Compare commits

..
Author SHA1 Message Date
github-actions[bot] 587356c778 Python: chore: upgrade dev dependencies 2026-03-16 15:36:52 +00:00
1b7940c91e Python: keep MCP cleanup on the owner task (#4687)
* Python: keep MCP cleanup on owner task

* Avoid MCP owner task deadlocks

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix MCP owner-task timeout tests

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-14 13:54:05 +00:00
Laveesh RohraandGitHub 2f4c4aa614 Python: Remove bad dependency (#4696)
* Remove bad dependency in requirements

* Remove bad dependency in requirements.txt
2026-03-13 23:15:56 +00:00
Eduard van ValkenburgandGitHub 052ba7be07 Python: normalize empty MCP tool output to null (#4683)
* Python: normalize empty MCP tool output to null

* Python: hardcode null for empty MCP output
2026-03-13 20:03:48 +00:00
Chris GillumandGitHub c67d3523ae .NET: [Durable Agents] Filter empty AIContent from durable agent state responses (#4670)
* Filter empty AIContent from durable agent state responses

Prevent opaque AIContent objects (e.g., with only RawRepresentation set)
from being stored in durable entity state, where they serialize to empty
JSON payloads. Base AIContent instances are kept only if they have
Annotations or AdditionalProperties.

Fixes https://github.com/microsoft/agent-framework/issues/4481

* Update CHANGELOG.md and fix linter violation
2026-03-13 18:16:46 +00:00
Shyju KrishnankuttyandGitHub 83ce6a9602 Sanitize user input in log statements for durable agent samples. (#4656) 2026-03-13 17:38:55 +00:00
50fdcbaf57 Python: chore(python): improve dependency range automation (#4343)
* chore(python): improve dependency range automation

- tighten dependency bounds and coding standards guidance\n- add dependency range validation workflow, reporting, and issue automation\n- update related tests and dependency pins for compatibility

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* updated text and pyarrow

* new lock

* fixed workflow

* updated deps

* fix tiktoken

* chore(python): refine dependency validation workflows

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* docs(python): add high-level dependency validation comments

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* WIP

* added additional comments and excludes

* added dev dependency handling and workflow and updates to package ranges

* added readme and simplified commands

* fix markers

* chore(python): address dependency review feedback

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Tighten dependency bounds, remove stale overrides, restore Python 3.10 support

- Apply dependency bound policy across all packages: stable >=1.0 deps use
  >=floor,<next_major; pre-1.0/prerelease deps use validated hard-bounded ranges
- Remove stale root tool.uv.override-dependencies (uvicorn, websockets, grpcio)
- Lower github_copilot requires-python to >=3.10 with github-copilot-sdk gated
  behind python_version >= 3.11 marker; import raises ImportError on 3.10
- Skip github_copilot pyright/mypy/test tasks on Python <3.11
- Use version-conditional pyrightconfig for samples on Python 3.10
- Add compatibility fix in core responses client for older openai typed dicts
- Normalize uv.lock prerelease mode and refresh dev dependencies
- Update CODING_STANDARD.md, DEV_SETUP.md, and package management skill docs

Closes #902

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* small tweaks

* add note in workflow

* fix workflows and several versions

* fix duplicate

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-13 12:32:37 +00:00
SergeyMenshykhandGitHub 67b0282813 Bump rollup from 7.5.9 to 7.5.11 (#4688) 2026-03-13 12:30:29 +00:00
Roger BarretoandGitHub 0009e330af Fix hosted agent samples Docker build failures due to experimental API warnings (#4641)
Add #pragma warning disable directives to suppress experimental API
diagnostics that cause build errors in Docker isolation (where repo-level
Directory.Build.props is not inherited):

- AgentWithHostedMCP: suppress MEAI001 (HostedMcpServerTool) and OPENAI001
  (GetResponsesClient)
- FoundrySingleAgent: suppress CA2252 (AIProjectClient preview features)
- FoundryMultiAgent: suppress CA2252 (AIProjectClient preview features)

Fixes #4365
2026-03-13 10:13:59 +00:00
a4b9539b62 [BREAKING] Python: clean up kwargs across agents, chat clients, tools, and sessions (#4581)
* Python: clean up kwargs across agents, chat clients, tools, and sessions (#3642)

Audit and refactor public **kwargs usage across core agents, chat clients,
tools, sessions, and provider packages per the migration strategy codified
in CODING_STANDARD.md.

Key changes:
- Add explicit runtime buckets: function_invocation_kwargs and client_kwargs
  on RawAgent.run() and chat client get_response() layers.
- Refactor FunctionTool to prefer explicit ctx: FunctionInvocationContext
  injection; legacy **kwargs tools still work via _forward_runtime_kwargs.
- Refactor Agent.as_tool() to use direct JSON schema, always-streaming
  wrapper, approval_mode parameter, and UserInputRequiredException
  propagation (integrates PR #4568 behavior).
- Remove implicit session bleeding into FunctionInvocationContext; tools
  that need a session must receive it via function_invocation_kwargs.
- Lower chat-client layers after FunctionInvocationLayer accept only
  compatibility **kwargs (client_kwargs flattened, function_invocation_kwargs
  ignored).
- Add layered docstring composition from Raw... implementations via
  _docstrings.py helper.
- Clean up provider constructors to use explicit additional_properties.
- Deprecation warnings on legacy direct kwargs paths.
- Update samples, tests, and typing across all 23 packages.

Resolves #3642

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* clarified docstring

* feedback fixes

* Add unit tests for _docstrings.py build/apply helpers

Tests cover: no docstring source, no extra kwargs, appending to existing
Keyword Args section, inserting after Args, inserting in plain docstrings,
multiline descriptions, ordering, and apply_layered_docstring.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add test for propagate_session TypeError on non-AgentSession values

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add tests for multi-content and empty UserInputRequiredException propagation

Cover the branching logic in _try_execute_function_calls for:
- Multiple user_input_request items in a single exception (extra_user_input_contents path)
- Empty contents list (fallback function_result path)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add tests for DurableAIAgent.get_session forwarding service_session_id

Verifies get_session correctly forwards service_session_id and session_id
to the executor's get_new_session, replacing the removed kwargs test.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Simplify ag-ui test stub to read session from client_kwargs only

Remove dual-mode detection (client_kwargs vs raw kwargs fallback) from
the test mock. Session is now read exclusively from client_kwargs,
matching the settled public calling convention.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* updated create and get sessions in durable

* fixed docstrings

* fix test

* updated session handling

* updated from main

* updated tests

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-13 08:58:32 +00:00
Eduard van ValkenburgandGitHub b7990908fe fix duplicate names between supplied tools and mcp servers (#4649) 2026-03-13 08:22:56 +00:00
84bae0f42a Python: Fix type hint for Case and Default (#3985)
* Fix type hint for `Case` and `Default`

* Add test

---------

Co-authored-by: Evan Mattson <35585003+moonbox3@users.noreply.github.com>
2026-03-13 08:17:24 +00:00
f696ac9b57 Python: A2AAgent defaults name/description from AgentCard (#4661)
* Python: A2AAgent defaults name/description from AgentCard

When an AgentCard is provided but name/description are not explicitly
set, A2AAgent now falls back to agent_card.name and agent_card.description.
This avoids redundant duplication when constructing A2AAgent instances,
especially in GroupChat orchestrations where name and description are
essential for routing decisions.

Explicit values still take precedence over card values.

Fixes #4630

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Use 'is None' checks instead of truthiness for name/description fallback

Ensures explicitly provided empty strings are not overridden by
agent_card values. Adds test for the empty string edge case.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-13 00:14:23 +00:00
5e33deff45 Python: Unify tool results as Content items with rich content support (#4331)
* feat(python): allow @tool functions to return rich content (images, audio)

Add support for tool functions to return Content objects that the model can perceive natively. Closes #4272

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Anthropic logging + mypy fix

* Address PR review: fix MCP ordering, fold helper into from_function_result, fix Chat client

- Preserve original content order in MCP tool results instead of text-first
- Move _build_function_result logic into Content.from_function_result()
- Chat Completions: inject user message for rich items (API only supports string tool content)
- Update tests for ordering and new from_function_result behavior

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Use native Responses API multi-part output, warn+omit for Chat client

- Responses client: put rich items directly in function_call_output's
  output field as list (native API support) instead of user message injection
- Chat client: warn and omit rich items (API doesn't support multi-part
  tool results), matching Ollama/Bedrock pattern
- Unify test image: use sample_image.jpg across all integration tests
- Add Azure OpenAI Responses integration test
- Assert model describes house image to verify perception

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix lint: remove print statement, wrap long line

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address review feedback: bug fixes, single-pass MCP, unit tests

- Add isinstance guard in from_function_result for non-Content lists
- Fix Anthropic empty tool_content fallback to string result
- Fix Content(type='text', text=None) edge case in parse_result
- Rewrite MCP _parse_tool_result_from_mcp as single-pass (no index counters)
- Add Anthropic unit tests: data image, uri image, unsupported media, all-unsupported
- Add OpenAI Chat unit test: rich items warning and omission
- Add OpenAI Responses unit tests: function_result with/without items
- Add test_types tests: only-rich-items list, non-Content list fallback

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix pyright errors: add type ignore comments for Any list iteration

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix mypy/pyright: ensure ToolExecutionException receives str

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix lint: remove duplicate test_prepare_options_excludes_conversation_id

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* refactor: unify all tool results into Content items

* addressed copilot comments

* pyright fix

* small fix

* comments

* fix: address Copilot review - warnings, blob safety, dedup

- Add warning logs when rich content is dropped in Claude agent and
  MCP server handlers (matching Chat/Bedrock/Ollama pattern)
- Defensive blob URI construction: wrap plain base64 in data: prefix
- Simplify Chat client _prepare_content_for_openai to use content.result
- Simplify Responses client text-only path, remove redundant nesting
- Add test for plain base64 blob without data: prefix

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix token double-counting in compaction and address review comments

- Exclude items from _serialize_content() to prevent double-counting
  tokens when items mirrors result in function_result content
- Add rich content warning in GitHub Copilot agent tool handler
- Replace raw Content debug log with concise item count/type summary
- Update stale test comments about FunctionTool.invoke return type

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-12 22:30:09 +00:00
b6a1315386 fix: omit toolConfig when tool_choice="none" in BedrockChatClient (#4535)
Bedrock's Converse API only accepts "auto", "any", or "tool" as valid
toolChoice keys. The previous code mapped tool_choice="none" to
{"none": {}}, which causes a botocore.exceptions.ParamValidationError.

When tool_choice="none" (set by FunctionInvocationLayer after exhausting
max iterations), the fix now omits toolConfig entirely so the model
won't attempt tool calls.

Added tests for tool_choice="none", "auto", and "required" modes.

Fixes #4529

Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
2026-03-12 18:49:08 +00:00
ed2fb3b9dd Python: Fix state snapshot to use deepcopy so nested mutations are detected in durable workflow activities (#4518)
* Use deepcopy for state snapshot to detect nested mutations (#4500)

Replace dict() shallow copy with copy.deepcopy() when snapshotting
workflow state before activity execution. The shallow copy shared
references to nested objects (dicts, lists), so in-place mutations by
executors were reflected in both the snapshot and live state, producing
an empty diff and preventing state updates from propagating to
downstream activities.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Python: Fix state snapshot to use deepcopy so nested mutations are detected in durable workflow activities

Fixes #4500

* Address PR review: remove report, extract testable helpers (#4500)

- Delete REPRODUCTION_REPORT.md (debugging artifact with local paths
  and raw LLM output)
- Extract _create_state_snapshot() and _compute_state_updates() as
  module-level helpers in _app.py so tests exercise the production
  code path
- Update TestStateSnapshotDiff to import and use production helpers
  instead of reimplementing snapshot/diff logic locally

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Apply pre-commit auto-fixes

* Add regression tests proving shallow copy bug and deep copy isolation (#4500)

Add two additional tests to TestStateSnapshotDiff:
- test_shallow_copy_would_miss_nested_mutations: reproduces the original
  bug by demonstrating that dict() (shallow copy) misses nested mutations
- test_create_state_snapshot_isolates_nested_objects: verifies the
  production _create_state_snapshot helper creates a true deep copy

These tests ensure a regression back to shallow copy would be caught.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add integration test exercising full activity code path (#4500)

Address PR review comment: add test_executor_activity_detects_nested_state_mutations
that captures the actual executor_activity function from _setup_executor_activity
and verifies it detects in-place nested mutations. This test would fail if
_app.py line 314 regressed from _create_state_snapshot() back to dict().

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address review feedback for #4518: review comment fixes

* Address PR review feedback for state snapshot diff

- Inline _compute_state_updates logic at call site to reuse precomputed
  original_keys/current_keys sets, avoiding redundant set allocations
- Fix test docstring to describe behavioral regression instead of
  hard-coding a specific line number
- Use SOURCE_ORCHESTRATOR constant in integration test instead of
  literal string

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Apply pre-commit auto-fixes

* fix: remove unused _compute_state_updates from _app.py (#4518)

The function was inlined per review comment, making the module-level
helper unused and triggering a pyright reportUnusedFunction error.
Move the helper into the test file where it is still needed for unit
testing the diffing logic.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-12 18:43:12 +00:00
Peter IbekweandGitHub aa2ff672fb .NET: Fix to emit WorkflowStartedEvent during workflow execution (#4514)
* Fix bug to emit WorkflowStartedEvent during workflow execution

* Updated based on PR comments
2026-03-12 15:45:17 +00:00
bcb55b4a98 .NET: Update A2A, MCP, and system package dependencies (#4647)
* .NET: Update A2A, MCP, and system package dependencies

Update dependency versions:
- A2A/A2A.AspNetCore: 0.3.3-preview → 0.3.4-preview
- ModelContextProtocol: 0.8.0-preview.1 → 1.1.0
- Microsoft.Bcl.AsyncInterfaces: 10.0.3 → 10.0.4
- System.Linq.AsyncEnumerable: 10.0.0 → 10.0.4
- Add Microsoft.Bcl.Memory 10.0.4

Remove internal polyfill extensions now provided by A2A SDK 0.3.4:
- A2AMetadataExtensions (source + tests)
- AdditionalPropertiesDictionaryExtensions (source + tests)

Update DefaultMcpToolHandler to match MCP SDK 1.1.0 API changes where
ImageContentBlock.Data and AudioContentBlock.Data changed from string
to ReadOnlyMemory<byte>.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* address pr review comments

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-12 14:16:36 +00:00
westeyandGitHub 921c5f9c17 .NET: Include ReasoningEncryptedContent by default when stored output disabled with Responses (#4623)
* Include ReasoningEncryptedContent by default when stored output disabled

* Fix formatting

* Fix formatter
2026-03-12 09:42:20 +00:00
dependabot[bot]GitHubdependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
fcdaaff9cd Bump rollup from 4.47.1 to 4.59.0 in /python/packages/devui/frontend (#4338)
Bumps [rollup](https://github.com/rollup/rollup) from 4.47.1 to 4.59.0.
- [Release notes](https://github.com/rollup/rollup/releases)
- [Changelog](https://github.com/rollup/rollup/blob/master/CHANGELOG.md)
- [Commits](https://github.com/rollup/rollup/compare/v4.47.1...v4.59.0)

---
updated-dependencies:
- dependency-name: rollup
  dependency-version: 4.59.0
  dependency-type: indirect
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
2026-03-12 02:42:59 +00:00
dependabot[bot]GitHubdependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
384291ba27 Bump minimatch from 3.1.2 to 3.1.5 in /python/packages/devui/frontend (#4337)
Bumps [minimatch](https://github.com/isaacs/minimatch) from 3.1.2 to 3.1.5.
- [Changelog](https://github.com/isaacs/minimatch/blob/main/changelog.md)
- [Commits](https://github.com/isaacs/minimatch/compare/v3.1.2...v3.1.5)

---
updated-dependencies:
- dependency-name: minimatch
  dependency-version: 3.1.5
  dependency-type: indirect
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
2026-03-12 02:42:02 +00:00
Tushar MudiGitHubREDMOND\tusharmudi <tusharmudi@microsoft.com>
378bee577e Fix CWE-863: Validate function approval responses in DevUI executor (#4598)
The DevUI /v1/responses endpoint accepts function_approval_response content
without verifying that the request_id corresponds to a real pending approval
request issued by the server. This allows forged approval responses to
execute arbitrary tools with attacker-controlled arguments, bypassing
approval_mode='always_require'.

Changes:
- Track outgoing approval requests in a server-side registry
  (_pending_approvals) keyed by request_id
- Validate incoming approval responses against this registry; reject
  any response whose request_id was not issued by the server
- Use server-stored function_call data (tool name, arguments, call_id)
  instead of client-supplied data when constructing the approval response
- Consume request_ids on use (pop from registry) to prevent replay attacks

Tests:
- 8 new tests covering forged rejection, server-data enforcement,
  anti-replay, multiple independent approvals, and edge cases

Co-authored-by: REDMOND\tusharmudi <tusharmudi@microsoft.com>
2026-03-12 02:34:31 +00:00
18e433fc6d Python: Validate approval responses against server-side pending request registry (#4548)
* Validate approval responses against server-side pending request registry

* improvements

* pin GHCP sdk version to non-breaking for now

* Pin CHCP sdk to LKG.

* really fix GHCP sdk pkg version

* Fix HITL approval validation security gaps and memory leak

- Validate rejected approval responses against pending_approvals registry,
  not just approved ones. Fabricated rejections without a prior request are
  now stripped from messages before reaching the LLM.
- Bound _pending_approvals with OrderedDict + LRU eviction (max 10k) to
  prevent unbounded memory growth from abandoned approval requests.
- Skip registration when function_call.name is None/empty; log warning
  when content.id or function_call is missing at registration time.
- Document pending_approvals parameter in run_agent_stream docstring.
- Add test for fabricated rejection attack scenario.
- Assert pending approval entry is preserved after function name mismatch.
- Pre-populate pending_approvals in rejection test for correct validation.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Apply pre-commit auto-fixes

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-11 23:21:29 +00:00
2f2495e196 Python: Fix function_approval_response extraction in AG-UI workflow path (#4550)
* Extract function_approval_response from workflow messages (#4546)

_extract_responses_from_messages now handles function_approval_response
content in addition to function_result content. Previously, approval
responses sent via the messages field were silently dropped because the
function only checked for content.type == "function_result".

The approval response is keyed by content.id and includes the approved
status, id, and serialized function_call — consistent with how
_coerce_content identifies approval response payloads.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Apply pre-commit auto-fixes

* Fix #4546: Update docstring and add integration tests for message-based approvals

- Update _extract_responses_from_messages docstring to reflect that it
  now handles function_approval_response content in addition to
  function_result content.
- Add integration tests for run_workflow_stream across two turns with
  approval responses provided via messages (function_approvals) rather
  than resume.interrupts, covering both approved and denied scenarios.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review feedback for #4546

- Use safer 'not .get("interrupt")' assertion instead of 'not in'
  to handle Pydantic v2 model_dump() including keys with None values
- Add unit test for mixed function_result and function_approval_response
  in the same message to TestExtractResponsesFromMessages

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-11 22:54:16 +00:00
Dmytro StrukandGitHub e5d6e8ca98 Fixed CA1873 warning (#4634) 2026-03-11 15:28:24 -07:00
Tao ChenandGitHub b1866bd279 Python: Fix missing status input for OpenAI responses API (#4626)
* Fix missing status input for OpenAI responses API

* Fix mypy

* Address comments

* Remove raw_rep restore

* Do not set status if it's None
2026-03-11 21:20:23 +00:00
3e03a305f6 Python: Implement annotation-based context compaction (#4469)
* Implement annotation-based context compaction

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Handle missing compaction attributes in BaseChatClient

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix CI typing and bandit issues

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Optimize incremental compaction annotation pass

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* refinement

* Python: add ToolResultCompactionStrategy and CompactionProvider

Add ToolResultCompactionStrategy that collapses older tool-call groups
into short summary messages (e.g. [Tool calls: get_weather]) while
keeping the most recent groups verbatim. This mirrors the .NET
ToolResultCompactionStrategy from PR #4533.

Add CompactionProvider as a context-provider that auto-applies compaction
before each agent turn and stores compacted history in session state
after each turn.

Includes tests and samples for both features.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* refinement and alignment with dotnet PR

* updated tool result compaction

* updated tool result compaction

* Python: add ToolResultCompactionStrategy, CompactionProvider, and skip_excluded

- ToolResultCompactionStrategy collapses older tool-call groups into
  [Tool results: func_name: result] summaries with bidirectional tracing
  (same pattern as SummarizationStrategy).
- CompactionProvider as BaseContextProvider with separate before_strategy
  and after_strategy parameters. before_strategy compacts loaded context;
  after_strategy compacts stored history via history_source_id.
- InMemoryHistoryProvider gains skip_excluded flag to filter out messages
  marked as excluded by compaction strategies.
- Tests, samples, and exports updated.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fixed checks

* fix mypy

* Fix: ensure summary messages from both strategies get full compaction annotations

SummarizationStrategy was not calling annotate_message_groups after
inserting its summary message, so the summary lacked core group
annotations (id, kind, index, has_reasoning, _excluded). Added the
missing call. ToolResultCompactionStrategy already had it.

Added tests verifying both strategies produce fully annotated summaries.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* updated propagation

* fix mypy

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-11 19:23:00 +00:00
Dmytro StrukandGitHub 565c0b1623 Updated package versions (#4632) 2026-03-11 19:05:27 +00:00
Dmytro StrukandGitHub 53b0753dfb Prepare RC4 release (#4631) 2026-03-11 18:53:38 +00:00
23ebfbc937 Python: Support skill scripts execution (#4558)
* support skill scripts execution

* fix mixed line endings

* address comments and fix syntax issues

* use few try/except instead of one

* change samples

* validate either script path or script resource is set not both

* fix: separate LLM args from runtime kwargs in skill script execution

* address pr review comments

* address PR review comments

* Update python/packages/core/agent_framework/_skills.py

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update python/packages/core/agent_framework/_skills.py

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update python/packages/core/agent_framework/_skills.py

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* 1. Fixing the caching bug where parameters_schema would re-inspect on every call when the result was None
   2. Updating the arguments tool description to be more generic (not CLI-specific)

* fix failing tests

* address pr review comments

* address pr review comments

* allow resource function returning any instead of sting

* address PR review comments

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
2026-03-11 18:28:30 +00:00
westeyandGitHub 2f8fd5f82f .NET: Add FinishReason to AgentResponses (#4617)
* Add FinishReason to AgentResponses

* Address PR comments
2026-03-11 14:22:56 +00:00
60d5093421 .NET: SDK Patch Bump (10.0.200) - Address false positive trigger of IL2026/IL3050 diagnostics in hosting projects (#4586)
* Suppress IL2026/IL3050 with targeted pragmas on affected methods

Add #pragma warning disable/restore for IL2026 and IL3050 only around
the specific methods where dotnet format incorrectly adds
[RequiresUnreferencedCode] and [RequiresDynamicCode] attributes despite
proper interceptors configuration in the csproj.

See https://github.com/dotnet/sdk/issues/51136

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Upgrade to .NET SDK 10.0.200 and remove IL2026/IL3050 workarounds

Bump global.json to SDK 10.0.200 which fixes the dotnet format bug
that incorrectly added [RequiresUnreferencedCode] and
[RequiresDynamicCode] attributes (https://github.com/dotnet/sdk/issues/51136).

Remove all #pragma warning disable IL2026/IL3050 workarounds from
source files and the --exclude-diagnostics flag from the CI format
workflow.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-11 10:47:08 +00:00
ChrisGitHubwesteyCopilotcopilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
d3f0c33180 .NET Compaction - Introducing compaction strategies and pipeline (#4533)
* Checkpoint

* Checkpoint

* Stable

* Strategies

* Updated

* Encoding

* Formatting

* Cleanup

* Formatting

* Tests

* Tuning

* Update tests

* Test update

* Remove working solution

* Add sample to solution

* Sample readyme

* Experimental

* Format

* Formatting

* Encoding

* Support IChatReducer

* Sample output formatting

* Initial plan

* Replace CompactingChatClient with MessageCompactionContextProvider

Co-authored-by: crickman <66376200+crickman@users.noreply.github.com>

* Boundary condition

* Fix encoding

* Fix cast

* Test coverage

* Namespace

* Improvements

* Efficiency

* Cleanup

* Detect service managed conversation

* Fix namespace

* Fix merge

* Fix test expectation

* Update dotnet/src/Microsoft.Agents.AI.Abstractions/InMemoryChatHistoryProvider.cs

Co-authored-by: westey <164392973+westey-m@users.noreply.github.com>

* Address PR comments (x1)

* Update comment

* Update comments

* Clean-up

* Format output

* Sync sample comment

* Fix condition

* Adjust data-flow

* Address comments (x2)

* Direct compaction

* Fix summarization content

* Argument check / fix count calculation

* Minor follow-up

* Diagnostics

* Minor updates

* Fix state test

* Fix sliding window perf

* Stable state keys

* Increase size computation

* Formatting

* Add README.md for Agent_Step18_CompactionPipeline sample (#4574)

* Sample comments

* Updated

* Update dotnet/src/Microsoft.Agents.AI/Compaction/MessageIndex.cs

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update dotnet/tests/Microsoft.Agents.AI.UnitTests/Compaction/CompactionProviderTests.cs

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update dotnet/src/Microsoft.Agents.AI/Compaction/MessageIndex.cs

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Address copilot comments

* Fix namespace

* Comments / convensions

* Prefix `MessageGroup` and `MessageIndex`

* Fix sliding window

* Update dotnet/src/Microsoft.Agents.AI/Compaction/SummarizationCompactionStrategy.cs

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update dotnet/src/Microsoft.Agents.AI.Abstractions/InMemoryChatHistoryProvider.cs

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Python alignment

* Fix merge

* Fix equality, readme, and sample

* Readme update and ToolResult fix

* Update dotnet/src/Microsoft.Agents.AI/Compaction/SummarizationCompactionStrategy.cs

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update dotnet/samples/02-agents/Agents/Agent_Step18_CompactionPipeline/README.md

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Simplify readme

* Update dotnet/samples/02-agents/Agents/Agent_Step18_CompactionPipeline/README.md

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Remove example

* Remove unused

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: westey <164392973+westey-m@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
2026-03-11 00:41:39 +00:00
CopilotGitHubcrickmancopilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
97b6c9951a Python: Fix broken link in purview README (504 on Microsoft 365 Dev Program URL) (#4610)
* Initial plan

* Fix broken link in purview README: replace 504-returning dev-program URL with stable learn.microsoft.com URL

Co-authored-by: crickman <66376200+crickman@users.noreply.github.com>

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: crickman <66376200+crickman@users.noreply.github.com>
2026-03-11 00:12:53 +00:00
e35f530f2e Python: Fix executor_completed event with non-copyable raw_representation in mixed workflows (#4493)
* Python: Fix `executor_completed` event with non-copyable raw_representation in mixed workflows

Fixes #4455

* fix(#4455): use class-level sets for deepcopy field exclusion

- SerializationMixin.__deepcopy__: check type(self).DEFAULT_EXCLUDE
  instead of hardcoding 'raw_representation'
- Content.__deepcopy__: add _SHALLOW_COPY_FIELDS class variable and
  check against it instead of hardcoding
- Fix tautological assertion in test (was always True)
- Add second excluded field to test to verify DEFAULT_EXCLUDE is
  respected generically

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Decouple __deepcopy__ from DEFAULT_EXCLUDE in SerializationMixin (#4455)

Introduce _SHALLOW_COPY_FIELDS class variable in SerializationMixin to
separate deep-copy semantics from serialization semantics. Previously,
__deepcopy__ used DEFAULT_EXCLUDE to decide which fields to shallow-copy,
conflating 'not serialized' with 'not safe to deep-copy'. A field added
to DEFAULT_EXCLUDE purely for serialization (e.g. additional_properties)
would be silently shared between original and copy.

- Add _SHALLOW_COPY_FIELDS (default {'raw_representation'}) to
  SerializationMixin, matching the pattern already used by Content
- Update __deepcopy__ to read from _SHALLOW_COPY_FIELDS instead of
  DEFAULT_EXCLUDE
- Add test verifying DEFAULT_EXCLUDE fields are deep-copied unless
  also in _SHALLOW_COPY_FIELDS
- Add test for Content._SHALLOW_COPY_FIELDS identity preservation
- Add test for ChatResponse deep-copying additional_properties

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add test for _SHALLOW_COPY_FIELDS and DEFAULT_EXCLUDE independence

Add test_deepcopy_shallow_copy_fields_override_default_exclude to verify
that a field in both DEFAULT_EXCLUDE and _SHALLOW_COPY_FIELDS is
shallow-copied (controlled by _SHALLOW_COPY_FIELDS), while a field in
DEFAULT_EXCLUDE only is still deep-copied. This addresses review comment
#11 ensuring the two class variables control independent concerns.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove unnecessary local variable in __deepcopy__

Inline cls._SHALLOW_COPY_FIELDS directly in the loop check instead of
assigning to a local variable first, per review feedback.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Apply pre-commit auto-fixes

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-10 22:20:05 +00:00
Peter IbekweandGitHub a3bfad4791 .NET: Added support for polymorphic type as workflow output (#4485)
* Added support for polymorphic type as workflow output

* Update Linq expression to avoid unnecessary allocations.

* Added caching as per PR comment
2026-03-10 19:45:01 +00:00
Ahmed MuhsinandGitHub 09b3e2e4f0 Python: Prevent pickle deserialization of untrusted HITL HTTP input (#4566)
* fix: prevent pickle deserialization of untrusted HITL input

Add strip_pickle_markers() to sanitize HTTP input before it reaches
pickle.loads() via the checkpoint decoding path. Applied as a 3-layer
defence-in-depth:

1. _app.py: sanitize req.get_json() at the HTTP boundary
2. _workflow.py: sanitize in _deserialize_hitl_response() before decode
3. _serialization.py: sanitize in reconstruct_to_type() as final guard

Any dict containing __pickled__ or __type__ markers from untrusted
sources is replaced with None, blocking arbitrary code execution via
crafted payloads to POST /workflow/respond/{instanceId}/{requestId}.

Includes 12 new unit tests covering the sanitizer and end-to-end
attack prevention.

* refactor: address review concerns for pickle fix

1. Remove deserialize_value() fallback in _deserialize_hitl_response
   untrusted HITL data now returns as-is when no type hint is available,
   never flowing into pickle.loads().

2. Move strip_pickle_markers() out of reconstruct_to_type()  the function
   is general-purpose again; untrusted-data callers are responsible for
   sanitizing first (documented with NOTE comment).

3. Define _PICKLE_MARKER/_TYPE_MARKER as local constants with import-time
   assertions against core's values  decouples from private names while
   failing loudly if core ever changes them.

4. Update tests to reflect new responsibility boundaries.

* fix: simplify warning message and fix ruff RUF001 lint

* fix: suppress pyright reportPrivateUsage on core marker imports

* Lower marker-strip log from warning to debug to avoid log flooding

* Replace assert with RuntimeError for marker sync checks (ruff S101)

* Fix pyright and ruff CI errors in security fix

- Use cast() for dict/list comprehensions in strip_pickle_markers (pyright)
- type: ignore for narrowed dict return in _workflow.py (pyright)
- Simplify marker imports: use core constants directly, remove local copies
- Remove duplicate pyright ignore comment

* Remove duplicate end-to-end test in TestStripPickleMarkers

* Suppress mypy redundant-cast on list cast needed by pyright
2026-03-10 19:29:33 +00:00
Tao ChenandGitHub 55fc882ca8 Python: Fix store=False not overriding client default (#4569)
* Fix store=False not overriding client default

* Address comments

* Fix unit tests

* Fix integration tests

* Fix tests
2026-03-10 18:44:59 +00:00
westeyandGitHub c15f075412 Cleanup unecessary usages of AsIChatClient (#4561) 2026-03-10 15:40:44 +00:00
CopilotGitHubSergeyMenshykhcopilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
fbcf1444ee Fix Strands Agents documentation links in ADR (#4584)
* Initial plan

* Fix broken Strands Agents documentation links in ADR 0001

Replace 5 broken strandsagents.com URLs (returning 404) with stable
GitHub source code links in docs/decisions/0001-agent-run-response.md.

