* Python: make FoundryToolbox.as_skills_provider() disable_caching effective
as_skills_provider() forwarded disable_caching to SkillsProvider, which
ignores it for a caller-supplied SkillsSource, so it was a no-op and the
toolbox re-read skill://index.json on every agent run.
Compose caching in as_skills_provider() instead: wrap the context-independent
_FoundryToolboxSkillsSource in DeduplicatingSkillsSource(CachingSkillsSource(...)).
Add a cache_refresh_interval param, fix the docstring, and add tests covering
cached, disabled, and refresh-interval behavior.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
* Clarify caller-invariant skill-set wording in as_skills_provider docs
Emphasize that the toolbox advertises the same skill set to every caller (the
per-request call-id governs execution/authorization, not which skills are
listed) rather than leaning on 'ignores SkillsSourceContext'.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
* Make MCP skills reconnect-safe via session_provider
Cached MCPSkill objects captured the MCP ClientSession at construction, so
after a FoundryToolbox reconnect (which replaces its session) load_skill and
read_skill_resource would fail against the closed session. This regressed once
as_skills_provider() started caching discovery by default.
Add an optional session_provider callable to MCPSkillsSource and MCPSkill
(exactly one of client or session_provider). When supplied, the session is
resolved on every fetch, mirroring how MCPTool resolves self.session live at
call time. _FoundryToolboxSkillsSource now passes a provider that returns the
toolbox's current session, so cached skills always use the live session.
The fixed client= path is unchanged and backward-compatible. Update core tests,
foundry_hosting tests, and core AGENTS.md.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
* Fix ty error: type captured session_provider as Callable in test
ty could not call the provider narrowed from \object\ (Top callable). Type the
captured value as Callable[[], object] and drop the redundant callable() assert.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
* Simplify _resolve_mcp_session_provider per review
Address review feedback: replace the dense (client is None) == (session_provider
is None) guard with explicit branches, and drop the cast by binding the narrowed
client to a typed local. Keeps strict 'exactly one' semantics (raises on both and
on neither), matching the codebase convention (e.g. security.py mcp_tool/url).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
* Add PR #7135 entries to the 1.12.0 changelog
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
* Drop redundant @pytest.mark.asyncio from MCP skills tests
asyncio_mode is 'auto', so the marker is unnecessary. Remove it from the whole
file for consistency with the async-by-default convention. Per review feedback.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 84150ec4-6f7c-4ef8-b9fb-12fa11652773
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Bump Python package versions for 1.12.0 release
Bump packages represented in the 1.12.0 changelog, promote Foundry Hosting, Azure Content Understanding, Gemini, Mistral, Monty, and Tools to beta, and apply the requested beta cohort date stamp. Root and core move to 1.12.0, released and RC packages use their selected increments, alpha packages including Hosting MCP use the 260721 stamp, and core floors are raised only for proven consumers.
Copilot-Session: 2dd9980a-b869-4c16-8642-75b7a6d6ebdf
* fix version in readme
* Add Responses conversation ID changes to release notes
Include the breaking Hosting Responses conversation ID helper changes from #7234 in the Python 1.12.0 changelog.
Copilot-Session: 2dd9980a-b869-4c16-8642-75b7a6d6ebdf
* Python: Fix PropertySchema.to_json_schema() not recursing into nested schemas
Nested array 'items' and object 'properties' kept the declarative 'kind'
key and empty 'enum' placeholders, producing JSON Schema OpenAI rejects
('schema must have a type key'). Recursively apply the same conversion the
top-level properties loop performs, including the serialized named-list
properties shape and nested required arrays.
Fixes#7198
(cherry picked from commit c156ffd05924fb5a1884625f2fc3d9bdc3e152b1)
* Python: Validate nested properties list before mutating to avoid partial conversion
Review feedback: the list-shaped properties branch popped name/required from
each element and returned on the first unexpected one, leaving earlier
elements half-converted. Validate the whole list first so an unexpected
shape leaves the node fully untouched.
