Python: Feature/hosted dwf by alliscode · Pull Request #5531 · microsoft/agent-framework

Ben Thomas (alliscode) · 2026-04-28T00:28:15Z

This pull request enhances support for passing conversation history into declarative workflow agents, ensuring that workflows receive the full message context (not just the latest user turn) when triggered via agents. It updates the input initialization logic, executor signatures, and hosting integration to handle lists of Message objects, and adds a regression test to guarantee correct population of System.LastMessageText for compatibility with existing YAML workflows.

…rt executor The declarative start executor (JoinExecutor) only advertised dict and str in its input_types, so WorkflowAgent.__init__ rejected it with 'Workflow's start executor cannot handle list[Message]'. Add list[Message] to the JoinExecutor handler annotation and add a matching branch in DeclarativeActionExecutor._ensure_state_initialized that extracts the last user-message text and falls through to the string-input initialization path, so =System.LastMessageText works end-to-end via as_agent(). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

When Workflow.as_agent() is invoked with a list[Message], the start executor now populates Conversation.messages / Conversation.history / System.conversations.{id}.messages with prior turns only (excluding the latest user message), and surfaces the latest user message via Inputs.input and System.LastMessage*. This matches InvokeAzureAgent's contract that the messages binding holds prior turns and the executor itself appends the new user input before invoking, avoiding double-append of the trailing user turn while preserving full history (incl. assistant/system/tool roles and multi-modal content) for downstream actions.

MessageRole and other str-subclass Enums passed isinstance(v, str) and were forwarded to pythonnet unchanged. pythonnet then raised 'MessageRole value cannot be converted to System.String' for every PowerFx primitive when ConditionGroup/Expr eval walked the symbol table containing Conversation.messages. Reduce Enum members to their underlying value before the primitive check so eval sees plain strings/ints.

_handle_inner_workflow only forwarded the latest user turn to WorkflowAgent.run, even though _handle_inner_agent already prepends history fetched from Foundry storage to the messages it sends a regular agent. Declarative workflows reset Conversation.messages on every run (state.initialize), so checkpoint replay alone does not give them prior turns - the host has to pass them in, the same way it does for non-workflow agents. Mirror that contract: fetch context.get_history() and pass [*history, *input_messages] to the workflow agent.

Evan Mattson (moonbox3) · 2026-04-28T00:30:56Z

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework/_workflows
_agent.py	358	71	80%	66, 74–80, 116–117, 210, 271, 284, 351, 362, 364, 424, 430, 447–448, 455, 457, 463, 525–526, 535, 543, 569, 602–604, 606, 608, 610, 615, 620, 674, 704, 721, 760–763, 769, 775, 779–780, 783–789, 793–794, 802, 863, 870, 876–877, 888, 920, 927, 948, 957, 961, 963–965, 972
_runner.py	174	2	98%	309–310
_workflow.py	278	19	93%	91, 270–272, 274–275, 293, 297, 461, 676, 697, 747, 759, 765, 770, 790–792, 843
packages/declarative/agent_framework_declarative/_workflows
_declarative_base.py	462	58	87%	47, 50, 154, 242, 256, 270, 281, 298, 425–426, 433–434, 449, 479, 562–564, 566–568, 570, 573, 580, 606, 632–638, 640–641, 643–651, 672–673, 675–676, 966–969, 993–994, 1001–1002, 1007, 1016–1017, 1023
_executors_control_flow.py	170	6	96%	82, 90, 153, 343–344, 348
packages/foundry_hosting/agent_framework_foundry_hosting
_responses.py	528	118	77%	151–152, 161–162, 166, 171, 185, 209, 235–237, 256–258, 260–262, 264–266, 270–273, 285–291, 301–302, 320–322, 327, 329, 336, 338–339, 341, 343, 349–352, 354–356, 360, 363, 368–374, 377–378, 380–381, 389–394, 466–467, 1112–1114, 1116, 1163–1164, 1166–1167, 1169–1170, 1172–1173, 1178, 1187, 1190–1192, 1194, 1212, 1215, 1250–1254, 1258–1264, 1271–1275, 1281–1287, 1294, 1300, 1303
TOTAL	30502	3560	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
6131	30 💤	0 ❌	0 🔥	1m 36s ⏱️

Copilot

Pull request overview

This PR improves hosted declarative workflow agents so they receive full conversation context (history + current input) when invoked via agent interfaces, aligning behavior with non-workflow agents and maintaining YAML compatibility via System.LastMessageText.