The Strands Agents docs site restructured from /api-reference/python/
to /api/python/, breaking the old links.

Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com>

* Update Strands Agents links to use official documentation site

Replace GitHub source links with official strandsagents.com/docs/api/python/
documentation URLs in docs/decisions/0001-agent-run-response.md.

Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com>

* Update Strands Agents links to use specific documentation URLs

- Streaming: strandsagents.com/docs/user-guide/concepts/streaming/
- Structured output: strandsagents.com/docs/user-guide/concepts/agents/structured-output/
- AgentResult/stop_reason: strandsagents.com/docs/api/python/strands.agent.agent_result/#agentresult

Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com>

* Deduplicate Strands AgentResult link in stop-reason row

Replaced the duplicate hyperlink on `stop_reason` with inline code,
keeping a single AgentResult link to the same URL.

Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com>

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com>
2026-03-10 15:17:23 +00:00
1b7668119d .NET: Enable Microsoft.Agents.AI.FoundryMemory for NuGet release (#4559)
* Enable Microsoft.Agents.AI.FoundryMemory for NuGet release

- Remove IsPackable=false override from .csproj to inherit IsPackable=true from nuget-package.props
- Add project to agent-framework-release.slnf for inclusion in build/sign/publish pipeline

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Update FoundryMemoryProvider and MemorySearch sample

- StoreAIContextAsync fires UpdateMemoriesAsync immediately (non-accumulation)
- WhenUpdatesCompletedAsync polls last updateId via GetUpdateResultAsync
- Updated FoundryAgents_Step22_MemorySearch sample to create/destroy memory store
  (matching features/foundry-agent-client pattern)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Update FoundryAgents_Step22_MemorySearch sample

- Sample now creates/destroys memory store (self-contained lifecycle)
- Uses WaitForMemoriesUpdateAsync for seeding memories
- Cleanup in finally block deletes both agent and memory store
- Matches features/foundry-agent-client pattern

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-10 13:52:45 +00:00
fd1c66121e .NET: Skip Azure Persistent (V1) flaky CodeInterpreter integration tests (#4583)
* Skip flaky CodeInterpreter integration tests in CI

The CreateAgent_CreatesAgentWithCodeInterpreter tests fail intermittently
because the Azure AI Code Interpreter service sometimes fails to read/execute
uploaded Python files. This causes all 4 integration test jobs to fail
consistently across both platforms (ubuntu/windows) and TFMs (net10.0/net472).

Mark both test variants with Skip to match the convention used by other
flaky tests in the suite (e.g., AzureAIAgentsPersistentStructuredOutputRunTests).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Skip flaky CodeInterpreter integration tests in CI

The CreateAgent_CreatesAgentWithCodeInterpreter tests fail intermittently
because the Azure AI Code Interpreter service sometimes fails to read/execute
uploaded Python files. This causes all 4 integration test jobs to fail
consistently across both platforms (ubuntu/windows) and TFMs (net10.0/net472).

Mark both test variants with Skip to match the convention used by other
flaky tests in the suite (e.g., AzureAIAgentsPersistentStructuredOutputRunTests).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-10 12:52:40 +00:00
ded32f3ff8 Python: Add A2A server sample (#4528)
* Python: Add A2A server sample and fix client streaming bug

Add a pure Python A2A server sample so testing the A2A client no longer
requires running the .NET server. The server uses the a2a-sdk's
A2AStarletteApplication with uvicorn and supports three agent types
(invoice, policy, logistics) backed by AzureOpenAIResponsesClient.

New files:
- a2a_server.py: Main server entry point with CLI args
- agent_executor.py: Bridges a2a-sdk AgentExecutor to Agent Framework
- agent_definitions.py: Agent and AgentCard factory definitions
- invoice_data.py: Mock invoice data and query tool functions
- a2a_server.http: REST Client requests for testing

Also fixes a streaming bug in agent_with_a2a.py where async with was
used on ResponseStream which does not support the async context manager
protocol. Changed to async for to match all other samples.

Closes #4045

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review: handle CancelledError and fix end_date filtering

- Re-raise asyncio.CancelledError before the broad exception handler
  so cooperative cancellation is not swallowed.
- Make end_date filter inclusive of the full day by comparing with
  < end + timedelta(days=1) instead of <= midnight.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-10 00:00:49 +00:00
e2f0bc814e Fix chat_response_cancellation sample to use Message objects (#4532)
The sample was passing raw strings in a list to get_response(), which
expects Message objects. This caused an AttributeError since strings
don't have a 'role' attribute.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-09 23:59:41 +00:00
6cb2289a16 Auto-finalize ResponseStream on iteration completion (#4478)
* Add multi-turn streaming sample and rename multi-turn samples

- Rename 03_multi_turn.py to 03a_multi_turn.py
- Add 03b_multi_turn_streaming.py showing streaming with session history
- The new sample demonstrates calling get_final_response() after
  iterating the stream to persist conversation history
- Update READMEs to reflect the new file names

Closes #4447

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Auto-finalize ResponseStream on iteration completion

When a ResponseStream is fully consumed via async iteration,
automatically trigger finalization (finalizer + result hooks).
This ensures session history is persisted in streaming multi-turn
conversations without requiring an explicit get_final_response() call.

- Add auto-finalize call in __anext__ on StopAsyncIteration
- Guard inner stream finalization to prevent double-execution
- Re-check _finalized after iteration in get_final_response()
- Add tests for auto-finalization and streaming session history
- Revert sample file renames from previous commit

Closes #4447

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* README fix

* Fix SIM102 lint: combine nested if statements

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-09 22:29:09 +00:00
2aaca50217 Python: Exclude conversation_id from chat completions API options (#4517)
* Python: Exclude conversation_id from chat completions options (#4315)

When a session with service_session_id is passed to an agent using the
Chat Completions client, conversation_id leaked through _prepare_options()
into AsyncCompletions.create(), causing an 'unexpected keyword argument'
error. The Responses client already excluded conversation_id but the Chat
Completions client did not.

Added conversation_id to the exclusion set in _prepare_options().

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Apply pre-commit auto-fixes

* Remove reproduction report artifact

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-09 17:03:50 +00:00
f74bda5a83 Python: Fix conversation-id propagation when chat_options is a dict (#4340)
* Fix #4305: Handle dict chat_options in _update_conversation_id

_update_conversation_id assumed chat_options had attribute access, but
ChatOptions is a TypedDict (dict). When a dict was passed, setting
.conversation_id raised AttributeError. Now checks isinstance(dict) and
uses key access for dicts, falling back to attribute access for objects.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR feedback: use Mapping ABC and add missing tests (#4305)

- Use collections.abc.Mapping instead of dict for isinstance check in
  _update_conversation_id, making it more robust for non-dict mapping types.
- Add test for object-style chat_options with optional options dict parameter.
- Add test verifying existing conversation_id gets overwritten (idempotent).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove unnecessary Mapping check in _update_conversation_id (#4305)

chat_options is always a dict, so the isinstance(chat_opts, Mapping)
check and the else branch for attribute-style access are dead code.
Simplify to direct dict key assignment and remove object-style tests.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-09 10:56:57 +00:00
23d6d91c8f Python: [Breaking] Upgrade to azure-ai-projects 2.0+ (#4536)
* Prepare azure-ai-projects 2.0 GA compatibility

Add allow_preview support for internal AIProjectClient creation, keep backward compatibility for renamed SDK model classes, and align Azure AI/core paths and tests for GA validation workflows.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* upgrade to ai-project==2.0.0

* Python: remove azure-ai-projects keyword-guard paths

Assume azure-ai-projects 2.0+ in Azure AI client/provider/responses code paths by removing _supports_keyword_argument gating and related fallback branching.

Also fix pyright typing in FoundryMemoryProvider memory store calls by using ResponseInputItemParam-typed items.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* check fixes

* Python: remove unsupported foundry_features option

Drop foundry_features from Azure AI client and provider surfaces because azure-ai-projects 2.0.0 does not expose that create_version parameter.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Python: add allow_preview to Foundry memory provider

Propagate allow_preview when FoundryMemoryProvider constructs an AIProjectClient and update tests accordingly.

Also finish wiring allow_preview through AzureAIClient-facing surfaces and related docs.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* aligning docstrings

* udpated lock

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-09 10:12:47 +00:00
d5e240b375 [BREAKING] Python: Update github-copilot-sdk integration to use ToolInvocation/ToolResult types (#4551)
* Update github_copilot package for github-copilot-sdk>=0.1.32 (#4549)

- Update requires-python from >=3.10 to >=3.11
- Remove Python 3.10 classifier
- Update mypy python_version to 3.11
- Update dependency to github-copilot-sdk>=0.1.32
- Fix ToolResult API: use snake_case kwargs (text_result_for_llm,
  result_type) instead of camelCase (textResultForLlm, resultType)
- Update test assertions to use attribute access on ToolResult
- Add ToolResult type assertions to tool handler tests

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix tests to use ToolInvocation dataclass instead of plain dict (#4549)

Update test_github_copilot_agent.py to pass ToolInvocation objects to tool
handlers instead of plain dicts, matching the github-copilot-sdk>=0.1.32 API
where ToolInvocation is a dataclass with an .arguments attribute.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add regression tests for ToolInvocation contract (#4549)

Add tests to lock in the new ToolInvocation-based calling convention:
- test_tool_handler_rejects_raw_dict_invocation: verifies passing a raw
  dict (old calling convention) raises TypeError/AttributeError
- test_tool_handler_with_empty_arguments: verifies ToolInvocation with
  empty arguments works correctly for no-arg tools

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Revert requires-python to >=3.10 to avoid breaking CI (#4549)

The repo CI runs with Python 3.10 (uv sync --all-packages) and all other
packages require >=3.10. Raising this package to >=3.11 would break the
shared install flow. The SDK dependency version constraint (>=0.1.32) will
enforce any Python version requirement from the SDK itself.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix min Python version for github_copilot package to >=3.11

github-copilot-sdk>=0.1.32 requires Python>=3.11, which conflicts
with the package's declared >=3.10 minimum, breaking uv sync.

* Bump py version for GH workflows to 3.11, exclude GHCP sdk from 3.10 items

* Fix uv command

* Fixes

* Update samples

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-03-09 09:57:51 +00:00
westeyandGitHub 1ca43f9643 .NET: Add security warnings to xml comments for core components (#4527)
* Add security warnings to xml comments for core components

* Address build errors.

* Fix formatting issue

* Fix formatting issue

* Supress formatting warning

* Supress format issue in ChatHistoryMemoryProvider

* Fix remarks paragraphs
2026-03-06 19:04:22 +00:00
CopilotGitHubrogerbarretocopilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
b98880df32 .NET: Update Anthropic to 12.8.0 and Anthropic.Foundry to 0.4.2 (#4475)
* Initial plan

* Update Anthropic to 12.8.0 and Anthropic.Foundry to 0.4.2

Co-authored-by: rogerbarreto <19890735+rogerbarreto@users.noreply.github.com>

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: rogerbarreto <19890735+rogerbarreto@users.noreply.github.com>
2026-03-06 18:57:02 +00:00
westeyandGitHub c8750cbe92 .NET: Create a sample to show bounded chat history with overflow into chat history memory (#4136)
* Create a sample to show bounded chat history with overflow into chat history memory

* Address PR comments.