* Python: Type nested-properties normalization for strict Pyright and drop unreachable dict branch
ObjectProperty always stores nested properties as a named list, so the
elif-dict branch in _normalize_nested_schemas was unreachable; remove it
and flatten the list conversion behind an early return. Cast the narrowed
items/props values so strict Pyright no longer reports unknown types.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Python: Emit additionalProperties: false on nested object nodes in PropertySchema.to_json_schema()
OpenAI strict structured outputs require additionalProperties: false on
every object node, but the chat clients only inject it at the schema
root, so declarative schemas with nested objects (e.g. array items)
failed with a schema-validation 400. Route the top-level properties loop
through _normalize_schema_node so all object nodes get the key, and add
a live OpenAI integration test covering the nested array-of-objects
response_format shape.
Verified live against the Responses API: the previous emission fails
with "In context=('properties', 'issues', 'items'),
'additionalProperties' is required to be supplied and to be false";
the new emission returns valid structured output.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Python: fix header_provider headers not reaching streamable HTTP requests
MCPStreamableHTTPTool.call_tool stores header_provider output in a
ContextVar, but the streamable HTTP transport sends requests from tasks
spawned at connect time, whose contexts never observe values set later.
The request hook therefore always read an empty dict on real connections
and the per-call headers (e.g. Authorization) were silently dropped.
Keep the ContextVar for in-context reads and add an instance-level
snapshot of the active call's headers that the request hook falls back
to across tasks.
* Python: serialize header_provider tool calls to prevent cross-call header mixing
Parallel tool invocations run concurrently per function-invocation batch,
so two call_tool invocations on the same MCPStreamableHTTPTool could
overwrite each other's active-header snapshot while requests were still
in flight, attaching the wrong per-call credentials. Hold a per-instance
lock for the duration of a header-bearing call, add a regression test
that fails without the lock, and normalize captured header casing in the
transport-task test.
* Python: forward GitHubCopilotOptions verbatim to create_session
Refactor the GitHub Copilot agent to forward the full options dict to the
Copilot SDK's create_session/resume_session instead of hand-mapping a fixed
subset. GitHubCopilotOptions stays as the curated, typed surface, but any
other create_session parameter (reasoning_effort, context_tier,
enable_citations, ...) is now passed through verbatim. Unknown keys surface
as TypeError from the SDK instead of being silently dropped.
De-duplicates the near-identical _create_session/_resume_session bodies into
a shared _build_session_kwargs helper.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: f9d016d5-4d8c-43f0-a8fb-f3cf3d1ad7eb
* Python: address review feedback on GHCP options passthrough
- Strip agent-internal/client-level keys (on_pre_tool_use, on_function_approval,
timeout, cli_path, log_level, base_directory) from the forwarded kwargs so they
cannot leak into create_session/resume_session and raise TypeError.
- Source caller tools from the merged options layer so tools supplied via
default_options are honored instead of silently dropped.
- Honor a caller-supplied native 'hooks' dict in _build_session_hooks (composing
with the on_pre_tool_use shortcut) instead of unconditionally overwriting it.
- Validate mock create_session/resume_session calls against the real SDK
signatures in tests so invalid kwargs surface as TypeError, and add regression
tests for the passthrough contract.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: f9d016d5-4d8c-43f0-a8fb-f3cf3d1ad7eb
* Python: avoid redundant re-read of model in _build_session_kwargs
model is popped from default_options into settings at init, so a per-run
model already lands in the merged kwargs. Keep that value when present and
only fall back to the resolved setting otherwise, instead of re-reading opts.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: f9d016d5-4d8c-43f0-a8fb-f3cf3d1ad7eb
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add agent-framework-azure-cosmos-memory context provider (draft)
Introduces CosmosMemoryContextProvider, a ContextProvider that wraps the azure-cosmos-agent-memory toolkit to give agents long-term, Cosmos DB-backed memory (fact/procedural recall + user summaries). Includes package scaffolding, unit tests (mocked client), live Azure integration tests (marked), samples, README, and AGENTS.md.