Changes:

Foundry hosting now prepends stored conversation history to workflow-agent inputs before invoking WorkflowAgent.run(...).
Declarative workflow state initialization now accepts list[Message], populating Conversation.* and System.LastMessage* accordingly, and improves PowerFx safety by coercing Enum values.
Adds a regression test ensuring Workflow.as_agent() populates System.LastMessageText correctly.

Show a summary per file

File	Description
python/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py	Passes Foundry-stored history + current request messages into workflow agent runs.
python/packages/declarative/agent_framework_declarative/_workflows/_declarative_base.py	Adds `list[Message]` trigger initialization and Enum-to-primitive coercion for PowerFx symbol conversion.
python/packages/declarative/agent_framework_declarative/_workflows/_executors_control_flow.py	Updates `JoinExecutor` trigger type to include `list[Message]` so workflows can be exposed as agents.
python/packages/declarative/tests/test_workflow_factory.py	Adds regression coverage for `Workflow.as_agent()` and `System.LastMessageText`.

Copilot's findings

Files reviewed: 4/4 changed files
Comments generated: 2

github-actions

Automated Code Review

Reviewers: 4 | Confidence: 89% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Design Approach

Automated review by alliscode's agents

…urn continuation Allow Workflow.run(message=..., checkpoint_id=...) so callers can restore prior workflow state from a checkpoint AND deliver a new message to the start executor in a single call. The existing reset_context logic already preserves shared state when checkpoint_id is set, so this gives us 'fresh start executor invocation with prior state intact' - exactly what hosted multi-turn declarative workflows need. - _workflow.py: drop the message+checkpoint_id mutual exclusion and update _execute_with_message_or_checkpoint to do both (restore then execute) when both are provided. - _agent.py: in _run_core's checkpoint branch, also forward input_messages so WorkflowAgent.run(messages, checkpoint_id=...) works end-to-end. Falls back to the legacy 'restore only' behavior when messages are absent. - _declarative_base.py: detect continuation in _ensure_state_initialized by checking whether DECLARATIVE_STATE_KEY already exists in shared state; if so, refresh inputs/LastMessage* and append non-user trigger messages instead of calling state.initialize() (which would wipe Conversation/Local/System). - foundry_hosting/_responses.py: collapse the host's two-call pattern (restore-only, then fresh run) into a single combined call now that the underlying APIs support it. - tests: drop the assertion that combined message+checkpoint_id raises. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Replace the prior 'combined message + checkpoint_id in one run()' approach with a cleaner default: Workflow.run no longer wipes shared state or runner- context messages between calls. Iteration counting and per-run kwargs still reset on a fresh-message run; checkpoint and responses runs are continuations that preserve everything. This lets a WorkflowAgent be invoked repeatedly on the same instance and maintain multi-turn context (e.g. accumulated Conversation.messages) without asking developers to opt in. Hosted-agent multi-turn pattern becomes two explicit calls: restore-from-checkpoint (drive to idle), then run-with-message. Key changes: - _workflow.py: drop _state.clear() and reset_for_new_run() from run(). Reset iteration count and run kwargs on fresh-message runs only. Restore 'Cannot provide both message and checkpoint_id' validation. Add async guard: fresh-message run with un-drained pending executor messages from a prior run is invalid. - _runner.py: clear _state before import_state in restore_from_checkpoint so restore is authoritative (import_state merges, not replaces). - _agent.py: revert checkpoint branch to restore-only (no message forward). - _responses.py (foundry_hosting): two-call host pattern - restore checkpoint silently, then run with new user input. - tests: state-preservation is the new default; rebuild Workflow for clean slate. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

