prompt")
+
+    spans = span_exporter.get_finished_spans()
+    prompt_spans = [s for s in spans if "prompts/get" in s.name]
+    assert len(prompt_spans) == 1
+    span = prompt_spans[0]
+    assert span.attributes.get(OtelAttr.ERROR_TYPE) == "McpError"
+    assert span.status.status_code == StatusCode.ERROR
+
+
+# endregion
+
+
+# region transport attributes
+
+
+def test_mcp_stdio_tool_transport_attributes():
+    """MCPStdioTool should have network.transport='pipe'."""
+    tool = MCPStdioTool(name="test", command="python")
+    attrs = tool._mcp_base_span_attributes()
+    assert attrs[OtelAttr.NETWORK_TRANSPORT] == "pipe"
+    assert OtelAttr.ADDRESS not in attrs
+
+
+def test_mcp_http_tool_transport_attributes():
+    """MCPStreamableHTTPTool should have tcp transport and URL-based server address/port."""
+    tool = MCPStreamableHTTPTool(name="test", url="https://api.example.com:8443/mcp")
+    attrs = tool._mcp_base_span_attributes()
+    assert attrs[OtelAttr.NETWORK_TRANSPORT] == "tcp"
+    assert attrs[OtelAttr.NETWORK_PROTOCOL_NAME] == "http"
+    assert attrs[OtelAttr.ADDRESS] == "api.example.com"
+    assert attrs[OtelAttr.PORT] == 8443
+
+
+def test_mcp_http_tool_default_port():
+    """MCPStreamableHTTPTool should default to 443 for https."""
+    tool = MCPStreamableHTTPTool(name="test", url="https://api.example.com/mcp")
+    attrs = tool._mcp_base_span_attributes()
+    assert attrs[OtelAttr.PORT] == 443
+
+
+def test_mcp_http_tool_http_default_port():
+    """MCPStreamableHTTPTool should default to 80 for http."""
+    tool = MCPStreamableHTTPTool(name="test", url="http://localhost/mcp")
+    attrs = tool._mcp_base_span_attributes()
+    assert attrs[OtelAttr.PORT] == 80
+
+
+def test_mcp_websocket_tool_transport_attributes():
+    """MCPWebsocketTool should have tcp transport and URL-based server address/port."""
+    tool = MCPWebsocketTool(name="test", url="wss://ws.example.com:9090/mcp")
+    attrs = tool._mcp_base_span_attributes()
+    assert attrs[OtelAttr.NETWORK_TRANSPORT] == "tcp"
+    assert attrs[OtelAttr.NETWORK_PROTOCOL_NAME] == "websocket"
+    assert attrs[OtelAttr.ADDRESS] == "ws.example.com"
+    assert attrs[OtelAttr.PORT] == 9090
+
+
+def test_mcp_websocket_tool_default_port():
+    """MCPWebsocketTool should default to 443 for wss."""
+    tool = MCPWebsocketTool(name="test", url="wss://ws.example.com/mcp")
+    attrs = tool._mcp_base_span_attributes()
+    assert attrs[OtelAttr.PORT] == 443
+
+
+# endregion
+
+
+# region observability disabled
+
+
+@pytest.mark.parametrize("enable_instrumentation", [False], indirect=True)
+async def test_mcp_spans_not_created_when_observability_disabled(span_exporter: InMemorySpanExporter):
+    """No MCP spans should be created when observability is disabled."""
+    tool = _make_connected_mcp_tool()
+    tool.session.list_tools = AsyncMock(return_value=_make_tool_list_result())
+    tool.session.call_tool = AsyncMock(return_value=_make_call_tool_result("ok"))
+
+    span_exporter.clear()
+    await tool.load_tools()
+    await tool.call_tool("get-weather", city="Seattle")
+
+    spans = span_exporter.get_finished_spans()
+    assert len(spans) == 0
+
+
+# endregion
@@ -25,6 +25,7 @@ from agent_framework import (
    prepend_agent_framework_to_user_agent,
    tool,
 )
+from agent_framework._serialization import make_json_safe
 from agent_framework.observability import (
    ROLE_EVENT_MAP,
    AgentTelemetryLayer,
@@ -3195,17 +3196,15 @@ def test_capture_messages_with_prepared_request_info_function_call_arguments(spa

    from opentelemetry import trace

-    from agent_framework import WorkflowAgent
-
    @dataclasses.dataclass
    class HandoffRequest:
        target_agent: str
        reason: str

-    arguments = WorkflowAgent.RequestInfoFunctionArgs(
-        request_id="call_dc",
-        data=HandoffRequest(target_agent="helper", reason="overflow"),
-    ).to_dict()
+    arguments = {
+        "request_id": "call_dc",
+        "data": make_json_safe(HandoffRequest(target_agent="helper", reason="overflow")),
+    }
    msg = Message(
        role="assistant",
        contents=[
@@ -699,3 +699,171 @@ async def test_resolve_executor_kwargs_empty_per_executor_does_not_fallback_to_g
    resolved = {"exec_a": {}, GLOBAL_KWARGS_KEY: {"global_key": "global_val"}}
    result = executor._resolve_executor_kwargs(resolved)  # pyright: ignore[reportPrivateUsage]
    assert result == {}
+
+
+# region Tool approval emission
+
+
+class _ApprovalEmittingAgent(BaseAgent):
+    """Agent that returns a single ``function_approval_request`` Content.
+
+    Used to verify that ``AgentExecutor`` does *not* surface the approval
+    payload via both an ``output`` event and a ``request_info`` event in the
+    same superstep — only the ``request_info`` event must carry it.
+    """
+
+    def __init__(
+        self,
+        *,
+        approval_request_id: str = "apr_1",
+        tool_name: str = "delete_file",
+        tool_arguments: dict[str, Any] | None = None,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self._approval_request_id = approval_request_id
+        self._tool_name = tool_name
+        self._tool_arguments: dict[str, Any] = tool_arguments or {"path": "/tmp/secret.txt"}
+        self.run_count = 0
+
+    def _build_approval_content(self) -> Content:
+        function_call = Content.from_function_call(
+            call_id=self._approval_request_id,
+            name=self._tool_name,
+            arguments=self._tool_arguments,
+        )
+        return Content.from_function_approval_request(id=self._approval_request_id, function_call=function_call)
+
+    @overload
+    def run(
+        self,
+        messages: AgentRunInputs | None = ...,
+        *,
+        stream: Literal[False] = ...,
+        session: AgentSession | None = ...,
+        **kwargs: Any,
+    ) -> Awaitable[AgentResponse[Any]]: ...
+
+    @overload
+    def run(
+        self,
+        messages: AgentRunInputs | None = ...,
+        *,
+        stream: Literal[True],
+        session: AgentSession | None = ...,
+        **kwargs: Any,
+    ) -> ResponseStream[AgentResponseUpdate, AgentResponse[Any]]: ...
+
+    def run(
+        self,
+        messages: AgentRunInputs | None = None,
+        *,
+        stream: bool = False,
+        session: AgentSession | None = None,
+        **kwargs: Any,
+    ) -> Awaitable[AgentResponse[Any]] | ResponseStream[AgentResponseUpdate, AgentResponse[Any]]:
+        self.run_count += 1
+        approval = self._build_approval_content()
+
+        if stream:
+
+            async def _stream() -> AsyncIterable[AgentResponseUpdate]:
+                yield AgentResponseUpdate(contents=[approval], role="assistant")
+
+            return ResponseStream(_stream(), finalizer=AgentResponse.from_updates)
+
+        async def _run() -> AgentResponse:
+            return AgentResponse(messages=[Message("assistant", [approval])])
+
+        return _run()
+
+
+def _has_approval_payload(event: WorkflowEvent[Any]) -> bool:
+    """Return True if the event's data carries a ``function_approval_request`` content."""
+    data: Any = event.data
+
+    def _contents_of(value: Any) -> list[Content]:
+        if isinstance(value, AgentResponseUpdate):
+            return list(value.contents)
+        if isinstance(value, AgentResponse):
+            return [c for m in value.messages for c in m.contents]
+        if isinstance(value, AgentExecutorResponse):
+            return [c for m in value.agent_response.messages for c in m.contents]
+        if isinstance(value, Message):
+            return list(value.contents)
+        if isinstance(value, Content):
+            return [value]
+        return []
+
+    return any(c.type == "function_approval_request" for c in _contents_of(data))
+
+
+async def test_agent_executor_does_not_double_emit_approval_non_streaming() -> None:
+    """Non-streaming: approval payload must only appear in the ``request_info`` event.
+
+    Regression test for the bug where ``AgentExecutor._run_agent`` first
+    ``yield_output``-ed the response (carrying the approval Content) and then
+    additionally emitted a ``request_info`` event for the same payload.
+    """
+    agent = _ApprovalEmittingAgent(id="approve_agent", name="ApproveAgent", approval_request_id="apr_ns_1")
+    executor = AgentExecutor(agent, id="approve_exec")
+    workflow = WorkflowBuilder(start_executor=executor).build()
+
+    request_info_events: list[WorkflowEvent[Any]] = []
+    output_events: list[WorkflowEvent[Any]] = []
+
+    for event in await workflow.run("please delete it"):
+        if event.type == "request_info":
+            request_info_events.append(event)
+        elif event.type == "output":
+            output_events.append(event)
+
+    assert len(request_info_events) == 1
+    assert _has_approval_payload(request_info_events[0])
+    # The approval payload must not also be surfaced as a workflow output.
+    assert not any(_has_approval_payload(e) for e in output_events)
+    assert agent.run_count == 1
+
+
+async def test_agent_executor_does_not_double_emit_approval_streaming() -> None:
+    """Streaming: per-update approval payload must not be ``yield_output``-ed."""
+    agent = _ApprovalEmittingAgent(id="approve_agent_s", name="ApproveAgentS", approval_request_id="apr_st_1")
+    executor = AgentExecutor(agent, id="approve_exec_s")
+    workflow = WorkflowBuilder(start_executor=executor).build()
+
+    request_info_events: list[WorkflowEvent[Any]] = []
+    output_events: list[WorkflowEvent[Any]] = []
+
+    async for event in workflow.run("please delete it", stream=True):
+        if event.type == "request_info":
+            request_info_events.append(event)
+        elif event.type == "output":
+            output_events.append(event)
+
+    assert len(request_info_events) == 1
+    assert _has_approval_payload(request_info_events[0])
+    assert not any(_has_approval_payload(e) for e in output_events)
+    assert agent.run_count == 1
+
+
+async def test_agent_executor_request_info_uses_user_input_request_id() -> None:
+    """``ctx.request_info`` must register the request under the agent's approval id.
+
+    This makes the workflow's pending-request id round-trip with the
+    ``function_approval_response.id`` the caller echoes back, so
+    ``Workflow._send_responses_internal`` can look it up directly.
+    """
+    agent = _ApprovalEmittingAgent(id="approve_agent_id", name="ApproveAgentId", approval_request_id="apr_match")
+    executor = AgentExecutor(agent, id="approve_exec_id")
+    workflow = WorkflowBuilder(start_executor=executor).build()
+
+    request_info_events: list[WorkflowEvent[Any]] = []
+    async for event in workflow.run("please delete it", stream=True):
+        if event.type == "request_info":
+            request_info_events.append(event)
+
+    assert len(request_info_events) == 1
+    assert request_info_events[0].request_id == "apr_match"
+
+
+# endregion Tool approval emission
@@ -1,6 +1,5 @@
 # Copyright (c) Microsoft. All rights reserved.