* Address PR comment and fix bug
2026-03-06 18:03:43 +00:00
332 changed files with 29434 additions and 4062 deletions
+18
View File
@@ -8,6 +8,10 @@ inputs:
os:
description: The operating system to set up
required: true
exclude-packages:
description: Space-separated list of packages to exclude from uv sync
required: false
default: ''
runs:
using: "composite"
@@ -19,6 +23,20 @@ runs:
enable-cache: true
cache-suffix: ${{ inputs.os }}-${{ inputs.python-version }}
cache-dependency-glob: "**/uv.lock"
- name: Exclude incompatible workspace packages
if: ${{ inputs.exclude-packages != '' }}
shell: bash
run: |
for pkg in ${{ inputs.exclude-packages }}; do
for f in python/packages/*/pyproject.toml; do
if grep -q "name = \"$pkg\"" "$f"; then
pkg_dir=$(dirname "$f" | sed 's|python/||')
echo "Excluding workspace package: $pkg ($pkg_dir)"
sed -i.bak '/\[tool\.uv\.workspace\]/a\exclude = ["'"$pkg_dir"'"]' python/pyproject.toml
sed -i.bak '/'"$pkg"' = { workspace = true }/d' python/pyproject.toml
fi
done
done
- name: Install the project
shell: bash
run: |
+1 -2
View File
@@ -86,11 +86,10 @@ jobs:
run: docker pull mcr.microsoft.com/dotnet/sdk:${{ matrix.dotnet }}
# This step will run dotnet format on each of the unique csproj files and fail if any changes are made
# exclude-diagnostics should be removed after fixes for IL2026 and IL3050 are out: https://github.com/dotnet/sdk/issues/51136
- name: Run dotnet format
if: steps.find-csproj.outputs.csproj_files != ''
run: |
for csproj in ${{ steps.find-csproj.outputs.csproj_files }}; do
echo "Running dotnet format on $csproj"
docker run --rm -v $(pwd):/app -w /app mcr.microsoft.com/dotnet/sdk:${{ matrix.dotnet }} /bin/sh -c "dotnet format $csproj --verify-no-changes --verbosity diagnostic --exclude-diagnostics IL2026 IL3050"
docker run --rm -v $(pwd):/app -w /app mcr.microsoft.com/dotnet/sdk:${{ matrix.dotnet }} /bin/sh -c "dotnet format $csproj --verify-no-changes --verbosity diagnostic"
done
+4 -4
View File
@@ -18,7 +18,7 @@ jobs:
strategy:
fail-fast: false
matrix:
python-version: ["3.10"]
python-version: ["3.11"]
runs-on: ubuntu-latest
continue-on-error: true
defaults:
@@ -55,7 +55,7 @@ jobs:
strategy:
fail-fast: false
matrix:
python-version: ["3.10"]
python-version: ["3.11"]
runs-on: ubuntu-latest
continue-on-error: true
defaults:
@@ -84,7 +84,7 @@ jobs:
strategy:
fail-fast: false
matrix:
python-version: ["3.10"]
python-version: ["3.11"]
runs-on: ubuntu-latest
continue-on-error: true
defaults:
@@ -117,7 +117,7 @@ jobs:
strategy:
fail-fast: false
matrix:
python-version: ["3.10"]
python-version: ["3.11"]
runs-on: ubuntu-latest
continue-on-error: true
defaults:
@@ -0,0 +1,216 @@
# Probe the highest allowed dependency versions, then open issues/PRs from the passing updates.
name: Python - Dependency Range Validation
on:
workflow_dispatch:
permissions:
contents: write
issues: write
pull-requests: write
env:
UV_CACHE_DIR: /tmp/.uv-cache
jobs:
dependency-range-validation:
name: Dependency Range Validation
runs-on: ubuntu-latest
env:
# For now only run 3.13, if we do encounter situations where there are mismatches between packages and python versions (other then 3.10 and 3.14 which are known to not be able to install everything)
# then we will have to reevaluate.
UV_PYTHON: "3.13"
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 0
- name: Set up python and install the project
uses: ./.github/actions/python-setup
with:
python-version: ${{ env.UV_PYTHON }}
os: ${{ runner.os }}
env:
UV_CACHE_DIR: /tmp/.uv-cache
- name: Run dependency range validation
id: validate_ranges
# Keep workflow running so we can still publish diagnostics from this run.
continue-on-error: true
run: uv run poe validate-dependency-bounds-project --mode upper --project "*"
working-directory: ./python
- name: Upload dependency range report
# Always publish the report so failures are inspectable even when validation fails.
if: always()
uses: actions/upload-artifact@v4
with:
name: dependency-range-results
path: python/scripts/dependencies/dependency-range-results.json
if-no-files-found: warn
- name: Create issues for failed dependency candidates
# Always process the report so failed candidates create actionable tracking issues.
if: always()
uses: actions/github-script@v8
with:
script: |
const fs = require("fs")
const reportPath = "python/scripts/dependencies/dependency-range-results.json"
if (!fs.existsSync(reportPath)) {
core.warning(`No dependency range report found at ${reportPath}`)
return
}
const report = JSON.parse(fs.readFileSync(reportPath, "utf8"))
const dependencyFailures = []
for (const packageResult of report.packages ?? []) {
for (const dependency of packageResult.dependencies ?? []) {
const candidateVersions = new Set(dependency.candidate_versions ?? [])
const failedAttempts = (dependency.attempts ?? []).filter(
(attempt) => attempt.status === "failed" && candidateVersions.has(attempt.trial_upper)
)
if (!failedAttempts.length) {
continue
}
const failuresByVersion = new Map()
for (const attempt of failedAttempts) {
const version = attempt.trial_upper || "unknown"
if (!failuresByVersion.has(version)) {
failuresByVersion.set(version, attempt.error || "No error output captured.")
}
}
dependencyFailures.push({
packageName: packageResult.package_name,
projectPath: packageResult.project_path,
dependencyName: dependency.name,
originalRequirements: dependency.original_requirements ?? [],
finalRequirements: dependency.final_requirements ?? [],
failedVersions: [...failuresByVersion.entries()].map(([version, error]) => ({ version, error })),
})
}
}
if (!dependencyFailures.length) {
core.info("No failing dependency candidates found.")
return
}
const owner = context.repo.owner
const repo = context.repo.repo
const openIssues = await github.paginate(github.rest.issues.listForRepo, {
owner,
repo,
state: "open",
per_page: 100,
})
const openIssueTitles = new Set(
openIssues.filter((issue) => !issue.pull_request).map((issue) => issue.title)
)
const formatError = (message) => String(message || "No error output captured.").replace(/```/g, "'''")
for (const failure of dependencyFailures) {
const title = `Dependency validation failed: ${failure.dependencyName} (${failure.packageName})`
if (openIssueTitles.has(title)) {
core.info(`Issue already exists: ${title}`)
continue
}
const visibleFailures = failure.failedVersions.slice(0, 5)
const omittedCount = failure.failedVersions.length - visibleFailures.length
const failureDetails = visibleFailures
.map(
(entry) =>
`- \`${entry.version}\`\n\n\`\`\`\n${formatError(entry.error).slice(0, 3500)}\n\`\`\``
)
.join("\n\n")
const body = [
"Automated dependency range validation found candidate versions that failed checks.",
"",
`- Package: \`${failure.packageName}\``,
`- Project path: \`${failure.projectPath}\``,
`- Dependency: \`${failure.dependencyName}\``,
`- Original requirements: ${
failure.originalRequirements.length
? failure.originalRequirements.map((value) => `\`${value}\``).join(", ")
: "_none_"
}`,
`- Final requirements after run: ${
failure.finalRequirements.length
? failure.finalRequirements.map((value) => `\`${value}\``).join(", ")
: "_none_"
}`,
"",
"### Failed versions and errors",
failureDetails,
omittedCount > 0 ? `\n_Additional failed versions omitted: ${omittedCount}_` : "",
"",
`Workflow run: ${context.serverUrl}/${owner}/${repo}/actions/runs/${context.runId}`,
].join("\n")
await github.rest.issues.create({
owner,
repo,
title,
body,
})
openIssueTitles.add(title)
core.info(`Created issue: ${title}`)
}
- name: Refresh lockfile
# Only refresh lockfile after a clean validation to avoid committing known-bad ranges.
if: steps.validate_ranges.outcome == 'success'
run: uv lock --upgrade
working-directory: ./python
- name: Commit and push dependency updates
id: commit_updates
if: steps.validate_ranges.outcome == 'success'
run: |
BRANCH="automation/python-dependency-range-updates"
git config user.name "github-actions[bot]"
git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
git checkout -B "${BRANCH}"
git add python/packages/*/pyproject.toml python/uv.lock
if git diff --cached --quiet; then
echo "has_changes=false" >> "$GITHUB_OUTPUT"
echo "No dependency updates to commit."
exit 0
fi
git commit -m "chore: update dependency ranges"
git push --force-with-lease --set-upstream origin "${BRANCH}"
echo "has_changes=true" >> "$GITHUB_OUTPUT"
- name: Create or update pull request with GitHub CLI
# Only open/update PRs for validated updates to keep automation branches trustworthy.
if: steps.validate_ranges.outcome == 'success' && steps.commit_updates.outputs.has_changes == 'true'
run: |
BRANCH="automation/python-dependency-range-updates"
PR_TITLE="Python: chore: update dependency ranges"
PR_BODY_FILE="$(mktemp)"
cat > "${PR_BODY_FILE}" <<'EOF'
This PR was generated by the dependency range validation workflow.
- Ran `uv run poe validate-dependency-bounds-project --mode upper --project "*"`
- Updated package dependency bounds
- Refreshed `python/uv.lock` with `uv lock --upgrade`
EOF
PR_NUMBER="$(gh pr list --head "${BRANCH}" --base main --state open --json number --jq '.[0].number')"
if [ -n "${PR_NUMBER}" ]; then
gh pr edit "${PR_NUMBER}" --title "${PR_TITLE}" --body-file "${PR_BODY_FILE}"
else
gh pr create --base main --head "${BRANCH}" --title "${PR_TITLE}" --body-file "${PR_BODY_FILE}"
fi
@@ -0,0 +1,91 @@
name: Python - Dev Dependency Upgrade
on:
workflow_dispatch:
permissions:
contents: write
pull-requests: write
env:
UV_CACHE_DIR: /tmp/.uv-cache
jobs:
upgrade-dev-dependencies:
name: Upgrade Dev Dependencies
runs-on: ubuntu-latest
env:
UV_PYTHON: "3.13"
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 0
- name: Set up python and install the project
uses: ./.github/actions/python-setup
with:
python-version: ${{ env.UV_PYTHON }}
os: ${{ runner.os }}
env:
UV_CACHE_DIR: /tmp/.uv-cache
- name: Upgrade dev dependencies and validate workspace
run: uv run poe upgrade-dev-dependencies
working-directory: ./python
- name: Commit and push dev dependency updates
id: commit_updates
run: |
BRANCH="automation/python-dev-dependency-updates"
git config user.name "github-actions[bot]"
git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
git checkout -B "${BRANCH}"
git add python/pyproject.toml python/packages/*/pyproject.toml python/uv.lock
if git diff --cached --quiet; then
echo "has_changes=false" >> "$GITHUB_OUTPUT"
echo "No dev dependency updates to commit."
exit 0
fi
git commit -F- <<'EOF'
Python: chore: upgrade dev dependencies
EOF
git push --force-with-lease --set-upstream origin "${BRANCH}"
echo "has_changes=true" >> "$GITHUB_OUTPUT"
- name: Create or update pull request with GitHub CLI
if: steps.commit_updates.outputs.has_changes == 'true'
run: |
BRANCH="automation/python-dev-dependency-updates"
PR_TITLE="Python: chore: upgrade dev dependencies"
PR_BODY_FILE="$(mktemp)"
cat > "${PR_BODY_FILE}" <<'EOF'
### Motivation and Context
This automated update refreshes Python dev dependency pins across the workspace and reruns the repo validation gates before opening a pull request.
### Description
- Ran `uv run poe upgrade-dev-dependencies`
- Refreshed dev dependency pins in workspace `pyproject.toml` files
- Refreshed `python/uv.lock` with `uv lock --upgrade`
- Reinstalled from the frozen lockfile and reran `check`, `typing`, and `test`
### Contribution Checklist
- [x] The code builds clean without any errors or warnings
- [x] The PR follows the [Contribution Guidelines](https://github.com/microsoft/agent-framework/blob/main/CONTRIBUTING.md)
- [x] All unit tests pass, and I have added new tests where possible
- [ ] **Is this a breaking change?** If yes, add "[BREAKING]" prefix to the title of the PR.
EOF
PR_NUMBER="$(gh pr list --head "${BRANCH}" --base main --state open --json number --jq '.[0].number')"
if [ -n "${PR_NUMBER}" ]; then
gh pr edit "${PR_NUMBER}" --title "${PR_TITLE}" --body-file "${PR_BODY_FILE}"
else
gh pr create --base main --head "${BRANCH}" --title "${PR_TITLE}" --body-file "${PR_BODY_FILE}"
fi
@@ -170,7 +170,7 @@ jobs:
environment: integration
timeout-minutes: 60
env:
UV_PYTHON: "3.10"
UV_PYTHON: "3.11"
OPENAI_CHAT_MODEL_ID: ${{ vars.OPENAI__CHATMODELID }}
OPENAI_RESPONSES_MODEL_ID: ${{ vars.OPENAI__RESPONSESMODELID }}
OPENAI_API_KEY: ${{ secrets.OPENAI__APIKEY }}
+4
View File
@@ -67,6 +67,7 @@ jobs:
with:
python-version: ${{ matrix.python-version }}
os: ${{ runner.os }}
exclude-packages: ${{ matrix.python-version == '3.10' && 'agent-framework-github-copilot' || '' }}
env:
# Configure a constant location for the uv cache
UV_CACHE_DIR: /tmp/.uv-cache
@@ -75,6 +76,9 @@ jobs:
- name: Run lab tests
run: cd packages/lab && uv run poe test
- name: Run resource-intensive lab tests
run: cd packages/lab && uv run pytest -m "resource_intensive and not integration" --junitxml=test-results-resource-intensive.xml
- name: Run lab lint
run: cd packages/lab && uv run poe lint
+1 -1
View File
@@ -288,7 +288,7 @@ jobs:
runs-on: ubuntu-latest
environment: integration
env:
UV_PYTHON: "3.10"
UV_PYTHON: "3.11"
OPENAI_CHAT_MODEL_ID: ${{ vars.OPENAI__CHATMODELID }}
OPENAI_RESPONSES_MODEL_ID: ${{ vars.OPENAI__RESPONSESMODELID }}
OPENAI_API_KEY: ${{ secrets.OPENAI__APIKEY }}
+1 -1
View File
@@ -20,7 +20,7 @@ jobs:
run:
working-directory: python
env:
UV_PYTHON: "3.10"
UV_PYTHON: "3.11"
steps:
- uses: actions/checkout@v6
# Save the PR number to a file since the workflow_run event
+2 -1
View File
@@ -34,12 +34,13 @@ jobs:
with:
python-version: ${{ matrix.python-version }}
os: ${{ runner.os }}
exclude-packages: ${{ matrix.python-version == '3.10' && 'agent-framework-github-copilot' || '' }}
env:
# Configure a constant location for the uv cache
UV_CACHE_DIR: /tmp/.uv-cache
# Unit tests
- name: Run all tests
run: uv run poe all-tests
run: uv run poe all-tests ${{ matrix.python-version == '3.10' && '--ignore-glob=packages/github_copilot/**' || '' }}
working-directory: ./python
# Surface failing tests
+3
View File
@@ -205,6 +205,9 @@ WARP.md
**/memory-bank/
**/projectBrief.md
**/tmpclaude*
# Dependency-bound validation reports
python/scripts/dependency-*-results.json
python/scripts/dependencies/dependency-*-results.json
# Azurite storage emulator files
*/__azurite_db_blob__.json*
+5 -5
View File
@@ -4,8 +4,8 @@ status: accepted
contact: westey-m
date: 2025-07-10 {YYYY-MM-DD when the decision was last updated}
deciders: sergeymenshykh, markwallace, rbarreto, dmytrostruk, westey-m, eavanvalkenburg, stephentoub
consulted:
informed:
consulted:
informed:
---
# Agent Run Responses Design
@@ -64,7 +64,7 @@ Approaches observed from the compared SDKs:
| AutoGen | **Approach 1** Separates messages into Agent-Agent (maps to Primary) and Internal (maps to Secondary) and these are returned as separate properties on the agent response object. See [types of messages](https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/tutorial/messages.html#types-of-messages) and [Response](https://microsoft.github.io/autogen/stable/reference/python/autogen_agentchat.base.html#autogen_agentchat.base.Response) | **Approach 2** Returns a stream of internal events and the last item is a Response object. See [ChatAgent.on_messages_stream](https://microsoft.github.io/autogen/stable/reference/python/autogen_agentchat.base.html#autogen_agentchat.base.ChatAgent.on_messages_stream) |
| OpenAI Agent SDK | **Approach 1** Separates new_items (Primary+Secondary) from final output (Primary) as separate properties on the [RunResult](https://github.com/openai/openai-agents-python/blob/main/src/agents/result.py#L39) | **Approach 1** Similar to non-streaming, has a way of streaming updates via a method on the response object which includes all data, and then a separate final output property on the response object which is populated only when the run is complete. See [RunResultStreaming](https://github.com/openai/openai-agents-python/blob/main/src/agents/result.py#L136) |
| Google ADK | **Approach 2** [Emits events](https://google.github.io/adk-docs/runtime/#step-by-step-breakdown) with [FinalResponse](https://github.com/google/adk-java/blob/main/core/src/main/java/com/google/adk/events/Event.java#L232) true (Primary) / false (Secondary) and callers have to filter out those with false to get just the final response message | **Approach 2** Similar to non-streaming except [events](https://google.github.io/adk-docs/runtime/#streaming-vs-non-streaming-output-partialtrue) are emitted with [Partial](https://github.com/google/adk-java/blob/main/core/src/main/java/com/google/adk/events/Event.java#L133) true to indicate that they are streaming messages. A final non partial event is also emitted. |
| AWS (Strands) | **Approach 3** Returns an [AgentResult](https://strandsagents.com/latest/documentation/docs/api-reference/python/agent/agent_result/) (Primary) with messages and a reason for the run's completion. | **Approach 2** [Streams events](https://strandsagents.com/latest/documentation/docs/api-reference/python/agent/agent/#strands.agent.agent.Agent.stream_async) (Primary+Secondary) including, response text, current_tool_use, even data from "callbacks" (strands plugins) |
| AWS (Strands) | **Approach 3** Returns an [AgentResult](https://strandsagents.com/docs/api/python/strands.agent.agent_result/) (Primary) with messages and a reason for the run's completion. | **Approach 2** [Streams events](https://strandsagents.com/docs/api/python/strands.agent.agent/) (Primary+Secondary) including, response text, current_tool_use, even data from "callbacks" (strands plugins) |
| LangGraph | **Approach 2** A mixed list of all [messages](https://langchain-ai.github.io/langgraph/agents/run_agents/#output-format) | **Approach 2** A mixed list of all [messages](https://langchain-ai.github.io/langgraph/agents/run_agents/#output-format) |
| Agno | **Combination of various approaches** Returns a [RunResponse](https://docs.agno.com/reference/agents/run-response) object with text content, messages (essentially chat history including inputs and instructions), reasoning and thinking text properties. Secondary events could potentially be extracted from messages. | **Approach 2** Returns [RunResponseEvent](https://docs.agno.com/reference/agents/run-response#runresponseevent-types-and-attributes) objects including tool call, memory update, etc, information, where the [RunResponseCompletedEvent](https://docs.agno.com/reference/agents/run-response#runresponsecompletedevent) has similar properties to RunResponse|
| A2A | **Approach 3** Returns a [Task or Message](https://a2aproject.github.io/A2A/latest/specification/#71-messagesend) where the message is the final result (Primary) and task is a reference to a long running process. | **Approach 2** Returns a [stream](https://a2aproject.github.io/A2A/latest/specification/#72-messagestream) that contains task updates (Secondary) and a final message (Primary) |
@@ -496,7 +496,7 @@ We need to decide what AIContent types, each agent response type will be mapped
|-|-|
| AutoGen | **Approach 1** Supports [configuring an agent](https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/tutorial/agents.html#structured-output) at agent creation. |
| Google ADK | **Approach 1** Both [input and output schemas can be specified for LLM Agents](https://google.github.io/adk-docs/agents/llm-agents/#structuring-data-input_schema-output_schema-output_key) at construction time. This option is specific to this agent type and other agent types do not necessarily support |
| AWS (Strands) | **Approach 2** Supports a special invocation method called [structured_output](https://strandsagents.com/latest/documentation/docs/api-reference/python/agent/agent/#strands.agent.agent.Agent.structured_output) |
| AWS (Strands) | **Approach 2** Supports a special invocation method called [structured_output](https://strandsagents.com/docs/api/python/strands.agent.agent/) |
| LangGraph | **Approach 1** Supports [configuring an agent](https://langchain-ai.github.io/langgraph/agents/agents/?h=structured#6-configure-structured-output) at agent construction time, and a [structured response](https://langchain-ai.github.io/langgraph/agents/run_agents/#output-format) can be retrieved as a special property on the agent response |
| Agno | **Approach 1** Supports [configuring an agent](https://docs.agno.com/input-output/structured-output/agent) at agent construction time |
| A2A | **Informal Approach 2** Doesn't formally support schema negotiation, but [hints can be provided via metadata](https://a2a-protocol.org/latest/specification/#97-structured-data-exchange-requesting-and-providing-json) at invocation time |
@@ -508,7 +508,7 @@ We need to decide what AIContent types, each agent response type will be mapped
|-|-|
| AutoGen | Supports a [stop reason](https://microsoft.github.io/autogen/stable/reference/python/autogen_agentchat.base.html#autogen_agentchat.base.TaskResult.stop_reason) which is a freeform text string |
| Google ADK | [No equivalent present](https://github.com/google/adk-python/blob/main/src/google/adk/events/event.py) |
| AWS (Strands) | Exposes a [stop_reason](https://strandsagents.com/latest/documentation/docs/api-reference/python/types/event_loop/#strands.types.event_loop.StopReason) property on the [AgentResult](https://strandsagents.com/latest/documentation/docs/api-reference/python/agent/agent_result/) class with options that are tied closely to LLM operations. |
| AWS (Strands) | Exposes a [stop_reason](https://strandsagents.com/docs/api/python/strands.types.event_loop/) property on the [AgentResult](https://strandsagents.com/docs/api/python/strands.agent.agent_result/) class with options that are tied closely to LLM operations. |
| LangGraph | No equivalent present, output contains only [messages](https://langchain-ai.github.io/langgraph/agents/run_agents/#output-format) |
| Agno | [No equivalent present](https://docs.agno.com/reference/agents/run-response) |
| A2A | No equivalent present, response only contains a [message](https://a2a-protocol.org/latest/specification/#64-message-object) or [task](https://a2a-protocol.org/latest/specification/#61-task-object). |
@@ -1240,3 +1240,10 @@ class AttributionAwareStrategy(CompactionStrategy):
- [ADR-0016: Unifying Context Management with ContextPlugin](0016-python-context-middleware.md) — Parent ADR that established `ContextProvider`, `HistoryProvider`, and `AgentSession` architecture.
- [Context Compaction Limitations Analysis](https://gist.github.com/victordibia/ec3f3baf97345f7e47da025cf55b999f) — Detailed analysis of why current architecture cannot support in-run compaction, with attempted solutions and their failure modes. Option 4 in this ADR corresponds to "Option A: Middleware Access to Mutable Message Source" from that analysis; Options 1-3 correspond to "Option B: Tool Loop Hook", adapted here to a `BaseChatClient` hook instead of `FunctionInvocationConfiguration`.
### Implementation Rollout Note
Implementation is split into two phases:
1. **Phase 1 (PR 1):** runtime compaction foundation in `agent_framework/_compaction.py`, in-run integration, and extensive core tests, plus in-run compaction samples (`basics`, `advanced`, `custom`).
2. **Phase 2 (PR 2):** history/storage compaction (`upsert`-based full replacement), provider support, storage tests, and storage-focused sample (`storage`).
+9 -7
View File
@@ -11,8 +11,8 @@
</PropertyGroup>
<ItemGroup>
<!-- Aspire.* -->
<PackageVersion Include="Anthropic" Version="12.3.0" />
<PackageVersion Include="Anthropic.Foundry" Version="0.4.1" />
<PackageVersion Include="Anthropic" Version="12.8.0" />
<PackageVersion Include="Anthropic.Foundry" Version="0.4.2" />
<PackageVersion Include="Aspire.Azure.AI.OpenAI" Version="13.0.0-preview.1.25560.3" />
<PackageVersion Include="Aspire.Hosting.AppHost" Version="$(AspireAppHostSdkVersion)" />
<PackageVersion Include="Aspire.Hosting.Azure.CognitiveServices" Version="$(AspireAppHostSdkVersion)" />
@@ -33,14 +33,15 @@
<!-- Newtonsoft.Json -->
<PackageVersion Include="Newtonsoft.Json" Version="13.0.4" />
<!-- System.* -->
<PackageVersion Include="Microsoft.Bcl.AsyncInterfaces" Version="10.0.3" />
<PackageVersion Include="Microsoft.Bcl.AsyncInterfaces" Version="10.0.4" />
<PackageVersion Include="Microsoft.Bcl.HashCode" Version="6.0.0" />
<PackageVersion Include="Microsoft.Bcl.Memory" Version="10.0.4" />
<PackageVersion Include="System.ClientModel" Version="1.9.0" />
<PackageVersion Include="System.CodeDom" Version="10.0.0" />
<PackageVersion Include="System.Collections.Immutable" Version="10.0.1" />
<PackageVersion Include="System.CommandLine" Version="2.0.0-rc.2.25502.107" />
<PackageVersion Include="System.Diagnostics.DiagnosticSource" Version="10.0.3" />
<PackageVersion Include="System.Linq.AsyncEnumerable" Version="10.0.0" />
<PackageVersion Include="System.Linq.AsyncEnumerable" Version="10.0.4" />
<PackageVersion Include="System.Net.Http.Json" Version="10.0.0" />
<PackageVersion Include="System.Net.ServerSentEvents" Version="10.0.3" />
<PackageVersion Include="System.Text.Json" Version="10.0.3" />
@@ -101,13 +102,14 @@
<PackageVersion Include="Microsoft.Agents.Authentication.Msal" Version="1.3.171-beta" />
<PackageVersion Include="Microsoft.Agents.Hosting.AspNetCore" Version="1.3.171-beta" />
<!-- A2A -->
<PackageVersion Include="A2A" Version="0.3.3-preview" />
<PackageVersion Include="A2A.AspNetCore" Version="0.3.3-preview" />
<PackageVersion Include="A2A" Version="0.3.4-preview" />
<PackageVersion Include="A2A.AspNetCore" Version="0.3.4-preview" />
<!-- MCP -->
<PackageVersion Include="ModelContextProtocol" Version="0.8.0-preview.1" />
<PackageVersion Include="ModelContextProtocol" Version="1.1.0" />
<!-- Inference SDKs -->
<PackageVersion Include="AWSSDK.Extensions.Bedrock.MEAI" Version="4.0.5.1" />
<PackageVersion Include="Microsoft.ML.OnnxRuntimeGenAI" Version="0.10.0" />
<PackageVersion Include="Microsoft.ML.Tokenizers" Version="2.0.0" />
<PackageVersion Include="OllamaSharp" Version="5.4.8" />
<PackageVersion Include="OpenAI" Version="2.8.0" />
<!-- Identity -->
+2
View File
@@ -56,6 +56,7 @@
<Project Path="samples/02-agents/Agents/Agent_Step15_DeepResearch/Agent_Step15_DeepResearch.csproj" />
<Project Path="samples/02-agents/Agents/Agent_Step16_Declarative/Agent_Step16_Declarative.csproj" />
<Project Path="samples/02-agents/Agents/Agent_Step17_AdditionalAIContext/Agent_Step17_AdditionalAIContext.csproj" />
<Project Path="samples/02-agents/Agents/Agent_Step18_CompactionPipeline/Agent_Step18_CompactionPipeline.csproj" />
</Folder>
<Folder Name="/Samples/02-agents/DeclarativeAgents/">
<Project Path="samples/02-agents/DeclarativeAgents/ChatClient/DeclarativeChatClientAgents.csproj" />
@@ -103,6 +104,7 @@
<Project Path="samples/02-agents/AgentWithMemory/AgentWithMemory_Step01_ChatHistoryMemory/AgentWithMemory_Step01_ChatHistoryMemory.csproj" />
<Project Path="samples/02-agents/AgentWithMemory/AgentWithMemory_Step02_MemoryUsingMem0/AgentWithMemory_Step02_MemoryUsingMem0.csproj" />
<Project Path="samples/02-agents/AgentWithMemory/AgentWithMemory_Step04_MemoryUsingFoundry/AgentWithMemory_Step04_MemoryUsingFoundry.csproj" />
<Project Path="samples/02-agents/AgentWithMemory/AgentWithMemory_Step05_BoundedChatHistory/AgentWithMemory_Step05_BoundedChatHistory.csproj" />
</Folder>
<Folder Name="/Samples/02-agents/AgentWithOpenAI/">
<File Path="samples/02-agents/AgentWithOpenAI/README.md" />
+1
View File
@@ -14,6 +14,7 @@
"src\\Microsoft.Agents.AI.Declarative\\Microsoft.Agents.AI.Declarative.csproj",
"src\\Microsoft.Agents.AI.DevUI\\Microsoft.Agents.AI.DevUI.csproj",
"src\\Microsoft.Agents.AI.DurableTask\\Microsoft.Agents.AI.DurableTask.csproj",
"src\\Microsoft.Agents.AI.FoundryMemory\\Microsoft.Agents.AI.FoundryMemory.csproj",
"src\\Microsoft.Agents.AI.Hosting.A2A.AspNetCore\\Microsoft.Agents.AI.Hosting.A2A.AspNetCore.csproj",
"src\\Microsoft.Agents.AI.Hosting.A2A\\Microsoft.Agents.AI.Hosting.A2A.csproj",
"src\\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj",
+1 -1
View File
@@ -1,6 +1,6 @@
{
"sdk": {
"version": "10.0.100",
"version": "10.0.200",
"rollForward": "minor",
"allowPrerelease": false
},
+4 -4
View File
@@ -2,11 +2,11 @@
<PropertyGroup>
<!-- Central version prefix - applies to all nuget packages. -->
<VersionPrefix>1.0.0</VersionPrefix>
<RCNumber>3</RCNumber>
<RCNumber>4</RCNumber>
<PackageVersion Condition="'$(IsReleaseCandidate)' == 'true'">$(VersionPrefix)-rc$(RCNumber)</PackageVersion>
<PackageVersion Condition="'$(IsReleaseCandidate)' != 'true' AND '$(VersionSuffix)' != ''">$(VersionPrefix)-$(VersionSuffix).260304.1</PackageVersion>
<PackageVersion Condition="'$(IsReleaseCandidate)' != 'true' AND '$(VersionSuffix)' == ''">$(VersionPrefix)-preview.260304.1</PackageVersion>
<GitTag>1.0.0-rc3</GitTag>
<PackageVersion Condition="'$(IsReleaseCandidate)' != 'true' AND '$(VersionSuffix)' != ''">$(VersionPrefix)-$(VersionSuffix).260311.1</PackageVersion>
<PackageVersion Condition="'$(IsReleaseCandidate)' != 'true' AND '$(VersionSuffix)' == ''">$(VersionPrefix)-preview.260311.1</PackageVersion>
<GitTag>1.0.0-rc4</GitTag>
<Configurations>Debug;Release;Publish</Configurations>
<IsPackable>true</IsPackable>
@@ -4,7 +4,6 @@ using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using OpenAI.Chat;
WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
@@ -27,7 +26,7 @@ ChatClient chatClient = new AzureOpenAIClient(
new DefaultAzureCredential())
.GetChatClient(deploymentName);
AIAgent agent = chatClient.AsIChatClient().AsAIAgent(
AIAgent agent = chatClient.AsAIAgent(
name: "AGUIAssistant",
instructions: "You are a helpful assistant.");
@@ -10,7 +10,6 @@
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
@@ -82,7 +82,7 @@ ChatClient chatClient = new AzureOpenAIClient(
new DefaultAzureCredential())
.GetChatClient(deploymentName);
ChatClientAgent agent = chatClient.AsIChatClient().AsAIAgent(
ChatClientAgent agent = chatClient.AsAIAgent(
name: "AGUIAssistant",
instructions: "You are a helpful assistant with access to restaurant information.",
tools: tools);
@@ -10,7 +10,6 @@
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
@@ -4,7 +4,6 @@ using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using OpenAI.Chat;
WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
@@ -27,7 +26,7 @@ ChatClient chatClient = new AzureOpenAIClient(
new DefaultAzureCredential())
.GetChatClient(deploymentName);
AIAgent agent = chatClient.AsIChatClient().AsAIAgent(
AIAgent agent = chatClient.AsAIAgent(
name: "AGUIAssistant",
instructions: "You are a helpful assistant.");
@@ -10,7 +10,6 @@
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
@@ -60,7 +60,7 @@ ChatClient openAIChatClient = new AzureOpenAIClient(
new DefaultAzureCredential())
.GetChatClient(deploymentName);
ChatClientAgent baseAgent = openAIChatClient.AsIChatClient().AsAIAgent(
ChatClientAgent baseAgent = openAIChatClient.AsAIAgent(
name: "AGUIAssistant",
instructions: "You are a helpful assistant in charge of approving expenses",
tools: tools);
@@ -10,7 +10,6 @@
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
@@ -4,7 +4,6 @@ using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.Options;
using OpenAI.Chat;
using RecipeAssistant;
@@ -37,7 +36,7 @@ ChatClient chatClient = new AzureOpenAIClient(
new DefaultAzureCredential())
.GetChatClient(deploymentName);
AIAgent baseAgent = chatClient.AsIChatClient().AsAIAgent(
AIAgent baseAgent = chatClient.AsAIAgent(
name: "RecipeAgent",
instructions: """
You are a helpful recipe assistant. When users ask you to create or suggest a recipe,
@@ -10,7 +10,6 @@
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
@@ -0,0 +1,22 @@
<Project Sdk="Microsoft.NET.Sdk">
<PropertyGroup>
<OutputType>Exe</OutputType>
<TargetFrameworks>net10.0</TargetFrameworks>
<Nullable>enable</Nullable>
<ImplicitUsings>enable</ImplicitUsings>
</PropertyGroup>
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
<PackageReference Include="Microsoft.SemanticKernel.Connectors.InMemory" />
</ItemGroup>
<ItemGroup>
<ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
</ItemGroup>
</Project>
@@ -0,0 +1,133 @@
// Copyright (c) Microsoft. All rights reserved.
using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.VectorData;
namespace SampleApp;
/// <summary>
/// A <see cref="ChatHistoryProvider"/> that keeps a bounded window of recent messages in session state
/// (via <see cref="InMemoryChatHistoryProvider"/>) and overflows older messages to a vector store
/// (via <see cref="ChatHistoryMemoryProvider"/>). When providing chat history, it searches the vector
/// store for relevant older messages and prepends them as a memory context message.
/// </summary>
/// <remarks>
/// Only non-system messages are counted towards the session state limit and overflow mechanism. System messages are always retained in session state and are not included in the vector store.
/// Function calls and function results are also dropped when truncation happens, both from in-memory state, and they are also not persisted to the vector store.
/// </remarks>
internal sealed class BoundedChatHistoryProvider : ChatHistoryProvider, IDisposable
{
private readonly InMemoryChatHistoryProvider _chatHistoryProvider;
private readonly ChatHistoryMemoryProvider _memoryProvider;
private readonly TruncatingChatReducer _reducer;
private readonly string _contextPrompt;
private IReadOnlyList<string>? _stateKeys;
/// <summary>
/// Initializes a new instance of the <see cref="BoundedChatHistoryProvider"/> class.
/// </summary>
/// <param name="maxSessionMessages">The maximum number of non-system messages to keep in session state before overflowing to the vector store.</param>
/// <param name="vectorStore">The vector store to use for storing and retrieving overflow chat history.</param>
/// <param name="collectionName">The name of the collection for storing overflow chat history in the vector store.</param>
/// <param name="vectorDimensions">The number of dimensions to use for the chat history vector store embeddings.</param>
/// <param name="stateInitializer">A delegate that initializes the memory provider state, providing the storage and search scopes.</param>
/// <param name="contextPrompt">Optional prompt to prefix memory search results. Defaults to a standard memory context prompt.</param>
public BoundedChatHistoryProvider(
int maxSessionMessages,
VectorStore vectorStore,
string collectionName,
int vectorDimensions,
Func<AgentSession?, ChatHistoryMemoryProvider.State> stateInitializer,
string? contextPrompt = null)
{
if (maxSessionMessages < 0)
{
throw new ArgumentOutOfRangeException(nameof(maxSessionMessages), "maxSessionMessages must be non-negative.");
}
this._reducer = new TruncatingChatReducer(maxSessionMessages);
this._chatHistoryProvider = new InMemoryChatHistoryProvider(new InMemoryChatHistoryProviderOptions
{
ChatReducer = this._reducer,
ReducerTriggerEvent = InMemoryChatHistoryProviderOptions.ChatReducerTriggerEvent.AfterMessageAdded,
StorageInputRequestMessageFilter = msgs => msgs,
});
this._memoryProvider = new ChatHistoryMemoryProvider(
vectorStore,
collectionName,
vectorDimensions,
stateInitializer,
options: new ChatHistoryMemoryProviderOptions
{
SearchInputMessageFilter = msgs => msgs,
StorageInputRequestMessageFilter = msgs => msgs,
});
this._contextPrompt = contextPrompt
?? "The following are memories from earlier in this conversation. Use them to inform your responses:";
}
/// <inheritdoc />
public override IReadOnlyList<string> StateKeys => this._stateKeys ??= this._chatHistoryProvider.StateKeys.Concat(this._memoryProvider.StateKeys).ToArray();
/// <inheritdoc />
protected override async ValueTask<IEnumerable<ChatMessage>> ProvideChatHistoryAsync(
InvokingContext context,
CancellationToken cancellationToken = default)
{
// Delegate to the inner provider's full lifecycle (retrieve, filter, stamp, merge with request messages).
var chatHistoryProviderInputContext = new InvokingContext(context.Agent, context.Session, []);
var allMessages = await this._chatHistoryProvider.InvokingAsync(chatHistoryProviderInputContext, cancellationToken).ConfigureAwait(false);
