[xuan (yuee3)]

Summary

Fixes a missed cancellation_token propagation in the MagenticOne orchestrator's progress-ledger model call for the structured_output branch, introduced when structured output support was added in #5863.

In _orchestrate_step, the ledger retry loop has three branches. The structured_output branch was the only model_client.create() call in MagenticOneOrchestrator that did not forward cancellation_token:

if self._model_client.model_info.get("structured_output", False):
    response = await self._model_client.create(
        self._get_compatible_context(context), json_output=LedgerEntry  # no cancellation_token
    )
elif self._model_client.model_info.get("json_output", False):
    response = await self._model_client.create(
        self._get_compatible_context(context), cancellation_token=cancellation_token, json_output=True
    )
else:
    response = await self._model_client.create(
        self._get_compatible_context(context), cancellation_token=cancellation_token
    )

Consequence: when a user cancels a MagenticOne run and the model client supports structured output (e.g. OpenAI), the in-flight progress-ledger LLM call is not interrupted, delaying cancellation by a full orchestration round. All other create() calls in the orchestrator already forward the token.

This PR also fixes an off-by-variable bug in the ledger parse error message in the same loop: it reported len(progress_ledger) (the parsed ledger dict) instead of len(output_json) (the extracted JSON objects actually being validated).

Changes

• _magentic_one_orchestrator.py: pass cancellation_token=cancellation_token in the structured_output ledger call; fix the error message to report len(output_json).
• test_magentic_one_group_chat.py: add test_magentic_one_group_chat_structured_output_cancellation_token — runs a MagenticOne team with a ReplayChatCompletionClient configured with structured_output=True and asserts every recorded create() call received the run's CancellationToken.

Testing

• New regression test fails on main (ledger call recorded cancellation_token=None) and passes with the fix.
• pytest tests/test_magentic_one_group_chat.py -> 8 passed.
• pytest tests/test_group_chat.py -> 91 passed; the 2 failures (test_declarative_groupchats_with_config) are pre-existing and unrelated to this change (missing OpenAI credentials; fail identically on pristine main).
• ruff check / ruff format --check clean on the changed files.

[xuan (yuee3)]

@microsoft-github-policy-service agree