-import json
 import uuid
 from collections.abc import Awaitable, Sequence
 from dataclasses import dataclass
@@ -30,6 +29,20 @@ from agent_framework import (
    handler,
    response_handler,
 )
+from agent_framework._workflows._typing_utils import deserialize_type
+
+
+@dataclass
+class HandoffRequest:
+    """Module-level dataclass used by request_info tests.
+
+    Defined at module scope (not nested inside a test method) so
+    ``serialize_type``/``deserialize_type`` can round-trip the request_type via
+    the importable qualified name ``tests.workflow.test_workflow_agent.HandoffRequest``.
+    """
+
+    target_agent: str
+    reason: str


 class SimpleExecutor(Executor):
@@ -240,52 +253,45 @@ class TestWorkflowAgent:
        # Should have received an approval request for the request info
        assert len(updates) > 0

-        approval_update: AgentResponseUpdate | None = None
+        request_update: AgentResponseUpdate | None = None
        for update in updates:
-            if any(content.type == "function_approval_request" for content in update.contents):
-                approval_update = update
+            if any(content.type == "function_call" for content in update.contents):
+                request_update = update
                break

-        assert approval_update is not None, "Should have received a request_info approval request"
+        assert request_update is not None, "Should have received a request_info wrapped in a function_call content"

-        function_call = next(content for content in approval_update.contents if content.type == "function_call")
-        approval_request = next(
-            content for content in approval_update.contents if content.type == "function_approval_request"
-        )
+        request_function_call = next(content for content in request_update.contents if content.type == "function_call")
+        assert request_function_call.call_id is not None

        # Verify the function call has expected structure
-        assert function_call.call_id is not None
-        assert function_call.name == "request_info"
-        assert isinstance(function_call.arguments, dict)
-        assert function_call.arguments.get("request_id") == approval_request.id
+        assert request_function_call.name == WorkflowAgent.REQUEST_INFO_FUNCTION_NAME
+        assert isinstance(request_function_call.arguments, dict)
+        assert request_function_call.arguments.get("request_id") is not None
+        assert request_function_call.arguments.get("request_event") is not None
+        request_event = request_function_call.arguments["request_event"]
+        assert request_event.get("type") == "request_info"
+        assert deserialize_type(request_event.get("response_type")) is str

-        # Approval request should reference the same function call
-        assert approval_request.id is not None
-        assert approval_request.function_call is not None
-        assert approval_request.function_call.call_id == function_call.call_id
-        assert approval_request.function_call.name == function_call.name
+        deserialized_args = WorkflowAgent.RequestInfoFunctionArgs.from_dict(request_function_call.arguments)
+        assert deserialized_args.request_id == request_function_call.call_id
+        assert isinstance(deserialized_args.request_event, WorkflowEvent)
+        assert deserialized_args.request_event.type == "request_info"
+        assert deserialized_args.request_event.data == "Mock request data"
+        assert deserialized_args.request_event.response_type is str

        # Verify the request is tracked in pending_requests
-        assert len(agent.pending_requests) == 1
-        assert function_call.call_id in agent.pending_requests
+        pending_requests = await workflow._runner_context.get_pending_request_info_events()
+        assert len(pending_requests) == 1
+        assert request_function_call.call_id in pending_requests

-        # Now provide an approval response with updated arguments to test continuation
-        response_args = WorkflowAgent.RequestInfoFunctionArgs(
-            request_id=approval_request.id,
-            data="User provided answer",
-        ).to_dict()
-
-        approval_response = Content.from_function_approval_response(
-            approved=True,
-            id=approval_request.id,
-            function_call=Content.from_function_call(
-                call_id=function_call.call_id,
-                name=function_call.name,
-                arguments=response_args,
-            ),
+        # Now provide a function result response with updated arguments to test continuation
+        function_result = Content.from_function_result(
+            call_id=request_function_call.call_id,
+            result="Mock response to request info",
        )

-        response_message = Message(role="user", contents=[approval_response])
+        response_message = Message(role="user", contents=[function_result])

        # Continue the workflow with the response
        continuation_result = await agent.run(response_message)
@@ -294,16 +300,11 @@ class TestWorkflowAgent:
        assert isinstance(continuation_result, AgentResponse)

        # Verify cleanup - pending requests should be cleared after function response handling
-        assert len(agent.pending_requests) == 0
+        pending_requests = await workflow._runner_context.get_pending_request_info_events()
+        assert len(pending_requests) == 0

    def test_request_info_dataclass_arguments_are_serialized_when_content_is_created(self) -> None:
        """Test WorkflowAgent prepares request_info arguments before observability captures messages."""
-
-        @dataclass
-        class HandoffRequest:
-            target_agent: str
-            reason: str
-
        executor = SimpleExecutor(id="executor1", response_text="Response")
        workflow = WorkflowBuilder(start_executor=executor).build()
        agent = WorkflowAgent(workflow=workflow, name="Request Test Agent")
@@ -314,14 +315,367 @@ class TestWorkflowAgent:
            response_type=str,
        )

-        function_call, approval_request = agent._process_request_info_event(event)  # pyright: ignore[reportPrivateUsage]
+        request_function_call = agent._process_request_info_event(event)  # pyright: ignore[reportPrivateUsage]