// Search the vector store for relevant older messages.
var aiContext = new AIContext { Messages = context.RequestMessages.ToList() };
var invokingContext = new AIContextProvider.InvokingContext(
context.Agent, context.Session, aiContext);
var result = await this._memoryProvider.InvokingAsync(invokingContext, cancellationToken).ConfigureAwait(false);
// Extract only the messages added by the memory provider (stamped with AIContextProvider source type).
var memoryMessages = result.Messages?
.Where(m => m.GetAgentRequestMessageSourceType() == AgentRequestMessageSourceType.AIContextProvider)
.ToList();
if (memoryMessages is { Count: > 0 })
{
var memoryText = string.Join("\n", memoryMessages.Select(m => m.Text).Where(t => !string.IsNullOrWhiteSpace(t)));
if (!string.IsNullOrWhiteSpace(memoryText))
{
var contextMessage = new ChatMessage(ChatRole.User, $"{this._contextPrompt}\n{memoryText}");
return new[] { contextMessage }.Concat(allMessages);
}
}
return allMessages;
}
/// <inheritdoc />
protected override async ValueTask StoreChatHistoryAsync(
InvokedContext context,
CancellationToken cancellationToken = default)
{
// Delegate storage to the in-memory provider. Its TruncatingChatReducer (AfterMessageAdded trigger)
// will automatically truncate to the configured maximum and expose any removed messages.
var innerContext = new InvokedContext(
context.Agent, context.Session, context.RequestMessages, context.ResponseMessages!);
await this._chatHistoryProvider.InvokedAsync(innerContext, cancellationToken).ConfigureAwait(false);
// Archive any messages that the reducer removed to the vector store.
if (this._reducer.RemovedMessages is { Count: > 0 })
{
var overflowContext = new AIContextProvider.InvokedContext(
context.Agent, context.Session, this._reducer.RemovedMessages, []);
await this._memoryProvider.InvokedAsync(overflowContext, cancellationToken).ConfigureAwait(false);
}
}
/// <inheritdoc/>
public void Dispose()
{
this._memoryProvider.Dispose();
}
}
@@ -0,0 +1,79 @@
// Copyright (c) Microsoft. All rights reserved.
// This sample shows how to create a bounded chat history provider that keeps a configurable number of
// recent messages in session state and automatically overflows older messages to a vector store.
// When the agent is invoked, it searches the vector store for relevant older messages and
// prepends them as a "memory" context message before the recent session history.
using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.VectorData;
using Microsoft.SemanticKernel.Connectors.InMemory;
using OpenAI.Chat;
using SampleApp;
var endpoint = Environment.GetEnvironmentVariable("AZURE_OPENAI_ENDPOINT") ?? throw new InvalidOperationException("AZURE_OPENAI_ENDPOINT is not set.");
var deploymentName = Environment.GetEnvironmentVariable("AZURE_OPENAI_DEPLOYMENT_NAME") ?? "gpt-4o-mini";
var embeddingDeploymentName = Environment.GetEnvironmentVariable("AZURE_OPENAI_EMBEDDING_DEPLOYMENT_NAME") ?? "text-embedding-3-large";
// WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
// In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
// latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
var credential = new DefaultAzureCredential();
// Create a vector store to store overflow chat messages.
// For demonstration purposes, we are using an in-memory vector store.
// Replace this with a persistent vector store implementation for production scenarios.
VectorStore vectorStore = new InMemoryVectorStore(new InMemoryVectorStoreOptions()
{
EmbeddingGenerator = new AzureOpenAIClient(new Uri(endpoint), credential)
.GetEmbeddingClient(embeddingDeploymentName)
.AsIEmbeddingGenerator()
});
var sessionId = Guid.NewGuid().ToString();
// Create the BoundedChatHistoryProvider with a maximum of 4 non-system messages in session state.
// It internally creates an InMemoryChatHistoryProvider with a TruncatingChatReducer and a
// ChatHistoryMemoryProvider with the correct configuration to ensure overflow messages are
// automatically archived to the vector store and recalled via semantic search.
var boundedProvider = new BoundedChatHistoryProvider(
maxSessionMessages: 4,
vectorStore,
collectionName: "chathistory-overflow",
vectorDimensions: 3072,
session => new ChatHistoryMemoryProvider.State(
storageScope: new() { UserId = "UID1", SessionId = sessionId },
searchScope: new() { UserId = "UID1" }));
// Create the agent with the bounded chat history provider.
AIAgent agent = new AzureOpenAIClient(new Uri(endpoint), credential)
.GetChatClient(deploymentName)
.AsAIAgent(new ChatClientAgentOptions
{
ChatOptions = new() { Instructions = "You are a helpful assistant. Answer questions concisely." },
Name = "Assistant",
ChatHistoryProvider = boundedProvider,
});
// Start a conversation. The first several exchanges will fill up the session state window.
AgentSession session = await agent.CreateSessionAsync();
Console.WriteLine("--- Filling the session window (4 messages max) ---\n");
Console.WriteLine(await agent.RunAsync("My favorite color is blue.", session));
Console.WriteLine(await agent.RunAsync("I have a dog named Max.", session));
// At this point the session state holds 4 messages (2 user + 2 assistant).
// The next exchange will push the oldest messages into the vector store.
Console.WriteLine("\n--- Next exchange will trigger overflow to vector store ---\n");
Console.WriteLine(await agent.RunAsync("What is the capital of France?", session));
// The oldest messages about favorite color have now been archived to the vector store.
// Ask the agent something that requires recalling the overflowed information.
Console.WriteLine("\n--- Asking about overflowed information (should recall from vector store) ---\n");
Console.WriteLine(await agent.RunAsync("What is my favorite color?", session));
@@ -0,0 +1,40 @@
# Bounded Chat History with Vector Store Overflow
This sample demonstrates how to create a custom `ChatHistoryProvider` that keeps a bounded window of recent messages in session state and automatically overflows older messages to a vector store. When the agent is invoked, it searches the vector store for relevant older messages and prepends them as memory context.
## Concepts
- **`TruncatingChatReducer`**: A custom `IChatReducer` that keeps the most recent N messages and exposes removed messages via a `RemovedMessages` property.
- **`BoundedChatHistoryProvider`**: A custom `ChatHistoryProvider` that composes:
- `InMemoryChatHistoryProvider` for fast session-state storage (bounded by the reducer)
- `ChatHistoryMemoryProvider` for vector-store overflow and semantic search of older messages
## Prerequisites
- [.NET 10 SDK](https://dotnet.microsoft.com/download/dotnet/10.0)
- An Azure OpenAI resource with:
- A chat deployment (e.g., `gpt-4o-mini`)
- An embedding deployment (e.g., `text-embedding-3-large`)
## Configuration
Set the following environment variables:
| Variable | Description | Default |
|---|---|---|
| `AZURE_OPENAI_ENDPOINT` | Your Azure OpenAI endpoint URL | *(required)* |
| `AZURE_OPENAI_DEPLOYMENT_NAME` | Chat model deployment name | `gpt-4o-mini` |
| `AZURE_OPENAI_EMBEDDING_DEPLOYMENT_NAME` | Embedding model deployment name | `text-embedding-3-large` |
## Running the Sample
```bash
dotnet run
```
## How it Works
1. The agent starts a conversation with a bounded session window of 4 non-system, non-function messages (i.e., user/assistant turns). System messages are always preserved, and function call/result messages are truncated and not preserved.
2. As messages accumulate beyond the limit, the `TruncatingChatReducer` removes the oldest messages.
3. The `BoundedChatHistoryProvider` detects the removed messages and stores them in a vector store via `ChatHistoryMemoryProvider`.
4. On subsequent invocations, the provider searches the vector store for relevant older messages and prepends them as memory context, allowing the agent to recall information from earlier in the conversation.
@@ -0,0 +1,65 @@
// Copyright (c) Microsoft. All rights reserved.
using Microsoft.Extensions.AI;
namespace SampleApp;
/// <summary>
/// A truncating chat reducer that keeps the most recent messages up to a configured maximum,
/// preserving any leading system message. Removed messages are exposed via <see cref="RemovedMessages"/>
/// so that a caller can archive them (e.g. to a vector store).
/// </summary>
internal sealed class TruncatingChatReducer : IChatReducer
{
private readonly int _maxMessages;
/// <summary>
/// Initializes a new instance of the <see cref="TruncatingChatReducer"/> class.
/// </summary>
/// <param name="maxMessages">The maximum number of non-system messages to retain.</param>
public TruncatingChatReducer(int maxMessages)
{
this._maxMessages = maxMessages > 0 ? maxMessages : throw new ArgumentOutOfRangeException(nameof(maxMessages));
}
/// <summary>
/// Gets the messages that were removed during the most recent call to <see cref="ReduceAsync"/>.
/// </summary>
public IReadOnlyList<ChatMessage> RemovedMessages { get; private set; } = [];
/// <inheritdoc />
public Task<IEnumerable<ChatMessage>> ReduceAsync(IEnumerable<ChatMessage> messages, CancellationToken cancellationToken)
{
_ = messages ?? throw new ArgumentNullException(nameof(messages));
ChatMessage? systemMessage = null;
Queue<ChatMessage> retained = new(capacity: this._maxMessages);
List<ChatMessage> removed = [];
foreach (var message in messages)
{
if (message.Role == ChatRole.System)
{
// Preserve the first system message outside the counting window.
systemMessage ??= message;
}
else if (!message.Contents.Any(c => c is FunctionCallContent or FunctionResultContent))
{
if (retained.Count >= this._maxMessages)
{
removed.Add(retained.Dequeue());
}
retained.Enqueue(message);
}
}
this.RemovedMessages = removed;
IEnumerable<ChatMessage> result = systemMessage is not null
? new[] { systemMessage }.Concat(retained)
: retained;
return Task.FromResult(result);
}
}
@@ -8,5 +8,6 @@ These samples show how to create an agent with the Agent Framework that uses Mem
|[Memory with MemoryStore](./AgentWithMemory_Step02_MemoryUsingMem0/)|This sample demonstrates how to create and run an agent that uses the Mem0 service to extract and retrieve individual memories.|
|[Custom Memory Implementation](../../01-get-started/04_memory/)|This sample demonstrates how to create a custom memory component and attach it to an agent.|
|[Memory with Azure AI Foundry](./AgentWithMemory_Step04_MemoryUsingFoundry/)|This sample demonstrates how to create and run an agent that uses Azure AI Foundry's managed memory service to extract and retrieve individual memories.|
|[Bounded Chat History with Overflow](./AgentWithMemory_Step05_BoundedChatHistory/)|This sample demonstrates how to create a bounded chat history provider that overflows older messages to a vector store and recalls them as memories.|
> **See also**: [Memory Search with Foundry Agents](../FoundryAgents/FoundryAgents_Step22_MemorySearch/) - demonstrates using the built-in Memory Search tool with Azure Foundry Agents.
@@ -0,0 +1,21 @@
<Project Sdk="Microsoft.NET.Sdk">
<PropertyGroup>
<OutputType>Exe</OutputType>
<TargetFrameworks>net10.0</TargetFrameworks>
<Nullable>enable</Nullable>
<ImplicitUsings>enable</ImplicitUsings>
</PropertyGroup>
<ItemGroup>
<PackageReference Include="Azure.AI.OpenAI" />
<PackageReference Include="Azure.Identity" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
<ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
</ItemGroup>
</Project>
@@ -0,0 +1,120 @@
// Copyright (c) Microsoft. All rights reserved.
// This sample demonstrates how to use a CompactionProvider with a compaction pipeline
// as an AIContextProvider for an agent's in-run context management. The pipeline chains multiple
// compaction strategies from gentle to aggressive:
// 1. ToolResultCompactionStrategy - Collapses old tool-call groups into concise summaries
// 2. SummarizationCompactionStrategy - LLM-compresses older conversation spans
// 3. SlidingWindowCompactionStrategy - Keeps only the most recent N user turns
// 4. TruncationCompactionStrategy - Emergency token-budget backstop
using System.ComponentModel;
using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Compaction;
using Microsoft.Extensions.AI;
var endpoint = Environment.GetEnvironmentVariable("AZURE_OPENAI_ENDPOINT") ?? throw new InvalidOperationException("AZURE_OPENAI_ENDPOINT is not set.");
var deploymentName = Environment.GetEnvironmentVariable("AZURE_OPENAI_DEPLOYMENT_NAME") ?? "gpt-4o-mini";
// WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
// In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
// latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
AzureOpenAIClient openAIClient = new(new Uri(endpoint), new DefaultAzureCredential());
// Create a chat client for the agent and a separate one for the summarization strategy.
// Using the same model for simplicity; in production, use a smaller/cheaper model for summarization.
IChatClient agentChatClient = openAIClient.GetChatClient(deploymentName).AsIChatClient();
IChatClient summarizerChatClient = openAIClient.GetChatClient(deploymentName).AsIChatClient();
// Define a tool the agent can use, so we can see tool-result compaction in action.
[Description("Look up the current price of a product by name.")]
static string LookupPrice([Description("The product name to look up.")] string productName) =>
productName.ToUpperInvariant() switch
{
"LAPTOP" => "The laptop costs $999.99.",
"KEYBOARD" => "The keyboard costs $79.99.",
"MOUSE" => "The mouse costs $29.99.",
_ => $"Sorry, I don't have pricing for '{productName}'."
};
// Configure the compaction pipeline with one of each strategy, ordered least to most aggressive.
PipelineCompactionStrategy compactionPipeline =
new(// 1. Gentle: collapse old tool-call groups into short summaries
new ToolResultCompactionStrategy(CompactionTriggers.MessagesExceed(7)),
// 2. Moderate: use an LLM to summarize older conversation spans into a concise message
new SummarizationCompactionStrategy(summarizerChatClient, CompactionTriggers.TokensExceed(0x500)),
// 3. Aggressive: keep only the last N user turns and their responses
new SlidingWindowCompactionStrategy(CompactionTriggers.TurnsExceed(4)),
// 4. Emergency: drop oldest groups until under the token budget
new TruncationCompactionStrategy(CompactionTriggers.TokensExceed(0x8000)));
// Create the agent with a CompactionProvider that uses the compaction pipeline.
AIAgent agent =
agentChatClient
.AsBuilder()
// Note: Adding the CompactionProvider at the builder level means it will be applied to all agents
// built from this builder and will manage context for both agent messages and tool calls.
.UseAIContextProviders(new CompactionProvider(compactionPipeline))
.BuildAIAgent(
new ChatClientAgentOptions
{
Name = "ShoppingAssistant",
ChatOptions = new()
{
Instructions =
"""
You are a helpful, but long winded, shopping assistant.
Help the user look up prices and compare products.
When responding, Be sure to be extra descriptive and use as
many words as possible without sounding ridiculous.
""",
Tools = [AIFunctionFactory.Create(LookupPrice)]
},
// Note: AIContextProviders may be specified here instead of ChatClientBuilder.UseAIContextProviders.
// Specifying compaction at the agent level skips compaction in the function calling loop.
//AIContextProviders = [new CompactionProvider(compactionPipeline)]
});
AgentSession session = await agent.CreateSessionAsync();
// Helper to print chat history size
void PrintChatHistory()
{
if (session.TryGetInMemoryChatHistory(out var history))
{
Console.ForegroundColor = ConsoleColor.Cyan;
Console.WriteLine($"\n[Messages: #{history.Count}]\n");
Console.ResetColor();
}
}
// Run a multi-turn conversation with tool calls to exercise the pipeline.
string[] prompts =
[
"What's the price of a laptop?",
"How about a keyboard?",
"And a mouse?",
"Which product is the cheapest?",
"Can you compare the laptop and the keyboard for me?",
"What was the first product I asked about?",
"Thank you!",
];
foreach (string prompt in prompts)
{
Console.ForegroundColor = ConsoleColor.Cyan;
Console.Write("\n[User] ");
Console.ResetColor();
Console.WriteLine(prompt);
Console.ForegroundColor = ConsoleColor.Cyan;
Console.Write("\n[Agent] ");
Console.ResetColor();
Console.WriteLine(await agent.RunAsync(prompt, session));
PrintChatHistory();
}
@@ -0,0 +1,132 @@
# Compaction Pipeline
This sample demonstrates how to use a `CompactionProvider` with a `PipelineCompactionStrategy` to manage long conversation histories in a token-efficient way. The pipeline chains four compaction strategies, ordered from gentle to aggressive, so that the least disruptive strategy runs first and more aggressive strategies only activate when necessary.
## What This Sample Shows
- **`CompactionProvider`** — an `AIContextProvider` that applies a compaction strategy before each agent invocation, keeping only the most relevant messages within the model's context window
- **`PipelineCompactionStrategy`** — chains multiple compaction strategies into an ordered pipeline; each strategy evaluates its own trigger independently and operates on the output of the previous one
- **`ToolResultCompactionStrategy`** — collapses older tool-call groups into concise inline summaries, activated by a message-count trigger
- **`SummarizationCompactionStrategy`** — uses an LLM to compress older conversation spans into a single summary message, activated by a token-count trigger
- **`SlidingWindowCompactionStrategy`** — retains only the most recent N user turns and their responses, activated by a turn-count trigger
- **`TruncationCompactionStrategy`** — emergency backstop that drops the oldest groups until the conversation fits within a hard token budget
- **`CompactionTriggers`** — factory methods (`MessagesExceed`, `TokensExceed`, `TurnsExceed`, `GroupsExceed`, `HasToolCalls`, `All`, `Any`) that control when each strategy activates
## Concepts
### Message groups
The compaction engine organizes messages into atomic *groups* that are treated as indivisible units during compaction. A group is either:
| Group kind | Contents |
|---|---|
| `System` | System prompt message(s) |
| `User` | A single user message |
| `ToolCall` | One assistant message with tool calls + the matching tool result messages |
| `AssistantText` | A single assistant text-only message |
| `Summary` | One or more messages summarizing earlier conversation spans, produced by compaction strategies |
`Summary` groups (`CompactionGroupKind.Summary`) are created by compaction strategies (for example, `SummarizationCompactionStrategy`) and do not originate directly from user or assistant messages.
Strategies exclude entire groups rather than individual messages, preserving the tool-call/result pairing required by most model APIs.
### Compaction triggers
A `CompactionTrigger` is a predicate evaluated against the current `MessageIndex`. When the trigger fires, the strategy performs compaction; when it does not fire, the strategy is skipped. Available triggers are:
| Trigger | Activates when… |
|---|---|
| `CompactionTriggers.Always` | Always (unconditional) |
| `CompactionTriggers.Never` | Never (disabled) |
| `CompactionTriggers.MessagesExceed(n)` | Included message count > n |
| `CompactionTriggers.TokensExceed(n)` | Included token count > n |
| `CompactionTriggers.TurnsExceed(n)` | Included user-turn count > n |
| `CompactionTriggers.GroupsExceed(n)` | Included group count > n |
| `CompactionTriggers.HasToolCalls()` | At least one included tool-call group exists |
| `CompactionTriggers.All(...)` | All supplied triggers fire (logical AND) |
| `CompactionTriggers.Any(...)` | Any supplied trigger fires (logical OR) |
### Pipeline ordering
Order strategies from **least aggressive** to **most aggressive**. The pipeline runs every strategy whose trigger is met. Earlier strategies reduce the conversation gently so that later, more destructive strategies may not need to activate at all.
```
1. ToolResultCompactionStrategy – gentle: replaces verbose tool results with a short label
2. SummarizationCompactionStrategy – moderate: LLM-summarizes older turns
3. SlidingWindowCompactionStrategy – aggressive: drops turns beyond the window
4. TruncationCompactionStrategy – emergency: hard token-budget enforcement
```
## Prerequisites
- .NET 10 SDK or later
- Azure OpenAI service endpoint and model deployment
- Azure CLI installed and authenticated
**Note**: This sample uses `DefaultAzureCredential`. Sign in with `az login` before running. For production, prefer a specific credential such as `ManagedIdentityCredential`. For more information, see the [Azure CLI authentication documentation](https://learn.microsoft.com/cli/azure/authenticate-azure-cli-interactively).
## Environment Variables
```powershell
$env:AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/" # Required
$env:AZURE_OPENAI_DEPLOYMENT_NAME="gpt-4o-mini" # Optional, defaults to gpt-4o-mini
```
## Running the Sample
```powershell
cd dotnet/samples/02-agents/Agents/Agent_Step18_CompactionPipeline
dotnet run
```
## Expected Behavior
The sample runs a seven-turn shopping-assistant conversation with tool calls. After each turn it prints the full message count so you can observe the pipeline compaction doesn't alter the source conversation.
Each of the four compaction strategies has a deliberately low threshold so that it activates during the short demonstration conversation. In a production scenario you would raise the thresholds to match your model's context window and cost requirements.
## Customizing the Pipeline
### Using a single strategy
If you only need one compaction strategy, pass it directly to `CompactionProvider` without wrapping it in a pipeline:
```csharp
CompactionProvider provider =
new(new SlidingWindowCompactionStrategy(CompactionTriggers.TurnsExceed(20)));
```
### Ad-hoc compaction outside the provider pipeline
`CompactionProvider.CompactAsync` applies a strategy to an arbitrary list of messages without an active agent session:
```csharp
IEnumerable<ChatMessage> compacted = await CompactionProvider.CompactAsync(
new TruncationCompactionStrategy(CompactionTriggers.TokensExceed(8000)),
existingMessages);
```
### Using a different model for summarization
The `SummarizationCompactionStrategy` accepts any `IChatClient`. Use a smaller, cheaper model to reduce summarization cost:
```csharp
IChatClient summarizerChatClient = openAIClient.GetChatClient("gpt-4o-mini").AsIChatClient();
new SummarizationCompactionStrategy(summarizerChatClient, CompactionTriggers.TokensExceed(4000))
```
### Registering through `ChatClientAgentOptions`
`CompactionProvider` can also be specified directly on `ChatClientAgentOptions` instead of calling `UseAIContextProviders` on the `ChatClientBuilder`:
```csharp
AIAgent agent = agentChatClient
.AsBuilder()
.BuildAIAgent(new ChatClientAgentOptions
{
AIContextProviders = [new CompactionProvider(compactionPipeline)]
});
```
This places the compaction provider at the agent level instead of the chat client level, which allows you to use different compaction strategies for different agents that share the same chat client.
> Note: In this mode the `CompactionProvider` is not engaged during the tool calling loop. Agent-level `AIContextProviders` run before chat history is stored, so any synthetic summary messages produced by `CompactionProvider` can become part of the persisted history when using `ChatHistoryProvider`. If you want to compact only the request context while preserving the original stored history, register `CompactionProvider` on the `ChatClientBuilder` via `UseAIContextProviders(...)` instead of on `ChatClientAgentOptions`.
@@ -44,6 +44,7 @@ Before you begin, ensure you have the following prerequisites:
|[Deep research with an agent](./Agent_Step15_DeepResearch/)|This sample demonstrates how to use the Deep Research Tool to perform comprehensive research on complex topics|
|[Declarative agent](./Agent_Step16_Declarative/)|This sample demonstrates how to declaratively define an agent.|
|[Providing additional AI Context to an agent using multiple AIContextProviders](./Agent_Step17_AdditionalAIContext/)|This sample demonstrates how to inject additional AI context into a ChatClientAgent using multiple custom AIContextProvider components that are attached to the agent.|
|[Using compaction pipeline with an agent](./Agent_Step18_CompactionPipeline/)|This sample demonstrates how to use a compaction pipeline to efficiently limit the size of the conversation history for an agent.|
## Running the samples from the console
@@ -12,11 +12,8 @@ using OpenAI.Responses;
string endpoint = Environment.GetEnvironmentVariable("AZURE_AI_PROJECT_ENDPOINT") ?? throw new InvalidOperationException("AZURE_AI_PROJECT_ENDPOINT is not set.");
string deploymentName = Environment.GetEnvironmentVariable("AZURE_AI_MODEL_DEPLOYMENT_NAME") ?? "gpt-4o-mini";
// Memory store configuration
// NOTE: Memory stores must be created beforehand via Azure Portal or Python SDK.
// The .NET SDK currently only supports using existing memory stores with agents.
string memoryStoreName = Environment.GetEnvironmentVariable("AZURE_AI_MEMORY_STORE_ID") ?? throw new InvalidOperationException("AZURE_AI_MEMORY_STORE_ID is not set.");
string embeddingModelName = Environment.GetEnvironmentVariable("AZURE_AI_EMBEDDING_DEPLOYMENT_NAME") ?? "text-embedding-ada-002";
string memoryStoreName = Environment.GetEnvironmentVariable("AZURE_AI_MEMORY_STORE_ID") ?? $"foundry-memory-sample-{Guid.NewGuid():N}";
const string AgentInstructions = """
You are a helpful assistant that remembers past conversations.
@@ -32,71 +29,57 @@ const string AgentNameNative = "MemorySearchAgent-NATIVE";
string userScope = $"user_{Environment.MachineName}";
// Get a client to create/retrieve/delete server side agents with Azure Foundry Agents.
AIProjectClient aiProjectClient = new(new Uri(endpoint), new AzureCliCredential());
DefaultAzureCredential credential = new();
AIProjectClient aiProjectClient = new(new Uri(endpoint), credential);
// Ensure the memory store exists and has memories to retrieve.
await EnsureMemoryStoreAsync();
// Create the Memory Search tool configuration
MemorySearchPreviewTool memorySearchTool = new(memoryStoreName, userScope)
{
// Optional: Configure how quickly new memories are indexed (in seconds)
UpdateDelay = 1,
// Optional: Configure search behavior
SearchOptions = new MemorySearchToolOptions
{
// Additional search options can be configured here if needed
}
};
MemorySearchPreviewTool memorySearchTool = new(memoryStoreName, userScope) { UpdateDelay = 0 };
// Create agent using Option 1 (MEAI) or Option 2 (Native SDK)
AIAgent agent = await CreateAgentWithMEAI();
// AIAgent agent = await CreateAgentWithNativeSDK();
Console.WriteLine("Agent created with Memory Search tool. Starting conversation...\n");
// Conversation 1: Share some personal information
Console.WriteLine("User: My name is Alice and I love programming in C#.");
AgentResponse response1 = await agent.RunAsync("My name is Alice and I love programming in C#.");
Console.WriteLine($"Agent: {response1.Messages.LastOrDefault()?.Text}\n");
// Allow time for memory to be indexed
await Task.Delay(2000);
// Conversation 2: Test if the agent remembers
Console.WriteLine("User: What's my name and what programming language do I prefer?");
AgentResponse response2 = await agent.RunAsync("What's my name and what programming language do I prefer?");
Console.WriteLine($"Agent: {response2.Messages.LastOrDefault()?.Text}\n");
// Inspect memory search results if available in raw response items
// Note: Memory search tool call results appear as AgentResponseItem types
foreach (var message in response2.Messages)
try
{
if (message.RawRepresentation is AgentResponseItem agentResponseItem &&
agentResponseItem is MemorySearchToolCallResponseItem memorySearchResult)
{
Console.WriteLine($"Memory Search Status: {memorySearchResult.Status}");
Console.WriteLine($"Memory Search Results Count: {memorySearchResult.Results.Count}");
Console.WriteLine("Agent created with Memory Search tool. Starting conversation...\n");
foreach (var result in memorySearchResult.Results)
// The agent uses the memory search tool to recall stored information.
Console.WriteLine("User: What's my name and what programming language do I prefer?");
AgentResponse response = await agent.RunAsync("What's my name and what programming language do I prefer?");
Console.WriteLine($"Agent: {response.Messages.LastOrDefault()?.Text}\n");
// Inspect memory search results if available in raw response items.
foreach (var message in response.Messages)
{
if (message.RawRepresentation is MemorySearchToolCallResponseItem memorySearchResult)
{
var memoryItem = result.MemoryItem;
Console.WriteLine($" - Memory ID: {memoryItem.MemoryId}");
Console.WriteLine($" Scope: {memoryItem.Scope}");
Console.WriteLine($" Content: {memoryItem.Content}");
Console.WriteLine($" Updated: {memoryItem.UpdatedAt}");
Console.WriteLine($"Memory Search Status: {memorySearchResult.Status}");
Console.WriteLine($"Memory Search Results Count: {memorySearchResult.Results.Count}");
foreach (var result in memorySearchResult.Results)
{
var memoryItem = result.MemoryItem;
Console.WriteLine($" - Memory ID: {memoryItem.MemoryId}");
Console.WriteLine($" Scope: {memoryItem.Scope}");
Console.WriteLine($" Content: {memoryItem.Content}");
Console.WriteLine($" Updated: {memoryItem.UpdatedAt}");
}
}
}
}
finally
{
// Cleanup: Delete the agent and memory store.
Console.WriteLine("\nCleaning up...");
await aiProjectClient.Agents.DeleteAgentAsync(agent.Name);
Console.WriteLine("Agent deleted.");
await aiProjectClient.MemoryStores.DeleteMemoryStoreAsync(memoryStoreName);
Console.WriteLine("Memory store deleted.");
}
// Cleanup: Delete the agent (memory store persists and should be cleaned up separately if needed)
Console.WriteLine("\nCleaning up agent...");
await aiProjectClient.Agents.DeleteAgentAsync(agent.Name);
Console.WriteLine("Agent deleted successfully.");
// NOTE: Memory stores are long-lived resources and are NOT deleted with the agent.
// To delete a memory store, use the Azure Portal or Python SDK:
// await project_client.memory_stores.delete(memory_store.name)
// --- Agent Creation Options ---
#pragma warning disable CS8321 // Local function is declared but never used
// Option 1 - Using MemorySearchTool wrapped as MEAI AITool
@@ -122,3 +105,36 @@ async Task<AIAgent> CreateAgentWithNativeSDK()
})
);
}
// Helpers — kept at the bottom so the main agent flow above stays clean.
async Task EnsureMemoryStoreAsync()
{
Console.WriteLine($"Creating memory store '{memoryStoreName}'...");
try
{
await aiProjectClient.MemoryStores.GetMemoryStoreAsync(memoryStoreName);
Console.WriteLine("Memory store already exists.");
}
catch (System.ClientModel.ClientResultException ex) when (ex.Status == 404)
{
MemoryStoreDefaultDefinition definition = new(deploymentName, embeddingModelName);
await aiProjectClient.MemoryStores.CreateMemoryStoreAsync(memoryStoreName, definition, "Sample memory store for Memory Search demo");
Console.WriteLine("Memory store created.");
}
Console.WriteLine("Storing memories from a prior conversation...");
MemoryUpdateOptions memoryOptions = new(userScope) { UpdateDelay = 0 };
memoryOptions.Items.Add(ResponseItem.CreateUserMessageItem("My name is Alice and I love programming in C#."));
MemoryUpdateResult updateResult = await aiProjectClient.MemoryStores.WaitForMemoriesUpdateAsync(
memoryStoreName: memoryStoreName,
options: memoryOptions,
pollingInterval: 500);
if (updateResult.Status == MemoryStoreUpdateStatus.Failed)
{
throw new InvalidOperationException($"Memory update failed: {updateResult.ErrorDetails}");
}
Console.WriteLine($"Memory update completed (status: {updateResult.Status}).\n");
}
@@ -17,7 +17,7 @@ internal sealed class Tools(ILogger<Tools> logger)
[Description("Starts a content generation workflow and returns the instance ID for tracking.")]
public string StartContentGenerationWorkflow([Description("The topic for content generation")] string topic)
{
this._logger.LogInformation("Starting content generation workflow for topic: {Topic}", topic);
this._logger.LogInformation("Starting content generation workflow for topic: {Topic}", SanitizeLogValue(topic));
const int MaxReviewAttempts = 3;
const float ApprovalTimeoutHours = 72;
@@ -34,7 +34,7 @@ internal sealed class Tools(ILogger<Tools> logger)
this._logger.LogInformation(
"Content generation workflow scheduled to be started for topic '{Topic}' with instance ID: {InstanceId}",
topic,
SanitizeLogValue(topic),
instanceId);
return $"Workflow started with instance ID: {instanceId}";
@@ -45,7 +45,7 @@ internal sealed class Tools(ILogger<Tools> logger)
[Description("The instance ID of the workflow to check")] string instanceId,
[Description("Whether to include detailed information")] bool includeDetails = true)
{
this._logger.LogInformation("Getting status for workflow instance: {InstanceId}", instanceId);
this._logger.LogInformation("Getting status for workflow instance: {InstanceId}", SanitizeLogValue(instanceId));
// Get the current agent context using the session-static property
OrchestrationMetadata? status = await DurableAgentContext.Current.GetOrchestrationStatusAsync(
@@ -54,7 +54,7 @@ internal sealed class Tools(ILogger<Tools> logger)
if (status is null)
{
this._logger.LogInformation("Workflow instance '{InstanceId}' not found.", instanceId);
this._logger.LogInformation("Workflow instance '{InstanceId}' not found.", SanitizeLogValue(instanceId));
return new
{
instanceId,
@@ -78,7 +78,16 @@ internal sealed class Tools(ILogger<Tools> logger)
[Description("The instance ID of the workflow to submit feedback for")] string instanceId,
[Description("Feedback to submit")] HumanApprovalResponse feedback)
{
this._logger.LogInformation("Submitting human approval for workflow instance: {InstanceId}", instanceId);
this._logger.LogInformation("Submitting human approval for workflow instance: {InstanceId}", SanitizeLogValue(instanceId));
await DurableAgentContext.Current.RaiseOrchestrationEventAsync(instanceId, "HumanApproval", feedback);
}
/// <summary>
/// Sanitizes a user-provided value for safe inclusion in log entries
/// by removing control characters that could be used for log forging.
/// </summary>
private static string SanitizeLogValue(string value) =>
value
.Replace("\r", string.Empty, StringComparison.Ordinal)
.Replace("\n", string.Empty, StringComparison.Ordinal);
}
@@ -157,8 +157,8 @@ public sealed class FunctionTriggers
this._logger.LogInformation(
"Resuming stream for conversation {ConversationId} from cursor: {Cursor}",
conversationId,
cursor ?? "(beginning)");
SanitizeLogValue(conversationId),
SanitizeLogValue(cursor) ?? "(beginning)");
// Check Accept header to determine response format
// text/plain = raw text output (ideal for terminals)
@@ -205,7 +205,7 @@ public sealed class FunctionTriggers
{
if (chunk.Error != null)
{
this._logger.LogWarning("Stream error for conversation {ConversationId}: {Error}", conversationId, chunk.Error);
this._logger.LogWarning("Stream error for conversation {ConversationId}: {Error}", SanitizeLogValue(conversationId), chunk.Error);
await WriteErrorAsync(httpContext.Response, chunk.Error, useSseFormat, cancellationToken);
break;
}
@@ -224,7 +224,7 @@ public sealed class FunctionTriggers
}
catch (OperationCanceledException)
{
this._logger.LogInformation("Client disconnected from stream {ConversationId}", conversationId);
this._logger.LogInformation("Client disconnected from stream {ConversationId}", SanitizeLogValue(conversationId));
}
return new EmptyResult();
@@ -316,4 +316,20 @@ public sealed class FunctionTriggers
await response.WriteAsync(sb.ToString());
}
/// <summary>
/// Sanitizes a user-provided value for safe inclusion in log entries
/// by removing control characters that could be used for log forging.
/// </summary>
private static string? SanitizeLogValue(string? value)
