mirror of
https://github.com/microsoft/agent-framework.git
synced 2026-06-16 21:04:09 +08:00
Eduard van Valkenburg
4609535e22
* Python: feat: add agent-framework-monty (Monty-backed CodeAct)
New alpha package that wraps pydantic-monty (a Rust-based Python
interpreter) behind the same CodeAct API surface as
agent-framework-hyperlight, so users can swap providers with minimal
code change.
Public API (agent_framework_monty):
- MontyCodeActProvider — ContextProvider that injects a run-scoped
execute_code tool plus dynamic CodeAct instructions.
- MontyExecuteCodeTool — standalone FunctionTool for mixed-tool agents
or manual static wiring.
- FileMount / FileMountInput / MountMode — public types mirroring the
Hyperlight names, with Monty's mode (read-only/read-write/overlay)
and write_bytes_limit on FileMount.
Constructor kwargs (both classes) mirror Hyperlight where possible:
tools, approval_mode, workspace_root, file_mounts; plus a Monty-only
resource_limits forwarding ResourceLimits to Monty.start().
Filesystem flow:
- workspace_root auto-mounts at /input (read-write), matching Hyperlight.
- file_mounts accepts string shorthand, (host, mount) tuple, or
FileMount with mode + write cap.
- Files written under read-write mounts are scanned post-execution and
returned as Content.from_data items (mirrors Hyperlight /output).
- overlay mounts buffer writes in-memory; read-only mounts reject writes.
Internals:
- _monty_bridge.InlineCodeBridge ports the inline (non-durable) bridge
from anthonychu/maf-codeact-monty-python; handles FunctionSnapshot /
FutureSnapshot pause/resume, dispatches direct typed calls + the
call_tool fallback, forwards mount/limits to Monty.start(...).
- generate_type_stubs emits per-tool stubs so Monty's `ty` type-checker
rejects bad calls before any host tool runs.
Alpha-policy compliance (per python-package-management skill):
- Added agent-framework-monty = { workspace = true } to root
pyproject.toml.
- Added row to python/PACKAGE_STATUS.md.
- Added monty entry under Experimental in python/AGENTS.md.
- NOT added to core[all]; NO agent_framework.monty lazy shim (deferred
to beta promotion).
Samples (three sets, import from agent_framework_monty directly):
- samples/02-agents/context_providers/code_act/monty_code_act.py
(provider pattern) + updated local README.
- samples/02-agents/tools/monty_code_interpreter/ (standalone +
manual-wiring + README).
- samples/04-hosting/foundry-hosted-agents/responses/11_monty_codeact/
(full hosted-agent layout with uv-based pyproject.toml + Dockerfile,
Azure Monitor wiring via APPLICATIONINSIGHTS_CONNECTION_STRING +
enable_instrumentation, ENABLE_INSTRUMENTATION and
ENABLE_SENSITIVE_DATA env vars). The alpha wheel is vendored into
./wheels/ (gitignored) via vendor-wheel.sh; new row added to the
parent Responses-API README.
Tests:
- 28 hermetic unit tests (stubbed pydantic_monty).
- 18 integration tests marked @pytest.mark.integration, auto-skipped
when pydantic_monty is unimportable; exercise the real Monty
runtime: print round-trip, last-expression value, direct typed
tool dispatch, call_tool fallback, async tool, asyncio.gather
parallelism, ty type-check rejection, OS blocked by default,
workspace_root read+write capture, read-only / overlay mount
semantics, resource_limits.max_duration_secs abort, approval
gating end-to-end, full Agent run with a scripted chat client.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: fix: monty FileMount test compares against the normalized POSIX path
The shorthand string mount goes through _normalize_mount_path, which
rewrites Windows drive letters like 'C:\\Users\\...' into
'/C:/Users/...' (POSIX-style). The Windows CI runners surfaced this
because tmp_path resolves to a backslashed Windows path; the test was
comparing against the raw str(host_a) instead of the normalized form.
Compare against _normalize_mount_path(str(host_a)) so the assertion is
platform-independent.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: fix: address PR #5915 review feedback
- _execute_code_tool docstring: clarify that the Monty backend supports
scoped filesystem access via workspace_root / file_mounts (blocked by
default).
- _to_monty_mount: import pydantic_monty lazily through load_monty so
missing-dependency errors surface as the same actionable RuntimeError
the rest of the package raises (not a bare ImportError at module load).
Renamed _load_monty -> load_monty for the same reason.
- _python_type_repr: emit None for type(None) instead of Any, and
normalize both typing.Union[...] and PEP-604 X | Y to PEP-604 syntax
so Optional[X] / Union[..., None] / -> None signatures round-trip
correctly through ty validation. Added a regression test.
- _PrintCollector: track a running character count instead of
recomputing sum(len(c) for c in self.chunks) per callback. Eliminates
the O(n^2) cost on print-heavy code.
- Instructions: mention that the value of the final expression is also
returned alongside captured stdout (matches actual behavior).
- 11_monty_codeact Dockerfile: pin ghcr.io/astral-sh/uv to 0.11.6
instead of :latest for reproducible builds.