-        assert function_call.arguments == {
-            "request_id": "request_123",
-            "data": {"target_agent": "helper", "reason": "overflow"},
-        }
-        assert approval_request.function_call is function_call
-        assert json.loads(json.dumps(function_call.arguments)) == function_call.arguments
+        assert request_function_call.call_id == "request_123"
+        assert isinstance(request_function_call.arguments, dict)
+        assert request_function_call.arguments.get("request_event") is not None
+        request_event = request_function_call.arguments["request_event"]
+        assert request_event.get("type") == "request_info"
+        assert request_event.get("request_id") == "request_123"
+        assert request_event.get("source_executor_id") == "executor1"
+        assert deserialize_type(request_event.get("response_type")) is str
+        assert request_event.get("data") == HandoffRequest(target_agent="helper", reason="overflow")
+
+        deserialized_args = WorkflowAgent.RequestInfoFunctionArgs.from_dict(request_function_call.arguments)
+        assert deserialized_args.request_id == "request_123"
+        assert isinstance(deserialized_args.request_event, WorkflowEvent)
+        assert deserialized_args.request_event.type == "request_info"
+        assert deserialized_args.request_event.data == HandoffRequest(target_agent="helper", reason="overflow")
+        assert deserialized_args.request_event.response_type is str
+
+    def test_process_request_info_event_passes_through_function_approval_request(self) -> None:
+        """If the event data is already a function approval request, it is forwarded unchanged.
+
+        Tool-approval requests emitted by an inner agent surface as ``Content``
+        objects with ``user_input_request=True``. ``WorkflowAgent`` must not
+        re-wrap these inside a synthesized ``request_info`` function call;
+        instead it should return the original content as-is so callers can
+        respond with a matching ``function_approval_response``.
+        """
+        executor = SimpleExecutor(id="executor1", response_text="Response")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Passthrough Agent")
+
+        approval_id = "approval-passthrough-1"
+        inner_function_call = Content.from_function_call(
+            call_id="tool-call-1",
+            name="delete_file",
+            arguments={"path": "/tmp/x"},
+        )
+        approval_request = Content.from_function_approval_request(
+            id=approval_id,
+            function_call=inner_function_call,
+        )
+        event = WorkflowEvent.request_info(
+            request_id=approval_id,
+            source_executor_id="executor1",
+            request_data=approval_request,
+            response_type=Content,
+        )
+
+        result = agent._process_request_info_event(event)  # pyright: ignore[reportPrivateUsage]
+
+        # The original FunctionApprovalRequestContent is returned as-is — same
+        # instance, with the original tool name preserved (NOT replaced by the
+        # synthesized REQUEST_INFO_FUNCTION_NAME).
+        assert result is approval_request
+        assert result.type == "function_approval_request"
+        assert result.id == approval_id
+        assert result.user_input_request is True
+        assert result.function_call is inner_function_call  # type: ignore[attr-defined]
+        assert result.function_call.name == "delete_file"  # type: ignore[attr-defined]
+        assert result.function_call.name != WorkflowAgent.REQUEST_INFO_FUNCTION_NAME  # type: ignore[attr-defined]
+
+    def test_extract_function_responses_passes_through_approval_response_approved(self) -> None:
+        """A function_approval_response with approved=True is keyed by content.id and forwarded as-is.
+
+        After the refactor, ``WorkflowAgent`` no longer unwraps a synthesized
+        ``request_info`` function call from approval responses — the response
+        content is routed straight back to the workflow under its own ``id``,
+        which matches the pending request id surfaced by
+        ``_process_request_info_event``.
+        """
+        executor = SimpleExecutor(id="executor1", response_text="Response")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Response Agent")
+
+        approval_id = "approval-response-approved-1"
+        inner_function_call = Content.from_function_call(
+            call_id="tool-call-1",
+            name="delete_file",
+            arguments={"path": "/tmp/x"},
+        )
+        approval_request = Content.from_function_approval_request(
+            id=approval_id,
+            function_call=inner_function_call,
+        )
+        approval_response = approval_request.to_function_approval_response(approved=True)  # type: ignore[attr-defined]
+        message = Message(role="user", contents=[approval_response])
+
+        responses = agent._extract_function_responses([message])  # pyright: ignore[reportPrivateUsage]
+
+        assert set(responses.keys()) == {approval_id}
+        assert responses[approval_id] is approval_response
+        assert responses[approval_id].approved is True  # type: ignore[attr-defined]
+
+    def test_extract_function_responses_passes_through_approval_response_denied(self) -> None:
+        """A function_approval_response with approved=False is forwarded the same way as an approval.
+
+        Only the ``approved`` flag changes — routing back to the workflow is
+        identical for accept and reject paths.
+        """
+        executor = SimpleExecutor(id="executor1", response_text="Response")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Response Agent")
+
+        approval_id = "approval-response-denied-1"
+        inner_function_call = Content.from_function_call(
+            call_id="tool-call-2",
+            name="send_email",
+            arguments={"to": "alice@example.com"},
+        )
+        approval_request = Content.from_function_approval_request(
+            id=approval_id,
+            function_call=inner_function_call,
+        )
+        approval_response = approval_request.to_function_approval_response(approved=False)  # type: ignore[attr-defined]
+        message = Message(role="user", contents=[approval_response])
+
+        responses = agent._extract_function_responses([message])  # pyright: ignore[reportPrivateUsage]
+
+        assert set(responses.keys()) == {approval_id}
+        assert responses[approval_id] is approval_response
+        assert responses[approval_id].approved is False  # type: ignore[attr-defined]
+
+    async def test_function_approval_request_flows_end_to_end_approved(self) -> None:
+        """End-to-end: an executor emits a function_approval_request, the agent
+        forwards it unchanged, and an ``approved=True`` response resumes the workflow.
+
+        This exercises the full pass-through path:
+        ``ctx.request_info(approval_content, ...)`` -> ``WorkflowAgent`` surfaces
+        the original ``FunctionApprovalRequestContent`` -> caller responds with a
+        ``FunctionApprovalResponseContent`` -> ``WorkflowAgent`` routes it back
+        to the workflow which delivers it to the executor's ``@response_handler``.
+        """
+        approval_id = "e2e-approval-1"
+        inner_function_call = Content.from_function_call(
+            call_id="tool-call-e2e-1",
+            name="delete_file",
+            arguments={"path": "/tmp/x"},
+        )
+        approval_request = Content.from_function_approval_request(
+            id=approval_id,
+            function_call=inner_function_call,
+        )
+
+        class ApprovalRequestingExecutor(Executor):
+            @handler
+            async def handle_message(self, _: list[Message], ctx: WorkflowContext) -> None:
+                await ctx.request_info(approval_request, Content, request_id=approval_id)
+
+            @response_handler
+            async def handle_response(
+                self,
+                original_request: Content,
+                response: Content,
+                ctx: WorkflowContext[Never, AgentResponse],
+            ) -> None:
+                assert response.type == "function_approval_response"
+                assert response.id == approval_id  # type: ignore[attr-defined]
+                approved = bool(response.approved)  # type: ignore[attr-defined]
+                tool_name = original_request.function_call.name  # type: ignore[attr-defined]
+                await ctx.yield_output(
+                    AgentResponse(
+                        messages=[
+                            Message(
+                                role="assistant",
+                                contents=[Content.from_text(text=f"{tool_name} approved={approved}")],
+                            )
+                        ]
+                    )
+                )
+
+        executor = ApprovalRequestingExecutor(id="approval_requester")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="E2E Approval Agent")
+
+        # First run: workflow pauses with the approval request.
+        first = await agent.run("please delete it")
+        assert isinstance(first, AgentResponse)
+
+        forwarded = next(
+            (
+                c
+                for m in first.messages
+                for c in m.contents
+                if c.type == "function_approval_request" and c.id == approval_id
+            ),
+            None,
+        )
+        assert forwarded is approval_request, "Approval request must surface unchanged"
+
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert approval_id in pending
+
+        # Respond with approved=True.
+        approval_response = approval_request.to_function_approval_response(approved=True)  # type: ignore[attr-defined]
+        final = await agent.run(Message(role="user", contents=[approval_response]))
+
+        assert isinstance(final, AgentResponse)
+        final_text = " ".join(m.text or "" for m in final.messages)
+        assert "delete_file approved=True" in final_text
+
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert approval_id not in pending
+
+    async def test_function_approval_request_flows_end_to_end_denied(self) -> None:
+        """End-to-end denied path: ``approved=False`` is delivered to the executor's
+        response handler so the workflow can branch on the rejection."""
+        approval_id = "e2e-approval-deny-1"
+        inner_function_call = Content.from_function_call(
+            call_id="tool-call-e2e-deny-1",
+            name="send_email",
+            arguments={"to": "alice@example.com"},
+        )
+        approval_request = Content.from_function_approval_request(
+            id=approval_id,
+            function_call=inner_function_call,
+        )
+
+        class ApprovalRequestingExecutor(Executor):
+            @handler
+            async def handle_message(self, _: list[Message], ctx: WorkflowContext) -> None:
+                await ctx.request_info(approval_request, Content, request_id=approval_id)
+
+            @response_handler
+            async def handle_response(
+                self,
+                original_request: Content,
+                response: Content,
+                ctx: WorkflowContext[Never, AgentResponse],
+            ) -> None:
+                assert response.type == "function_approval_response"
+                assert response.id == approval_id  # type: ignore[attr-defined]
+                approved = bool(response.approved)  # type: ignore[attr-defined]
+                tool_name = original_request.function_call.name  # type: ignore[attr-defined]
+                await ctx.yield_output(
+                    AgentResponse(
+                        messages=[
+                            Message(
+                                role="assistant",
+                                contents=[Content.from_text(text=f"{tool_name} approved={approved}")],
+                            )
+                        ]
+                    )
+                )
+
+        executor = ApprovalRequestingExecutor(id="approval_requester_deny")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="E2E Approval Deny Agent")
+
+        first = await agent.run("please send")
+        assert isinstance(first, AgentResponse)
+        forwarded = next(
+            (
+                c
+                for m in first.messages
+                for c in m.contents
+                if c.type == "function_approval_request" and c.id == approval_id
+            ),
+            None,
+        )
+        assert forwarded is approval_request
+
+        # Respond with approved=False.
+        approval_response = approval_request.to_function_approval_response(approved=False)  # type: ignore[attr-defined]
+        final = await agent.run(Message(role="user", contents=[approval_response]))
+
+        assert isinstance(final, AgentResponse)
+        final_text = " ".join(m.text or "" for m in final.messages)
+        assert "send_email approved=False" in final_text
+
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert approval_id not in pending
+
+    async def test_request_info_non_approval_flows_end_to_end(self) -> None:
+        """End-to-end: when request data is not a function approval content, the
+        agent surfaces a synthesized ``function_call`` (name=REQUEST_INFO_FUNCTION_NAME)
+        and routes a matching ``function_result`` back to the executor.
+        """
+        captured: dict[str, Any] = {}
+
+        class HandoffRequestingExecutor(Executor):
+            @handler
+            async def handle_message(self, _: list[Message], ctx: WorkflowContext) -> None:
+                await ctx.request_info(
+                    HandoffRequest(target_agent="helper", reason="overflow"),
+                    str,
+                )
+
+            @response_handler
+            async def handle_response(
+                self,
+                original_request: HandoffRequest,
+                response: str,
+                ctx: WorkflowContext[Never, AgentResponse],
+            ) -> None:
+                captured["original"] = original_request
+                captured["response"] = response
+                await ctx.yield_output(
+                    AgentResponse(
+                        messages=[
+                            Message(
+                                role="assistant",
+                                contents=[
+                                    Content.from_text(text=f"handoff to {original_request.target_agent}: {response}")
+                                ],
+                            )
+                        ]
+                    )
+                )
+
+        executor = HandoffRequestingExecutor(id="handoff_requester")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="E2E Handoff Agent")
+
+        # First run: workflow pauses with a synthesized request_info function_call.
+        first = await agent.run("start handoff")
+        assert isinstance(first, AgentResponse)
+
+        function_call = next(
+            (
+                c
+                for m in first.messages
+                for c in m.contents
+                if c.type == "function_call" and c.name == WorkflowAgent.REQUEST_INFO_FUNCTION_NAME
+            ),
+            None,
+        )
+        assert function_call is not None, "Expected a synthesized request_info function_call"
+        assert function_call.call_id is not None
+        assert isinstance(function_call.arguments, dict)
+        request_id = function_call.arguments["request_id"]
+        assert function_call.call_id == request_id
+        request_payload = function_call.arguments["request_event"]
+        assert request_payload.get("type") == "request_info"
+        assert request_payload.get("data") == HandoffRequest(target_agent="helper", reason="overflow")
+
+        deserialized_args = WorkflowAgent.RequestInfoFunctionArgs.from_dict(function_call.arguments)
+        assert deserialized_args.request_id == request_id
+        assert isinstance(deserialized_args.request_event, WorkflowEvent)
+        assert deserialized_args.request_event.type == "request_info"
+        assert deserialized_args.request_event.data == HandoffRequest(target_agent="helper", reason="overflow")
+        assert deserialized_args.request_event.response_type is str
+
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert request_id in pending
+
+        # Respond with a function_result keyed by the call_id.
+        function_result = Content.from_function_result(call_id=request_id, result="ok-do-it")
+        final = await agent.run(Message(role="user", contents=[function_result]))
+
+        assert isinstance(final, AgentResponse)
+        final_text = " ".join(m.text or "" for m in final.messages)
+        assert "handoff to helper: ok-do-it" in final_text
+
+        # The executor's response handler received the original request and the response.
+        assert isinstance(captured.get("original"), HandoffRequest)
+        assert captured["original"].target_agent == "helper"
+        assert captured["response"] == "ok-do-it"
+
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert request_id not in pending

    def test_workflow_as_agent_method(self) -> None:
        """Test that Workflow.as_agent() creates a properly configured WorkflowAgent."""
@@ -1592,3 +1946,406 @@ class TestWorkflowAgentMergeUpdates:

        # Order: text (user), text (assistant), function_result (orphan at end)
        assert content_types == ["text", "text", "function_result"]
+
+
+class _ToolApprovalMockAgent(SupportsAgentRun):
+    """Mock agent whose first run returns a FunctionApprovalRequestContent.
+
+    Subsequent runs (after receiving an approval response in the input messages)
+    return a final assistant text response that echoes the approved arguments.
+
+    This mirrors a real agent whose tool invocation requires user approval.
+    """
+
+    def __init__(
+        self,
+        name: str,
+        *,
+        tool_name: str = "delete_file",
+        tool_arguments: dict[str, Any] | None = None,
+        approval_request_ids: Sequence[str] | None = None,
+    ) -> None:
+        self.id = str(uuid.uuid4())
+        self.name = name
+        self.description: str | None = None
+        self._tool_name = tool_name
+        self._tool_arguments = tool_arguments or {"path": "/tmp/example"}
+        # Pre-allocated request ids so the test can verify what the WorkflowAgent forwards.
+        self._approval_request_ids: list[str] = list(approval_request_ids) if approval_request_ids else []
+        self.run_count = 0
+        # Inputs received on the most recent (continuation) run, for assertions.
+        self.last_run_messages: list[Message] = []
+
+    def create_session(self, **kwargs: Any) -> AgentSession:
+        return AgentSession()
+
+    def get_session(self, *, service_session_id: str, **kwargs: Any) -> AgentSession:
+        return AgentSession()
+
+    def _next_request_id(self) -> str:
+        if self._approval_request_ids:
+            return self._approval_request_ids.pop(0)
+        return str(uuid.uuid4())
+
+    def _build_approval_request(self) -> Content:
+        request_id = self._next_request_id()
+        function_call = Content.from_function_call(
+            call_id=request_id,
+            name=self._tool_name,
+            arguments=self._tool_arguments,
+        )
+        return Content.from_function_approval_request(id=request_id, function_call=function_call)
+
+    @overload
+    def run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = ...,
+        *,
+        stream: Literal[False] = ...,
+        session: AgentSession | None = ...,
+        **kwargs: Any,
+    ) -> Awaitable[AgentResponse[Any]]: ...
+
+    @overload
+    def run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = ...,
+        *,
+        stream: Literal[True],
+        session: AgentSession | None = ...,
+        **kwargs: Any,
+    ) -> ResponseStream[AgentResponseUpdate, AgentResponse[Any]]: ...
+
+    def run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = None,
+        *,
+        stream: bool = False,
+        session: AgentSession | None = None,
+        **kwargs: Any,
+    ) -> Awaitable[AgentResponse] | ResponseStream[AgentResponseUpdate, AgentResponse]:
+        if stream:
+            return self._run_stream(messages=messages, session=session, **kwargs)
+        return self._run(messages=messages, session=session, **kwargs)
+
+    def _normalize(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None,
+    ) -> list[Message]:
+        if messages is None:
+            return []
+        if isinstance(messages, str):
+            return [Message(role="user", contents=[Content.from_text(text=messages)])]
+        if isinstance(messages, Message):
+            return [messages]
+        if isinstance(messages, Content):
+            return [Message(role="user", contents=[messages])]
+        result: list[Message] = []
+        for item in messages:
+            if isinstance(item, Message):
+                result.append(item)
+            elif isinstance(item, Content):
+                result.append(Message(role="user", contents=[item]))
+            else:
+                result.append(Message(role="user", contents=[Content.from_text(text=item)]))
+        return result
+
+    def _approval_responses_in(self, messages: list[Message]) -> list[Content]:
+        approvals: list[Content] = []
+        for msg in messages:
+            for content in msg.contents:
+                if content.type == "function_approval_response":
+                    approvals.append(content)
+        return approvals
+
+    async def _run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = None,
+        *,
+        session: AgentSession | None = None,
+        **kwargs: Any,
+    ) -> AgentResponse:
+        normalized = self._normalize(messages)
+        self.last_run_messages = normalized
+        self.run_count += 1
+
+        approvals = self._approval_responses_in(normalized)
+        if approvals:
+            # Continuation: reflect approved arguments in the final response text.
+            approved_text = "; ".join(
+                f"approved={a.approved} id={a.id}"  # type: ignore[attr-defined]
+                for a in approvals
+            )
+            return AgentResponse(messages=[Message("assistant", [Content.from_text(text=f"done ({approved_text})")])])
+
+        # First run: ask for tool approval.
+        approval = self._build_approval_request()
+        return AgentResponse(messages=[Message("assistant", [approval])])
+
+    def _run_stream(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = None,
+        *,
+        session: AgentSession | None = None,
+        **kwargs: Any,
+    ) -> ResponseStream[AgentResponseUpdate, AgentResponse]:
+        normalized = self._normalize(messages)
+        self.last_run_messages = normalized
+        self.run_count += 1
+        approvals = self._approval_responses_in(normalized)
+
+        async def _iter():
+            if approvals:
+                approved_text = "; ".join(
+                    f"approved={a.approved} id={a.id}"  # type: ignore[attr-defined]
+                    for a in approvals
+                )
+                yield AgentResponseUpdate(
+                    contents=[Content.from_text(text=f"done ({approved_text})")],
+                    role="assistant",
+                    author_name=self.name,
+                )
+                return
+            approval = self._build_approval_request()
+            yield AgentResponseUpdate(
+                contents=[approval],
+                role="assistant",
+                author_name=self.name,
+            )
+
+        return ResponseStream(_iter(), finalizer=AgentResponse.from_updates)
+
+
+class TestWorkflowAgentToolApproval:
+    """Tests for tool-approval requests bubbling through WorkflowAgent.
+
+    Covers the case where a workflow contains an AgentExecutor whose underlying
+    agent emits a FunctionApprovalRequestContent (tool needing user approval).
+    The WorkflowAgent must:
+      * forward the original FunctionApprovalRequestContent unchanged (no
+        wrapping inside a synthesized 'request_info' function call), and
+      * route a subsequent FunctionApprovalResponseContent back to the
+        AgentExecutor so the agent can resume.
+    """
+
+    def _find_approval_request(
+        self,
+        contents: Sequence[Content],
+        tool_name: str,
+    ) -> Content | None:
+        for content in contents:
+            if (
+                content.type == "function_approval_request"
+                and getattr(content.function_call, "name", None) == tool_name  # type: ignore[attr-defined]
+            ):
+                return content
+        return None
+
+    async def test_tool_approval_request_forwarded_unchanged(self) -> None:
+        """The agent's FunctionApprovalRequestContent surfaces verbatim (not re-wrapped)."""
+        approval_id = "approval-abc-123"
+        mock_agent = _ToolApprovalMockAgent(
+            name="approval-agent",
+            tool_name="delete_file",
+            tool_arguments={"path": "/tmp/secret.txt"},
+            approval_request_ids=[approval_id],
+        )
+
+        @executor
+        async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+            await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+        workflow = WorkflowBuilder(start_executor=start).add_edge(start, mock_agent).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Test Agent")
+
+        result = await agent.run("please delete the file")
+
+        assert isinstance(result, AgentResponse)
+
+        # Locate the approval request emitted by the WorkflowAgent.
+        all_contents: list[Content] = [c for m in result.messages for c in m.contents]
+        approval = self._find_approval_request(all_contents, tool_name="delete_file")
+        assert approval is not None, "WorkflowAgent did not forward the tool approval request"
+
+        # The id and inner function_call must match what the underlying agent produced
+        # — i.e. the WorkflowAgent must NOT have re-wrapped it inside a synthesized
+        # 'request_info' approval request.
+        assert approval.id == approval_id
+        function_call = approval.function_call  # type: ignore[attr-defined]
+        assert function_call is not None
+        assert function_call.name == "delete_file"
+        assert function_call.name != WorkflowAgent.REQUEST_INFO_FUNCTION_NAME
+        assert function_call.arguments == {"path": "/tmp/secret.txt"}
+
+        # The agent must be paused awaiting the approval response.
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert approval_id in pending
+
+    async def test_tool_approval_request_forwarded_unchanged_streaming(self) -> None:
+        """Streaming variant: the approval request is forwarded as-is in updates."""
+        approval_id = "approval-stream-1"
+        mock_agent = _ToolApprovalMockAgent(
+            name="approval-agent-stream",
+            tool_name="send_email",
+            tool_arguments={"to": "alice@example.com"},
+            approval_request_ids=[approval_id],
+        )
+
+        @executor
+        async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+            await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+        workflow = WorkflowBuilder(start_executor=start).add_edge(start, mock_agent).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Stream Agent")
+
+        updates: list[AgentResponseUpdate] = []
+        async for update in agent.run("hi", stream=True):
+            updates.append(update)
+
+        approval_updates = [u for u in updates if any(c.type == "function_approval_request" for c in u.contents)]
+        assert approval_updates, "Streaming did not surface a tool approval request"
+
+        approval = self._find_approval_request(approval_updates[-1].contents, tool_name="send_email")
+        assert approval is not None
+        assert approval.id == approval_id
+        function_call = approval.function_call  # type: ignore[attr-defined]
+        assert function_call is not None
+        assert function_call.name == "send_email"
+        assert function_call.name != WorkflowAgent.REQUEST_INFO_FUNCTION_NAME
+        assert function_call.arguments == {"to": "alice@example.com"}
+
+    async def test_tool_approval_response_resumes_agent(self) -> None:
+        """Sending the approval response back resumes the agent and clears pending requests."""
+        approval_id = "approval-resume-1"
+        mock_agent = _ToolApprovalMockAgent(
+            name="approval-resume-agent",
+            tool_name="delete_file",
+            tool_arguments={"path": "/tmp/x"},
+            approval_request_ids=[approval_id],
+        )
+
+        @executor