Draft: uv.lock is intentionally left unchanged. This package depends on azure-cosmos-agent-memory (requires Python >=3.11), which is unsatisfiable against the workspace's current >=3.10 floor, so adding it to the shared lock requires a workspace decision (raise floor to 3.11 or exclude from workspace). Test coverage to be expanded.
* ci: exclude azure-cosmos-memory from uv workspace resolution
The package depends on azure-cosmos-agent-memory which requires Python
>=3.11 and a prompty pre-release (>=2.0.0a9). Both are unsatisfiable
against the workspace's >=3.10 floor and pre-release policy, causing
uv sync to fail in every Python CI job. Exclude the package from the
shared workspace so it is resolved and tested as a standalone package.
* ci: fix code-quality failures for azure-cosmos-memory
- Strip trailing whitespace from package files (pre-commit trailing-whitespace hook)
- Exclude the package README from markdown-code-lint: the package is excluded
from the uv workspace, so its README snippets import a module that is not
installed in the workspace env and Pyright cannot resolve it
* Exclude azure-cosmos-memory README from markdown-code-lint task
* Address PR review comments on cosmos-memory context provider
- Wire credential into Cosmos and AI Foundry clients; let toolkit own
DefaultAzureCredential when none supplied (remove dead import).
- Honor auto_extract=False by zeroing extraction/summary cadence thresholds.
- Skip whitespace-only conversation turns and store stripped content.
- Show confidence 0.0 and coerce confidence to float in _format_memories.
- Register both 'integration' and 'azure' pytest markers accurately.
- Fix duplicated install block in README.
- Update and extend unit tests for new credential wiring and fixes.
* Include azure-cosmos-memory in the uv workspace
Follow the github_copilot pattern for a package with a Python 3.11-only
dependency: lower requires-python to >=3.10 and gate azure-cosmos-agent-memory
behind a python_version >= '3.11' marker. Add a direct, gated prompty
pre-release dependency so the workspace's if-necessary-or-explicit prerelease
policy permits the toolkit's transitive prompty requirement. Guard the test
modules with pytest.importorskip so the 3.10 CI leg skips cleanly. Remove the
workspace exclude and the markdown-code-lint exclude, and regenerate uv.lock.
* Address review feedback on cosmos-memory provider
Rename provider parameters to match Agent Framework conventions:
foundry_endpoint (was ai_foundry_endpoint) and embedding_model/chat_model
(were *_deployment_name). Move DEFAULT_* to module-level constants, type
memory_types as a Literal, use DEFAULT_CONTEXT_PROMPT as the default value,
and add ProcessorConfig/CosmosMemorySettings TypedDicts. Resolve connection
settings via agent_framework load_settings with required-field validation,
replacing the manual getenv/raise blocks. Scope user_id/thread_id to the
provider state and drop the unpreventable first-turn warning.
Rewrite the samples around Agent (not raw SessionContext), provider-scoped
state, and session-id threading; use PEP 723 inline dependencies instead of a
samples dependency group; use a plain input() loop; remove the dead custom
processor stub. Update README/AGENTS for the renamed parameters and env vars.
Add a samples ruff per-file-ignores entry now that the package is linted in CI.
* Add emulator-backed vector search integration test
Bump azure-cosmos-agent-memory to >=0.2.0b2 (adds the embeddings/chat client
injection seam) and add tests/test_emulator.py: an integration (not azure)
suite that exercises real Cosmos vector search with a quantizedFlat index
against a local Cosmos DB emulator, using deterministic in-memory fakes for
embeddings and chat so no Azure AI Foundry account or LLM is required.
To run on a stock emulator the fixture strips the toolkit's full-text index
(the provider only does pure vector search) and requests provisioned autoscale
throughput instead of serverless. The suite skips cleanly when no emulator is
reachable.