{
if (value is null)
{
return null;
}
return value
.Replace("\r", string.Empty, StringComparison.Ordinal)
.Replace("\n", string.Empty, StringComparison.Ordinal);
}
}
@@ -10,7 +10,7 @@ using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;
using ChatClient = OpenAI.Chat.ChatClient;
using OpenAI.Chat;
namespace AGUIDojoServer;
@@ -36,7 +36,7 @@ internal static class ChatClientAgentFactory
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
return chatClient.AsIChatClient().AsAIAgent(
return chatClient.AsAIAgent(
name: "AgenticChat",
description: "A simple chat agent using Azure OpenAI");
}
@@ -45,7 +45,7 @@ internal static class ChatClientAgentFactory
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
return chatClient.AsIChatClient().AsAIAgent(
return chatClient.AsAIAgent(
name: "BackendToolRenderer",
description: "An agent that can render backend tools using Azure OpenAI",
tools: [AIFunctionFactory.Create(
@@ -59,7 +59,7 @@ internal static class ChatClientAgentFactory
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
return chatClient.AsIChatClient().AsAIAgent(
return chatClient.AsAIAgent(
name: "HumanInTheLoopAgent",
description: "An agent that involves human feedback in its decision-making process using Azure OpenAI");
}
@@ -68,7 +68,7 @@ internal static class ChatClientAgentFactory
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
return chatClient.AsIChatClient().AsAIAgent(
return chatClient.AsAIAgent(
name: "ToolBasedGenerativeUIAgent",
description: "An agent that uses tools to generate user interfaces using Azure OpenAI");
}
@@ -76,7 +76,7 @@ internal static class ChatClientAgentFactory
public static AIAgent CreateAgenticUI(JsonSerializerOptions options)
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
var baseAgent = chatClient.AsIChatClient().AsAIAgent(new ChatClientAgentOptions
var baseAgent = chatClient.AsAIAgent(new ChatClientAgentOptions
{
Name = "AgenticUIAgent",
Description = "An agent that generates agentic user interfaces using Azure OpenAI",
@@ -119,7 +119,7 @@ internal static class ChatClientAgentFactory
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
var baseAgent = chatClient.AsIChatClient().AsAIAgent(
var baseAgent = chatClient.AsAIAgent(
name: "SharedStateAgent",
description: "An agent that demonstrates shared state patterns using Azure OpenAI");
@@ -130,7 +130,7 @@ internal static class ChatClientAgentFactory
{
ChatClient chatClient = s_azureOpenAIClient!.GetChatClient(s_deploymentName!);
var baseAgent = chatClient.AsIChatClient().AsAIAgent(new ChatClientAgentOptions
var baseAgent = chatClient.AsAIAgent(new ChatClientAgentOptions
{
Name = "PredictiveStateUpdatesAgent",
Description = "An agent that demonstrates predictive state updates using Azure OpenAI",
@@ -74,7 +74,7 @@ AzureOpenAIClient azureOpenAIClient = new AzureOpenAIClient(
ChatClient chatClient = azureOpenAIClient.GetChatClient(deploymentName);
// Create AI agent
ChatClientAgent agent = chatClient.AsIChatClient().AsAIAgent(
ChatClientAgent agent = chatClient.AsAIAgent(
name: "ChatAssistant",
instructions: "You are a helpful assistant.");
@@ -162,7 +162,7 @@ dotnet run
Edit the instructions in `Server/Program.cs`:
```csharp
ChatClientAgent agent = chatClient.AsIChatClient().AsAIAgent(
ChatClientAgent agent = chatClient.AsAIAgent(
name: "ChatAssistant",
instructions: "You are a helpful coding assistant specializing in C# and .NET.");
```
@@ -6,7 +6,6 @@ using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using OpenAI.Chat;
WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
@@ -28,7 +27,7 @@ AzureOpenAIClient azureOpenAIClient = new(
ChatClient chatClient = azureOpenAIClient.GetChatClient(deploymentName);
ChatClientAgent agent = chatClient.AsIChatClient().AsAIAgent(
ChatClientAgent agent = chatClient.AsAIAgent(
name: "ChatAssistant",
instructions: "You are a helpful assistant.");
@@ -12,6 +12,7 @@ using Microsoft.AspNetCore.Authentication.JwtBearer;
using Microsoft.AspNetCore.Authorization;
using Microsoft.Extensions.AI;
using OpenAI;
using OpenAI.Chat;
WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
@@ -89,7 +90,6 @@ builder.Services.AddSingleton<AIAgent>(sp =>
return new OpenAIClient(apiKey)
.GetChatClient(model)
.AsIChatClient()
.AsAIAgent(
name: "ExpenseApprovalAgent",
instructions: "You are an expense approval assistant. You can list pending expenses "
@@ -10,7 +10,6 @@
<ItemGroup>
<PackageReference Include="Microsoft.AspNetCore.Authentication.JwtBearer" />
<PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
</ItemGroup>
<ItemGroup>
@@ -4,6 +4,9 @@
// In this case the OpenAI responses service will invoke any MCP tools as required. MCP tools are not invoked by the Agent Framework.
// The sample demonstrates how to use MCP tools with auto approval by setting ApprovalMode to NeverRequire.
#pragma warning disable MEAI001 // HostedMcpServerTool, HostedMcpServerToolApprovalMode are experimental
#pragma warning disable OPENAI001 // GetResponsesClient is experimental
using Azure.AI.AgentServer.AgentFramework.Extensions;
using Azure.AI.OpenAI;
using Azure.Identity;
@@ -3,6 +3,8 @@
// This sample demonstrates a multi-agent workflow with Writer and Reviewer agents
// using Azure AI Foundry AIProjectClient and the Agent Framework WorkflowBuilder.
#pragma warning disable CA2252 // AIProjectClient and Agents API require opting into preview features
using Azure.AI.AgentServer.AgentFramework.Extensions;
using Azure.AI.Projects;
using Azure.Identity;
@@ -4,6 +4,8 @@
// Uses Microsoft Agent Framework with Azure AI Foundry.
// Ready for deployment to Foundry Hosted Agent service.
#pragma warning disable CA2252 // AIProjectClient and Agents API require opting into preview features
using System.ComponentModel;
using System.Globalization;
using System.Text;
@@ -127,6 +127,7 @@ public sealed class A2AAgent : AIAgent
{
AgentId = this.Id,
ResponseId = message.MessageId,
FinishReason = ChatFinishReason.Stop,
RawRepresentation = message,
Messages = [message.ToChatMessage()],
AdditionalProperties = message.Metadata?.ToAdditionalProperties(),
@@ -141,6 +142,7 @@ public sealed class A2AAgent : AIAgent
{
AgentId = this.Id,
ResponseId = agentTask.Id,
FinishReason = MapTaskStateToFinishReason(agentTask.Status.State),
RawRepresentation = agentTask,
Messages = agentTask.ToChatMessages() ?? [],
ContinuationToken = CreateContinuationToken(agentTask.Id, agentTask.Status.State),
@@ -328,6 +330,7 @@ public sealed class A2AAgent : AIAgent
{
AgentId = this.Id,
ResponseId = message.MessageId,
FinishReason = ChatFinishReason.Stop,
RawRepresentation = message,
Role = ChatRole.Assistant,
MessageId = message.MessageId,
@@ -342,6 +345,7 @@ public sealed class A2AAgent : AIAgent
{
AgentId = this.Id,
ResponseId = task.Id,
FinishReason = MapTaskStateToFinishReason(task.Status.State),
RawRepresentation = task,
Role = ChatRole.Assistant,
Contents = task.ToAIContents(),
@@ -365,7 +369,16 @@ public sealed class A2AAgent : AIAgent
responseUpdate.Contents = artifactUpdateEvent.Artifact.ToAIContents();
responseUpdate.RawRepresentation = artifactUpdateEvent;
}
else if (taskUpdateEvent is TaskStatusUpdateEvent statusUpdateEvent)
{
responseUpdate.FinishReason = MapTaskStateToFinishReason(statusUpdateEvent.Status.State);
}
return responseUpdate;
}
private static ChatFinishReason? MapTaskStateToFinishReason(TaskState state)
{
return state == TaskState.Completed ? ChatFinishReason.Stop : null;
}
}
@@ -1,36 +0,0 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Text.Json;
using Microsoft.Extensions.AI;
namespace A2A;
/// <summary>
/// Extension methods for A2A metadata dictionary.
/// </summary>
internal static class A2AMetadataExtensions
{
/// <summary>
/// Converts a dictionary of metadata to an <see cref="AdditionalPropertiesDictionary"/>.
/// </summary>
/// <remarks>
/// This method can be replaced by the one from A2A SDK once it is public.
/// </remarks>
/// <param name="metadata">The metadata dictionary to convert.</param>
/// <returns>The converted <see cref="AdditionalPropertiesDictionary"/>, or null if the input is null or empty.</returns>
internal static AdditionalPropertiesDictionary? ToAdditionalProperties(this Dictionary<string, JsonElement>? metadata)
{
if (metadata is not { Count: > 0 })
{
return null;
}
var additionalProperties = new AdditionalPropertiesDictionary();
foreach (var kvp in metadata)
{
additionalProperties[kvp.Key] = kvp.Value;
}
return additionalProperties;
}
}
@@ -1,44 +0,0 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Text.Json;
using Microsoft.Agents.AI;
namespace Microsoft.Extensions.AI;
/// <summary>
/// Extension methods for AdditionalPropertiesDictionary.
/// </summary>
internal static class AdditionalPropertiesDictionaryExtensions
{
/// <summary>
/// Converts an <see cref="AdditionalPropertiesDictionary"/> to a dictionary of <see cref="JsonElement"/> values suitable for A2A metadata.
/// </summary>
/// <remarks>
/// This method can be replaced by the one from A2A SDK once it is available.
/// </remarks>
/// <param name="additionalProperties">The additional properties dictionary to convert, or <c>null</c>.</param>
/// <returns>A dictionary of JSON elements representing the metadata, or <c>null</c> if the input is null or empty.</returns>
internal static Dictionary<string, JsonElement>? ToA2AMetadata(this AdditionalPropertiesDictionary? additionalProperties)
{
if (additionalProperties is not { Count: > 0 })
{
return null;
}
var metadata = new Dictionary<string, JsonElement>();
foreach (var kvp in additionalProperties)
{
if (kvp.Value is JsonElement)
{
metadata[kvp.Key] = (JsonElement)kvp.Value!;
continue;
}
metadata[kvp.Key] = JsonSerializer.SerializeToElement(kvp.Value, A2AJsonUtilities.DefaultOptions.GetTypeInfo(typeof(object)));
}
return metadata;
}
}
@@ -20,6 +20,19 @@ namespace Microsoft.Agents.AI;
/// <see cref="AIAgent"/> serves as the foundational class for implementing AI agents that can participate in conversations
/// and process user requests. An agent instance may participate in multiple concurrent conversations, and each conversation
/// may involve multiple agents working together.
/// <para>
/// <strong>Security considerations:</strong> An <see cref="AIAgent"/> orchestrates data flow across trust boundaries —
/// messages are sent to external AI services, context providers, chat history stores, and function tools. Agent Framework
/// passes messages through as-is without validation or sanitization. Developers must be aware that:
/// <list type="bullet">
/// <item><description>User-supplied messages may contain prompt injection attempts designed to manipulate LLM behavior.</description></item>
/// <item><description>LLM responses should be treated as untrusted output — they may contain hallucinations, malicious payloads (e.g., scripts, SQL),
/// or content influenced by indirect prompt injection. Always validate and sanitize LLM output before rendering in HTML, executing as code,
/// or using in database queries.</description></item>
/// <item><description>Messages with different roles carry different trust levels: <c>system</c> messages have the highest trust and must be developer-controlled;
/// <c>user</c>, <c>assistant</c>, and <c>tool</c> messages should be treated as untrusted.</description></item>
/// </list>
/// </para>
/// </remarks>
[DebuggerDisplay("{DebuggerDisplay,nq}")]
public abstract partial class AIAgent
@@ -165,6 +178,11 @@ public abstract partial class AIAgent
/// This method enables saving conversation sessions to persistent storage,
/// allowing conversations to resume across application restarts or be migrated between
/// different agent instances. Use <see cref="DeserializeSessionAsync"/> to restore the session.
/// <para>
/// <strong>Security consideration:</strong> Serialized sessions may contain conversation content, session identifiers,
/// and other potentially sensitive data including PII. Ensure that serialized session data is stored securely with
/// appropriate access controls and encryption at rest.
/// </para>
/// </remarks>
public ValueTask<JsonElement> SerializeSessionAsync(AgentSession session, JsonSerializerOptions? jsonSerializerOptions = null, CancellationToken cancellationToken = default)
=> this.SerializeSessionCoreAsync(session, jsonSerializerOptions, cancellationToken);
@@ -194,6 +212,12 @@ public abstract partial class AIAgent
/// This method enables restoration of conversation sessions from previously saved state,
/// allowing conversations to resume across application restarts or be migrated between
/// different agent instances.
/// <para>
/// <strong>Security consideration:</strong> Restoring a session from an untrusted source is equivalent to accepting untrusted input.
/// Serialized sessions may contain conversation content, session identifiers, and potentially sensitive data. A compromised
/// storage backend could alter message roles to escalate trust, or inject adversarial content that influences LLM behavior.
/// Treat serialized session data as sensitive and ensure it is stored and transmitted securely.
/// </para>
/// </remarks>
public ValueTask<AgentSession> DeserializeSessionAsync(JsonElement serializedState, JsonSerializerOptions? jsonSerializerOptions = null, CancellationToken cancellationToken = default)
=> this.DeserializeSessionCoreAsync(serializedState, jsonSerializerOptions, cancellationToken);
@@ -301,6 +325,11 @@ public abstract partial class AIAgent
/// The messages are processed in the order provided and become part of the conversation history.
/// The agent's response will also be added to <paramref name="session"/> if one is provided.
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Agent Framework does not validate or sanitize message content — it is passed through
/// to the underlying AI service as-is. If input messages include untrusted user content, developers should be aware of prompt injection risks.
/// System-role messages must be developer-controlled and should never contain end-user input.
/// </para>
/// </remarks>
public Task<AgentResponse> RunAsync(
IEnumerable<ChatMessage> messages,
@@ -426,6 +455,11 @@ public abstract partial class AIAgent
/// Each <see cref="AgentResponseUpdate"/> represents a portion of the complete response, allowing consumers
/// to display partial results, implement progressive loading, or provide immediate feedback to users.
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Agent Framework does not validate or sanitize message content — it is passed through
/// to the underlying AI service as-is. If input messages include untrusted user content, developers should be aware of prompt injection risks.
/// System-role messages must be developer-controlled and should never contain end-user input.
/// </para>
/// </remarks>
public async IAsyncEnumerable<AgentResponseUpdate> RunStreamingAsync(
IEnumerable<ChatMessage> messages,
@@ -28,6 +28,14 @@ namespace Microsoft.Agents.AI;
/// <see cref="InvokingAsync"/> to provide context, and optionally called at the end of invocation via
/// <see cref="InvokedAsync"/> to process results.
/// </para>
/// <para>
/// <strong>Security considerations:</strong> Context providers may inject messages with any role, including <c>system</c>, which
/// has the highest trust level and directly shapes LLM behavior. Developers must ensure that all providers attached to an agent
/// are trusted. Agent Framework does not validate or filter the data returned by providers — it is accepted as-is and merged into
/// the request context. If a provider retrieves data from an external source (e.g., a vector database or memory service), be aware
/// that a compromised data source could introduce adversarial content designed to manipulate LLM behavior via indirect prompt injection.
/// Implementers should validate and sanitize data retrieved from external sources before returning it.
/// </para>
/// </remarks>
public abstract class AIContextProvider
{
@@ -96,6 +104,11 @@ public abstract class AIContextProvider
/// <item><description>Injecting contextual messages from conversation history</description></item>
/// </list>
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Data retrieved from external sources (e.g., vector databases, memory services, or
/// knowledge bases) may contain adversarial content designed to influence LLM behavior via indirect prompt injection.
/// Implementers should validate data integrity and consider the trustworthiness of the data source.
/// </para>
/// </remarks>
public ValueTask<AIContext> InvokingAsync(InvokingContext context, CancellationToken cancellationToken = default)
=> this.InvokingCoreAsync(Throw.IfNull(context), cancellationToken);
@@ -195,6 +208,11 @@ public abstract class AIContextProvider
/// In contrast with <see cref="InvokingCoreAsync"/>, this method only returns additional context to be merged with the input,
/// while <see cref="InvokingCoreAsync"/> is responsible for returning the full merged <see cref="AIContext"/> for the invocation.
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Any messages, tools, or instructions returned by this method will be merged into the
/// AI request context. If data is retrieved from external or untrusted sources, implementers should validate and sanitize it
/// to prevent indirect prompt injection attacks.
/// </para>
/// </remarks>
/// <param name="context">Contains the request context including the caller provided messages that will be used by the agent for this invocation.</param>
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests. The default is <see cref="CancellationToken.None"/>.</param>
@@ -299,6 +317,10 @@ public abstract class AIContextProvider
/// <para>
/// The default implementation of <see cref="InvokedCoreAsync"/> only calls this method if the invocation succeeded.
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Messages being processed/stored may contain PII and sensitive conversation content.
/// Implementers should ensure appropriate encryption at rest and access controls for the storage backend.
/// </para>
/// </remarks>
protected virtual ValueTask StoreAIContextAsync(InvokedContext context, CancellationToken cancellationToken = default) =>
default;
@@ -61,6 +61,7 @@ public class AgentResponse
this.AdditionalProperties = response.AdditionalProperties;
this.CreatedAt = response.CreatedAt;
this.FinishReason = response.FinishReason;
this.Messages = response.Messages;
this.RawRepresentation = response;
this.ResponseId = response.ResponseId;
@@ -84,6 +85,7 @@ public class AgentResponse
this.AdditionalProperties = response.AdditionalProperties;
this.CreatedAt = response.CreatedAt;
this.FinishReason = response.FinishReason;
this.Messages = response.Messages;
this.RawRepresentation = response;
this.ResponseId = response.ResponseId;
@@ -190,6 +192,21 @@ public class AgentResponse
/// </remarks>
public DateTimeOffset? CreatedAt { get; set; }
/// <summary>
/// Gets or sets the reason for the agent response finishing.
/// </summary>
/// <value>
/// A <see cref="ChatFinishReason"/> value indicating why the response finished (e.g., stop, length, content filter, tool calls),
/// or <see langword="null"/> if the finish reason is not available.
/// </value>
/// <remarks>
/// <para>
/// This property is particularly useful for detecting non-normal completions, such as content filtering
/// or token limit truncation, which may require special handling by the caller.
/// </para>
/// </remarks>
public ChatFinishReason? FinishReason { get; set; }
/// <summary>
/// Gets or sets the resource usage information for generating this response.
/// </summary>
@@ -276,6 +293,7 @@ public class AgentResponse
RawRepresentation = message.RawRepresentation,
Role = message.Role,
FinishReason = this.FinishReason,
AgentId = this.AgentId,
ResponseId = this.ResponseId,
MessageId = message.MessageId,
@@ -38,6 +38,7 @@ public static class AgentResponseExtensions
{
AdditionalProperties = response.AdditionalProperties,
CreatedAt = response.CreatedAt,
FinishReason = response.FinishReason,
Messages = response.Messages,
RawRepresentation = response,
ResponseId = response.ResponseId,
@@ -71,6 +72,7 @@ public static class AgentResponseExtensions
AuthorName = responseUpdate.AuthorName,
Contents = responseUpdate.Contents,
CreatedAt = responseUpdate.CreatedAt,
FinishReason = responseUpdate.FinishReason,
MessageId = responseUpdate.MessageId,
RawRepresentation = responseUpdate,
ResponseId = responseUpdate.ResponseId,
@@ -70,6 +70,7 @@ public class AgentResponseUpdate
this.AuthorName = chatResponseUpdate.AuthorName;
this.Contents = chatResponseUpdate.Contents;
this.CreatedAt = chatResponseUpdate.CreatedAt;
this.FinishReason = chatResponseUpdate.FinishReason;
this.MessageId = chatResponseUpdate.MessageId;
this.RawRepresentation = chatResponseUpdate;
this.ResponseId = chatResponseUpdate.ResponseId;
@@ -153,6 +154,15 @@ public class AgentResponseUpdate
/// </remarks>
public ResponseContinuationToken? ContinuationToken { get; set; }
/// <summary>
/// Gets or sets the reason for the agent response finishing.
/// </summary>
/// <value>
/// A <see cref="ChatFinishReason"/> value indicating why the response finished (e.g., stop, length, content filter, tool calls),
/// or <see langword="null"/> if the finish reason is not available or not yet determined (mid-stream).
/// </value>
public ChatFinishReason? FinishReason { get; set; }
/// <inheritdoc/>
public override string ToString() => this.Text;
@@ -42,6 +42,15 @@ namespace Microsoft.Agents.AI;
/// <see cref="JsonElement"/> and the <see cref="AIAgent.DeserializeSessionAsync(JsonElement, JsonSerializerOptions?, System.Threading.CancellationToken)"/> method
/// can be used to deserialize the session.
/// </para>
/// <para>
/// <strong>Security considerations:</strong> Serialized sessions may contain conversation content, session identifiers,
/// and other potentially sensitive data including PII. Developers should:
/// <list type="bullet">
/// <item><description>Treat serialized session data as sensitive and store it securely with appropriate access controls and encryption at rest.</description></item>
/// <item><description>Treat restoring a session from an untrusted source as equivalent to accepting untrusted input. A compromised storage backend
/// could alter message roles to escalate trust, or inject adversarial content that influences LLM behavior.</description></item>
/// </list>
/// </para>
/// </remarks>
/// <seealso cref="AIAgent"/>
/// <seealso cref="AIAgent.CreateSessionAsync(System.Threading.CancellationToken)"/>
@@ -67,6 +76,11 @@ public abstract class AgentSession
/// <summary>
/// Gets any arbitrary state associated with this session.
/// </summary>
/// <remarks>
/// Data stored in the <see cref="StateBag"/> will be included when the session is serialized.
/// Avoid storing secrets, credentials, or highly sensitive data in the state bag without appropriate encryption,
/// as this data may be persisted to external storage.
/// </remarks>
[JsonPropertyName("stateBag")]
public AgentSessionStateBag StateBag { get; protected set; } = new();
@@ -37,6 +37,14 @@ namespace Microsoft.Agents.AI;
/// A <see cref="ChatHistoryProvider"/> is only relevant for scenarios where the underlying AI service that the agent is using
/// does not use in-service chat history storage.
/// </para>
/// <para>
/// <strong>Security considerations:</strong> Agent Framework does not validate or filter the messages returned by the provider
/// during load — they are accepted as-is and treated identically to user-supplied messages. Implementers must ensure that only
/// trusted data is returned. If the underlying storage is compromised, adversarial content could influence LLM behavior via
/// indirect prompt injection — for example, injected messages could alter the conversation context or impersonate different roles.
/// Messages stored in chat history may contain PII and sensitive conversation content; implementers should consider encryption
/// at rest and appropriate access controls for the storage backend.
/// </para>
/// </remarks>
public abstract class ChatHistoryProvider
{
@@ -159,6 +167,11 @@ public abstract class ChatHistoryProvider
/// Messages are returned in chronological order to maintain proper conversation flow and context for the agent.
/// The oldest messages appear first in the collection, followed by more recent messages.
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Messages loaded from storage should be treated with the same caution as user-supplied
/// messages. A compromised storage backend could alter message roles to escalate trust (e.g., changing <c>user</c> messages to
/// <c>system</c> messages) or inject adversarial content that influences LLM behavior.
/// </para>
/// </remarks>
/// <param name="context">Contains the request context including the caller provided messages that will be used by the agent for this invocation.</param>
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests. The default is <see cref="CancellationToken.None"/>.</param>
@@ -273,6 +286,10 @@ public abstract class ChatHistoryProvider
/// <para>
/// The default implementation of <see cref="InvokedCoreAsync"/> only calls this method if the invocation succeeded.
/// </para>
/// <para>
/// <strong>Security consideration:</strong> Messages being stored may contain PII and sensitive conversation content.
/// Implementers should ensure appropriate encryption at rest and access controls for the storage backend.
/// </para>
/// </remarks>
protected virtual ValueTask StoreChatHistoryAsync(InvokedContext context, CancellationToken cancellationToken = default) =>
default;
@@ -79,20 +79,21 @@ public sealed class InMemoryChatHistoryProvider : ChatHistoryProvider
/// <exception cref="ArgumentNullException"><paramref name="messages"/> is <see langword="null"/>.</exception>
public void SetMessages(AgentSession? session, List<ChatMessage> messages)
{
_ = Throw.IfNull(messages);
Throw.IfNull(messages);
var state = this._sessionState.GetOrInitializeState(session);
State state = this._sessionState.GetOrInitializeState(session);
state.Messages = messages;
}
/// <inheritdoc />
protected override async ValueTask<IEnumerable<ChatMessage>> ProvideChatHistoryAsync(InvokingContext context, CancellationToken cancellationToken = default)
{
var state = this._sessionState.GetOrInitializeState(context.Session);
State state = this._sessionState.GetOrInitializeState(context.Session);
if (this.ReducerTriggerEvent is InMemoryChatHistoryProviderOptions.ChatReducerTriggerEvent.BeforeMessagesRetrieval && this.ChatReducer is not null)
{
state.Messages = (await this.ChatReducer.ReduceAsync(state.Messages, cancellationToken).ConfigureAwait(false)).ToList();
// Apply pre-retrieval reduction if configured
await ReduceMessagesAsync(this.ChatReducer, state, cancellationToken).ConfigureAwait(false);
}
return state.Messages;
@@ -101,7 +102,7 @@ public sealed class InMemoryChatHistoryProvider : ChatHistoryProvider
/// <inheritdoc />
protected override async ValueTask StoreChatHistoryAsync(InvokedContext context, CancellationToken cancellationToken = default)
{
var state = this._sessionState.GetOrInitializeState(context.Session);
State state = this._sessionState.GetOrInitializeState(context.Session);
// Add request and response messages to the provider
var allNewMessages = context.RequestMessages.Concat(context.ResponseMessages ?? []);
@@ -109,10 +110,16 @@ public sealed class InMemoryChatHistoryProvider : ChatHistoryProvider
if (this.ReducerTriggerEvent is InMemoryChatHistoryProviderOptions.ChatReducerTriggerEvent.AfterMessageAdded && this.ChatReducer is not null)
{
state.Messages = (await this.ChatReducer.ReduceAsync(state.Messages, cancellationToken).ConfigureAwait(false)).ToList();
// Apply pre-write reduction strategy if configured
await ReduceMessagesAsync(this.ChatReducer, state, cancellationToken).ConfigureAwait(false);
}
}
private static async Task ReduceMessagesAsync(IChatReducer reducer, State state, CancellationToken cancellationToken = default)
{
state.Messages = [.. await reducer.ReduceAsync(state.Messages, cancellationToken).ConfigureAwait(false)];
}
/// <summary>
/// Represents the state of a <see cref="InMemoryChatHistoryProvider"/> stored in the <see cref="AgentSession.StateBag"/>.
/// </summary>
@@ -17,6 +17,24 @@ namespace Microsoft.Agents.AI;
/// <summary>
/// Provides a Cosmos DB implementation of the <see cref="ChatHistoryProvider"/> abstract class.
/// </summary>
/// <remarks>
/// <para>
/// <strong>Security considerations:</strong>
/// <list type="bullet">
/// <item><description><strong>PII and sensitive data:</strong> Chat history stored in Cosmos DB may contain PII, sensitive conversation
/// content, and system instructions. Ensure the Cosmos DB account is configured with appropriate access controls, encryption at rest,
/// and network security (e.g., private endpoints, virtual network rules). The <see cref="MessageTtlSeconds"/> property can be used to
/// automatically expire messages and limit data retention.</description></item>
/// <item><description><strong>Compromised store risks:</strong> Agent Framework does not validate or filter messages loaded from the
/// store — they are accepted as-is. If the Cosmos DB store is compromised, adversarial content could be injected into the conversation
/// context, potentially influencing LLM behavior via indirect prompt injection. Altered message roles (e.g., changing <c>user</c> to
/// <c>system</c>) could escalate trust levels.</description></item>
/// <item><description><strong>Authentication:</strong> Agent Framework does not manage authentication or encryption for the Cosmos DB
/// connection — these are the responsibility of the <see cref="CosmosClient"/> configuration. Use managed identity
/// or token-based authentication where possible, and avoid embedding connection strings with keys in source code.</description></item>
/// </list>
/// </para>
/// </remarks>
[RequiresUnreferencedCode("The CosmosChatHistoryProvider uses JSON serialization which is incompatible with trimming.")]
[RequiresDynamicCode("The CosmosChatHistoryProvider uses JSON serialization which is incompatible with NativeAOT.")]
public sealed class CosmosChatHistoryProvider : ChatHistoryProvider, IDisposable
@@ -4,6 +4,12 @@
### Changed
- Filter empty `AIContent` from durable agent state responses ([#4670](https://github.com/microsoft/agent-framework/pull/4670))
## v1.0.0-preview.260311.1
### Changed
- Added TTL configuration for durable agent entities ([#2679](https://github.com/microsoft/agent-framework/pull/2679))
- Switch to new "Run" method name ([#2843](https://github.com/microsoft/agent-framework/pull/2843))
- Removed AgentThreadMetadata and used AgentSessionId directly instead ([#3067](https://github.com/microsoft/agent-framework/pull/3067));
@@ -16,6 +22,8 @@
- Marked all `RunAsync<T>` overloads as `new`, added missing ones, and added support for primitives and arrays ([#3803](https://github.com/microsoft/agent-framework/pull/3803))
- Improve session cast error message quality and consistency ([#3973](https://github.com/microsoft/agent-framework/pull/3973))
NOTE: Some of the above changes may have been part of earlier releases not mentioned in this file.
## v1.0.0-preview.251204.1
- Added orchestration ID to durable agent entity state ([#2137](https://github.com/microsoft/agent-framework/pull/2137))
@@ -1,6 +1,7 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Text.Json.Serialization;
using Microsoft.Extensions.AI;
namespace Microsoft.Agents.AI.DurableTask.State;
@@ -28,7 +29,10 @@ internal sealed class DurableAgentStateResponse : DurableAgentStateEntry
{
CorrelationId = correlationId,
CreatedAt = response.CreatedAt ?? response.Messages.Max(m => m.CreatedAt) ?? DateTimeOffset.UtcNow,
Messages = response.Messages.Select(DurableAgentStateMessage.FromChatMessage).ToList(),
Messages = response.Messages
.Where(HasSerializableContent)
.Select(DurableAgentStateMessage.FromChatMessage)
.ToList(),
Usage = DurableAgentStateUsage.FromUsage(response.Usage)
};
}
@@ -46,4 +50,18 @@ internal sealed class DurableAgentStateResponse : DurableAgentStateEntry
Usage = this.Usage?.ToUsageDetails(),
};
}
// Checks whether a ChatMessage has any content that will produce meaningful serialized data.
// Known derived AIContent types (TextContent, FunctionCallContent, etc.) are always serializable.
// Base AIContent instances only carry RawRepresentation (which is [JsonIgnore]), Annotations, and
// AdditionalProperties. We keep the message if any base AIContent has annotations or additional
// properties set. NOTE: if AIContent gains new serializable properties in the future, this check
// should be updated accordingly.
private static bool HasSerializableContent(ChatMessage message)
{
return message.Contents.Any(c =>
c.GetType() != typeof(AIContent) ||
c.Annotations?.Count > 0 ||
c.AdditionalProperties?.Count > 0);
}
}
@@ -13,10 +13,6 @@
</PropertyGroup>
<Import Project="$(RepoRoot)/dotnet/nuget/nuget-package.props" />
<PropertyGroup>
<!-- Disable packing until we are ready to release this as a nuget -->
<IsPackable>false</IsPackable>
</PropertyGroup>
<ItemGroup>
<ProjectReference Include="..\Microsoft.Agents.AI.Abstractions\Microsoft.Agents.AI.Abstractions.csproj" />
@@ -1,36 +0,0 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Text.Json;
using Microsoft.Extensions.AI;
namespace Microsoft.Agents.AI.Hosting.A2A.Converters;
/// <summary>
/// Extension methods for A2A metadata dictionary.
/// </summary>
internal static class A2AMetadataExtensions
{
/// <summary>
/// Converts a dictionary of metadata to an <see cref="AdditionalPropertiesDictionary"/>.
/// </summary>
/// <remarks>
/// This method can be replaced by the one from A2A SDK once it is public.
/// </remarks>
/// <param name="metadata">The metadata dictionary to convert.</param>
/// <returns>The converted <see cref="AdditionalPropertiesDictionary"/>, or null if the input is null or empty.</returns>
internal static AdditionalPropertiesDictionary? ToAdditionalProperties(this Dictionary<string, JsonElement>? metadata)
{
if (metadata is not { Count: > 0 })
{
return null;
}
var additionalProperties = new AdditionalPropertiesDictionary();
foreach (var kvp in metadata)
{
additionalProperties[kvp.Key] = kvp.Value;
}
return additionalProperties;
}
}
@@ -1,44 +0,0 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Text.Json;
using Microsoft.Extensions.AI;
namespace Microsoft.Agents.AI.Hosting.A2A.Converters;
/// <summary>
/// Extension methods for AdditionalPropertiesDictionary.
/// </summary>
internal static class AdditionalPropertiesDictionaryExtensions
{
/// <summary>
/// Converts an <see cref="AdditionalPropertiesDictionary"/> to a dictionary of <see cref="JsonElement"/> values suitable for A2A metadata.
/// </summary>
/// <remarks>
/// This method can be replaced by the one from A2A SDK once it is available.
/// </remarks>
/// <param name="additionalProperties">The additional properties dictionary to convert, or <c>null</c>.</param>
/// <returns>A dictionary of JSON elements representing the metadata, or <c>null</c> if the input is null or empty.</returns>
internal static Dictionary<string, JsonElement>? ToA2AMetadata(this AdditionalPropertiesDictionary? additionalProperties)
{
if (additionalProperties is not { Count: > 0 })
{
return null;
}
var metadata = new Dictionary<string, JsonElement>();
foreach (var kvp in additionalProperties)
{
if (kvp.Value is JsonElement)
{
metadata[kvp.Key] = (JsonElement)kvp.Value!;
continue;
}
metadata[kvp.Key] = JsonSerializer.SerializeToElement(kvp.Value, A2AHostingJsonUtilities.DefaultOptions.GetTypeInfo(typeof(object)));
}
return metadata;
}
}
@@ -72,9 +72,7 @@ internal static class AIAgentChatCompletionsProcessor
await foreach (var agentResponseUpdate in agent.RunStreamingAsync(chatMessages, options: options, cancellationToken: cancellationToken).WithCancellation(cancellationToken))
{