+        async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+            await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+        workflow = WorkflowBuilder(start_executor=start).add_edge(start, mock_agent).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Resume Agent")
+
+        first_result = await agent.run("delete it")
+        approval = self._find_approval_request(
+            [c for m in first_result.messages for c in m.contents],
+            tool_name="delete_file",
+        )
+        assert approval is not None
+        assert mock_agent.run_count == 1
+
+        # Build the approval response. NOTE: the inner function_call's name is the
+        # original tool name ('delete_file'), NOT 'request_info'. This exercises the
+        # branch in WorkflowAgent._extract_function_responses that routes raw
+        # tool-approval responses straight through using content.id.
+        approval_response = approval.to_function_approval_response(approved=True)  # type: ignore[attr-defined]
+        response_message = Message(role="user", contents=[approval_response])
+
+        final_result = await agent.run(response_message)
+        assert isinstance(final_result, AgentResponse)
+
+        # The mock agent should have been invoked a second time and seen the
+        # approval response in its inputs.
+        assert mock_agent.run_count == 2
+        approvals_seen = [
+            c for m in mock_agent.last_run_messages for c in m.contents if c.type == "function_approval_response"
+        ]
+        assert len(approvals_seen) == 1
+        assert approvals_seen[0].id == approval_id  # type: ignore[attr-defined]
+        assert approvals_seen[0].approved is True  # type: ignore[attr-defined]
+
+        # The pending approval should now be cleared.
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert approval_id not in pending
+
+        # The final assistant message reflects the resumption.
+        final_text = " ".join(m.text or "" for m in final_result.messages)
+        assert "done" in final_text
+        assert approval_id in final_text
+
+    async def test_tool_approval_response_rejected_resumes_agent(self) -> None:
+        """Rejection path: ``approved=False`` is forwarded to the inner agent and clears the pending request.
+
+        The WorkflowAgent must route a rejection response back to the paused
+        ``AgentExecutor`` exactly the same way as an approval — only the
+        ``approved`` flag differs. The inner agent decides what to do with it.
+        """
+        approval_id = "approval-reject-1"
+        mock_agent = _ToolApprovalMockAgent(
+            name="approval-reject-agent",
+            tool_name="delete_file",
+            tool_arguments={"path": "/tmp/x"},
+            approval_request_ids=[approval_id],
+        )
+
+        @executor
+        async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+            await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+        workflow = WorkflowBuilder(start_executor=start).add_edge(start, mock_agent).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Reject Agent")
+
+        first_result = await agent.run("delete it")
+        approval = self._find_approval_request(
+            [c for m in first_result.messages for c in m.contents],
+            tool_name="delete_file",
+        )
+        assert approval is not None
+        assert mock_agent.run_count == 1
+
+        # Reject the tool invocation.
+        approval_response = approval.to_function_approval_response(approved=False)  # type: ignore[attr-defined]
+        response_message = Message(role="user", contents=[approval_response])
+
+        final_result = await agent.run(response_message)
+        assert isinstance(final_result, AgentResponse)
+
+        # The inner agent must have been resumed and seen ``approved=False``.
+        assert mock_agent.run_count == 2
+        approvals_seen = [
+            c for m in mock_agent.last_run_messages for c in m.contents if c.type == "function_approval_response"
+        ]
+        assert len(approvals_seen) == 1
+        assert approvals_seen[0].id == approval_id  # type: ignore[attr-defined]
+        assert approvals_seen[0].approved is False  # type: ignore[attr-defined]
+
+        # Pending approval cleared regardless of approve/reject.
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        assert approval_id not in pending
+
+        # The final assistant message reflects the rejection.
+        final_text = " ".join(m.text or "" for m in final_result.messages)
+        assert "approved=False" in final_text
+        assert approval_id in final_text
+
+    async def test_tool_approval_request_id_matches_pending_request(self) -> None:
+        """The approval request id surfaced by WorkflowAgent matches the workflow's pending request id.
+
+        This guards the AgentExecutor change that forwards
+        request_id=user_input_request.id to ctx.request_info(...), which is what
+        allows the response routed back via WorkflowAgent to resolve the pending
+        request without an id-mismatch error.
+        """
+        approval_id = "approval-id-match-1"
+        mock_agent = _ToolApprovalMockAgent(
+            name="approval-id-match-agent",
+            approval_request_ids=[approval_id],
+        )
+
+        @executor
+        async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+            await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+        workflow = WorkflowBuilder(start_executor=start).add_edge(start, mock_agent).build()
+        agent = WorkflowAgent(workflow=workflow, name="Approval Id Agent")
+
+        await agent.run("go")
+
+        pending = await workflow._runner_context.get_pending_request_info_events()
+        # The agent's approval id is used as the workflow's pending request id.
+        assert list(pending.keys()) == [approval_id]
@@ -0,0 +1,149 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Tests for the ``Workflow.status`` property."""
+
+from dataclasses import dataclass
+
+import pytest
+
+from agent_framework import (
+    Executor,
+    Workflow,
+    WorkflowBuilder,
+    WorkflowContext,
+    WorkflowEvent,
+    WorkflowRunState,
+    handler,
+    response_handler,
+)
+from agent_framework._workflows._executor import Executor as _Executor
+from agent_framework._workflows._request_info_mixin import RequestInfoMixin
+
+
+class PassThroughExecutor(Executor):
+    """Executor that yields its input as a workflow output and stops."""
+
+    @handler
+    async def passthrough(self, msg: str, ctx: WorkflowContext[str, str]) -> None:
+        await ctx.yield_output(msg)
+
+
+class FailingExecutor(Executor):
+    """Executor that raises at runtime to drive the FAILED status."""
+
+    @handler
+    async def fail(self, msg: int, ctx: WorkflowContext) -> None:  # pragma: no cover - invoked via workflow
+        raise RuntimeError("boom")
+
+
+@dataclass
+class _ApprovalRequest:
+    prompt: str
+    request_id: str = ""
+
+    def __post_init__(self) -> None:
+        if not self.request_id:
+            import uuid
+
+            self.request_id = str(uuid.uuid4())
+
+
+class ApprovalExecutor(_Executor, RequestInfoMixin):
+    """Executor that issues a single request_info call and finalizes on response."""
+
+    def __init__(self, id: str = "approval"):
+        super().__init__(id=id)
+
+    @handler
+    async def start(self, message: str, ctx: WorkflowContext[str, str]) -> None:
+        await ctx.request_info(_ApprovalRequest(prompt=message), bool)
+
+    @response_handler
+    async def on_response(
+        self, original_request: _ApprovalRequest, approved: bool, ctx: WorkflowContext[str, str]
+    ) -> None:
+        await ctx.yield_output(f"approved={approved}")
+
+
+def _build_passthrough_workflow() -> Workflow:
+    executor = PassThroughExecutor(id="p")
+    return WorkflowBuilder(start_executor=executor, output_from=[executor]).build()
+
+
+def _build_failing_workflow() -> Workflow:
+    # FailingExecutor has no workflow_output_types, so we leave designation
+    # implicit; the deprecation warning is filtered at call sites that need it.
+    return WorkflowBuilder(start_executor=FailingExecutor(id="f")).build()
+
+
+def _build_approval_workflow() -> Workflow:
+    executor = ApprovalExecutor(id="approval")
+    return WorkflowBuilder(start_executor=executor, output_from=[executor]).build()
+
+
+async def test_status_default_is_idle_before_first_run():
+    wf = _build_passthrough_workflow()
+    assert wf.status is WorkflowRunState.IDLE
+
+
+async def test_status_is_idle_after_successful_run():
+    wf = _build_passthrough_workflow()
+    await wf.run("hello")
+    assert wf.status is WorkflowRunState.IDLE
+
+
+async def test_status_is_failed_after_failure():
+    wf = _build_failing_workflow()
+    with pytest.raises(RuntimeError, match="boom"):
+        await wf.run(0)
+    assert wf.status is WorkflowRunState.FAILED
+
+
+async def test_status_transitions_during_streaming_run():
+    """Workflow.status mirrors the most recent emitted status event."""
+    wf = _build_passthrough_workflow()
+    observed: list[WorkflowRunState] = []
+
+    async for event in wf.run("hi", stream=True):
+        if isinstance(event, WorkflowEvent) and event.type == "status":
+            # By the time a status event surfaces to the consumer, the property
+            # must already reflect that state (updated in lockstep with emission).
+            assert wf.status == event.state
+            observed.append(event.state)  # type: ignore
+
+    # IN_PROGRESS must precede IDLE; both must appear.
+    assert WorkflowRunState.IN_PROGRESS in observed
+    assert observed[-1] is WorkflowRunState.IDLE
+    assert wf.status is WorkflowRunState.IDLE
+
+
+async def test_status_idle_with_pending_requests_then_resolves_to_idle():
+    wf = _build_approval_workflow()
+
+    request_event: WorkflowEvent | None = None
+    async for event in wf.run("please approve", stream=True):
+        if isinstance(event, WorkflowEvent) and event.type == "request_info":
+            request_event = event
+
+    assert request_event is not None
+    assert wf.status is WorkflowRunState.IDLE_WITH_PENDING_REQUESTS
+
+    async for _ in wf.run(stream=True, responses={request_event.request_id: True}):
+        pass
+
+    assert wf.status is WorkflowRunState.IDLE
+
+
+async def test_status_in_progress_pending_requests_observed_mid_run():
+    """While streaming, status reaches IN_PROGRESS_PENDING_REQUESTS after a request_info event."""
+    wf = _build_approval_workflow()
+    seen_states: list[WorkflowRunState] = []
+
+    async for event in wf.run("please approve", stream=True):
+        if isinstance(event, WorkflowEvent) and event.type == "status":
+            seen_states.append(event.state)  # type: ignore
+
+    assert WorkflowRunState.IN_PROGRESS in seen_states
+    assert WorkflowRunState.IN_PROGRESS_PENDING_REQUESTS in seen_states
+    assert seen_states[-1] is WorkflowRunState.IDLE_WITH_PENDING_REQUESTS
+    assert wf.status is WorkflowRunState.IDLE_WITH_PENDING_REQUESTS
@@ -4,7 +4,7 @@ description = "Microsoft Foundry integrations for Microsoft Agent Framework."
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.8.0"
+version = "1.8.1"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -23,10 +23,10 @@ classifiers = [
  "Typing :: Typed",
 ]
 dependencies = [
-    "agent-framework-core>=1.8.0,<2",
-    "agent-framework-openai>=1.8.0,<2",
+    "agent-framework-core>=1.8.1,<2",
+    "agent-framework-openai>=1.8.1,<2",
    "azure-ai-inference>=1.0.0b9,<1.0.0b10",
-    "azure-ai-projects>=2.1.0,<3.0",
+    "azure-ai-projects>=2.2.0,<3.0",
 ]