- _workflow.py: collapse nested if (SIM102), drop redundant assignment (RET504) - _declarative_base.py: remove unused last_user_msg = tail assignment whose Message | None type clashed with the prior Message-typed branch Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Copilot

Copilot's findings

Files reviewed: 8/8 changed files
Comments generated: 4

- _declarative_base.py: continuation branch was writing 'Inputs.input' via state.set, which routes to the Custom namespace and never updates the PowerFx-visible Workflow.Inputs.input. Update state_data['Inputs'] in place via get_state_data / set_state_data so =Workflow.Inputs.input and =inputs.input see the new turn's user text on continuation. - _declarative_base.py: refresh docstring to clarify that on a list[Message] trigger, Conversation.messages excludes the current user message at the start of the turn (agent executors append it before invoking the inner agent). - _responses.py: when previous_response_id is supplied (no conversation_id), the prior checkpoint lives under <storage>/<previous_response_id> but new checkpoints must land under <storage>/<current_response_id> for the next turn to find them. Hold onto restore_storage from the get_latest lookup and pass it to the restore-only run; pass write_storage (current id) to the message-delivery run and to checkpoint cleanup. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

- Replace state._state.get(...) protected access with new public is_initialized() method on DeclarativeWorkflowState (also clearer intent for the continuation detection use case). - Add narrow pyright ignores for the Any-typed trigger paths that pyright cannot fully narrow (the list[Message] isinstance loop and the fallback-DefaultTransform branch). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

github-actions

Automated Code Review

Reviewers: 1 | Confidence: 95%

✓ Design Approach

The patch fixes the immediate continuation bugs, but two parts of the new design still mis-handle multi-turn state. In declarative workflows, the continuation path now replays the entire incoming session history back into already-restored Conversation.* state, so chat history will grow with duplicates on every turn. In foundry hosting, the new unconditional restore-only call can leave WorkflowAgent.pending_requests populated, causing the subsequent user-message call to be interpreted as request responses instead of a fresh turn. Both issues stem from layering new behavior on top of existing stateful abstractions without accounting for what those layers already persist.

Automated review by alliscode's agents

* Add Workflow.reset() public method as recovery escape hatch when an in-flight run aborted (e.g. WorkflowConvergenceException) and the workflow is not checkpointed. Update the in-flight messages guard's error message to point callers at it. * Add test_workflow_run_inflight_messages_guard exercising both the guard (sync + streaming) and the reset() recovery path. * Add test_workflow_reset_rejects_concurrent_runs to lock down the in-progress guard on reset. * Add test_as_agent_continuation_preserves_prior_state covering the is_continuation branch in _ensure_state_initialized: stamps a marker between calls and asserts it survives, while Inputs.input and System.LastMessageText refresh to the new turn. * Add test_powerfx_safe.py regression tests for the Enum branch in _make_powerfx_safe (str-subclass, int-subclass, plain Enum, and Enums nested in dict/list). * Drop redundant @pytest.mark.asyncio on test_as_agent_round_trip_with_last_message_text (asyncio_mode='auto'). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Address Copilot review on _responses.py: the restore-only checkpoint replay populates self._agent.pending_requests for any request_info events captured in the checkpoint. The follow-up run(input_messages) call would then route through WorkflowAgent._process_pending_requests, which expects function-response content and rejects plain text input as 'unexpected content while awaiting request info responses'. Workflows resumed from a checkpoint that was idle-with-pending-requests would therefore fail every subsequent plain-text user turn. Inspect the loaded checkpoint and skip the pre-pass when its pending_request_info_events dict is non-empty. Workflows that don't use request_info (the current sample set) are unaffected; workflows that do will fall through to a fresh-message run rather than silently corrupting the routing state. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

The exact-version pins on azure-ai-agentserver-{core,responses,invocations} forced foundry-hosting consumers to upgrade in lockstep with every beta bump from upstream. Switch to '>=current,<next-major' so we pick up patch and feature updates within the same major series without a coordinated release. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