* Fix CI typing and package checks for azure-cosmos-memory
The package recently joined the uv workspace, so its source and tests are now covered by the Test Typing Checks and Package Checks gates for the first time.
tests: rename stale constructor kwargs to the current provider API (foundry_endpoint/embedding_model/chat_model); use a typed _STUB_AGENT for the unused agent param so pyright/pyrefly/ty/zuban all accept it; make processor_config values ints; assert non-None memory_client in the emulator tests.
source: relax reportUnknown*/reportOptional* for this package only (the toolkit ships no py.typed; mirrors the hosting-telegram precedent); decouple the conditional toolkit import from the annotation type; use settings.get(); fix memory_types list invariance; drop a redundant None guard; read role via getattr.
* Apply pyupgrade: single-arg AsyncGenerator in test_integration
* Make Cosmos memory extraction drain transparently on provider exit
The provider now drains in-flight background memory extraction in __aexit__, so applications no longer need to call flush() in their own control flow; the client's close() would otherwise cancel pending extraction tasks. flush() is hardened against clients that expose no usable background-task registry.
sample: interactive_chat reads input via asyncio.to_thread so the event loop stays free and background extraction runs during the session; removes the manual flush now that the provider drains on exit.
tests: add explicit transparent-extraction integration tests (emulator: after_run schedules extraction and __aexit__ drains it; live Azure: a fact is extracted and recalled in a later session with no manual flush). Emulator tests reuse a single fixed database to avoid exhausting the emulator's partition budget across runs.
* Add custom extraction-prompt seam and sample to cosmos-memory provider
Adds a prompts_dir option to CosmosMemoryContextProvider that points the Agent Memory Toolkit pipeline at a caller-supplied directory of Prompty templates, so callers can override extract_memories.prompty to control what the extraction LLM produces. The toolkit exposes no public prompts-directory seam, so the provider contains the one internal touch (swapping the pipeline's template loader after the store connects); applies to both provider-built and supplied clients.
sample: interactive_chat_custom_extraction.py - the interactive chat wired with a custom coding-assistant extraction rubric. It derives a complete prompts directory at runtime (copies the bundled templates and augments extract_memories.prompty) so it stays schema-compatible with the installed toolkit.
tests: unit tests assert the provider redirects the pipeline loader only when prompts_dir is set; an emulator integration test proves end to end that a unique marker in a custom extract_memories.prompty reaches the extraction LLM call.
* docs: document prompts_dir custom-extraction seam in cosmos-memory README
Replaces the stale, non-functional CustomMemoryProcessor snippet with the working prompts_dir approach, lists the new interactive_chat_custom_extraction.py sample, and corrects the interactive-sample feature list.
* Address review: rename _new_session, drop defensive toolkit import guard
Sample (comment): rename _new_thread to _new_session in both interactive samples (a new session is the new thread).
Provider (comment): replace the _memory_toolkit_available flag + __init__ ImportError guard with a plain guarded import that re-raises a clear ImportError, matching the github_copilot package's pattern for its 3.11-only SDK. Kept requires-python >=3.10 (bumping this one workspace member to 3.11 would force the entire uv workspace lock floor to 3.11). Tests now run importorskip before importing the package, mirroring github_copilot.
* Pass cadence via cadence_thresholds instead of mutating os.environ
* Mark package alpha and drop private naming in samples
* Require Python 3.11 and inject user summary as untrusted context
* CI: exclude azure-cosmos-memory from uv sync on Python 3.10
* Re-trigger CI (flaky external link check)
* Require chat/embedding models instead of silent defaults
* Fix pyright: narrow resolved chat/embedding models to str
---------
Co-authored-by: Theo van Kraay <thvankra@microsoft.com>
TokenBudgetComposedStrategy estimates tokens by feeding a JSON-serialized
message to the tokenizer, but _serialize_message() used ensure_ascii=True.
That escapes non-ASCII text into \uXXXX sequences, so CJK and other
non-Latin content is token-counted as the escape sequences rather than the
characters the model actually sees, inflating the estimate (~1.6x for mixed
Japanese, more for pure CJK) and skewing compaction/token-budget decisions.