- 11_monty_codeact README: replace the bare "see parent README" pointer
with sample-specific steps (./vendor-wheel.sh + uv sync + uv run),
since the sample uses pyproject.toml + a vendored wheel rather than
requirements.txt.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: sample: 11_monty_codeact installs agent-framework-monty from PyPI
Drop the vendored-wheel scaffolding now that agent-framework-monty is on
PyPI as an alpha (1.0.0a*) release:
- pyproject.toml: remove [tool.uv.sources] override; keep [tool.uv]
prerelease = "allow" so uv pulls the alpha automatically.
- Dockerfile: drop the COPY wheels/ step.
- README: drop the ./vendor-wheel.sh setup step and the
not-yet-on-PyPI warning.
- Delete vendor-wheel.sh and the gitignored wheels/ directory.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: fix(monty): harden post-execution file capture against symlink escape
Same class of issue as the MSRC-reported Hyperlight finding: the
post-execution capture walked workspace_root with Path.rglob() +
is_file() + read_bytes() - all of which follow symlinks. An attacker
who controls the workspace (cloned repo, extracted archive, shared
workspace) could pre-place `workspace/leak.txt -> /etc/passwd` or
`workspace/outside_dir -> /etc/` and have host files surface as
captured Content items.
Monty's mount layer already rejects symlink reads from inside the
sandbox across all three modes (verified empirically), so the runtime
path was safe. This commit closes the post-execution scan path.
Changes:
- New `_iter_real_files(root)` walker that uses iterdir() +
is_symlink() to skip symlinks at every directory level and yields
only real files. Replaces the previous `host_root.rglob("*")` calls
in both `_snapshot_writable_mounts` and `_capture_written_files`.
- Use `Path.lstat()` instead of `Path.stat()` so size/mtime can never
be taken from a symlink target.
- Three new integration tests reproducing the MSRC attack shape
against the workspace_root flow: symlink-to-file outside workspace,
symlink-to-directory outside workspace, and a guard ensuring
legitimate sandbox writes are still captured when symlinks are
present.
Per user request, hyperlight is untouched in this commit (separate fix).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: fix(monty): skip symlink regression tests when unsupported
Apply the same Windows-CI safety guard as the hyperlight fix in PR #5919:
the three symlink integration tests create symlinks via Path.symlink_to(),
which fails with OSError / NotImplementedError on unprivileged Windows
runners. Add a local _symlinks_supported helper (mirroring the one in
packages/core/tests/core/test_skills.py) and pytest.skip when symlinks
aren't available, so the tests no longer fail for environment reasons.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: fix(monty): address PR #5915 follow-up review feedback
- _invoke_tool: drop the inspect.iscoroutinefunction(...) branch and
always `await self.tool_map[name](**kwargs)`. Every entry in
tool_map is `partial(FunctionTool.invoke, skip_parsing=True)` and
FunctionTool.invoke is `async def`, so the branching was dead code -
and on Python versions affected by cpython#98590,
iscoroutinefunction(partial(bound_async_method, ...)) returns False,
causing the bridge to take the asyncio.to_thread path, return an
unawaited coroutine, and surface it as a JSON-serialization failure
for every tool call. Added a regression test
test_invoke_tool_awaits_partial_wrapped_async_method.
- generate_type_stubs: skip tools whose name is not a valid Python
identifier or is a Python keyword. FunctionTool.name has no upstream
validation, so a name like "weird-name" produced a syntax error in
the stubs and a name like "broken\n pass\nasync def injected"
would inject arbitrary stub source. Non-identifier names stay
reachable via `call_tool("weird-name", ...)` at runtime; they just
don't get type-checked stubs. Added regression test
test_generate_type_stubs_skips_non_identifier_tool_names.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
4609535e22
·
2026-05-20 00:35:23 +00:00
History
CodeAct context providers
Demonstrates the provider-owned CodeAct flow with two backends:
| File | Backend | Notes |
|---|---|---|
code_act.py |
Hyperlight WASM sandbox via HyperlightCodeActProvider |
Hardened sandbox with WASM isolation; sandbox tools called via call_tool(...). |
monty_code_act.py |
Monty Rust-based Python interpreter via MontyCodeActProvider (alpha) |
Cross-platform pure interpreter; sandbox tools can be called as typed async functions (await compute(...)) or via call_tool(...). |
Both providers inject an execute_code tool into the agent and keep the
registered sandbox tools (compute, fetch_data) hidden from the model — the
model invokes them from inside the sandbox.
Installation
pip install agent-framework agent-framework-hyperlight --pre # Hyperlight sample
pip install agent-framework agent-framework-monty --pre # Monty sample
The Hyperlight Wasm backend is currently published only for
linux/x86_64andwin32/AMD64with Python<3.14. On other platformsexecute_codewill fail at runtime when it tries to create the sandbox.Monty is cross-platform and has no hypervisor/WASM backend dependency, but it interprets a Python subset (e.g.
os/network/subprocess access is blocked).agent-framework-montyis an alpha package and is not yet part ofagent-framework[all]; install it explicitly with--pre.
Prerequisites
- An Azure AI Foundry project endpoint (
FOUNDRY_PROJECT_ENDPOINT) - A deployed model (
FOUNDRY_MODEL) - Azure CLI authenticated (
az login)
Run
python code_act.py # Hyperlight
python monty_code_act.py # Monty
See the source files for the full annotated examples.