[Vedant Sonkar (vedantsonkar)]

Motivation & Context

Call-level compaction (compaction_strategy on an Agent or client.as_agent()) works mid-run: it marks older self-generated tool-call/tool-result messages as excluded and inserts a summary the model sees. However, the summary messages never reached the final AgentResponse - only the exclusion flags did, because the flags are in-place mutations on Message objects shared between the model-input list (prepared_messages) and the returned transcript (function_call_messages), while the inserted summaries live only in the model-input list.

As a result, HistoryProvider.after_run() persisted that asymmetric response, and with InMemoryHistoryProvider(skip_excluded=True) - the documented way to combine compaction with history - the next get_messages() silently dropped the excluded tool groups with no summary replacing them. Any agent using tools + call-level compaction + a history provider hits this.

Fixes #8099

Description & Review Guide

• What are the major changes?

  • Generalized the existing private _reconcile_compaction_summaries helper in _compaction.py so both ChatMiddlewareLayer and the function-invocation loop use the same ownership-aware, fixed-point reconciliation algorithm.
  • Two reconciliation call sites in FunctionInvocationLayer._get_response_with_function_invocation (_tools.py): after each Phase-2 model call and after the Phase-3 final no-tools call, before the transcript is returned or persisted.
  • The function-loop transcript's message identities define the ownership boundary. Nested summary dependencies resolve to a fixed point, summaries are inserted before their transcript-owned source messages, and summaries involving caller-owned input are rejected so persisting the response does not duplicate caller history.
  • Centralized non-streaming terminal transcript assembly so ordinary completion, approval requests, user-input requests, and middleware termination all prepend exactly the accumulated transcript prefix without duplicating the current call/result messages.
  • Regression tests cover the Phase-2 terminal path, max-iterations path, nested summary-of-summary dependencies, mixed caller/transcript ownership, approval/user-input/middleware early exits, and agent-level persistence with InMemoryHistoryProvider(skip_excluded=True).
  • Spec traceability (normative bullet, matrix row, related-issue entry) in docs/specs/004-python-function-calling-loop.md, per python/AGENTS.md.
  • A separate chore commit applies the one-line ruff formatting to test_agent_executor.py that the locked formatter / pre-commit hook now requires (unrelated to the fix, isolated for clarity).

• What is the impact of these changes?

  • Non-streaming runs that compact tool groups in-run now return the inserted summaries in ChatResponse.messages / AgentResponse.messages, positioned before the group each summary replaces - so history loaded with skip_excluded keeps the summarized content instead of losing it silently.
  • Streaming is untouched (its final response is assembled from accumulated updates, a mechanism that never carried compaction summaries or exclusion flags; existing streaming compaction tests remain green).
  • Scope is limited to the non-streaming function-invocation loop; no public API changes, no breaking change.

• What do you want reviewers to focus on?

  • Reusing _reconcile_compaction_summaries with transcript identities as the ownership boundary, especially fixed-point handling of nested summaries and rejection of summaries involving caller-owned input.
  • Centralized terminal transcript assembly across ordinary completion, approval requests, user-input requests, and middleware termination.
  • Why reconciliation is unconditional rather than gated on the compaction_strategy parameter: compaction may come from the inner client's own default strategy, which the invocation layer does not see as a parameter.

Validated with uv run poe test -P core (5136 passed, 38 skipped, 2 xfailed), uv run poe syntax -P core, uv run poe pyright -P core, and uv run poe test-typing -P core - all green.

Related Issue

Fixes #8099

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) - a workflow keeps the label and title prefix in sync automatically.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Vedant Sonkar (vedantsonkar)]

@microsoft-github-policy-service agree

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Vedant Sonkar (vedantsonkar)]

Eduard van Valkenburg (@eavanvalkenburg) Please do review and let me know!

[Vedant Sonkar (vedantsonkar)]

Eduard van Valkenburg (@eavanvalkenburg) I investigated the failing Python 3.14 Ubuntu job locally. The new #8099 regression cases pass under Python 3.14 (9/9), including the terminal-exit and nested-summary cases. The job’s public annotation only reports a generic exit code 1, while the remaining matrix jobs were cancelled due to fail-fast. Could someone with Actions-log access share the failing node ID and traceback? At present this looks likely to be an unrelated or CI-only failure rather than a reproducible regression from this PR.

[Eduard van Valkenburg (eavanvalkenburg)]

Thanks for the contribution. The current head is not green: Python - Tests and Merge Gatekeeper are failing, with failures and cancellations across the Python test matrix. Could you please get the checks passing, then re-request review?

[Vedant Sonkar (vedantsonkar)]

Eduard van Valkenburg (@eavanvalkenburg) Investigated the Python test failure. The Responses tests were using the real local file-backed AgentServer session store while pytest-xdist ran workers in parallel. Workers could write the same agent_sessions.json, causing intermittent storage errors and a failed response status.
The fix preserves the default file-backed store and instead gives each xdist worker its own temporary AGENTSERVER_STATE_ROOT. The change is scoped to test_responses.py.
Local validation passed:

• Responses suite under xdist: passed
• Python 3.10: 12750 passed, 74 skipped, 2 xfailed (with the workflow’s expected two-package exclusions)
• Python 3.11: 13000 passed, 66 skipped, 2 xfailed
• Python 3.12: 13000 passed, 66 skipped, 2 xfailed
• Python 3.13: 13000 passed, 66 skipped, 2 xfailed
• Python 3.14: 11910 passed, 66 skipped, 2 xfailed
• Ruff and pre-commit checks: passed

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Vedant Sonkar (vedantsonkar)]

Eduard van Valkenburg (@eavanvalkenburg) All tests seems to be passing now, can we merge this one ?