Serialize with ensure_ascii=False, matching the ensure_ascii=False already
used elsewhere in this module. Only affects token estimation; the serialized
string is never stored or transmitted.
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
* Python: preserve explicit null arguments in auto function calling
FunctionTool.invoke dumped validated arguments with model_dump(exclude_none=True),
which strips any argument the model set to null. A required nullable parameter
(e.g. unit: Literal["C","F"] | None) that the model deliberately sets to null was
therefore dropped, and the function failed to invoke on the missing argument.
Use exclude_unset instead: keep the arguments the model actually provided (null
included) and omit only the ones it left out, so the function's own defaults still
apply. Because the input model is generated from the function signature, its field
defaults match the signature defaults, so omitted optionals are unchanged.
Fixes#5934
Signed-off-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
* Python: extend null-arg fix to the auto function-calling path
The earlier change fixed FunctionTool.invoke, but _auto_invoke_function
(the path a model-emitted function_call actually takes) still ran
model_dump(exclude_none=True), so an explicit null for a required
nullable argument was still dropped and the call failed with a missing
argument. Switch it to exclude_unset to match invoke, and add a
regression test that drives _auto_invoke_function with an explicit null.
---------
Signed-off-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
* feat(durabletask): surface HITL respond-URL addressing to workflow executors
Let a workflow notify a human reviewer (e.g. email an approval link) from inside the
graph, without the caller threading the instanceId/requestId by hand.
- durabletask: the orchestrator injects host_context {instance_id, workflow_name,
request_path_prefix} into each activity input; CapturingRunnerContext surfaces it as
host_metadata. No new core API.
- azurefunctions: WorkflowHitlContext.from_context(ctx) builds the canonical
respond/status URLs (returns None in-process so callers degrade gracefully).
Re-exported through the agent_framework.azure lazy namespace.
- Nested sub-workflows: the address context (root instance + workflow name +
accumulated {executor}~{ordinal}~ prefix) propagates down call_sub_orchestrator via a
new SUBWORKFLOW_ADDRESS_KEY marker, so an executor at any depth builds a URL that
targets the addressable top-level instance with a qualified request id. The per-child
ordinal matches the read-side enumerate() index used by the status/respond endpoints.
The marker is stripped from untrusted input alongside SUBWORKFLOW_INPUT_KEY
(confused-deputy / info-leak guard).
- Samples 12 and 13 reworked into the retry-safe two-step notify pattern: the emitter
generates an explicit request id and a downstream NotifyExecutor builds the URL and
notifies, so failed upstream retries never produce a dead link.
Tests: unit coverage for the metadata round-trip, address/ordinal agreement (fan-out at
depth and nested prefix accumulation), marker stripping, and URL building; integration
tests assert the helper-built URL equals the server respondUrl and resumes the run, for
both the flat (12) and nested (13) samples.
* refactor(durabletask): read back request_info id instead of generating one in samples
Add WorkflowHitlContext.pending_request_id(ctx), an async helper that returns the
id request_info just generated (read from the runner context's pending request-info
events). This works on any host via the core RunnerContext protocol method, so it
needs no core change.
Samples 12 and 13 now call request_info() and read the id back to forward to the
NotifyExecutor, instead of minting a uuid by hand and passing request_id=. The
read-back happens in the same activity execution that generated the id, so the
pending request event and the notify message still commit together with the same id
(retry-safe; failed upstream retries notify no one).
* docs(azurefunctions): document request_info id read-back and notify safety
Tighten pending_request_id docstring to require calling it immediately after request_info, and explain why that is safe on the durable host (each executor runs in its own activity with its own runner context, so the pending set only holds this executor's requests and the newest is the one just emitted). Document the two-step notify pattern in the 12 and 13 sample READMEs, including the downstream-notifier retry safety and the nested address-prefix propagation.
* fix(python): resolve ty typing error and address PR review comments
- test_subworkflow_orchestration: replace the mypy-only type:ignore[arg-type] with a cast so the ty checker passes too (the other four checkers already honored the ignore).