var finishReason = (agentResponseUpdate.RawRepresentation is ChatResponseUpdate { FinishReason: not null } chatResponseUpdate)
? chatResponseUpdate.FinishReason.ToString()
: "stop";
var finishReason = agentResponseUpdate.FinishReason?.ToString() ?? "stop";
var choiceChunks = new List<ChatCompletionChoiceChunk>();
CompletionUsage? usageDetails = null;
@@ -34,9 +34,7 @@ internal static class AgentResponseExtensions
var chatCompletionChoices = new List<ChatCompletionChoice>();
var index = 0;
var finishReason = (agentResponse.RawRepresentation is ChatResponse { FinishReason: not null } chatResponse)
? chatResponse.FinishReason.ToString()
: "stop"; // "stop" is a natural stop point; returning this by-default
var finishReason = agentResponse.FinishReason?.ToString() ?? ChatFinishReason.Stop.Value; // "stop" is a natural stop point; returning this by-default
foreach (var message in agentResponse.Messages)
{
@@ -13,16 +13,38 @@ using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Mem0;
#pragma warning disable IDE0001 // Simplify Names - Microsoft.Extensions.Logging.LogLevel.Trace doesn't get found in net472 when removing the namespace.
/// <summary>
/// Provides a Mem0 backed <see cref="MessageAIContextProvider"/> that persists conversation messages as memories
/// and retrieves related memories to augment the agent invocation context.
/// </summary>
/// <remarks>
/// <para>
/// The provider stores user, assistant and system messages as Mem0 memories and retrieves relevant memories
/// for new invocations using a semantic search endpoint. Retrieved memories are injected as user messages
/// to the model, prefixed by a configurable context prompt.
/// </para>
/// <para>
/// <strong>Security considerations:</strong>
/// <list type="bullet">
/// <item><description><strong>External service trust:</strong> This provider communicates with an external Mem0 service over HTTP.
/// Agent Framework does not manage authentication, encryption, or connection details for this service — these are the responsibility
/// of the <see cref="HttpClient"/> configuration. Ensure the HTTP client is configured with appropriate authentication
/// and uses HTTPS to protect data in transit.</description></item>
/// <item><description><strong>PII and sensitive data:</strong> Conversation messages (including user inputs, LLM responses, and system
/// instructions) are sent to the external Mem0 service for storage. These messages may contain PII or sensitive information.
/// Ensure the Mem0 service is configured with appropriate data retention policies and access controls.</description></item>
/// <item><description><strong>Indirect prompt injection:</strong> Memories retrieved from the Mem0 service are injected into the LLM
/// context as user messages. If the memory store is compromised, adversarial content could influence LLM behavior. The data
/// returned from the service is accepted as-is without validation or sanitization.</description></item>
/// <item><description><strong>Trace logging:</strong> When <see cref="Microsoft.Extensions.Logging.LogLevel.Trace"/> is enabled,
/// full memory content (including search queries and results) may be logged. This data may contain PII and should not be enabled
/// in production environments.</description></item>
/// </list>
/// </para>
/// </remarks>
public sealed class Mem0Provider : MessageAIContextProvider
#pragma warning restore IDE0001 // Simplify Names
{
private const string DefaultContextPrompt = "## Memories\nConsider the following memories when answering user questions:";
@@ -100,15 +100,23 @@ public static class OpenAIResponseClientExtensions
/// This corresponds to setting the "store" property in the JSON representation to false.
/// </remarks>
/// <param name="responseClient">The client.</param>
/// <param name="includeReasoningEncryptedContent">
/// Includes an encrypted version of reasoning tokens in reasoning item outputs.
/// This enables reasoning items to be used in multi-turn conversations when using the Responses API statelessly
/// (like when the store parameter is set to false, or when an organization is enrolled in the zero data retention program).
/// Defaults to <see langword="true"/>.
/// </param>
/// <returns>An <see cref="IChatClient"/> that can be used to converse via the <see cref="ResponsesClient"/> that does not store responses for later retrieval.</returns>
/// <exception cref="ArgumentNullException"><paramref name="responseClient"/> is <see langword="null"/>.</exception>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public static IChatClient AsIChatClientWithStoredOutputDisabled(this ResponsesClient responseClient)
public static IChatClient AsIChatClientWithStoredOutputDisabled(this ResponsesClient responseClient, bool includeReasoningEncryptedContent = true)
{
return Throw.IfNull(responseClient)
.AsIChatClient()
.AsBuilder()
.ConfigureOptions(x => x.RawRepresentationFactory = _ => new CreateResponseOptions() { StoredOutputEnabled = false })
.ConfigureOptions(x => x.RawRepresentationFactory = _ => includeReasoningEncryptedContent
? new CreateResponseOptions() { StoredOutputEnabled = false, IncludedProperties = { IncludedResponseProperty.ReasoningEncryptedContent } }
: new CreateResponseOptions() { StoredOutputEnabled = false })
.Build();
}
}
@@ -5,6 +5,7 @@ using System.Collections.Generic;
using System.Globalization;
using System.Linq;
using System.Net.Http;
using System.Text;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.AI;
@@ -222,31 +223,36 @@ public sealed class DefaultMcpToolHandler : IMcpToolHandler, IAsyncDisposable
}
}
private static AIContent ConvertContentBlock(ContentBlock block)
internal static AIContent ConvertContentBlock(ContentBlock block)
{
return block switch
{
TextContentBlock text => new TextContent(text.Text),
ImageContentBlock image => CreateDataContentFromBase64(image.Data, image.MimeType ?? "image/*"),
AudioContentBlock audio => CreateDataContentFromBase64(audio.Data, audio.MimeType ?? "audio/*"),
ImageContentBlock image => CreateDataContent(image.Data, image.MimeType ?? "image/*"),
AudioContentBlock audio => CreateDataContent(audio.Data, audio.MimeType ?? "audio/*"),
_ => new TextContent(block.ToString() ?? string.Empty),
};
}
private static DataContent CreateDataContentFromBase64(string? base64Data, string mediaType)
private static DataContent CreateDataContent(ReadOnlyMemory<byte> base64Utf8Data, string mediaType)
{
if (string.IsNullOrEmpty(base64Data))
if (base64Utf8Data.IsEmpty)
{
return new DataContent($"data:{mediaType};base64,", mediaType);
}
#if NET8_0_OR_GREATER
string base64 = Encoding.UTF8.GetString(base64Utf8Data.Span);
#else
string base64 = Encoding.UTF8.GetString(base64Utf8Data.ToArray());
#endif
// If it's already a data URI, use it directly
if (base64Data.StartsWith("data:", StringComparison.OrdinalIgnoreCase))
if (base64.StartsWith("data:", StringComparison.OrdinalIgnoreCase))
{
return new DataContent(base64Data, mediaType);
return new DataContent(base64, mediaType);
}
// Otherwise, construct a data URI from the base64 data
return new DataContent($"data:{mediaType};base64,{base64Data}", mediaType);
return new DataContent($"data:{mediaType};base64,{base64}", mediaType);
}
}
@@ -72,6 +72,9 @@ internal sealed class LockstepRunEventStream : IRunEventStream
this.RunStatus = RunStatus.Running;
runActivity?.AddEvent(new ActivityEvent(EventNames.WorkflowStarted));
// Emit WorkflowStartedEvent to the event stream for consumers
eventSink.Enqueue(new WorkflowStartedEvent());
do
{
while (this._stepRunner.HasUnprocessedMessages &&
@@ -88,9 +88,16 @@ internal sealed class StreamingRunEventStream : IRunEventStream
// Run all available supersteps continuously
// Events are streamed out in real-time as they happen via the event handler
while (this._stepRunner.HasUnprocessedMessages && !linkedSource.Token.IsCancellationRequested)
if (this._stepRunner.HasUnprocessedMessages)
{
await this._stepRunner.RunSuperStepAsync(linkedSource.Token).ConfigureAwait(false);
// Emit WorkflowStartedEvent only when there's actual work to process
// This avoids spurious events on timeout-only loop iterations
await this._eventChannel.Writer.WriteAsync(new WorkflowStartedEvent(), linkedSource.Token).ConfigureAwait(false);
while (this._stepRunner.HasUnprocessedMessages && !linkedSource.Token.IsCancellationRequested)
{
await this._stepRunner.RunSuperStepAsync(linkedSource.Token).ConfigureAwait(false);
}
}
// Update status based on what's waiting
@@ -3,6 +3,7 @@
#pragma warning disable CS0618 // Type or member is obsolete - Internal use of obsolete types for backward compatibility
using System;
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
@@ -133,7 +134,25 @@ internal sealed class ExecutorProtocol(MessageRouter router, ISet<Type> sendType
public bool CanHandle(Type type) => router.CanHandle(type);
public bool CanOutput(Type type) => this._yieldTypes.Contains(new(type));
private readonly ConcurrentDictionary<Type, bool> _canOutputCache = new();
public bool CanOutput(Type type)
{
return this._canOutputCache.GetOrAdd(type, this.CanOutputCore);
}
private bool CanOutputCore(Type type)
{
foreach (TypeId yieldType in this._yieldTypes)
{
if (yieldType.IsMatchPolymorphic(type))
{
return true;
}
}
return false;
}
public ProtocolDescriptor Describe() => new(this.Router.IncomingTypes, yieldTypes, sendTypes, this.Router.HasCatchAll);
}
@@ -124,6 +124,7 @@ internal sealed class MessageMerger
List<ChatMessage> messages = [];
Dictionary<string, AgentResponse> responses = [];
HashSet<string> agentIds = [];
HashSet<ChatFinishReason> finishReasons = [];
foreach (string responseId in this._mergeStates.Keys)
{
@@ -156,6 +157,11 @@ internal sealed class MessageMerger
createdTimes.Add(response.CreatedAt.Value);
}
if (response.FinishReason.HasValue)
{
finishReasons.Add(response.FinishReason.Value);
}
usage = MergeUsage(usage, response.Usage);
additionalProperties = MergeProperties(additionalProperties, response.AdditionalProperties);
}
@@ -182,6 +188,7 @@ internal sealed class MessageMerger
AgentId = primaryAgentId
?? primaryAgentName
?? (agentIds.Count == 1 ? agentIds.First() : null),
FinishReason = finishReasons.Count == 1 ? finishReasons.First() : null,
CreatedAt = DateTimeOffset.UtcNow,
Usage = usage,
AdditionalProperties = additionalProperties
@@ -207,6 +214,7 @@ internal sealed class MessageMerger
AgentId = incoming.AgentId ?? current.AgentId,
AdditionalProperties = MergeProperties(current.AdditionalProperties, incoming.AdditionalProperties),
CreatedAt = incoming.CreatedAt ?? current.CreatedAt,
FinishReason = incoming.FinishReason ?? current.FinishReason,
Messages = current.Messages.Concat(incoming.Messages).ToList(),
ResponseId = current.ResponseId,
RawRepresentation = rawRepresentation,
@@ -17,6 +17,25 @@ namespace Microsoft.Agents.AI;
/// <summary>
/// Provides an <see cref="AIAgent"/> that delegates to an <see cref="IChatClient"/> implementation.
/// </summary>
/// <remarks>
/// <para>
/// <strong>Security considerations:</strong> The <see cref="ChatClientAgent"/> orchestrates data flow across trust boundaries.
/// The underlying AI service is an external endpoint and LLM responses should be treated as untrusted output. Developers should be aware of:
/// <list type="bullet">
/// <item><description><strong>Hallucination:</strong> LLMs may generate plausible-sounding but factually incorrect information.
/// Do not treat LLM output as authoritative without verification.</description></item>
/// <item><description><strong>Indirect prompt injection:</strong> Data retrieved by tools, AI context providers, or chat history providers may
/// contain adversarial content designed to influence LLM behavior or exfiltrate data through tool calls.</description></item>
/// <item><description><strong>Malicious payloads:</strong> LLM output may contain content that is harmful if rendered or executed without
/// sanitization — for example, HTML/JavaScript for cross-site scripting, SQL for injection, or shell commands.</description></item>
/// <item><description><strong>Tool invocation:</strong> By default, all tools provided to the agent are invoked without user approval.
/// The AI selects which functions to call and with what arguments. Function arguments should be treated as untrusted input.
/// Developers should require explicit approval for tools with side effects, data sensitivity, or irreversibility.</description></item>
/// </list>
/// Developers should validate and sanitize LLM output before rendering it in HTML, executing it as code, using it in database queries,
/// or passing it to any security-sensitive context. Apply defense-in-depth by combining tool approval requirements with output validation.
/// </para>
/// </remarks>
public sealed partial class ChatClientAgent : AIAgent
{
private readonly ChatClientAgentOptions? _agentOptions;
@@ -44,6 +63,9 @@ public sealed partial class ChatClientAgent : AIAgent
/// Optional collection of tools that the agent can invoke during conversations.
/// These tools augment any tools that may be provided to the agent via <see cref="ChatOptions.Tools"/> when
/// the agent is run.
/// By default, all provided tools are invoked without user approval. The AI selects which functions to call and chooses
/// the arguments — these arguments should be treated as untrusted input. Developers should require explicit approval
/// for tools that have side effects, access sensitive data, or perform irreversible operations.
/// </param>
/// <param name="loggerFactory">
/// Optional logger factory for creating loggers used by the agent and its components.
@@ -55,7 +55,7 @@ public static class ChatClientExtensions
if (chatClient.GetService<FunctionInvokingChatClient>() is null)
{
_ = chatBuilder.Use((innerClient, services) =>
chatBuilder.Use((innerClient, services) =>
{
var loggerFactory = services.GetService<ILoggerFactory>();
@@ -0,0 +1,159 @@
// Copyright (c) Microsoft. All rights reserved.
using System;
using System.Collections.Generic;
using Microsoft.Extensions.AI;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Content-based equality comparison for <see cref="ChatMessage"/> instances.
/// </summary>
internal static class ChatMessageContentEquality
{
/// <summary>
/// Determines whether two <see cref="ChatMessage"/> instances represent the same message by content.
/// </summary>
/// <remarks>
/// When both messages define a <see cref="ChatMessage.MessageId"/>, identity is determined solely
/// by that identifier. Otherwise, the comparison falls through to <see cref="ChatMessage.Role"/>,
/// <see cref="ChatMessage.AuthorName"/>, and each item in <see cref="ChatMessage.Contents"/>.
/// </remarks>
internal static bool ContentEquals(this ChatMessage? message, ChatMessage? other)
{
if (ReferenceEquals(message, other))
{
return true;
}
if (message is null || other is null)
{
return false;
}
// A matching MessageId is sufficient.
if (message.MessageId is not null && other.MessageId is not null)
{
return string.Equals(message.MessageId, other.MessageId, StringComparison.Ordinal);
}
if (message.Role != other.Role)
{
return false;
}
if (!string.Equals(message.AuthorName, other.AuthorName, StringComparison.Ordinal))
{
return false;
}
return ContentsEqual(message.Contents, other.Contents);
}
private static bool ContentsEqual(IList<AIContent> left, IList<AIContent> right)
{
if (left.Count != right.Count)
{
return false;
}
for (int i = 0; i < left.Count; i++)
{
if (!ContentItemEquals(left[i], right[i]))
{
return false;
}
}
return true;
}
private static bool ContentItemEquals(AIContent left, AIContent right)
{
if (ReferenceEquals(left, right))
{
return true;
}
if (left.GetType() != right.GetType())
{
return false;
}
return (left, right) switch
{
(TextContent a, TextContent b) => TextContentEquals(a, b),
(TextReasoningContent a, TextReasoningContent b) => TextReasoningContentEquals(a, b),
(DataContent a, DataContent b) => DataContentEquals(a, b),
(UriContent a, UriContent b) => UriContentEquals(a, b),
(ErrorContent a, ErrorContent b) => ErrorContentEquals(a, b),
(FunctionCallContent a, FunctionCallContent b) => FunctionCallContentEquals(a, b),
(FunctionResultContent a, FunctionResultContent b) => FunctionResultContentEquals(a, b),
(HostedFileContent a, HostedFileContent b) => HostedFileContentEquals(a, b),
(AIContent a, AIContent b) => a.GetType() == b.GetType(),
};
}
private static bool TextContentEquals(TextContent a, TextContent b) =>
string.Equals(a.Text, b.Text, StringComparison.Ordinal);
private static bool TextReasoningContentEquals(TextReasoningContent a, TextReasoningContent b) =>
string.Equals(a.Text, b.Text, StringComparison.Ordinal) &&
string.Equals(a.ProtectedData, b.ProtectedData, StringComparison.Ordinal);
private static bool DataContentEquals(DataContent a, DataContent b) =>
string.Equals(a.MediaType, b.MediaType, StringComparison.Ordinal) &&
string.Equals(a.Name, b.Name, StringComparison.Ordinal) &&
a.Data.Span.SequenceEqual(b.Data.Span);
private static bool UriContentEquals(UriContent a, UriContent b) =>
Equals(a.Uri, b.Uri) &&
string.Equals(a.MediaType, b.MediaType, StringComparison.Ordinal);
private static bool ErrorContentEquals(ErrorContent a, ErrorContent b) =>
string.Equals(a.Message, b.Message, StringComparison.Ordinal) &&
string.Equals(a.ErrorCode, b.ErrorCode, StringComparison.Ordinal) &&
Equals(a.Details, b.Details);
private static bool FunctionCallContentEquals(FunctionCallContent a, FunctionCallContent b) =>
string.Equals(a.CallId, b.CallId, StringComparison.Ordinal) &&
string.Equals(a.Name, b.Name, StringComparison.Ordinal) &&
ArgumentsEqual(a.Arguments, b.Arguments);
private static bool FunctionResultContentEquals(FunctionResultContent a, FunctionResultContent b) =>
string.Equals(a.CallId, b.CallId, StringComparison.Ordinal) &&
Equals(a.Result, b.Result);
private static bool ArgumentsEqual(IDictionary<string, object?>? left, IDictionary<string, object?>? right)
{
if (ReferenceEquals(left, right))
{
return true;
}
if (left is null || right is null)
{
return false;
}
if (left.Count != right.Count)
{
return false;
}
foreach (KeyValuePair<string, object?> entry in left)
{
if (!right.TryGetValue(entry.Key, out object? value) || !Equals(entry.Value, value))
{
return false;
}
}
return true;
}
private static bool HostedFileContentEquals(HostedFileContent a, HostedFileContent b) =>
string.Equals(a.FileId, b.FileId, StringComparison.Ordinal) &&
string.Equals(a.MediaType, b.MediaType, StringComparison.Ordinal) &&
string.Equals(a.Name, b.Name, StringComparison.Ordinal);
}
@@ -0,0 +1,82 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.Logging;
using Microsoft.Shared.DiagnosticIds;
using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A compaction strategy that delegates to an <see cref="IChatReducer"/> to reduce the conversation's
/// included messages.
/// </summary>
/// <remarks>
/// <para>
/// This strategy bridges the <see cref="IChatReducer"/> abstraction from <c>Microsoft.Extensions.AI</c>
/// into the compaction pipeline. It collects the currently included messages from the
/// <see cref="CompactionMessageIndex"/>, passes them to the reducer, and rebuilds the index from the
/// reduced message list when the reducer produces fewer messages.
/// </para>
/// <para>
/// The <see cref="CompactionTrigger"/> controls when reduction is attempted.
/// Use <see cref="CompactionTriggers"/> for common trigger conditions such as token or message thresholds.
/// </para>
/// <para>
/// Use this strategy when you have an existing <see cref="IChatReducer"/> implementation
/// (such as <c>MessageCountingChatReducer</c>) and want to apply it as part of a
/// <see cref="CompactionStrategy"/> pipeline or as an in-run compaction strategy.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class ChatReducerCompactionStrategy : CompactionStrategy
{
/// <summary>
/// Initializes a new instance of the <see cref="ChatReducerCompactionStrategy"/> class.
/// </summary>
/// <param name="chatReducer">
/// The <see cref="IChatReducer"/> that performs the message reduction.
/// </param>
/// <param name="trigger">
/// The <see cref="CompactionTrigger"/> that controls when compaction proceeds.
/// </param>
public ChatReducerCompactionStrategy(IChatReducer chatReducer, CompactionTrigger trigger)
: base(trigger)
{
this.ChatReducer = Throw.IfNull(chatReducer);
}
/// <summary>
/// Gets the chat reducer used to reduce messages.
/// </summary>
public IChatReducer ChatReducer { get; }
/// <inheritdoc/>
protected override async ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
{
// No need to short-circuit on empty conversations, this is handled by <see cref="CompactionStrategy.CompactAsync"/>.
List<ChatMessage> includedMessages = [.. index.GetIncludedMessages()];
IEnumerable<ChatMessage> reduced = await this.ChatReducer.ReduceAsync(includedMessages, cancellationToken).ConfigureAwait(false);
IList<ChatMessage> reducedMessages = reduced as IList<ChatMessage> ?? [.. reduced];
if (reducedMessages.Count >= includedMessages.Count)
{
return false;
}
// Rebuild the index from the reduced messages
CompactionMessageIndex rebuilt = CompactionMessageIndex.Create(reducedMessages, index.Tokenizer);
index.Groups.Clear();
foreach (CompactionMessageGroup group in rebuilt.Groups)
{
index.Groups.Add(group);
}
return true;
}
}
@@ -0,0 +1,55 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Diagnostics.CodeAnalysis;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Identifies the kind of a <see cref="CompactionMessageGroup"/>.
/// </summary>
/// <remarks>
/// Message groups are used to classify logically related messages that must be kept together
/// during compaction operations. For example, an assistant message containing tool calls
/// and its corresponding tool result messages form an atomic <see cref="ToolCall"/> group.
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public enum CompactionGroupKind
{
/// <summary>
/// A system message group containing one or more system messages.
/// </summary>
System,
/// <summary>
/// A user message group containing a single user message.
/// </summary>
User,
/// <summary>
/// An assistant message group containing a single assistant text response (no tool calls).
/// </summary>
AssistantText,
/// <summary>
/// An atomic tool call group containing an assistant message with tool calls
/// followed by the corresponding tool result messages.
/// </summary>
/// <remarks>
/// This group must be treated as an atomic unit during compaction. Removing the assistant
/// message without its tool results (or vice versa) will cause LLM API errors.
/// </remarks>
ToolCall,
#pragma warning disable IDE0001 // Simplify Names
/// <summary>
/// A summary message group produced by a compaction strategy (e.g., <c>SummarizationCompactionStrategy</c>).
/// </summary>
/// <remarks>
/// Summary groups replace previously compacted messages with a condensed representation.
/// They are identified by the <see cref="CompactionMessageGroup.SummaryPropertyKey"/> metadata entry
/// on the underlying <see cref="Microsoft.Extensions.AI.ChatMessage"/>.
/// </remarks>
#pragma warning restore IDE0001 // Simplify Names
Summary,
}
@@ -0,0 +1,112 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Diagnostics.CodeAnalysis;
using Microsoft.Extensions.Logging;
namespace Microsoft.Agents.AI.Compaction;
#pragma warning disable SYSLIB1006 // Multiple logging methods cannot use the same event id within a class
/// <summary>
/// Extensions for logging compaction diagnostics.
/// </summary>
/// <remarks>
/// This extension uses the <see cref="LoggerMessageAttribute"/> to
/// generate logging code at compile time to achieve optimized code.
/// </remarks>
[ExcludeFromCodeCoverage]
internal static partial class CompactionLogMessages
{
/// <summary>
/// Logs when compaction is skipped because the trigger condition was not met.
/// </summary>
[LoggerMessage(
Level = LogLevel.Trace,
Message = "Compaction skipped for {StrategyName}: trigger condition not met or insufficient groups.")]
public static partial void LogCompactionSkipped(
this ILogger logger,
string strategyName);
/// <summary>
/// Logs compaction completion with before/after metrics.
/// </summary>
[LoggerMessage(
Level = LogLevel.Debug,
Message = "Compaction completed: {StrategyName} in {DurationMs}ms — Messages {BeforeMessages}→{AfterMessages}, Groups {BeforeGroups}→{AfterGroups}, Tokens {BeforeTokens}→{AfterTokens}")]
public static partial void LogCompactionCompleted(
this ILogger logger,
string strategyName,
long durationMs,
int beforeMessages,
int afterMessages,
int beforeGroups,
int afterGroups,
int beforeTokens,
int afterTokens);
/// <summary>
/// Logs when the compaction provider skips compaction.
/// </summary>
[LoggerMessage(
Level = LogLevel.Trace,
Message = "CompactionProvider skipped: {Reason}.")]
public static partial void LogCompactionProviderSkipped(
this ILogger logger,
string reason);
/// <summary>
/// Logs when the compaction provider begins applying a compaction strategy.
/// </summary>
[LoggerMessage(
Level = LogLevel.Debug,
Message = "CompactionProvider applying compaction to {MessageCount} messages using {StrategyName}.")]
public static partial void LogCompactionProviderApplying(
this ILogger logger,
int messageCount,
string strategyName);
/// <summary>
/// Logs when the compaction provider has applied compaction with result metrics.
/// </summary>
[LoggerMessage(
Level = LogLevel.Debug,
Message = "CompactionProvider compaction applied: messages {BeforeMessages}→{AfterMessages}.")]
public static partial void LogCompactionProviderApplied(
this ILogger logger,
int beforeMessages,
int afterMessages);
/// <summary>
/// Logs when a summarization LLM call is starting.
/// </summary>
[LoggerMessage(
Level = LogLevel.Debug,
Message = "Summarization starting for {GroupCount} groups ({MessageCount} messages) using {ChatClientType}.")]
public static partial void LogSummarizationStarting(
this ILogger logger,
int groupCount,
int messageCount,
string chatClientType);
/// <summary>
/// Logs when a summarization LLM call has completed.
/// </summary>
[LoggerMessage(
Level = LogLevel.Debug,
Message = "Summarization completed: summary length {SummaryLength} characters, inserted at index {InsertIndex}.")]
public static partial void LogSummarizationCompleted(
this ILogger logger,
int summaryLength,
int insertIndex);
/// <summary>
/// Logs when a summarization LLM call fails and groups are restored.
/// </summary>
[LoggerMessage(
Level = LogLevel.Warning,
Message = "Summarization failed for {GroupCount} groups; restoring excluded groups and continuing without compaction. Error: {ErrorMessage}")]
public static partial void LogSummarizationFailed(
this ILogger logger,
int groupCount,
string errorMessage);
}
@@ -0,0 +1,116 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Diagnostics.CodeAnalysis;
using System.Text.Json.Serialization;
using Microsoft.Extensions.AI;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Represents a logical group of <see cref="ChatMessage"/> instances that must be kept or removed together during compaction.
/// </summary>
/// <remarks>
/// <para>
/// Message groups ensure atomic preservation of related messages. For example, an assistant message
/// containing tool calls and its corresponding tool result messages form a <see cref="CompactionGroupKind.ToolCall"/>
/// group — removing one without the other would cause LLM API errors.
/// </para>
/// <para>
/// Groups also support exclusion semantics: a group can be marked as excluded (with an optional reason)
/// to indicate it should not be included in the messages sent to the model, while still being preserved
/// for diagnostics, storage, or later re-inclusion.
/// </para>
/// <para>
/// Each group tracks its <see cref="MessageCount"/>, <see cref="ByteCount"/>, and <see cref="TokenCount"/>
/// so that <see cref="CompactionMessageIndex"/> can efficiently aggregate totals across all or only included groups.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class CompactionMessageGroup
{
/// <summary>
/// The <see cref="ChatMessage.AdditionalProperties"/> key used to identify a message as a compaction summary.
/// </summary>
/// <remarks>
/// When this key is present with a value of <see langword="true"/>, the message is classified as
/// <see cref="CompactionGroupKind.Summary"/> by <see cref="CompactionMessageIndex.Create"/>.
/// </remarks>
public static readonly string SummaryPropertyKey = "_is_summary";
/// <summary>
/// Initializes a new instance of the <see cref="CompactionMessageGroup"/> class.
/// </summary>
/// <param name="kind">The kind of message group.</param>
/// <param name="messages">The messages in this group. The list is captured as a read-only snapshot.</param>
/// <param name="byteCount">The total UTF-8 byte count of the text content in the messages.</param>
/// <param name="tokenCount">The token count for the messages, computed by a tokenizer or estimated.</param>
/// <param name="turnIndex">
/// The user turn this group belongs to, or <see langword="null"/> for <see cref="CompactionGroupKind.System"/>.
/// </param>
[JsonConstructor]
internal CompactionMessageGroup(CompactionGroupKind kind, IReadOnlyList<ChatMessage> messages, int byteCount, int tokenCount, int? turnIndex = null)
{
this.Kind = kind;
this.Messages = messages;
this.MessageCount = messages.Count;
this.ByteCount = byteCount;
this.TokenCount = tokenCount;
this.TurnIndex = turnIndex;
}
/// <summary>
/// Gets the kind of this message group.
/// </summary>
public CompactionGroupKind Kind { get; }
/// <summary>
/// Gets the messages in this group.
/// </summary>
public IReadOnlyList<ChatMessage> Messages { get; }
/// <summary>
/// Gets the number of messages in this group.
/// </summary>
public int MessageCount { get; }
/// <summary>
/// Gets the total UTF-8 byte count of the text content in this group's messages.
/// </summary>
public int ByteCount { get; }
/// <summary>
/// Gets the estimated or actual token count for this group's messages.
/// </summary>
public int TokenCount { get; }
/// <summary>
/// Gets user turn index this group belongs to, or <see langword="null"/> for groups
/// that precede the first user message (e.g., system messages). A turn index of 0
/// corresponds with any non-system message that precedes the first user message,
/// turn index 1 corresponds with the first user message and its subsequent non-user
/// messages, and so on...
/// </summary>
/// <remarks>
/// A turn starts with a <see cref="CompactionGroupKind.User"/> group and includes all subsequent
/// non-user, non-system groups until the next user group or end of conversation. System messages
/// (<see cref="CompactionGroupKind.System"/>) are always assigned a <see langword="null"/> turn index
/// since they never belong to a user turn.
/// </remarks>
public int? TurnIndex { get; }
/// <summary>
/// Gets or sets a value indicating whether this group is excluded from the projected message list.
/// </summary>
/// <remarks>
/// Excluded groups are preserved in the collection for diagnostics or storage purposes
/// but are not included when calling <see cref="CompactionMessageIndex.GetIncludedMessages"/>.
/// </remarks>
public bool IsExcluded { get; set; }
/// <summary>
/// Gets or sets an optional reason explaining why this group was excluded.
/// </summary>
public string? ExcludeReason { get; set; }
}
@@ -0,0 +1,529 @@
// Copyright (c) Microsoft. All rights reserved.
using System;
using System.Collections.Generic;
using System.Diagnostics.CodeAnalysis;
using System.Linq;
using System.Text;
using Microsoft.Extensions.AI;
using Microsoft.ML.Tokenizers;
using Microsoft.Shared.DiagnosticIds;
using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A collection of <see cref="CompactionMessageGroup"/> instances and derived metrics based on a flat list of <see cref="ChatMessage"/> objects.
/// </summary>
/// <remarks>
/// <see cref="CompactionMessageIndex"/> provides structural grouping of messages into logical <see cref="CompactionMessageGroup"/> units. Individual
/// groups can be marked as excluded without being removed, allowing compaction strategies to toggle visibility while preserving
/// the full history for diagnostics or storage. Metrics are provided both including and excluding excluded groups,
/// allowing strategies to make informed decisions based on the impact of potential exclusions.
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class CompactionMessageIndex
{
private int _currentTurn;
private ChatMessage? _lastProcessedMessage;
/// <summary>
/// Gets the list of message groups in this collection.
/// </summary>
public IList<CompactionMessageGroup> Groups { get; }
/// <summary>
/// Gets the tokenizer used for computing token counts, or <see langword="null"/> if token counts are estimated.
/// </summary>
public Tokenizer? Tokenizer { get; }
/// <summary>
/// Initializes a new instance of the <see cref="CompactionMessageIndex"/> class with the specified groups.
/// </summary>
/// <param name="groups">The message groups.</param>
/// <param name="tokenizer">An optional tokenizer retained for computing token counts when adding new groups.</param>
public CompactionMessageIndex(IList<CompactionMessageGroup> groups, Tokenizer? tokenizer = null)
{
this.Groups = Throw.IfNull(groups, nameof(groups));
this.Tokenizer = tokenizer;
// Restore turn counter and last processed message from the groups
for (int index = groups.Count - 1; index >= 0; --index)
{
if (this._lastProcessedMessage is null && this.Groups[index].Kind != CompactionGroupKind.Summary)
{
IReadOnlyList<ChatMessage> groupMessages = this.Groups[index].Messages;
this._lastProcessedMessage = groupMessages[^1];
}
if (this.Groups[index].TurnIndex.HasValue)
{
this._currentTurn = this.Groups[index].TurnIndex!.Value;
// Both values restored — no need to keep scanning
if (this._lastProcessedMessage is not null)
{
break;
}
}
}
}
/// <summary>
/// Creates a <see cref="CompactionMessageIndex"/> from a flat list of <see cref="ChatMessage"/> instances.
/// </summary>
/// <param name="messages">The messages to group.</param>
/// <param name="tokenizer">
/// An optional <see cref="Tokenizer"/> for computing token counts on each group.
/// When <see langword="null"/>, token counts are estimated as <c>ByteCount / 4</c>.
/// </param>
/// <returns>A new <see cref="CompactionMessageIndex"/> with messages organized into logical groups.</returns>
/// <remarks>
/// The grouping algorithm:
/// <list type="bullet">
/// <item><description>System messages become <see cref="CompactionGroupKind.System"/> groups.</description></item>
/// <item><description>User messages become <see cref="CompactionGroupKind.User"/> groups.</description></item>
/// <item><description>Assistant messages with tool calls, followed by their corresponding tool result messages, become <see cref="CompactionGroupKind.ToolCall"/> groups.</description></item>
/// <item><description>Assistant messages marked with <see cref="CompactionMessageGroup.SummaryPropertyKey"/> become <see cref="CompactionGroupKind.Summary"/> groups.</description></item>
/// <item><description>Assistant messages without tool calls become <see cref="CompactionGroupKind.AssistantText"/> groups.</description></item>
/// </list>
/// </remarks>
internal static CompactionMessageIndex Create(IList<ChatMessage> messages, Tokenizer? tokenizer = null)
{
CompactionMessageIndex instance = new([], tokenizer);
instance.AppendFromMessages(messages, 0);
return instance;
}
/// <summary>
/// Incrementally updates the groups with new messages from the conversation.
/// </summary>
/// <param name="allMessages">
/// The full list of messages for the conversation. This must be the same list (or a replacement with the same
/// prefix) that was used to create or last update this instance.
/// </param>
/// <remarks>
/// <para>
/// Uses equality on the last processed message to detect changes. Only the messages after that position are
/// processed and appended as new groups. Existing groups and their compaction state (exclusions) are preserved.