The in-flight-messages guard prevented silent misbehavior, but the companion Workflow.reset() escape hatch only cleared _messages while leaving iteration count, executor-local state, and shared State mutations in an indeterminate condition after a mid-run failure. That gave a false sense of recovery. Recovery from a mid-run failure is supported only via checkpoint restoration. Keep the guard and reframe its error message accordingly; remove reset() and its tests. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Resolve foundry_hosting/pyproject.toml conflict by keeping the loosened azure-ai-agentserver-* pins (>=X.Y.ZbN,<NEXT_MAJOR) while taking main's agent-framework-core>=1.2.1,<2 bump. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

alliscodeand others added 4 commits April 27, 2026 12:53

CopilotAI review requested due to automatic review settings April 28, 2026 00:28

Evan Mattson (moonbox3) added the python Usage: [Issues, PRs], Target: Python label Apr 28, 2026

github-actionsBot changed the title ~~Feature/hosted dwf~~Python: Feature/hosted dwfApr 28, 2026

Copilotstarted reviewing on behalf of Ben Thomas (alliscode)April 28, 2026 00:28 View session

CopilotAI reviewed Apr 28, 2026
View reviewed changes

Comment threadpython/packages/declarative/agent_framework_declarative/_workflows/_declarative_base.py

github-actionsBot reviewed Apr 28, 2026
View reviewed changes

Evan Mattson (moonbox3) approved these changes Apr 28, 2026
View reviewed changes

alliscodeand others added 2 commits April 28, 2026 08:49

Ben Thomas (alliscode)force-pushed the feature/hosted-dwf branch from b355a43 to e8dfcc9Compare April 28, 2026 18:10

Microsoft (microsoft) deleted a comment from CopilotAIApr 28, 2026

Ben Thomas (alliscode) requested a review from Copilot April 28, 2026 18:14

Copilotstarted reviewing on behalf of Ben Thomas (alliscode)April 28, 2026 18:15 View session

CopilotAI reviewed Apr 28, 2026
View reviewed changes

alliscodeand others added 2 commits April 28, 2026 11:27

Ben Thomas (alliscode) marked this pull request as ready for review April 28, 2026 18:45

github-actionsBot reviewed Apr 28, 2026
View reviewed changes

Comment threadpython/packages/foundry_hosting/agent_framework_foundry_hosting/_responses.py

alliscodeand others added 5 commits April 28, 2026 12:43

Tao Chen (TaoChenOSU) reviewed Apr 28, 2026
View reviewed changes

dependabotBot mentioned this pull request Jul 26, 2026
Bump the nuget-all group with 6 updates dotnet-presentations/ai-workshop#547
Merged

Navigation Menu

Uh oh!

Python: Feature/hosted dwf - #5531

Python: Feature/hosted dwf#5531
Evan Mattson (moonbox3) merged 17 commits into
mainfrom
feature/hosted-dwf

Ben Thomas (alliscode) commented Apr 28, 2026

Uh oh!

Evan Mattson (moonbox3) commented Apr 28, 2026•
edited
Loading

Uh oh!

CopilotAI left a comment

Uh oh!

Uh oh!

github-actionsBot left a comment

Uh oh!

CopilotAI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actionsBot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

Ben Thomas (alliscode) commented Apr 28, 2026

Uh oh!

Evan Mattson (moonbox3) commented Apr 28, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Python Unit Test Overview

Uh oh!

CopilotAI left a comment

Choose a reason for hiding this comment

Pull request overview

Copilot's findings

Uh oh!

Uh oh!

github-actionsBot left a comment

Choose a reason for hiding this comment

Automated Code Review

Uh oh!

CopilotAI left a comment

Choose a reason for hiding this comment

Copilot's findings

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actionsBot left a comment

Choose a reason for hiding this comment

Automated Code Review

✓ Design Approach

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Evan Mattson (moonbox3) commented Apr 28, 2026•
edited
Loading