- samples 12/13 README: guard the notify snippet against None before build_respond_url to match the documented graceful-degradation behavior.
- integration tests 12/13: reword comments that implied request_info now generates an explicit uuid4; it generates the id internally by default.
* fix(python): honor configurable Functions route prefix and address HITL PR review
- Resolve the route prefix from host.json (extensions.http.routePrefix, default api) in a new azurefunctions _routes module, used by both the server endpoints and WorkflowHitlContext, so a custom or empty routePrefix no longer 404s respond/status URLs (was hardcoded /api/ in four places).
- Extract respond/status URL construction into one shared builder called from _app.py and _hitl_context.py, removing the sync-by-test duplication.
- Broaden loopback detection (localhost, 127.0.0.0/8, 0.0.0.0, ::1, [::1]) via a _is_loopback helper so local links use http.
- Pin the host_context key names as shared constants in durabletask so producer and azurefunctions consumer cannot drift.
- Reword a stale base_url comment to reference WEBSITE_HOSTNAME.
- Add unit tests for the route module and loopback handling.
* refactor(azurefunctions): derive server-side HITL URLs from the request URL
The run and status endpoints now derive the base URL and route prefix from the incoming request URL (the value the host actually routed) via split_request_url, so the caller-visible respond/status URLs no longer depend on reading host.json on the server. The in-workflow helper keeps reading host.json since it has no request context. Replaces strip_route_prefix and updates its tests.
* test(durabletask): enforce sub-workflow ordinal and read-index agreement
Extract the read-side subworkflows grouping into a shared _index_subworkflows helper (used by the orchestrator) and add test_readside_index_matches_dispatch_ordinal, which round-trips a fan-out through that helper and asserts subworkflows[executor][ordinal] resolves to the child the dispatch stamped that ordinal onto. Turns the previously comment-only write-ordinal / read-index invariant into a shared, CI-enforced one.
* fix(azurefunctions): suppress bandit B104 on loopback host set
* Python: fix per-run additional_beta_flags leaking into Anthropic request kwargs
_prepare_options copied every key from the caller-supplied options dict
into run_options except "instructions" and "response_format". A
per-run additional_beta_flags value is correctly folded into the betas
set by _prepare_betas, but the raw key was never excluded, so it
survived into run_options and was forwarded straight through to
AsyncMessages.create(), which rejects it with TypeError: got an
unexpected keyword argument 'additional_beta_flags'. Add it to the
exclusion set alongside the other framework-level keys.
Fixes#5764
* Exclude additional_beta_flags from filtered_kwargs too
Copilot's review on the original fix pointed out the exclusion only
covered the options-dict copy, not kwargs passed directly to
_prepare_options — so additional_beta_flags supplied as a raw kwarg
would still leak through and reproduce the same TypeError. Add the
same exclusion to filtered_kwargs for consistency, with a regression
test covering the kwarg path.
---------
Co-authored-by: Chris Brown <albatrossflyon1@gmail.com>
* Python: feat: cross-session origin attribution on context messages
Add an optional origin_session_id parameter to SessionContext.extend_messages
that propagates into the existing _attribution payload on
Message.additional_properties. Downstream context observers can use it to
detect when a provider injects content stored under a different session than
the requesting one.
Populate the field from the harness memory consolidation pipeline
(_harness/_memory.py) when injected topics include contributions from
sessions other than the current one. Add a self-contained sample observer
under samples/02-agents/context_providers/cross_session_observer.py
demonstrating how to subscribe to the signal.
Backward-compatible: omitting the parameter preserves the existing
attribution shape exactly. Tests added in test_sessions.py and
test_harness_memory.py cover the new parameter, the harness cross-session
case, and the same-session case.
Motivated by Dai et al., Stateful Agent Backdoor (arXiv:2605.06158, May
2026), which specifically surveys MAF in section 6.1 / Table 10.
See #5914 for design discussion.
Surfaced during independent audit conducted by @finnoybu (Ken Tannenbaum, AEGIS Initiative); [MEDIUM, python/packages/core].