/// </para>
/// <para>
/// If the last processed message is not found (e.g., the message list was replaced entirely
/// or a sliding window shifted past it), all groups are cleared and rebuilt from scratch.
/// </para>
/// <para>
/// If the last message in <paramref name="allMessages"/> matches the last
/// processed message, no work is performed.
/// </para>
/// </remarks>
internal void Update(IList<ChatMessage> allMessages)
{
if (allMessages.Count == 0)
{
this.Groups.Clear();
this._currentTurn = 0;
this._lastProcessedMessage = null;
return;
}
// If the last message is unchanged and the list hasn't shrunk, there is nothing new to process.
if (this._lastProcessedMessage is not null &&
allMessages.Count >= this.RawMessageCount &&
allMessages[allMessages.Count - 1].ContentEquals(this._lastProcessedMessage))
{
return;
}
// Walk backwards to locate where we left off.
int foundIndex = -1;
if (this._lastProcessedMessage is not null)
{
for (int i = allMessages.Count - 1; i >= 0; --i)
{
if (allMessages[i].ContentEquals(this._lastProcessedMessage))
{
foundIndex = i;
break;
}
}
}
if (foundIndex < 0)
{
// Last processed message not found — total rebuild.
this.Groups.Clear();
this._currentTurn = 0;
this.AppendFromMessages(allMessages, 0);
return;
}
// Guard against a sliding window that removed messages from the front:
// the number of messages up to (and including) the found position must
// match the number of messages already represented by existing groups.
if (foundIndex + 1 < this.RawMessageCount)
{
// Front of the message list was trimmed — rebuild.
this.Groups.Clear();
this._currentTurn = 0;
this.AppendFromMessages(allMessages, 0);
return;
}
// Process only the delta messages.
this.AppendFromMessages(allMessages, foundIndex + 1);
}
private void AppendFromMessages(IList<ChatMessage> messages, int startIndex)
{
int index = startIndex;
while (index < messages.Count)
{
ChatMessage message = messages[index];
if (message.Role == ChatRole.System)
{
// System messages are not part of any turn
this.Groups.Add(CreateGroup(CompactionGroupKind.System, [message], this.Tokenizer, turnIndex: null));
index++;
}
else if (message.Role == ChatRole.User)
{
this._currentTurn++;
this.Groups.Add(CreateGroup(CompactionGroupKind.User, [message], this.Tokenizer, this._currentTurn));
index++;
}
else if (message.Role == ChatRole.Assistant && HasToolCalls(message))
{
List<ChatMessage> groupMessages = [message];
index++;
// Collect all subsequent tool result messages and reasoning-only assistant messages
while (index < messages.Count &&
(messages[index].Role == ChatRole.Tool ||
(messages[index].Role == ChatRole.Assistant && HasOnlyReasoning(messages[index]))))
{
groupMessages.Add(messages[index]);
index++;
}
this.Groups.Add(CreateGroup(CompactionGroupKind.ToolCall, groupMessages, this.Tokenizer, this._currentTurn));
}
else if (message.Role == ChatRole.Assistant && IsSummaryMessage(message))
{
this.Groups.Add(CreateGroup(CompactionGroupKind.Summary, [message], this.Tokenizer, this._currentTurn));
index++;
}
else if (message.Role == ChatRole.Assistant && HasOnlyReasoning(message))
{
// Reasoning-only assistant messages that precede a tool-call assistant message
// are part of the same atomic tool-call group. Look ahead past consecutive
// reasoning messages to find a possible tool-call message.
int lookahead = index + 1;
while (lookahead < messages.Count &&
messages[lookahead].Role == ChatRole.Assistant &&
HasOnlyReasoning(messages[lookahead]))
{
lookahead++;
}
if (lookahead < messages.Count && messages[lookahead].Role == ChatRole.Assistant && HasToolCalls(messages[lookahead]))
{
// Group all reasoning messages + the tool-call message together
List<ChatMessage> groupMessages = [];
for (int j = index; j <= lookahead; j++)
{
groupMessages.Add(messages[j]);
}
index = lookahead + 1;
// Collect all subsequent tool result messages and reasoning-only assistant messages
while (index < messages.Count &&
(messages[index].Role == ChatRole.Tool ||
(messages[index].Role == ChatRole.Assistant && HasOnlyReasoning(messages[index]))))
{
groupMessages.Add(messages[index]);
index++;
}
this.Groups.Add(CreateGroup(CompactionGroupKind.ToolCall, groupMessages, this.Tokenizer, this._currentTurn));
}
else
{
this.Groups.Add(CreateGroup(CompactionGroupKind.AssistantText, [message], this.Tokenizer, this._currentTurn));
index++;
}
}
else
{
this.Groups.Add(CreateGroup(CompactionGroupKind.AssistantText, [message], this.Tokenizer, this._currentTurn));
index++;
}
}
if (messages.Count > 0)
{
this._lastProcessedMessage = messages[^1];
}
}
/// <summary>
/// Creates a new <see cref="CompactionMessageGroup"/> with byte and token counts computed using this collection's
/// <see cref="Tokenizer"/>, and adds it to the <see cref="Groups"/> list at the specified index.
/// </summary>
/// <param name="index">The zero-based index at which the group should be inserted.</param>
/// <param name="kind">The kind of message group.</param>
/// <param name="messages">The messages in the group.</param>
/// <param name="turnIndex">The optional turn index to assign to the new group.</param>
/// <returns>The newly created <see cref="CompactionMessageGroup"/>.</returns>
public CompactionMessageGroup InsertGroup(int index, CompactionGroupKind kind, IReadOnlyList<ChatMessage> messages, int? turnIndex = null)
{
CompactionMessageGroup group = CreateGroup(kind, messages, this.Tokenizer, turnIndex);
this.Groups.Insert(index, group);
return group;
}
/// <summary>
/// Creates a new <see cref="CompactionMessageGroup"/> with byte and token counts computed using this collection's
/// <see cref="Tokenizer"/>, and appends it to the end of the <see cref="Groups"/> list.
/// </summary>
/// <param name="kind">The kind of message group.</param>
/// <param name="messages">The messages in the group.</param>
/// <param name="turnIndex">The optional turn index to assign to the new group.</param>
/// <returns>The newly created <see cref="CompactionMessageGroup"/>.</returns>
public CompactionMessageGroup AddGroup(CompactionGroupKind kind, IReadOnlyList<ChatMessage> messages, int? turnIndex = null)
{
CompactionMessageGroup group = CreateGroup(kind, messages, this.Tokenizer, turnIndex);
this.Groups.Add(group);
return group;
}
/// <summary>
/// Returns only the messages from groups that are not excluded.
/// </summary>
/// <returns>A list of <see cref="ChatMessage"/> instances from included groups, in order.</returns>
public IEnumerable<ChatMessage> GetIncludedMessages() =>
this.Groups.Where(group => !group.IsExcluded).SelectMany(group => group.Messages);
/// <summary>
/// Returns all messages from all groups, including excluded ones.
/// </summary>
/// <returns>A list of all <see cref="ChatMessage"/> instances, in order.</returns>
public IEnumerable<ChatMessage> GetAllMessages() => this.Groups.SelectMany(group => group.Messages);
/// <summary>
/// Gets the total number of groups, including excluded ones.
/// </summary>
public int TotalGroupCount => this.Groups.Count;
/// <summary>
/// Gets the total number of messages across all groups, including excluded ones.
/// </summary>
public int TotalMessageCount => this.Groups.Sum(group => group.MessageCount);
/// <summary>
/// Gets the total UTF-8 byte count across all groups, including excluded ones.
/// </summary>
public int TotalByteCount => this.Groups.Sum(group => group.ByteCount);
/// <summary>
/// Gets the total token count across all groups, including excluded ones.
/// </summary>
public int TotalTokenCount => this.Groups.Sum(group => group.TokenCount);
/// <summary>
/// Gets the total number of groups that are not excluded.
/// </summary>
public int IncludedGroupCount => this.Groups.Count(group => !group.IsExcluded);
/// <summary>
/// Gets the total number of messages across all included (non-excluded) groups.
/// </summary>
public int IncludedMessageCount => this.Groups.Where(group => !group.IsExcluded).Sum(group => group.MessageCount);
/// <summary>
/// Gets the total UTF-8 byte count across all included (non-excluded) groups.
/// </summary>
public int IncludedByteCount => this.Groups.Where(group => !group.IsExcluded).Sum(group => group.ByteCount);
/// <summary>
/// Gets the total token count across all included (non-excluded) groups.
/// </summary>
public int IncludedTokenCount => this.Groups.Where(group => !group.IsExcluded).Sum(group => group.TokenCount);
/// <summary>
/// Gets the total number of user turns across all groups (including those with excluded groups).
/// </summary>
public int TotalTurnCount => this.Groups.Select(group => group.TurnIndex).Distinct().Count(turnIndex => turnIndex is not null && turnIndex > 0);
/// <summary>
/// Gets the number of user turns that have at least one non-excluded group.
/// </summary>
public int IncludedTurnCount => this.Groups.Where(group => !group.IsExcluded && group.TurnIndex is not null && group.TurnIndex > 0).Select(group => group.TurnIndex).Distinct().Count();
/// <summary>
/// Gets the total number of groups across all included (non-excluded) groups that are not <see cref="CompactionGroupKind.System"/>.
/// </summary>
public int IncludedNonSystemGroupCount => this.Groups.Count(group => !group.IsExcluded && group.Kind != CompactionGroupKind.System);
/// <summary>
/// Gets the total number of original messages (that are not summaries).
/// </summary>
public int RawMessageCount => this.Groups.Where(group => group.Kind != CompactionGroupKind.Summary).Sum(group => group.MessageCount);
/// <summary>
/// Returns all groups that belong to the specified user turn.
/// </summary>
/// <param name="turnIndex">The desired turn index.</param>
/// <returns>The groups belonging to the turn, in order.</returns>
public IEnumerable<CompactionMessageGroup> GetTurnGroups(int turnIndex) => this.Groups.Where(group => group.TurnIndex == turnIndex);
/// <summary>
/// Computes the UTF-8 byte count for a set of messages across all content types.
/// </summary>
/// <param name="messages">The messages to compute byte count for.</param>
/// <returns>The total UTF-8 byte count of all message content.</returns>
internal static int ComputeByteCount(IReadOnlyList<ChatMessage> messages)
{
int total = 0;
for (int i = 0; i < messages.Count; i++)
{
IList<AIContent> contents = messages[i].Contents;
for (int j = 0; j < contents.Count; j++)
{
total += ComputeContentByteCount(contents[j]);
}
}
return total;
}
/// <summary>
/// Computes the token count for a set of messages using the specified tokenizer.
/// </summary>
/// <param name="messages">The messages to compute token count for.</param>
/// <param name="tokenizer">The tokenizer to use for counting tokens.</param>
/// <returns>The total token count across all message content.</returns>
/// <remarks>
/// Text-bearing content (<see cref="TextContent"/> and <see cref="TextReasoningContent"/>)
/// is tokenized directly. All other content types estimate tokens as <c>byteCount / 4</c>.
/// </remarks>
internal static int ComputeTokenCount(IReadOnlyList<ChatMessage> messages, Tokenizer tokenizer)
{
int total = 0;
for (int i = 0; i < messages.Count; i++)
{
IList<AIContent> contents = messages[i].Contents;
for (int j = 0; j < contents.Count; j++)
{
AIContent content = contents[j];
switch (content)
{
case TextContent text:
if (text.Text is { Length: > 0 } t)
{
total += tokenizer.CountTokens(t);
}
break;
case TextReasoningContent reasoning:
if (reasoning.Text is { Length: > 0 } rt)
{
total += tokenizer.CountTokens(rt);
}
if (reasoning.ProtectedData is { Length: > 0 } pd)
{
total += tokenizer.CountTokens(pd);
}
break;
default:
total += ComputeContentByteCount(content) / 4;
break;
}
}
}
return total;
}
private static int ComputeContentByteCount(AIContent content)
{
switch (content)
{
case TextContent text:
return GetStringByteCount(text.Text);
case TextReasoningContent reasoning:
return GetStringByteCount(reasoning.Text) + GetStringByteCount(reasoning.ProtectedData);
case DataContent data:
return data.Data.Length + GetStringByteCount(data.MediaType) + GetStringByteCount(data.Name);
case UriContent uri:
return (uri.Uri is Uri uriValue ? GetStringByteCount(uriValue.OriginalString) : 0) + GetStringByteCount(uri.MediaType);
case FunctionCallContent call:
int callBytes = GetStringByteCount(call.CallId) + GetStringByteCount(call.Name);
if (call.Arguments is not null)
{
foreach (KeyValuePair<string, object?> arg in call.Arguments)
{
callBytes += GetStringByteCount(arg.Key);
callBytes += GetStringByteCount(arg.Value?.ToString());
}
}
return callBytes;
case FunctionResultContent result:
return GetStringByteCount(result.CallId) + GetStringByteCount(result.Result?.ToString());
case ErrorContent error:
return GetStringByteCount(error.Message) + GetStringByteCount(error.ErrorCode) + GetStringByteCount(error.Details);
case HostedFileContent file:
return GetStringByteCount(file.FileId) + GetStringByteCount(file.MediaType) + GetStringByteCount(file.Name);
default:
return 0;
}
}
private static int GetStringByteCount(string? value) =>
value is { Length: > 0 } ? Encoding.UTF8.GetByteCount(value) : 0;
private static CompactionMessageGroup CreateGroup(CompactionGroupKind kind, IReadOnlyList<ChatMessage> messages, Tokenizer? tokenizer, int? turnIndex)
{
int byteCount = ComputeByteCount(messages);
int tokenCount = tokenizer is not null
? ComputeTokenCount(messages, tokenizer)
: byteCount / 4;
return new CompactionMessageGroup(kind, messages, byteCount, tokenCount, turnIndex);
}
private static bool HasToolCalls(ChatMessage message)
{
foreach (AIContent content in message.Contents)
{
if (content is FunctionCallContent)
{
return true;
}
}
return false;
}
private static bool HasOnlyReasoning(ChatMessage message) =>
message.Contents.All(content => content is TextReasoningContent);
private static bool IsSummaryMessage(ChatMessage message) =>
message.AdditionalProperties?.TryGetValue(CompactionMessageGroup.SummaryPropertyKey, out object? value) is true
&& value is true;
}
@@ -0,0 +1,186 @@
// Copyright (c) Microsoft. All rights reserved.
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Diagnostics.CodeAnalysis;
using System.Text.Json.Serialization;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.Logging;
using Microsoft.Extensions.Logging.Abstractions;
using Microsoft.Shared.DiagnosticIds;
using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A <see cref="AIContextProvider"/> that applies a <see cref="CompactionStrategy"/> to compact
/// the message list before each agent invocation.
/// </summary>
/// <remarks>
/// <para>
/// This provider performs in-run compaction by organizing messages into atomic groups (preserving
/// tool-call/result pairings) before applying compaction logic. Only included messages are forwarded
/// to the agent's underlying chat client.
/// </para>
/// <para>
/// The <see cref="CompactionProvider"/> can be added to an agent's context provider pipeline
/// via <see cref="ChatClientAgentOptions.AIContextProviders"/> or via <c>UseAIContextProviders</c>
/// on a <see cref="ChatClientBuilder"/> or <see cref="AIAgentBuilder"/>.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class CompactionProvider : AIContextProvider
{
private readonly CompactionStrategy _compactionStrategy;
private readonly ProviderSessionState<State> _sessionState;
private readonly ILoggerFactory? _loggerFactory;
/// <summary>
/// Initializes a new instance of the <see cref="CompactionProvider"/> class.
/// </summary>
/// <param name="compactionStrategy">The compaction strategy to apply before each invocation.</param>
/// <param name="stateKey">
/// An optional key used to store the provider state in the <see cref="AgentSession.StateBag"/>. Provide
/// an explicit value if configuring multiple agents with different compaction strategies that will interact
/// in the same session.
/// </param>
/// <param name="loggerFactory">
/// An optional <see cref="ILoggerFactory"/> used to create a logger for provider diagnostics.
/// When <see langword="null"/>, logging is disabled.
/// </param>
/// <exception cref="ArgumentNullException"><paramref name="compactionStrategy"/> is <see langword="null"/>.</exception>
public CompactionProvider(CompactionStrategy compactionStrategy, string? stateKey = null, ILoggerFactory? loggerFactory = null)
{
this._compactionStrategy = Throw.IfNull(compactionStrategy);
stateKey ??= this._compactionStrategy.GetType().Name;
this.StateKeys = [stateKey];
this._sessionState = new ProviderSessionState<State>(
_ => new State(),
stateKey,
AgentJsonUtilities.DefaultOptions);
this._loggerFactory = loggerFactory;
}
/// <inheritdoc />
public override IReadOnlyList<string> StateKeys { get; }
/// <summary>
/// Applies compaction strategy to the provided message list and returns the compacted messages.
/// This can be used for ad-hoc compaction outside of the provider pipeline.
/// </summary>
/// <param name="compactionStrategy">The compaction strategy to apply before each invocation.</param>
/// <param name="messages">The messages to compact</param>
/// <param name="logger">An optional <see cref="ILogger"/> for emitting compaction diagnostics.</param>
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests.</param>
/// <returns>An enumeration of the compacted <see cref="ChatMessage"/> instances.</returns>
public static async Task<IEnumerable<ChatMessage>> CompactAsync(CompactionStrategy compactionStrategy, IEnumerable<ChatMessage> messages, ILogger? logger = null, CancellationToken cancellationToken = default)
{
Throw.IfNull(compactionStrategy);
Throw.IfNull(messages);
List<ChatMessage> messageList = messages as List<ChatMessage> ?? [.. messages];
CompactionMessageIndex messageIndex = CompactionMessageIndex.Create(messageList);
await compactionStrategy.CompactAsync(messageIndex, logger, cancellationToken).ConfigureAwait(false);
return messageIndex.GetIncludedMessages();
}
/// <summary>
/// Applies the compaction strategy to the accumulated message list before forwarding it to the agent.
/// </summary>
/// <param name="context">Contains the request context including all accumulated messages.</param>
/// <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 an <see cref="AIContext"/>
/// with the compacted message list.
/// </returns>
protected override async ValueTask<AIContext> InvokingCoreAsync(InvokingContext context, CancellationToken cancellationToken = default)
{
using Activity? activity = CompactionTelemetry.ActivitySource.StartActivity(CompactionTelemetry.ActivityNames.CompactionProviderInvoke);
ILoggerFactory loggerFactory = this.GetLoggerFactory(context.Agent);
ILogger logger = loggerFactory.CreateLogger<CompactionProvider>();
AgentSession? session = context.Session;
IEnumerable<ChatMessage>? allMessages = context.AIContext.Messages;
if (session is null || allMessages is null)
{
logger.LogCompactionProviderSkipped("no session or no messages");
return context.AIContext;
}
ChatClientAgentSession? chatClientSession = session.GetService<ChatClientAgentSession>();
if (chatClientSession is not null &&
!string.IsNullOrWhiteSpace(chatClientSession.ConversationId))
{
logger.LogCompactionProviderSkipped("session managed by remote service");
return context.AIContext;
}
List<ChatMessage> messageList = allMessages as List<ChatMessage> ?? [.. allMessages];
State state = this._sessionState.GetOrInitializeState(session);
CompactionMessageIndex messageIndex;
if (state.MessageGroups.Count > 0)
{
// Update existing index with any new messages appended since the last call.
messageIndex = new([.. state.MessageGroups]);
messageIndex.Update(messageList);
}
else
{
// First pass — initialize the message index from scratch.
messageIndex = CompactionMessageIndex.Create(messageList);
}
string strategyName = this._compactionStrategy.GetType().Name;
int beforeMessages = messageIndex.IncludedMessageCount;
logger.LogCompactionProviderApplying(beforeMessages, strategyName);
// Apply compaction
await this._compactionStrategy.CompactAsync(
messageIndex,
loggerFactory.CreateLogger(this._compactionStrategy.GetType()),
cancellationToken).ConfigureAwait(false);
int afterMessages = messageIndex.IncludedMessageCount;
if (afterMessages < beforeMessages)
{
logger.LogCompactionProviderApplied(beforeMessages, afterMessages);
}
// Persist the index
state.MessageGroups.Clear();
state.MessageGroups.AddRange(messageIndex.Groups);
return new AIContext
{
Instructions = context.AIContext.Instructions,
Messages = messageIndex.GetIncludedMessages(),
Tools = context.AIContext.Tools
};
}
private ILoggerFactory GetLoggerFactory(AIAgent agent) =>
this._loggerFactory ??
agent.GetService<IChatClient>()?.GetService<ILoggerFactory>() ??
NullLoggerFactory.Instance;
/// <summary>
/// Represents the persisted state of a <see cref="CompactionProvider"/> stored in the <see cref="AgentSession.StateBag"/>.
/// </summary>
internal sealed class State
{
/// <summary>
/// Gets or sets the message index groups used for incremental compaction updates.
/// </summary>
[JsonPropertyName("messagegroups")]
public List<CompactionMessageGroup> MessageGroups { get; set; } = [];
}
}
@@ -0,0 +1,164 @@
// Copyright (c) Microsoft. All rights reserved.
using System;
using System.Diagnostics;
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.Logging;
using Microsoft.Extensions.Logging.Abstractions;
using Microsoft.Shared.DiagnosticIds;
using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Base class for strategies that compact a <see cref="CompactionMessageIndex"/> to reduce context size.
/// </summary>
/// <remarks>
/// <para>
/// Compaction strategies operate on <see cref="CompactionMessageIndex"/> instances, which organize messages
/// into atomic groups that respect the tool-call/result pairing constraint. Strategies mutate the collection
/// in place by marking groups as excluded, removing groups, or replacing message content (e.g., with summaries).
/// </para>
/// <para>
/// Every strategy requires a <see cref="CompactionTrigger"/> that determines whether compaction should
/// proceed based on current <see cref="CompactionMessageIndex"/> metrics (token count, message count, turn count, etc.).
/// The base class evaluates this trigger at the start of <see cref="CompactAsync"/> and skips compaction when
/// the trigger returns <see langword="false"/>.
/// </para>
/// <para>
/// An optional <b>target</b> condition controls when compaction stops. Strategies incrementally exclude
/// groups and re-evaluate the target after each exclusion, stopping as soon as the target returns
/// <see langword="true"/>. When no target is specified, it defaults to the inverse of the trigger —
/// meaning compaction stops when the trigger condition would no longer fire.
/// </para>
/// <para>
/// Strategies can be applied at three lifecycle points:
/// <list type="bullet">
/// <item><description><b>In-run</b>: During the tool loop, before each LLM call, to keep context within token limits.</description></item>
/// <item><description><b>Pre-write</b>: Before persisting messages to storage via <see cref="ChatHistoryProvider"/>.</description></item>
/// <item><description><b>On existing storage</b>: As a maintenance operation to compact stored history.</description></item>
/// </list>
/// </para>
/// <para>
/// Multiple strategies can be composed by applying them sequentially to the same <see cref="CompactionMessageIndex"/>
/// via <see cref="PipelineCompactionStrategy"/>.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public abstract class CompactionStrategy
{
/// <summary>
/// Initializes a new instance of the <see cref="CompactionStrategy"/> class.
/// </summary>
/// <param name="trigger">
/// The <see cref="CompactionTrigger"/> that determines whether compaction should proceed.
/// </param>
/// <param name="target">
/// An optional target condition that controls when compaction stops. Strategies re-evaluate
/// this predicate after each incremental exclusion and stop when it returns <see langword="true"/>.
/// When <see langword="null"/>, defaults to the inverse of the <paramref name="trigger"/> — compaction
/// stops as soon as the trigger condition would no longer fire.
/// </param>
protected CompactionStrategy(CompactionTrigger trigger, CompactionTrigger? target = null)
{
this.Trigger = Throw.IfNull(trigger);
this.Target = target ?? (index => !trigger(index));
}
/// <summary>
/// Gets the trigger predicate that controls when compaction proceeds.
/// </summary>
protected CompactionTrigger Trigger { get; }
/// <summary>
/// Gets the target predicate that controls when compaction stops.
/// Strategies re-evaluate this after each incremental exclusion and stop when it returns <see langword="true"/>.
/// </summary>
protected CompactionTrigger Target { get; }
/// <summary>
/// Applies the strategy-specific compaction logic to the specified message index.
/// </summary>
/// <remarks>
/// This method is called by <see cref="CompactAsync"/> only when the <see cref="Trigger"/>
/// returns <see langword="true"/>. Implementations do not need to evaluate the trigger or
/// report metrics — the base class handles both. Implementations should use <see cref="Target"/>
/// to determine when to stop compacting incrementally.
/// </remarks>
/// <param name="index">The message index to compact. The strategy mutates this collection in place.</param>
/// <param name="logger">The <see cref="ILogger"/> for emitting compaction diagnostics.</param>
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests.</param>
/// <returns>A task whose result is <see langword="true"/> if any compaction was performed, <see langword="false"/> otherwise.</returns>
protected abstract ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken);
/// <summary>
/// Evaluates the <see cref="Trigger"/> and, when it fires, delegates to
/// <see cref="CompactCoreAsync"/> and reports compaction metrics.
/// </summary>
/// <param name="index">The message index to compact. The strategy mutates this collection in place.</param>
/// <param name="logger">An optional <see cref="ILogger"/> for emitting compaction diagnostics. When <see langword="null"/>, logging is disabled.</param>
/// <param name="cancellationToken">The <see cref="CancellationToken"/> to monitor for cancellation requests.</param>
/// <returns>A task representing the asynchronous operation. The task result is <see langword="true"/> if compaction occurred, <see langword="false"/> otherwise.</returns>
public async ValueTask<bool> CompactAsync(CompactionMessageIndex index, ILogger? logger = null, CancellationToken cancellationToken = default)
{
string strategyName = this.GetType().Name;
logger ??= NullLogger.Instance;
using Activity? activity = CompactionTelemetry.ActivitySource.StartActivity(CompactionTelemetry.ActivityNames.Compact);
activity?.SetTag(CompactionTelemetry.Tags.Strategy, strategyName);
if (index.IncludedNonSystemGroupCount <= 1 || !this.Trigger(index))
{
activity?.SetTag(CompactionTelemetry.Tags.Triggered, false);
logger.LogCompactionSkipped(strategyName);
return false;
}
activity?.SetTag(CompactionTelemetry.Tags.Triggered, true);
int beforeTokens = index.IncludedTokenCount;
int beforeGroups = index.IncludedGroupCount;
int beforeMessages = index.IncludedMessageCount;
Stopwatch stopwatch = Stopwatch.StartNew();
bool compacted = await this.CompactCoreAsync(index, logger, cancellationToken).ConfigureAwait(false);
stopwatch.Stop();
activity?.SetTag(CompactionTelemetry.Tags.Compacted, compacted);
if (compacted)
{
activity?
.SetTag(CompactionTelemetry.Tags.BeforeTokens, beforeTokens)
.SetTag(CompactionTelemetry.Tags.AfterTokens, index.IncludedTokenCount)
.SetTag(CompactionTelemetry.Tags.BeforeMessages, beforeMessages)
.SetTag(CompactionTelemetry.Tags.AfterMessages, index.IncludedMessageCount)
.SetTag(CompactionTelemetry.Tags.BeforeGroups, beforeGroups)
.SetTag(CompactionTelemetry.Tags.AfterGroups, index.IncludedGroupCount)
.SetTag(CompactionTelemetry.Tags.DurationMs, stopwatch.ElapsedMilliseconds);
logger.LogCompactionCompleted(
strategyName,
stopwatch.ElapsedMilliseconds,
beforeMessages,
index.IncludedMessageCount,
beforeGroups,
index.IncludedGroupCount,
beforeTokens,
index.IncludedTokenCount);
}
return compacted;
}
/// <summary>
/// Ensures the provided value is not a negative number.
/// </summary>
/// <param name="value">The target value.</param>
/// <returns>0 if negative; otherwise the value</returns>
protected static int EnsureNonNegative(int value) => Math.Max(0, value);
}
@@ -0,0 +1,45 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Provides shared telemetry infrastructure for compaction operations.
/// </summary>
internal static class CompactionTelemetry
{
/// <summary>
/// The <see cref="ActivitySource"/> used to create activities for compaction operations.
/// </summary>
public static readonly ActivitySource ActivitySource = new(OpenTelemetryConsts.DefaultSourceName);
/// <summary>
/// Activity names used by compaction tracing.
/// </summary>
public static class ActivityNames
{
public const string Compact = "compaction.compact";
public const string CompactionProviderInvoke = "compaction.provider.invoke";
public const string Summarize = "compaction.summarize";
}
/// <summary>
/// Tag names used on compaction activities.
/// </summary>
public static class Tags
{
public const string Strategy = "compaction.strategy";
public const string Triggered = "compaction.triggered";
public const string Compacted = "compaction.compacted";
public const string BeforeTokens = "compaction.before.tokens";
public const string AfterTokens = "compaction.after.tokens";
public const string BeforeMessages = "compaction.before.messages";
public const string AfterMessages = "compaction.after.messages";
public const string BeforeGroups = "compaction.before.groups";
public const string AfterGroups = "compaction.after.groups";
public const string DurationMs = "compaction.duration_ms";
public const string GroupsSummarized = "compaction.groups_summarized";
public const string SummaryLength = "compaction.summary_length";
}
}
@@ -0,0 +1,15 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Diagnostics.CodeAnalysis;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Defines a condition based on <see cref="CompactionMessageIndex"/> metrics used by a <see cref="CompactionStrategy"/>
/// to determine when to trigger compaction and when the target compaction threshold has been met.
/// </summary>
/// <param name="index">An index over conversation messages that provides group, token, message, and turn metrics.</param>
/// <returns><see langword="true"/> to indicate the condition has been met; otherwise <see langword="false"/>.</returns>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public delegate bool CompactionTrigger(CompactionMessageIndex index);
@@ -0,0 +1,134 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Diagnostics.CodeAnalysis;
using System.Linq;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// Factory to create <see cref="CompactionTrigger"/> predicates.
/// </summary>
/// <remarks>
/// <para>
/// A <see cref="CompactionTrigger"/> defines a condition based on <see cref="CompactionMessageIndex"/> metrics used
/// by a <see cref="CompactionStrategy"/> to determine when to trigger compaction and when the target
/// compaction threshold has been met.
/// </para>
/// <para>
/// Combine triggers with <see cref="All"/> or <see cref="Any"/> for compound conditions.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public static class CompactionTriggers
{
/// <summary>
/// Always trigger, regardless of the message index state.
/// </summary>
public static readonly CompactionTrigger Always =
_ => true;
/// <summary>
/// Never trigger, regardless of the message index state.
/// </summary>
public static readonly CompactionTrigger Never =
_ => false;
/// <summary>
/// Creates a trigger that fires when the included token count is below the specified maximum.
/// </summary>
/// <param name="maxTokens">The token threshold.</param>
/// <returns>A <see cref="CompactionTrigger"/> that evaluates included token count.</returns>
public static CompactionTrigger TokensBelow(int maxTokens) =>
index => index.IncludedTokenCount < maxTokens;
/// <summary>
/// Creates a trigger that fires when the included token count exceeds the specified maximum.
/// </summary>
/// <param name="maxTokens">The token threshold.</param>
/// <returns>A <see cref="CompactionTrigger"/> that evaluates included token count.</returns>
public static CompactionTrigger TokensExceed(int maxTokens) =>
index => index.IncludedTokenCount > maxTokens;
/// <summary>
/// Creates a trigger that fires when the included message count exceeds the specified maximum.
/// </summary>
/// <param name="maxMessages">The message threshold.</param>
/// <returns>A <see cref="CompactionTrigger"/> that evaluates included message count.</returns>
public static CompactionTrigger MessagesExceed(int maxMessages) =>
index => index.IncludedMessageCount > maxMessages;
/// <summary>
/// Creates a trigger that fires when the included user turn count exceeds the specified maximum.
/// </summary>
/// <param name="maxTurns">The turn threshold.</param>
/// <returns>A <see cref="CompactionTrigger"/> that evaluates included turn count.</returns>
/// <remarks>
/// <para>
/// A user turn starts with a <see cref="CompactionGroupKind.User"/> group and includes all subsequent
/// non-user, non-system groups until the next user group or end of conversation. Each group is assigned
/// a <see cref="CompactionMessageGroup.TurnIndex"/> indicating which user turn it belongs to.
/// System messages (<see cref="CompactionGroupKind.System"/>) are always assigned a <see langword="null"/>
/// <see cref="CompactionMessageGroup.TurnIndex"/> since they never belong to a user turn.
/// </para>
/// <para>
/// The turn count is the number of distinct values defined by <see cref="CompactionMessageGroup.TurnIndex"/>.
/// </para>
/// </remarks>
public static CompactionTrigger TurnsExceed(int maxTurns) =>
index => index.IncludedTurnCount > maxTurns;
/// <summary>
/// Creates a trigger that fires when the included group count exceeds the specified maximum.
/// </summary>
/// <param name="maxGroups">The group threshold.</param>
/// <returns>A <see cref="CompactionTrigger"/> that evaluates included group count.</returns>
public static CompactionTrigger GroupsExceed(int maxGroups) =>
index => index.IncludedGroupCount > maxGroups;
/// <summary>
/// Creates a trigger that fires when the included message index contains at least one
/// non-excluded <see cref="CompactionGroupKind.ToolCall"/> group.
/// </summary>
/// <returns>A <see cref="CompactionTrigger"/> that evaluates included tool call presence.</returns>
public static CompactionTrigger HasToolCalls() =>
index => index.Groups.Any(g => !g.IsExcluded && g.Kind == CompactionGroupKind.ToolCall);
/// <summary>
/// Creates a compound trigger that fires only when <b>all</b> of the specified triggers fire.
/// </summary>
/// <param name="triggers">The triggers to combine with logical AND.</param>
/// <returns>A <see cref="CompactionTrigger"/> that requires all conditions to be met.</returns>
public static CompactionTrigger All(params CompactionTrigger[] triggers) =>
index =>
{
for (int i = 0; i < triggers.Length; i++)
{
if (!triggers[i](index))
{
return false;
}
}
return true;
};
/// <summary>
/// Creates a compound trigger that fires when <b>any</b> of the specified triggers fire.
/// </summary>
/// <param name="triggers">The triggers to combine with logical OR.</param>
/// <returns>A <see cref="CompactionTrigger"/> that requires at least one condition to be met.</returns>
public static CompactionTrigger Any(params CompactionTrigger[] triggers) =>
index =>
{
for (int i = 0; i < triggers.Length; i++)
{
if (triggers[i](index))
{
return true;
}
}
return false;
};
}
@@ -0,0 +1,62 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.Logging;
using Microsoft.Shared.DiagnosticIds;
using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A compaction strategy that executes a sequential pipeline of <see cref="CompactionStrategy"/> instances