 [tool.uv]
@@ -567,7 +567,7 @@ class ResponsesHostServer(ResponsesAgentServerHost):
        by the hosting infrastructure or files will be preserved upon deactivation.
        """
        input_items = await context.get_input_items()
-        input_messages = await _items_to_messages(input_items)
+        input_messages = await _items_to_messages(input_items, approval_storage=self._approval_storage)
        is_streaming_request = request.stream is not None and request.stream is True

        _, are_options_set = _to_chat_options(request)
@@ -664,7 +664,11 @@ class ResponsesHostServer(ResponsesAgentServerHost):
                checkpoint_storage=write_storage,
            )

-            async for item in _to_outputs_for_messages(response_event_stream, response.messages):
+            async for item in _to_outputs_for_messages(
+                response_event_stream,
+                response.messages,
+                approval_storage=self._approval_storage,
+            ):
                yield item

            await self._delete_not_latest_checkpoints(write_storage, self._agent.workflow.name)
@@ -685,7 +689,9 @@ class ResponsesHostServer(ResponsesAgentServerHost):
                for event in tracker.handle(content):
                    yield event
                if tracker.needs_async:
-                    async for item in _to_outputs(response_event_stream, content):
+                    async for item in _to_outputs(
+                        response_event_stream, content, approval_storage=self._approval_storage
+                    ):
                        yield item
                    tracker.needs_async = False

@@ -4,7 +4,7 @@ description = "Foundry Hosting integration for Microsoft Agent Framework."
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.0.0a260604"
+version = "1.0.0a260609"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -23,7 +23,7 @@ classifiers = [
  "Typing :: Typed",
 ]
 dependencies = [
-    "agent-framework-core>=1.8.0,<2",
+    "agent-framework-core>=1.8.1,<2",
    "azure-ai-agentserver-core>=2.0.0b3,<3",
    "azure-ai-agentserver-responses>=1.0.0b7,<2",
    "azure-ai-agentserver-invocations>=1.0.0b3,<2",
@@ -11,24 +11,33 @@ the registered _handle_create handler.
 from __future__ import annotations

 import json
-from collections.abc import AsyncIterator, Callable
+import uuid
+from collections.abc import AsyncIterator, Awaitable, Callable, Sequence
 from dataclasses import dataclass
+from typing import Literal, overload
 from unittest.mock import AsyncMock, MagicMock, patch

 import httpx
 import pytest
 from agent_framework import (
+    AgentExecutorRequest,
    AgentResponse,
    AgentResponseUpdate,
+    AgentSession,
    Content,
    FileCheckpointStorage,
    HistoryProvider,
    Message,
    RawAgent,
    ResponseStream,
+    SupportsAgentRun,
+    WorkflowAgent,
+    WorkflowBuilder,
    WorkflowCheckpoint,
    WorkflowCheckpointException,
+    WorkflowContext,
    WorkflowMessage,
+    executor,
 )
 from azure.ai.agentserver.responses import InMemoryResponseProvider
 from mcp import McpError
@@ -102,7 +111,7 @@ def _make_agent(
    return agent


-def _make_server(agent: MagicMock, **kwargs: Any) -> ResponsesHostServer:
+def _make_server(agent: Any, **kwargs: Any) -> ResponsesHostServer:
    """Create a ResponsesHostServer with an in-memory store."""
    return ResponsesHostServer(agent, store=InMemoryResponseProvider(), **kwargs)

@@ -3469,3 +3478,498 @@ class TestOAuthConsentSurfacing:


 # endregion
+
+# region Workflow agent hosting (end-to-end)
+
+
+class _ToolApprovalWorkflowAgentMock(SupportsAgentRun):
+    """Inner agent for a hosted ``WorkflowAgent`` whose first run emits a
+    ``FunctionApprovalRequestContent`` and whose follow-up run (after
+    receiving a ``FunctionApprovalResponseContent`` in its inputs) returns a
+    final assistant text response.
+
+    Mirrors a real agent whose tool invocation requires user approval. Used
+    here to exercise the full HTTP pipeline through ``ResponsesHostServer``
+    when the hosted agent is a ``WorkflowAgent`` containing a tool-approval
+    flow.
+    """
+
+    def __init__(
+        self,
+        name: str,
+        *,
+        tool_name: str = "delete_file",
+        tool_arguments: dict[str, Any] | None = None,
+        approval_request_ids: Sequence[str] | None = None,
+        final_text: str = "done",
+    ) -> None:
+        self.id = str(uuid.uuid4())
+        self.name = name
+        self.description: str | None = None
+        self._tool_name = tool_name
+        self._tool_arguments = tool_arguments or {"path": "/tmp/example"}
+        self._approval_request_ids: list[str] = list(approval_request_ids) if approval_request_ids else []
+        self._final_text = final_text
+        self.run_count = 0
+        self.last_run_messages: list[Message] = []
+
+    def create_session(self, **kwargs: Any) -> AgentSession:
+        return AgentSession()
+
+    def get_session(self, *, service_session_id: str, **kwargs: Any) -> AgentSession:
+        return AgentSession()
+
+    def _next_request_id(self) -> str:
+        # Stable across calls: when the workflow checkpoint round-trips through
+        # restore, ``AgentExecutor`` re-invokes the inner agent during replay.
+        # We must surface the *same* approval request id on each invocation so
+        # the workflow's pending-request id matches the id the test echoes
+        # back as ``mcp_approval_response``.
+        if self._approval_request_ids:
+            return self._approval_request_ids[0]
+        return str(uuid.uuid4())
+
+    def _build_approval_request(self) -> Content:
+        request_id = self._next_request_id()
+        function_call = Content.from_function_call(
+            call_id=request_id,
+            name=self._tool_name,
+            arguments=self._tool_arguments,
+            additional_properties={"server_label": "test_server"},
+        )
+        return Content.from_function_approval_request(id=request_id, function_call=function_call)
+
+    @overload
+    def run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = ...,
+        *,
+        stream: Literal[False] = ...,
+        session: AgentSession | None = ...,
+        **kwargs: Any,
+    ) -> Awaitable[AgentResponse[Any]]: ...
+
+    @overload
+    def run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = ...,
+        *,
+        stream: Literal[True],
+        session: AgentSession | None = ...,
+        **kwargs: Any,
+    ) -> ResponseStream[AgentResponseUpdate, AgentResponse[Any]]: ...
+
+    def run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = None,
+        *,
+        stream: bool = False,
+        session: AgentSession | None = None,
+        **kwargs: Any,
+    ) -> Awaitable[AgentResponse] | ResponseStream[AgentResponseUpdate, AgentResponse]:
+        if stream:
+            return self._run_stream(messages=messages, **kwargs)
+        return self._run(messages=messages, **kwargs)
+
+    @staticmethod
+    def _normalize(
+        messages: str | Content | Message | Sequence[str | Content | Message] | None,
+    ) -> list[Message]:
+        if messages is None:
+            return []
+        if isinstance(messages, str):
+            return [Message(role="user", contents=[Content.from_text(text=messages)])]
+        if isinstance(messages, Message):
+            return [messages]
+        if isinstance(messages, Content):
+            return [Message(role="user", contents=[messages])]
+        result: list[Message] = []
+        for item in messages:
+            if isinstance(item, Message):
+                result.append(item)
+            elif isinstance(item, Content):
+                result.append(Message(role="user", contents=[item]))
+            else:
+                result.append(Message(role="user", contents=[Content.from_text(text=item)]))
+        return result
+
+    @staticmethod
+    def _approval_responses_in(messages: list[Message]) -> list[Content]:
+        return [c for m in messages for c in m.contents if c.type == "function_approval_response"]
+
+    async def _run(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = None,
+        **kwargs: Any,
+    ) -> AgentResponse:
+        normalized = self._normalize(messages)
+        self.last_run_messages = normalized
+        self.run_count += 1
+        if self._approval_responses_in(normalized):
+            return AgentResponse(messages=[Message("assistant", [Content.from_text(text=self._final_text)])])
+        approval = self._build_approval_request()
+        return AgentResponse(messages=[Message("assistant", [approval])])
+
+    def _run_stream(
+        self,
+        messages: str | Content | Message | Sequence[str | Content | Message] | None = None,
+        **kwargs: Any,
+    ) -> ResponseStream[AgentResponseUpdate, AgentResponse]:
+        normalized = self._normalize(messages)
+        self.last_run_messages = normalized
+        self.run_count += 1
+        approvals = self._approval_responses_in(normalized)
+
+        async def _iter() -> AsyncIterator[AgentResponseUpdate]:
+            if approvals:
+                yield AgentResponseUpdate(
+                    contents=[Content.from_text(text=self._final_text)],
+                    role="assistant",
+                    author_name=self.name,
+                )
+                return
+            yield AgentResponseUpdate(
+                contents=[self._build_approval_request()],
+                role="assistant",
+                author_name=self.name,
+            )
+
+        return ResponseStream(_iter(), finalizer=AgentResponse.from_updates)
+
+
+def _build_text_workflow_agent(text: str) -> WorkflowAgent:
+    """Build a minimal ``WorkflowAgent`` whose inner agent emits a fixed text."""
+
+    class _TextAgent(SupportsAgentRun):
+        def __init__(self, name: str, text: str) -> None:
+            self.id = str(uuid.uuid4())
+            self.name = name
+            self.description: str | None = None
+            self._text = text
+
+        def create_session(self, **kwargs: Any) -> AgentSession:
+            return AgentSession()
+
+        def get_session(self, *, service_session_id: str, **kwargs: Any) -> AgentSession:
+            return AgentSession()
+
+        @overload
+        def run(
+            self,
+            messages: Any = ...,
+            *,
+            stream: Literal[False] = ...,
+            session: AgentSession | None = ...,
+            **kwargs: Any,
+        ) -> Awaitable[AgentResponse[Any]]: ...
+
+        @overload
+        def run(
+            self,
+            messages: Any = ...,
+            *,
+            stream: Literal[True],
+            session: AgentSession | None = ...,
+            **kwargs: Any,
+        ) -> ResponseStream[AgentResponseUpdate, AgentResponse[Any]]: ...
+
+        def run(
+            self,
+            messages: Any = None,
+            *,
+            stream: bool = False,
+            session: AgentSession | None = None,
+            **kwargs: Any,
+        ) -> Awaitable[AgentResponse] | ResponseStream[AgentResponseUpdate, AgentResponse]:
+            text = self._text
+            name = self.name
+
+            async def _aresult() -> AgentResponse:
+                return AgentResponse(messages=[Message("assistant", [Content.from_text(text=text)])])
+
+            async def _aiter() -> AsyncIterator[AgentResponseUpdate]:
+                yield AgentResponseUpdate(
+                    contents=[Content.from_text(text=text)],
+                    role="assistant",
+                    author_name=name,
+                )
+
+            if stream:
+                return ResponseStream(_aiter(), finalizer=AgentResponse.from_updates)
+            return _aresult()
+
+    inner = _TextAgent("text-agent", text)
+
+    @executor
+    async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+        await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+    workflow = WorkflowBuilder(start_executor=start).add_edge(start, inner).build()
+    return WorkflowAgent(workflow=workflow, name="Text Workflow Agent")
+
+
+def _build_approval_workflow_agent(
+    *,
+    approval_request_id: str,
+    tool_name: str = "delete_file",
+    tool_arguments: dict[str, Any] | None = None,
+    final_text: str = "done",
+) -> tuple[WorkflowAgent, _ToolApprovalWorkflowAgentMock]:
+    """Build a ``WorkflowAgent`` whose inner agent emits a tool approval request."""
+    mock_agent = _ToolApprovalWorkflowAgentMock(
+        name="approval-agent",
+        tool_name=tool_name,
+        tool_arguments=tool_arguments or {"path": "/tmp/secret.txt"},
+        approval_request_ids=[approval_request_id],
+        final_text=final_text,
+    )
+
+    @executor
+    async def start(messages: list[Message], ctx: WorkflowContext[AgentExecutorRequest]) -> None:
+        await ctx.send_message(AgentExecutorRequest(messages=messages, should_respond=True))
+
+    workflow = WorkflowBuilder(start_executor=start).add_edge(start, mock_agent).build()
+    workflow_agent = WorkflowAgent(workflow=workflow, name="Approval Workflow Agent")
+    return workflow_agent, mock_agent
+
+
+class TestWorkflowAgentHosting:
+    """End-to-end HTTP tests for ``ResponsesHostServer`` hosting a ``WorkflowAgent``.
+
+    These tests drive ``_handle_inner_workflow`` through the ASGI stack:
+    they exercise checkpoint write/restore (multi-turn) and the
+    tool-approval round-trip path, which is the primary differentiator
+    relative to the regular agent path.
+    """
+
+    async def test_basic_text_response(self) -> None:
+        workflow_agent = _build_text_workflow_agent("hello from workflow")
+        server = _make_server(workflow_agent)
+
+        resp = await _post(server, input_text="hi", stream=False)
+        assert resp.status_code == 200
+        body = resp.json()
+        assert body["status"] == "completed"
+
+        text_found = any(
+            part.get("type") == "output_text" and part.get("text") == "hello from workflow"
+            for item in body["output"]
+            if item["type"] == "message"
+            for part in item.get("content", [])
+        )
+        assert text_found, f"Expected workflow output text in {body['output']}"
+
+    async def test_basic_text_response_streaming(self) -> None:
+        workflow_agent = _build_text_workflow_agent("hello stream")
+        server = _make_server(workflow_agent)
+
+        resp = await _post(server, input_text="hi", stream=True)
+        assert resp.status_code == 200
+        events = _parse_sse_events(resp.text)
+        types = _sse_event_types(events)
+        assert types[0] == "response.created"
+        assert types[-1] == "response.completed"
+        assert "response.output_text.delta" in types
+        text_done = [e for e in events if e["event"] == "response.output_text.done"]
+        assert any(e["data"]["text"] == "hello stream" for e in text_done)
+
+    async def test_non_streaming_emits_mcp_approval_request_and_persists_to_storage(self) -> None:
+        workflow_agent, mock_agent = _build_approval_workflow_agent(approval_request_id="apr_wf_ns")
+        server = _make_server(workflow_agent)
+
+        resp = await _post(server, stream=False)
+        assert resp.status_code == 200
+        body = resp.json()
+        assert body["status"] == "completed"
+
+        approval_items = [it for it in body["output"] if it["type"] == "mcp_approval_request"]
+        assert len(approval_items) == 1
+        assert approval_items[0]["name"] == "delete_file"
+        assert approval_items[0]["server_label"] == "test_server"
+        approval_request_id = approval_items[0]["id"]
+
+        # The id surfaced over the wire is generated by the response stream
+        # builder; the original approval ``Content`` (carrying the inner
+        # ``function_call``) must be persisted under that id so the next
+        # turn can reconstruct it.
+        loaded = await server._approval_storage.load_approval_request(  # pyright: ignore[reportPrivateUsage]
+            approval_request_id
+        )
+        assert loaded.type == "function_approval_request"
+        assert loaded.function_call.name == "delete_file"  # type: ignore[attr-defined]
+        assert mock_agent.run_count == 1
+
+    async def test_streaming_emits_mcp_approval_request_and_persists_to_storage(self) -> None:
+        workflow_agent, mock_agent = _build_approval_workflow_agent(approval_request_id="apr_wf_st")
+        server = _make_server(workflow_agent)
+
+        resp = await _post(server, stream=True)
+        assert resp.status_code == 200
+
+        events = _parse_sse_events(resp.text)
+        types = _sse_event_types(events)
+        assert types[0] == "response.created"
+        assert types[-1] == "response.completed"
+
+        approval_request_id: str | None = None
+        for e in events:
+            if e["event"] != "response.output_item.added":
+                continue
+            item = e["data"].get("item") or {}
+            if item.get("type") == "mcp_approval_request":
+                approval_request_id = item.get("id")
+                break
+        assert approval_request_id is not None
+
+        loaded = await server._approval_storage.load_approval_request(  # pyright: ignore[reportPrivateUsage]
+            approval_request_id
+        )
+        assert loaded.type == "function_approval_request"
+        assert mock_agent.run_count == 1
+
+    async def test_round_trip_approval_response_resumes_workflow_agent(self) -> None:
+        """Two-turn HTTP round-trip:
+
+        Turn 1 emits ``mcp_approval_request`` and writes a workflow
+        checkpoint under the response id. Turn 2 sends the
+        ``mcp_approval_response`` with ``previous_response_id`` set, so the
+        host restores the checkpoint, the WorkflowAgent routes the
+        approval response back to the paused inner agent, and the inner
+        agent emits the final assistant text.
+        """
+        workflow_agent, mock_agent = _build_approval_workflow_agent(
+            approval_request_id="apr_wf_rt",
+            final_text="done with approval",
+        )
+        server = _make_server(workflow_agent)
+
+        first = await _post(server, stream=False)
+        assert first.status_code == 200
+        first_body = first.json()
+        first_response_id = first_body["id"]
+        approval_items = [it for it in first_body["output"] if it["type"] == "mcp_approval_request"]
+        assert len(approval_items) == 1
+        approval_request_id = approval_items[0]["id"]
+        assert mock_agent.run_count == 1
+
+        second_payload: dict[str, Any] = {
+            "model": "test-model",
+            "input": [
+                {
+                    "type": "mcp_approval_response",
+                    "approval_request_id": approval_request_id,
+                    "approve": True,
+                }
+            ],
+            "stream": False,
+            "previous_response_id": first_response_id,
+        }
+        second = await _post_json(server, second_payload)
+        assert second.status_code == 200
+        second_body = second.json()
+        assert second_body["status"] == "completed"
+
+        # The inner agent must have been resumed (restore replay + new turn).
+        # Restore call is a no-op for the mock (no input); the new-turn call
+        # delivers the approval response, so run_count grows by at least 1.
+        assert mock_agent.run_count >= 2
+
+        # The final assistant text from the resumed inner agent surfaces in
+        # the HTTP output.
+        text_pieces = [
+            part.get("text", "")
+            for item in second_body["output"]
+            if item["type"] == "message"
+            for part in item.get("content", [])
+            if part.get("type") == "output_text"
+        ]
+        assert any("done with approval" in t for t in text_pieces), (
+            f"expected resumed workflow output, got {second_body['output']}"
+        )
+
+        # The new-turn invocation of the inner agent must have received the
+        # approval response routed back through WorkflowAgent.
+        approval_responses = [
+            c for m in mock_agent.last_run_messages for c in m.contents if c.type == "function_approval_response"
+        ]
+        assert len(approval_responses) == 1
+        assert approval_responses[0].approved is True  # type: ignore[attr-defined]
+
+    async def test_round_trip_approval_response_streaming(self) -> None:
+        """Streaming variant of the round-trip: turn 2 is requested with
+        ``stream=true`` and surfaces the resumed text as SSE events."""
+        workflow_agent, mock_agent = _build_approval_workflow_agent(
+            approval_request_id="apr_wf_rt_st",
+            final_text="streamed-done",
+        )
+        server = _make_server(workflow_agent)
+
+        first = await _post(server, stream=False)
+        first_body = first.json()
+        first_response_id = first_body["id"]
+        approval_request_id = next(it["id"] for it in first_body["output"] if it["type"] == "mcp_approval_request")
+
+        second = await _post_json(
+            server,
+            {
+                "model": "test-model",
+                "input": [
+                    {
+                        "type": "mcp_approval_response",
+                        "approval_request_id": approval_request_id,
+                        "approve": True,
+                    }
+                ],
+                "stream": True,
+                "previous_response_id": first_response_id,
+            },
+        )
+        assert second.status_code == 200
+        events = _parse_sse_events(second.text)
+        types = _sse_event_types(events)
+        assert types[0] == "response.created"
+        assert types[-1] == "response.completed"
+
+        text_done = [e for e in events if e["event"] == "response.output_text.done"]
+        assert any("streamed-done" in e["data"]["text"] for e in text_done)
+        assert mock_agent.run_count >= 2
+
+    async def test_round_trip_approval_response_rejected(self) -> None:
+        """Sending ``approve=False`` must surface as ``approved=False`` to the
+        inner agent on resume."""
+        workflow_agent, mock_agent = _build_approval_workflow_agent(
+            approval_request_id="apr_wf_reject",
+            final_text="acknowledged",
+        )
+        server = _make_server(workflow_agent)
+
+        first = await _post(server, stream=False)
+        first_body = first.json()
+        first_response_id = first_body["id"]
+        approval_request_id = next(it["id"] for it in first_body["output"] if it["type"] == "mcp_approval_request")
+
+        second = await _post_json(
+            server,
+            {
+                "model": "test-model",
+                "input": [
+                    {
+                        "type": "mcp_approval_response",
+                        "approval_request_id": approval_request_id,
+                        "approve": False,
+                    }
+                ],
+                "stream": False,
+                "previous_response_id": first_response_id,
+            },
+        )
+        assert second.status_code == 200
+
+        approval_responses = [
+            c for m in mock_agent.last_run_messages for c in m.contents if c.type == "function_approval_response"
+        ]
+        assert len(approval_responses) == 1
+        assert approval_responses[0].approved is False  # type: ignore[attr-defined]
+
+
+# endregion
@@ -10,6 +10,10 @@ pip install agent-framework-gemini --pre