* Address cross-session attribution review feedback
* Address follow-up review feedback
* Address follow-up review comments
* Address origin attribution review feedback
---------
Co-authored-by: finnoybu <21694570+finnoybu@users.noreply.github.com>
* Python: Make foundry toolbox MCP skills sample self-contained
Rework sample 12 (foundry_toolbox_mcp_skills) so users can build it from
zero with azd, mirroring samples 04 and 09:
- Bundle two single-file SKILL.md skills (support-style, escalation-policy)
and a skills-only toolbox.yaml (with one connectionless code_interpreter
tool, required by `azd ai toolbox create`).
- Rewrite the README as an azd-native, from-zero guide (create skills ->
create toolbox -> set TOOLBOX_ENDPOINT -> run) and fix the stale
MCPSkillsSource API description to match main.py.
- Switch config from TOOLBOX_NAME to the versioned TOOLBOX_ENDPOINT
(.env.example, agent.yaml, agent.manifest.yaml); add .azdignore.
Also enable the sample to run unattended behind ResponsesHostServer:
- Forward disable_load_skill_approval / disable_read_skill_resource_approval
/ disable_run_skill_script_approval from FoundryToolbox.as_skills_provider()
to the underlying SkillsProvider, so load_skill needs no approval round-trip
(the Responses host runs without an AgentSession, which the default approval
flow requires). main.py now uses as_skills_provider(disable_load_skill_approval=True).
- Add unit tests covering the default and overridden approval behaviour.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 4f14f83d-1868-45c1-be1a-12f49a58ac36
* Python: Address PR review on toolbox MCP skills sample
- Remove the unused parameters section from agent.manifest.yaml (TOOLBOX_ENDPOINT
is supplied via environment_variables, matching sample 04).
- README: state the sample is self-contained directly instead of contrasting
with the C# sample.
- README: describe skill discovery behaviour without naming the internal
MCPSkillsSource class.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 4f14f83d-1868-45c1-be1a-12f49a58ac36
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix AG-UI workflow handoff replay results
Decisions:
- Reconcile finalized function results only for call IDs exposed in the current run, and skip results already emitted or never exposed.
- Treat message-derived function results as workflow responses only when their IDs match pending interrupts.
Files:
- Updated _workflow_run.py reconciliation and resume filtering.
- Added runner and public two-turn handoff acceptance coverage.
- Expanded finalized-response call-ID, privacy, and deduplication tests.
Verification:
- uv run poe test -P ag-ui
- uv run poe syntax -P ag-ui -C
- uv run poe pyright -P ag-ui
- uv run poe test-typing -P ag-ui
Notes:
- No blockers. The handoff sample remains unchanged; local PRD and issue files are not included.
* Prevent duplicate AG-UI workflow tool results
* Preserve finalized AG-UI tool results
* Use AG-UI text emission controls
* Python: Fix Magentic manager duplicating conversation history
_complete() reused one persistent AgentSession, so the default history provider
re-injected prior turns on top of the full prompt the manager already rebuilds
each call — duplicating task/facts/plan and compounding every round. Use a fresh
session per call; keep self._session only for
checkpointing. GroupChatOrchestrator is unaffected. Add a regression test and
update the session-propagation test.
* Python: Clean up Magentic manager per Copilot review (drop dead _session)
---------
Co-authored-by: Evan Mattson <35585003+moonbox3@users.noreply.github.com>
* Python: quiet A2AExecutor logging for unmapped content types
Tool-use responses include function_call/function_result content that the A2A executor does not surface, causing a WARNING per tool call. Log these at DEBUG and skip instead, matching the outbound content-conversion convention used across the Python chat clients.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: ee74cc55-44df-4fcf-b38f-1f79f2600dfc
* Address PR review: assert debug log args and drop redundant cast
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: ee74cc55-44df-4fcf-b38f-1f79f2600dfc
* fix: preserve function-call name when merging streaming deltas
`Content._add_function_call_content` built the merged name with
`getattr(self, "name", getattr(other, "name", None))`. Because
`Content.__init__` always sets `self.name` (defaulting to `None`), the
attribute is never missing, so the `getattr` default is never consulted
and `other.name` is ignored. When two function_call contents are merged
and only the second carries the name -- e.g. a streaming delta where the
function name arrives after the first chunk -- the name was silently
dropped.