/// against the same <see cref="CompactionMessageIndex"/>.
/// </summary>
/// <remarks>
/// <para>
/// Each strategy in the pipeline operates on the result of the previous one, enabling composed behaviors
/// such as summarizing older messages first and then truncating to fit a token budget.
/// </para>
/// <para>
/// The pipeline itself always executes while each child strategy evaluates its own
/// <see cref="CompactionStrategy.Trigger"/> independently to decide whether it should compact.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class PipelineCompactionStrategy : CompactionStrategy
{
/// <summary>
/// Initializes a new instance of the <see cref="PipelineCompactionStrategy"/> class.
/// </summary>
/// <param name="strategies">The ordered sequence of strategies to execute.</param>
public PipelineCompactionStrategy(params IEnumerable<CompactionStrategy> strategies)
: base(CompactionTriggers.Always)
{
this.Strategies = [.. Throw.IfNull(strategies)];
}
/// <summary>
/// Gets the ordered list of strategies in this pipeline.
/// </summary>
public IReadOnlyList<CompactionStrategy> Strategies { get; }
/// <inheritdoc/>
protected override async ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
{
bool anyCompacted = false;
foreach (CompactionStrategy strategy in this.Strategies)
{
bool compacted = await strategy.CompactAsync(index, logger, cancellationToken).ConfigureAwait(false);
if (compacted)
{
anyCompacted = true;
}
}
return anyCompacted;
}
}
@@ -0,0 +1,140 @@
// Copyright (c) Microsoft. All rights reserved.
using System;
using System.Collections.Generic;
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.Logging;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A compaction strategy that removes the oldest user turns and their associated response groups
/// to bound conversation length.
/// </summary>
/// <remarks>
/// <para>
/// This strategy always preserves system messages. It identifies user turns in the
/// conversation (via <see cref="CompactionMessageGroup.TurnIndex"/>) and excludes the oldest turns
/// one at a time until the <see cref="CompactionStrategy.Target"/> condition is met.
/// </para>
/// <para>
/// <see cref="MinimumPreservedTurns"/> is a hard floor: even if the <see cref="CompactionStrategy.Target"/>
/// has not been reached, compaction will not touch the last <see cref="MinimumPreservedTurns"/> turns
/// (by <see cref="CompactionMessageGroup.TurnIndex"/>). Groups with a <see cref="CompactionMessageGroup.TurnIndex"/>
/// of <c>0</c> or <see langword="null"/> are always preserved regardless of this setting.
/// </para>
/// <para>
/// This strategy is more predictable than token-based truncation for bounding conversation
/// length, since it operates on logical turn boundaries rather than estimated token counts.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class SlidingWindowCompactionStrategy : CompactionStrategy
{
/// <summary>
/// The default minimum number of most-recent turns to preserve.
/// </summary>
public const int DefaultMinimumPreserved = 1;
/// <summary>
/// Initializes a new instance of the <see cref="SlidingWindowCompactionStrategy"/> class.
/// </summary>
/// <param name="trigger">
/// The <see cref="CompactionTrigger"/> that controls when compaction proceeds.
/// Use <see cref="CompactionTriggers.TurnsExceed"/> for turn-based thresholds.
/// </param>
/// <param name="minimumPreservedTurns">
/// The minimum number of most-recent turns (by <see cref="CompactionMessageGroup.TurnIndex"/>) to preserve.
/// This is a hard floor — compaction will not exclude turns within this range, regardless of the target condition.
/// Groups with <see cref="CompactionMessageGroup.TurnIndex"/> of <c>0</c> or <see langword="null"/> are always preserved.
/// </param>
/// <param name="target">
/// An optional target condition that controls when compaction stops. When <see langword="null"/>,
/// defaults to the inverse of the <paramref name="trigger"/> — compaction stops as soon as the trigger would no longer fire.
/// </param>
public SlidingWindowCompactionStrategy(CompactionTrigger trigger, int minimumPreservedTurns = DefaultMinimumPreserved, CompactionTrigger? target = null)
: base(trigger, target)
{
this.MinimumPreservedTurns = EnsureNonNegative(minimumPreservedTurns);
}
/// <summary>
/// Gets the minimum number of most-recent turns (by <see cref="CompactionMessageGroup.TurnIndex"/>) that are always preserved.
/// This is a hard floor that compaction cannot exceed, regardless of the target condition.
/// Groups with <see cref="CompactionMessageGroup.TurnIndex"/> of <c>0</c> or <see langword="null"/> are always preserved
/// independently of this value.
/// </summary>
public int MinimumPreservedTurns { get; }
/// <inheritdoc/>
protected override ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
{
// Forward pass: pre-index non-system included groups by TurnIndex.
Dictionary<int, List<int>> turnGroups = [];
List<int> turnOrder = [];
for (int i = 0; i < index.Groups.Count; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (!group.IsExcluded && group.Kind != CompactionGroupKind.System && group.TurnIndex is int turnIndex)
{
if (!turnGroups.TryGetValue(turnIndex, out List<int>? indices))
{
indices = [];
turnGroups[turnIndex] = indices;
turnOrder.Add(turnIndex);
}
indices.Add(i);
}
}
// Backward pass: identify protected turns by TurnIndex.
// TurnIndex = 0 is always protected (non-system messages before first user message).
// TurnIndex = null is always protected (system messages, already excluded from turn tracking).
HashSet<int> protectedTurnIndices = [];
if (turnGroups.ContainsKey(0))
{
protectedTurnIndices.Add(0);
}
// Protect the last MinimumPreservedTurns distinct turns.
int turnsToProtect = Math.Min(this.MinimumPreservedTurns, turnOrder.Count);
for (int i = turnOrder.Count - turnsToProtect; i < turnOrder.Count; i++)
{
protectedTurnIndices.Add(turnOrder[i]);
}
// Exclude turns oldest-first, skipping protected turns, checking target after each turn.
bool compacted = false;
for (int t = 0; t < turnOrder.Count; t++)
{
int currentTurnIndex = turnOrder[t];
if (protectedTurnIndices.Contains(currentTurnIndex))
{
continue;
}
List<int> groupIndices = turnGroups[currentTurnIndex];
for (int g = 0; g < groupIndices.Count; g++)
{
int idx = groupIndices[g];
index.Groups[idx].IsExcluded = true;
index.Groups[idx].ExcludeReason = $"Excluded by {nameof(SlidingWindowCompactionStrategy)}";
}
compacted = true;
if (this.Target(index))
{
break;
}
}
return new ValueTask<bool>(compacted);
}
}
@@ -0,0 +1,207 @@
// Copyright (c) Microsoft. All rights reserved.
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.Logging;
using Microsoft.Shared.DiagnosticIds;
using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A compaction strategy that uses an LLM to summarize older portions of the conversation,
/// replacing them with a single summary message that preserves key facts and context.
/// </summary>
/// <remarks>
/// <para>
/// This strategy protects system messages and the most recent <see cref="MinimumPreservedGroups"/>
/// non-system groups. All older groups are collected and sent to the <see cref="IChatClient"/>
/// for summarization. The resulting summary replaces those messages as a single assistant message
/// with <see cref="CompactionGroupKind.Summary"/>.
/// </para>
/// <para>
/// <see cref="MinimumPreservedGroups"/> is a hard floor: even if the <see cref="CompactionStrategy.Target"/>
/// has not been reached, compaction will not touch the last <see cref="MinimumPreservedGroups"/> non-system groups.
/// </para>
/// <para>
/// The <see cref="CompactionTrigger"/> predicate controls when compaction proceeds. Use
/// <see cref="CompactionTriggers"/> for common trigger conditions such as token thresholds.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class SummarizationCompactionStrategy : CompactionStrategy
{
/// <summary>
/// The default summarization prompt used when none is provided.
/// </summary>
public const string DefaultSummarizationPrompt =
"""
You are a conversation summarizer. Produce a concise summary of the conversation that preserves:
- Key facts, decisions, and user preferences
- Important context needed for future turns
- Tool call outcomes and their significance
Omit pleasantries and redundant exchanges. Be factual and brief.
""";
/// <summary>
/// The default minimum number of most-recent non-system groups to preserve.
/// </summary>
public const int DefaultMinimumPreserved = 8;
/// <summary>
/// Initializes a new instance of the <see cref="SummarizationCompactionStrategy"/> class.
/// </summary>
/// <param name="chatClient">The <see cref="IChatClient"/> to use for generating summaries. A smaller, faster model is recommended.</param>
/// <param name="trigger">
/// The <see cref="CompactionTrigger"/> that controls when compaction proceeds.
/// </param>
/// <param name="minimumPreservedGroups">
/// The minimum number of most-recent non-system message groups to preserve.
/// This is a hard floor — compaction will not summarize groups beyond this limit,
/// regardless of the target condition. Defaults to 8, preserving the current and recent exchanges.
/// </param>
/// <param name="summarizationPrompt">
/// An optional custom system prompt for the summarization LLM call. When <see langword="null"/>,
/// <see cref="DefaultSummarizationPrompt"/> is used.
/// </param>
/// <param name="target">
/// An optional target condition that controls when compaction stops. When <see langword="null"/>,
/// defaults to the inverse of the <paramref name="trigger"/> — compaction stops as soon as the trigger would no longer fire.
/// </param>
public SummarizationCompactionStrategy(
IChatClient chatClient,
CompactionTrigger trigger,
int minimumPreservedGroups = DefaultMinimumPreserved,
string? summarizationPrompt = null,
CompactionTrigger? target = null)
: base(trigger, target)
{
this.ChatClient = Throw.IfNull(chatClient);
this.MinimumPreservedGroups = EnsureNonNegative(minimumPreservedGroups);
this.SummarizationPrompt = summarizationPrompt ?? DefaultSummarizationPrompt;
}
/// <summary>
/// Gets the chat client used for generating summaries.
/// </summary>
public IChatClient ChatClient { get; }
/// <summary>
/// Gets the minimum number of most-recent non-system groups that are always preserved.
/// This is a hard floor that compaction cannot exceed, regardless of the target condition.
/// </summary>
public int MinimumPreservedGroups { get; }
/// <summary>
/// Gets the prompt used when requesting summaries from the chat client.
/// </summary>
public string SummarizationPrompt { get; }
/// <inheritdoc/>
protected override async ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
{
// Count non-system, non-excluded groups to determine which are protected
int nonSystemIncludedCount = 0;
for (int i = 0; i < index.Groups.Count; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (!group.IsExcluded && group.Kind != CompactionGroupKind.System)
{
nonSystemIncludedCount++;
}
}
int protectedFromEnd = Math.Min(this.MinimumPreservedGroups, nonSystemIncludedCount);
int maxSummarizable = nonSystemIncludedCount - protectedFromEnd;
if (maxSummarizable <= 0)
{
return false;
}
// Mark oldest non-system groups for summarization one at a time until the target is met.
// Track which groups were excluded so we can restore them if the LLM call fails.
List<ChatMessage> summarizationMessages = [new ChatMessage(ChatRole.System, this.SummarizationPrompt)];
List<CompactionMessageGroup> excludedGroups = [];
int insertIndex = -1;
for (int i = 0; i < index.Groups.Count && excludedGroups.Count < maxSummarizable; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (group.IsExcluded || group.Kind == CompactionGroupKind.System)
{
continue;
}
if (insertIndex < 0)
{
insertIndex = i;
}
// Collect messages from this group for summarization
summarizationMessages.AddRange(group.Messages);
group.IsExcluded = true;
group.ExcludeReason = $"Summarized by {nameof(SummarizationCompactionStrategy)}";
excludedGroups.Add(group);
// Stop marking when target condition is met
if (this.Target(index))
{
break;
}
}
// Generate summary using the chat client (single LLM call for all marked groups)
int summarized = excludedGroups.Count;
if (logger.IsEnabled(LogLevel.Debug))
{
logger.LogSummarizationStarting(summarized, summarizationMessages.Count - 1, this.ChatClient.GetType().Name);
}
using Activity? summarizeActivity = CompactionTelemetry.ActivitySource.StartActivity(CompactionTelemetry.ActivityNames.Summarize);
summarizeActivity?.SetTag(CompactionTelemetry.Tags.GroupsSummarized, summarized);
ChatResponse response;
try
{
response = await this.ChatClient.GetResponseAsync(
summarizationMessages,
cancellationToken: cancellationToken).ConfigureAwait(false);
}
catch (Exception ex) when (ex is not OperationCanceledException)
{
// Restore excluded groups so the conversation is not left in an inconsistent state
for (int i = 0; i < excludedGroups.Count; i++)
{
excludedGroups[i].IsExcluded = false;
excludedGroups[i].ExcludeReason = null;
}
logger.LogSummarizationFailed(summarized, ex.Message);
return false;
}
string summaryText = string.IsNullOrWhiteSpace(response.Text) ? "[Summary unavailable]" : response.Text;
summarizeActivity?.SetTag(CompactionTelemetry.Tags.SummaryLength, summaryText.Length);
// Insert a summary group at the position of the first summarized group
ChatMessage summaryMessage = new(ChatRole.Assistant, $"[Summary]\n{summaryText}");
(summaryMessage.AdditionalProperties ??= [])[CompactionMessageGroup.SummaryPropertyKey] = true;
index.InsertGroup(insertIndex, CompactionGroupKind.Summary, [summaryMessage]);
logger.LogSummarizationCompleted(summaryText.Length, insertIndex);
return true;
}
}
@@ -0,0 +1,234 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.Logging;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A compaction strategy that collapses old tool call groups into single concise assistant
/// messages, removing the detailed tool results while preserving a record of which tools were called
/// and what they returned.
/// </summary>
/// <remarks>
/// <para>
/// This is the gentlest compaction strategy — it does not remove any user messages or
/// plain assistant responses. It only targets <see cref="CompactionGroupKind.ToolCall"/>
/// groups outside the protected recent window, replacing each multi-message group
/// (assistant call + tool results) with a single assistant message in a YAML-like format:
/// <code>
/// [Tool Calls]
/// get_weather:
/// - Sunny and 72°F
/// search_docs:
/// - Found 3 docs
/// </code>
/// </para>
/// <para>
/// <see cref="MinimumPreservedGroups"/> is a hard floor: even if the <see cref="CompactionStrategy.Target"/>
/// has not been reached, compaction will not touch the last <see cref="MinimumPreservedGroups"/> non-system groups.
/// </para>
/// <para>
/// The <see cref="CompactionTrigger"/> predicate controls when compaction proceeds. Use
/// <see cref="CompactionTriggers"/> for common trigger conditions such as token thresholds.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class ToolResultCompactionStrategy : CompactionStrategy
{
/// <summary>
/// The default minimum number of most-recent non-system groups to preserve.
/// </summary>
public const int DefaultMinimumPreserved = 16;
/// <summary>
/// Initializes a new instance of the <see cref="ToolResultCompactionStrategy"/> class.
/// </summary>
/// <param name="trigger">
/// The <see cref="CompactionTrigger"/> that controls when compaction proceeds.
/// </param>
/// <param name="minimumPreservedGroups">
/// The minimum number of most-recent non-system message groups to preserve.
/// This is a hard floor — compaction will not collapse groups beyond this limit,
/// regardless of the target condition.
/// Defaults to <see cref="DefaultMinimumPreserved"/>, ensuring the current turn's tool interactions remain visible.
/// </param>
/// <param name="target">
/// An optional target condition that controls when compaction stops. When <see langword="null"/>,
/// defaults to the inverse of the <paramref name="trigger"/> — compaction stops as soon as the trigger would no longer fire.
/// </param>
public ToolResultCompactionStrategy(CompactionTrigger trigger, int minimumPreservedGroups = DefaultMinimumPreserved, CompactionTrigger? target = null)
: base(trigger, target)
{
this.MinimumPreservedGroups = EnsureNonNegative(minimumPreservedGroups);
}
/// <summary>
/// Gets the minimum number of most-recent non-system groups that are always preserved.
/// This is a hard floor that compaction cannot exceed, regardless of the target condition.
/// </summary>
public int MinimumPreservedGroups { get; }
/// <inheritdoc/>
protected override ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
{
// Identify protected groups: the N most-recent non-system, non-excluded groups
List<int> nonSystemIncludedIndices = [];
for (int i = 0; i < index.Groups.Count; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (!group.IsExcluded && group.Kind != CompactionGroupKind.System)
{
nonSystemIncludedIndices.Add(i);
}
}
int protectedStart = EnsureNonNegative(nonSystemIncludedIndices.Count - this.MinimumPreservedGroups);
HashSet<int> protectedGroupIndices = [];
for (int i = protectedStart; i < nonSystemIncludedIndices.Count; i++)
{
protectedGroupIndices.Add(nonSystemIncludedIndices[i]);
}
// Collect eligible tool groups in order (oldest first)
List<int> eligibleIndices = [];
for (int i = 0; i < index.Groups.Count; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (!group.IsExcluded && group.Kind == CompactionGroupKind.ToolCall && !protectedGroupIndices.Contains(i))
{
eligibleIndices.Add(i);
}
}
if (eligibleIndices.Count == 0)
{
return new ValueTask<bool>(false);
}
// Collapse one tool group at a time from oldest, re-checking target after each
bool compacted = false;
int offset = 0;
for (int e = 0; e < eligibleIndices.Count; e++)
{
int idx = eligibleIndices[e] + offset;
CompactionMessageGroup group = index.Groups[idx];
string summary = BuildToolCallSummary(group);
// Exclude the original group and insert a collapsed replacement
group.IsExcluded = true;
group.ExcludeReason = $"Collapsed by {nameof(ToolResultCompactionStrategy)}";
ChatMessage summaryMessage = new(ChatRole.Assistant, summary);
(summaryMessage.AdditionalProperties ??= [])[CompactionMessageGroup.SummaryPropertyKey] = true;
index.InsertGroup(idx + 1, CompactionGroupKind.Summary, [summaryMessage], group.TurnIndex);
offset++; // Each insertion shifts subsequent indices by 1
compacted = true;
// Stop when target condition is met
if (this.Target(index))
{
break;
}
}
return new ValueTask<bool>(compacted);
}
/// <summary>
/// Builds a concise summary string for a tool call group, including tool names,
/// results, and deduplication counts for repeated tool names.
/// </summary>
private static string BuildToolCallSummary(CompactionMessageGroup group)
{
// Collect function calls (callId, name) and results (callId → result text)
List<(string CallId, string Name)> functionCalls = [];
Dictionary<string, string> resultsByCallId = new();
List<string> plainTextResults = [];
foreach (ChatMessage message in group.Messages)
{
if (message.Contents is null)
{
continue;
}
bool hasFunctionResult = false;
foreach (AIContent content in message.Contents)
{
if (content is FunctionCallContent fcc)
{
functionCalls.Add((fcc.CallId, fcc.Name));
}
else if (content is FunctionResultContent frc && frc.CallId is not null)
{
resultsByCallId[frc.CallId] = frc.Result?.ToString() ?? string.Empty;
hasFunctionResult = true;
}
}
// Collect plain text from Tool-role messages that lack FunctionResultContent
if (!hasFunctionResult && message.Role == ChatRole.Tool && message.Text is string text)
{
plainTextResults.Add(text);
}
}
// Match function calls to their results using CallId or positional fallback,
// grouping by tool name while preserving first-seen order.
int plainTextIdx = 0;
List<string> orderedNames = [];
Dictionary<string, List<string>> groupedResults = new();
foreach ((string callId, string name) in functionCalls)
{
if (!groupedResults.TryGetValue(name, out _))
{
orderedNames.Add(name);
groupedResults[name] = [];
}
string? result = null;
if (resultsByCallId.TryGetValue(callId, out string? matchedResult))
{
result = matchedResult;
}
else if (plainTextIdx < plainTextResults.Count)
{
result = plainTextResults[plainTextIdx++];
}
if (!string.IsNullOrEmpty(result))
{
groupedResults[name].Add(result);
}
}
// Format as YAML-like block with [Tool Calls] header
List<string> lines = ["[Tool Calls]"];
foreach (string name in orderedNames)
{
List<string> results = groupedResults[name];
lines.Add($"{name}:");
if (results.Count > 0)
{
foreach (string result in results)
{
lines.Add($" - {result}");
}
}
}
return string.Join("\n", lines);
}
}
@@ -0,0 +1,110 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Diagnostics.CodeAnalysis;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Extensions.Logging;
using Microsoft.Shared.DiagnosticIds;
namespace Microsoft.Agents.AI.Compaction;
/// <summary>
/// A compaction strategy that removes the oldest non-system message groups,
/// keeping at least <see cref="MinimumPreservedGroups"/> most-recent groups intact.
/// </summary>
/// <remarks>
/// <para>
/// This strategy preserves system messages and removes the oldest non-system message groups first.
/// It respects atomic group boundaries — an assistant message with tool calls and its
/// corresponding tool result messages are always removed together.
/// </para>
/// <para>
/// <see cref="MinimumPreservedGroups"/> is a hard floor: even if the <see cref="CompactionStrategy.Target"/>
/// has not been reached, compaction will not touch the last <see cref="MinimumPreservedGroups"/> non-system groups.
/// </para>
/// <para>
/// The <see cref="CompactionTrigger"/> controls when compaction proceeds.
/// Use <see cref="CompactionTriggers"/> for common trigger conditions such as token or group thresholds.
/// </para>
/// </remarks>
[Experimental(DiagnosticIds.Experiments.AgentsAIExperiments)]
public sealed class TruncationCompactionStrategy : CompactionStrategy
{
/// <summary>
/// The default minimum number of most-recent non-system groups to preserve.
/// </summary>
public const int DefaultMinimumPreserved = 32;
/// <summary>
/// Initializes a new instance of the <see cref="TruncationCompactionStrategy"/> class.
/// </summary>
/// <param name="trigger">
/// The <see cref="CompactionTrigger"/> that controls when compaction proceeds.
/// </param>
/// <param name="minimumPreservedGroups">
/// The minimum number of most-recent non-system message groups to preserve.
/// This is a hard floor — compaction will not remove groups beyond this limit,
/// regardless of the target condition.
/// </param>
/// <param name="target">
/// An optional target condition that controls when compaction stops. When <see langword="null"/>,
/// defaults to the inverse of the <paramref name="trigger"/> — compaction stops as soon as the trigger would no longer fire.
/// </param>
public TruncationCompactionStrategy(CompactionTrigger trigger, int minimumPreservedGroups = DefaultMinimumPreserved, CompactionTrigger? target = null)
: base(trigger, target)
{
this.MinimumPreservedGroups = EnsureNonNegative(minimumPreservedGroups);
}
/// <summary>
/// Gets the minimum number of most-recent non-system message groups that are always preserved.
/// This is a hard floor that compaction cannot exceed, regardless of the target condition.
/// </summary>
public int MinimumPreservedGroups { get; }
/// <inheritdoc/>
protected override ValueTask<bool> CompactCoreAsync(CompactionMessageIndex index, ILogger logger, CancellationToken cancellationToken)
{
// Count removable (non-system, non-excluded) groups
int removableCount = 0;
for (int i = 0; i < index.Groups.Count; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (!group.IsExcluded && group.Kind != CompactionGroupKind.System)
{
removableCount++;
}
}
int maxRemovable = removableCount - this.MinimumPreservedGroups;
if (maxRemovable <= 0)
{
return new ValueTask<bool>(false);
}
// Exclude oldest non-system groups one at a time, re-checking target after each
bool compacted = false;
int removed = 0;
for (int i = 0; i < index.Groups.Count && removed < maxRemovable; i++)
{
CompactionMessageGroup group = index.Groups[i];
if (group.IsExcluded || group.Kind == CompactionGroupKind.System)
{
continue;
}
group.IsExcluded = true;
group.ExcludeReason = $"Truncated by {nameof(TruncationCompactionStrategy)}";
removed++;
compacted = true;
// Stop when target condition is met
if (this.Target(index))
{
break;
}
}
return new ValueTask<bool>(compacted);
}
}
@@ -13,6 +13,7 @@ using Microsoft.Shared.Diagnostics;
namespace Microsoft.Agents.AI;
#pragma warning disable IDE0001 // Simplify Names - Microsoft.Extensions.Logging.LogLevel.Trace doesn't get found in net472 when removing the namespace.
/// <summary>
/// A context provider that stores all chat history in a vector store and is able to
/// retrieve related chat history later to augment the current conversation.
@@ -33,8 +34,25 @@ namespace Microsoft.Agents.AI;
/// exposes a function tool that the model can invoke to retrieve relevant memories on demand instead of
/// injecting them automatically on each invocation.
/// </para>
/// <para>
/// <strong>Security considerations:</strong>
/// <list type="bullet">
/// <item><description><strong>Indirect prompt injection:</strong> Messages retrieved from the vector store via semantic search
/// are injected into the LLM context. If the vector store is compromised, adversarial content could influence LLM behavior.
/// The data returned from the store is accepted as-is without validation or sanitization.</description></item>
/// <item><description><strong>PII and sensitive data:</strong> Conversation messages (including user inputs and LLM responses)
/// are stored as vectors in the underlying store. These messages may contain PII or sensitive information. Ensure the vector
/// store is configured with appropriate access controls and encryption at rest.</description></item>
/// <item><description><strong>On-demand search tool:</strong> When using <see cref="ChatHistoryMemoryProviderOptions.SearchBehavior.OnDemandFunctionCalling"/>,
/// the AI model controls when and what to search for. The search query is AI-generated and should be treated as untrusted input
/// by the vector store implementation.</description></item>
/// <item><description><strong>Trace logging:</strong> When <see cref="Microsoft.Extensions.Logging.LogLevel.Trace"/> is enabled,
/// full search queries and results may be logged. This data may contain PII.</description></item>
/// </list>
/// </para>
/// </remarks>
public sealed class ChatHistoryMemoryProvider : MessageAIContextProvider, IDisposable
#pragma warning restore IDE0001 // Simplify Names
{
private const string DefaultContextPrompt = "## Memories\nConsider the following memories when answering user questions:";
private const int DefaultMaxResults = 3;
@@ -18,10 +18,14 @@
<ItemGroup>
<ProjectReference Include="..\Microsoft.Agents.AI.Abstractions\Microsoft.Agents.AI.Abstractions.csproj" />
</ItemGroup>
<ItemGroup>
<PackageReference Include="Microsoft.Extensions.AI" />
<PackageReference Include="Microsoft.Extensions.VectorData.Abstractions" />
<PackageReference Include="Microsoft.Extensions.DependencyInjection.Abstractions" />
<PackageReference Include="Microsoft.Extensions.Logging.Abstractions" />
<PackageReference Include="Microsoft.ML.Tokenizers" />
<PackageReference Include="System.Diagnostics.DiagnosticSource" />
</ItemGroup>
@@ -36,7 +40,7 @@
<InternalsVisibleTo Include="Microsoft.Agents.AI.UnitTests" />
<InternalsVisibleTo Include="Microsoft.Agents.AI.Declarative.UnitTests" />
<InternalsVisibleTo Include="Microsoft.Agents.AI.Hosting.UnitTests"/>
<InternalsVisibleTo Include="Microsoft.Agents.AI.Hosting.UnitTests" />
</ItemGroup>
</Project>
@@ -70,6 +70,12 @@ public sealed class OpenTelemetryAgent : DelegatingAIAgent, IDisposable
/// and outputs, such as message content, function call arguments, and function call results.
/// The default value can be overridden by setting the <c>OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT</c>
/// environment variable to "true". Explicitly setting this property will override the environment variable.
/// <para>
/// <strong>Security consideration:</strong> When sensitive data capture is enabled, the full text of chat messages —
/// including user inputs, LLM responses, function call arguments, and function results — is emitted as telemetry.
/// This data may contain PII or other sensitive information. Ensure that your telemetry pipeline is configured
/// with appropriate access controls and data retention policies.
/// </para>
/// </remarks>
public bool EnableSensitiveData
{
@@ -31,6 +31,18 @@ namespace Microsoft.Agents.AI;
/// to the current request messages when forming the search input. This can improve search relevance by providing
/// multi-turn context to the retrieval layer without permanently altering the conversation history.
/// </para>
/// <para>
/// <strong>Security considerations:</strong> Search results retrieved from external sources are injected into the LLM context and may
/// contain adversarial content designed to manipulate LLM behavior via indirect prompt injection. Developers should be aware that:
/// <list type="bullet">
/// <item><description>The search query may be constructed from user input or LLM-generated content, both of which are untrusted.
/// Implementers of the search delegate should validate search inputs and apply appropriate access controls to search results.</description></item>
/// <item><description>Retrieved documents are formatted and injected as messages in the AI request context. If the external data source
/// is compromised, adversarial content could influence the LLM's responses.</description></item>
/// <item><description>When using <see cref="TextSearchProviderOptions.TextSearchBehavior.OnDemandFunctionCalling"/>, the AI model controls
/// when and what to search for — the search query text is AI-generated and should be treated as untrusted input by the search implementation.</description></item>
/// </list>
/// </para>
/// </remarks>
public sealed class TextSearchProvider : MessageAIContextProvider
{
@@ -14,6 +14,8 @@ namespace AzureAIAgentsPersistent.IntegrationTests;
public class AzureAIAgentsPersistentCreateTests
{
private const string SkipCodeInterpreterReason = "Azure AI Code Interpreter intermittently fails to execute uploaded files in CI";
private readonly PersistentAgentsClient _persistentAgentsClient = new(TestConfiguration.GetRequiredValue(TestSettings.AzureAIProjectEndpoint), TestAzureCliCredentials.CreateAzureCliCredential());
[Theory]
@@ -131,11 +133,11 @@ public class AzureAIAgentsPersistentCreateTests
}
}
[Fact]
[Fact(Skip = SkipCodeInterpreterReason)]
public Task CreateAgent_CreatesAgentWithCodeInterpreter_ChatClientAgentOptionsAsync()
=> this.CreateAgent_CreatesAgentWithCodeInterpreterAsync("CreateWithChatClientAgentOptionsAsync");
[RetryFact(Constants.RetryCount, Constants.RetryDelay)]
[Fact(Skip = SkipCodeInterpreterReason)]
public Task CreateAgent_CreatesAgentWithCodeInterpreter_FoundryOptionsAsync()
=> this.CreateAgent_CreatesAgentWithCodeInterpreterAsync("CreateWithFoundryOptionsAsync");
@@ -126,6 +126,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.Single(result.Messages);
Assert.Equal(ChatRole.Assistant, result.Messages[0].Role);
Assert.Equal("Hello! How can I help you today?", result.Messages[0].Text);
Assert.Equal(ChatFinishReason.Stop, result.FinishReason);
}
[Fact]
@@ -249,8 +250,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.Equal("stream-1", updates[0].MessageId);
Assert.Equal(this._agent.Id, updates[0].AgentId);
Assert.Equal("stream-1", updates[0].ResponseId);
Assert.NotNull(updates[0].RawRepresentation);
Assert.Equal(ChatFinishReason.Stop, updates[0].FinishReason);
Assert.IsType<AgentMessage>(updates[0].RawRepresentation);
Assert.Equal("stream-1", ((AgentMessage)updates[0].RawRepresentation!).MessageId);
}
@@ -501,8 +501,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.NotNull(result);
Assert.Equal(this._agent.Id, result.AgentId);
Assert.Equal("task-789", result.ResponseId);
Assert.NotNull(result.RawRepresentation);
Assert.Null(result.FinishReason);
Assert.IsType<AgentTask>(result.RawRepresentation);
Assert.Equal("task-789", ((AgentTask)result.RawRepresentation).Id);
@@ -552,6 +551,15 @@ public sealed class A2AAgentTests : IDisposable
{
Assert.Null(result.ContinuationToken);
}
if (taskState is TaskState.Completed)
{
Assert.Equal(ChatFinishReason.Stop, result.FinishReason);
}
else
{
Assert.Null(result.FinishReason);
}
}
[Fact]
@@ -661,6 +669,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.Equal(MessageId, update0.ResponseId);
Assert.Equal(this._agent.Id, update0.AgentId);
Assert.Equal(MessageText, update0.Text);
Assert.Equal(ChatFinishReason.Stop, update0.FinishReason);
Assert.IsType<AgentMessage>(update0.RawRepresentation);
Assert.Equal(MessageId, ((AgentMessage)update0.RawRepresentation!).MessageId);
}
@@ -702,6 +711,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.Equal(ChatRole.Assistant, update0.Role);
Assert.Equal(TaskId, update0.ResponseId);
Assert.Equal(this._agent.Id, update0.AgentId);
Assert.Null(update0.FinishReason);
Assert.IsType<AgentTask>(update0.RawRepresentation);
Assert.Equal(TaskId, ((AgentTask)update0.RawRepresentation!).Id);
@@ -741,6 +751,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.Equal(ChatRole.Assistant, update0.Role);
Assert.Equal(TaskId, update0.ResponseId);
Assert.Equal(this._agent.Id, update0.AgentId);
Assert.Null(update0.FinishReason);
Assert.IsType<TaskStatusUpdateEvent>(update0.RawRepresentation);
// Assert - session should be updated with context and task IDs
@@ -784,6 +795,7 @@ public sealed class A2AAgentTests : IDisposable
Assert.Equal(ChatRole.Assistant, update0.Role);
Assert.Equal(TaskId, update0.ResponseId);
Assert.Equal(this._agent.Id, update0.AgentId);
Assert.Null(update0.FinishReason);
Assert.IsType<TaskArtifactUpdateEvent>(update0.RawRepresentation);
// Assert - artifact content should be in the update
@@ -1,67 +0,0 @@
// Copyright (c) Microsoft. All rights reserved.
using System.Collections.Generic;
using System.Text.Json;
using A2A;
namespace Microsoft.Agents.AI.A2A.UnitTests;
/// <summary>
/// Unit tests for the <see cref="A2AMetadataExtensions"/> class.
/// </summary>
public sealed class A2AMetadataExtensionsTests
{
[Fact]
public void ToAdditionalProperties_WithNullMetadata_ReturnsNull()
{
// Arrange
Dictionary<string, JsonElement>? metadata = null;
// Act
var result = metadata.ToAdditionalProperties();
// Assert
Assert.Null(result);
}
[Fact]
public void ToAdditionalProperties_WithEmptyMetadata_ReturnsNull()
{
// Arrange
var metadata = new Dictionary<string, JsonElement>();
// Act
var result = metadata.ToAdditionalProperties();
// Assert
Assert.Null(result);
}
[Fact]
public void ToAdditionalProperties_WithMultipleProperties_ReturnsAdditionalPropertiesDictionaryWithAllProperties()
{
// Arrange
var metadata = new Dictionary<string, JsonElement>
{
{ "stringKey", JsonSerializer.SerializeToElement("stringValue") },
{ "numberKey", JsonSerializer.SerializeToElement(42) },
{ "booleanKey", JsonSerializer.SerializeToElement(true) }
};
// Act
var result = metadata.ToAdditionalProperties();
// Assert
Assert.NotNull(result);
Assert.Equal(3, result.Count);
Assert.True(result.ContainsKey("stringKey"));
Assert.Equal("stringValue", ((JsonElement)result["stringKey"]!).GetString());
Assert.True(result.ContainsKey("numberKey"));
Assert.Equal(42, ((JsonElement)result["numberKey"]!).GetInt32());
Assert.True(result.ContainsKey("booleanKey"));
Assert.True(((JsonElement)result["booleanKey"]!).GetBoolean());
}
}

Some files were not shown because too many files have changed in this diff Show More