 The Gemini integration enables Microsoft Agent Framework applications to call Google Gemini models with familiar chat abstractions, including streaming, tool/function calling, and structured output.

+## Structured Output
+
+Gemini structured output can be configured with either a Pydantic model in `response_format`, a JSON schema mapping in `response_format`, or a Gemini-specific `response_schema`. Declarative agents that define `outputSchema` pass that schema through `response_format`.
+
 ## Authentication

 The connector supports both `google-genai` authentication modes.
@@ -109,8 +109,8 @@ class GeminiChatOptions(ChatOptions[ResponseModelT], Generic[ResponseModelT], to
            or ``types.Tool`` objects returned by ``get_code_interpreter_tool``, ``get_web_search_tool``,
            ``get_mcp_tool``, ``get_file_search_tool``, or ``get_maps_grounding_tool``.
        tool_choice: How the model picks a tool. One of ``'auto'``, ``'none'``, or ``'required'``.
-        response_format: Pydantic model type for structured JSON output. The response text is
-            parsed into the model and exposed via ``ChatResponse.value``.
+        response_format: Pydantic model type or JSON schema mapping for structured JSON output.
+            The response text is parsed and exposed via ``ChatResponse.value``.
        instructions: Extra system-level instructions prepended to the system message.

    Not supported, and passing these raises a type error:
@@ -255,6 +255,29 @@ _OPTION_CONSUMED_KEYS: frozenset[str] = frozenset({

 _OPTION_EXCLUDE_KEYS: frozenset[str] = _OPTION_EXPLICIT_KEYS | _OPTION_CONSUMED_KEYS

+_JSON_SCHEMA_TYPES: frozenset[str] = frozenset({
+    "array",
+    "boolean",
+    "integer",
+    "null",
+    "number",
+    "object",
+    "string",
+})
+
+_JSON_SCHEMA_KEYWORDS: frozenset[str] = frozenset({
+    "$defs",
+    "additionalProperties",
+    "allOf",
+    "anyOf",
+    "enum",
+    "items",
+    "oneOf",
+    "properties",
+    "required",
+    "type",
+})
+
 _FINISH_REASON_MAP: dict[str, FinishReasonLiteral] = {
    "STOP": "stop",
    "MAX_TOKENS": "length",
@@ -747,9 +770,13 @@ class RawGeminiChatClient(
                continue
            kwargs[_OPTION_TRANSLATIONS.get(key, key)] = value

-        if options.get("response_format") or options.get("response_schema"):
+        response_format = options.get("response_format")
+        response_schema = options.get("response_schema")
+        if response_format is not None or response_schema is not None:
            kwargs["response_mime_type"] = "application/json"
-        if schema := options.get("response_schema"):
+        if response_schema is not None:
+            kwargs["response_schema"] = response_schema
+        elif (schema := self._extract_response_schema(response_format)) is not None:
            kwargs["response_schema"] = schema
        if tools := self._prepare_tools(options):
            kwargs["tools"] = tools
@@ -762,6 +789,48 @@ class RawGeminiChatClient(

        return types.GenerateContentConfig(**kwargs)

+    @staticmethod
+    def _extract_response_schema(response_format: Any) -> dict[str, Any] | None:
+        """Extract a Gemini response schema from supported mapping response_format shapes."""
+        if not isinstance(response_format, Mapping):
+            return None
+        mapping = cast("Mapping[str, Any]", response_format)
+
+        if (nested := RawGeminiChatClient._extract_response_schema(mapping.get("format"))) is not None:
+            return nested
+
+        json_schema = mapping.get("json_schema")
+        if isinstance(json_schema, Mapping):
+            schema = cast("Mapping[str, Any]", json_schema).get("schema")
+            if isinstance(schema, Mapping):
+                return dict(cast("Mapping[str, Any]", schema))
+
+        schema = mapping.get("schema")
+        if isinstance(schema, Mapping):
+            return dict(cast("Mapping[str, Any]", schema))
+
+        if RawGeminiChatClient._is_json_schema_mapping(mapping):
+            return dict(mapping)
+
+        return None
+
+    @staticmethod
+    def _is_json_schema_mapping(value: Mapping[str, Any]) -> bool:
+        """Return True when a mapping appears to be a JSON Schema rather than a response-format envelope."""
+        if not any(keyword in value for keyword in _JSON_SCHEMA_KEYWORDS):
+            return False
+
+        schema_type = value.get("type")
+        if schema_type is None:
+            return True
+        if isinstance(schema_type, str):
+            return schema_type in _JSON_SCHEMA_TYPES
+        if isinstance(schema_type, Sequence) and not isinstance(schema_type, (str, bytes)):
+            entries = cast("Sequence[object]", schema_type)
+            return all(isinstance(item, str) and item in _JSON_SCHEMA_TYPES for item in entries)
+
+        return False
+
    def _prepare_tools(self, options: Mapping[str, Any]) -> list[types.Tool] | None:
        """Translate the framework tool list into Gemini API tool objects.

@@ -4,7 +4,7 @@ description = "Google Gemini integration for Microsoft Agent Framework."
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.0.0a260521"
+version = "1.0.0a260609"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -24,7 +24,7 @@ classifiers = [
  "Typing :: Typed",
 ]
 dependencies = [
-    "agent-framework-core>=1.6.0,<2.0",
+    "agent-framework-core>=1.8.1,<2.0",
    "google-genai>=1.65.0,<2.0.0",
 ]

--- a/Show More
+++ b/Show More