Use the same "either side" pattern already used for the sibling
`exception` field on the next line: `getattr(self, "name", None) or
getattr(other, "name", None)`. Extend the existing merge test to cover
the late-name and both-None cases.
* test: construct nameless function-call deltas via Content(...) directly
Per review (pyright `reportArgumentType`): `Content.from_function_call`
annotates `name: str`, so passing `name=None` to model a streaming delta
with no name yet tripped the typing gate. Build those nameless deltas
with the `Content("function_call", ...)` constructor instead (its `name`
param is `str | None`) — the factory just wraps that same constructor, so
the runtime objects and the merge assertions are unchanged.
---------
Co-authored-by: Evan Mattson <35585003+moonbox3@users.noreply.github.com>
* Python: bridge AG-UI request state into sessions
Decisions:
- Project resolved AG-UI Shared State into the per-run AgentSession without typed restoration.
- Preserve existing local/service session identifiers and keep AG-UI state out of provider metadata.
Files changed:
- packages/ag-ui/agent_framework_ag_ui/_agent_run.py
- packages/ag-ui/tests/ag_ui/test_endpoint.py
Verification:
- uv run poe check -P ag-ui
- uv run poe test -P ag-ui (912 passed)
Notes:
- Scoped cross-run Session Continuation State remains for the next dependent issue.
* Python: persist scoped AG-UI session continuity
Decisions:
- Store private Session Continuation State atomically in scoped thread snapshots and restore it through the core AgentSession contract.
- Exclude Shared State keys, all HistoryProvider buckets, and tool approval state; request overlays evict colliding private values.
- Finalize interrupted response streams before snapshotting so provider after_run mutations are included.
Files changed:
- packages/ag-ui/agent_framework_ag_ui/_agent_run.py
- packages/ag-ui/agent_framework_ag_ui/_snapshots.py
- packages/ag-ui/tests/ag_ui/test_endpoint.py
- packages/ag-ui/tests/ag_ui/test_snapshots.py
- packages/ag-ui/AGENTS.md
Verification:
- uv run poe test -P ag-ui (921 passed)
- uv run poe check -P ag-ui
- uv run poe typing -P ag-ui
Notes:
- Lifecycle, isolation, and broader storage guidance remain for the next dependent issue.
* Python: document AG-UI session continuity lifecycle
Decisions:
- Keep scoped thread snapshots as the single reset and continuity boundary, with missing request Shared State preserving private continuation.
- Document trusted typed-restoration storage, State Authorities, custom-store round trips, and one-active-run last-writer-wins consistency.
- Verify failure, hydration privacy, scope/thread isolation, and reset mechanics through public endpoint and store seams.
Files changed:
- packages/ag-ui/README.md
- packages/ag-ui/tests/ag_ui/test_endpoint.py
- packages/ag-ui/tests/ag_ui/test_snapshots.py
Verification:
- uv run pytest -q <focused lifecycle tests> (6 passed)
- uv run poe test -P ag-ui
- uv run poe syntax -P ag-ui -C
- uv run poe typing -P ag-ui
- uv run poe check -P ag-ui
- uv run poe markdown-code-lint
Notes:
- No runtime capability probe, secondary state store, locking, or configuration flag was added.
- No blockers remain for this lifecycle and guidance slice.
* Python: harden AG-UI session continuity
* Python: isolate AG-UI request state
* Fix: Ollama parallel tool calls collide on same call_id
* fix(ollama): use uuid4 for tool call IDs and support colons in tool names
---------
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
* docs: add env example files for durabletask samples
* docs: clarify env example values and comments
* docs: set default Redis URL in streaming sample env example
Tao Chen