Python: fix Gemini finish reason fallback and usage-attach cascade

[Jeremy Schoemaker (shoemoney)]

Motivation & Context

#7105 gave chat clients a fallback so an unmapped provider finish reason value passes through as the raw string instead of vanishing to None. It touched ag, bedrock, claude, core, github_copilot, ollama, and openai. It did not touch gemini.

Bedrock's version of the same method after #7105:

def _map_finish_reason(self, reason: str | None) -> str | None:
    if not reason:
        return None
    return FINISH_REASON_MAP.get(reason.lower(), reason)

Gemini's still has no fallback:

def _map_finish_reason(self, reason: str | None) -> FinishReasonLiteral | None:
    if not reason:
        return None
    return _FINISH_REASON_MAP.get(reason)

_FINISH_REASON_MAP covers 13 of the 18 members of google.genai.types.FinishReason in the package's own pinned dependency range (google-genai>=1.69.0,<3.0.0). Missing: OTHER, TOO_MANY_TOOL_CALLS, NO_IMAGE, IMAGE_OTHER (plus FINISH_REASON_UNSPECIFIED, which is a legitimate absent case and correctly stays None).

The bug cascades. In _process_chunk:

if finish_reason and (usage := self._parse_usage(chunk.usage_metadata)):

Usage is attached to a streamed chunk only when finish_reason is truthy, so an unmapped reason drops the finish reason and the whole turn's token/billing accounting. ChatTelemetryLayer in observability.py skips recording the terminal span state the same way.

Concrete scenario: a Gemini call inside an agentic tool loop trips Gemini's own tool-call-count guardrail and returns finish_reason=TOO_MANY_TOOL_CALLS. The caller sees finish_reason=None, no usage, and no way to tell the run stopped abnormally instead of completing normally.

Description & Review Guide

• What are the major changes?
  • _map_finish_reason now falls back to the raw reason string, wrapped as FinishReason(reason), instead of None. FINISH_REASON_UNSPECIFIED and an absent reason still map to None.
  • The return type widens from FinishReasonLiteral | None to FinishReasonLiteral | FinishReason | None. This is not the same pattern bedrock uses (str | None with no wrapping) — bedrock's file carries a blanket # type: ignore at the top of the module, gemini's does not, so a bare str fails this package's strict Pyright config. FinishReason(reason) is the same construct ollama and openai used for the same fallback in Python: Normalize chat finish reasons #7105.
• What is the impact of these changes?
  • An unmapped-but-real Gemini finish reason (currently OTHER, TOO_MANY_TOOL_CALLS, NO_IMAGE, IMAGE_OTHER) is now surfaced to the caller instead of silently disappearing, and the final streamed chunk's usage/token accounting is no longer dropped alongside it.
• What do you want reviewers to focus on?
  • Whether FINISH_REASON_UNSPECIFIED is the only value that should keep mapping to None, or whether other values should be excluded from the fallback too.

Related Issue

Fixes #7836

Testing

Ran the gemini package's unit test suite only (not the full monorepo suite):

python -m pytest packages/gemini/tests -m "not integration" -v

GREEN (with the fix): 157 passed, 8 deselected (integration tests, no credentials configured).

Reverted only the source change (kept the new/updated tests) to confirm RED:

5 failed, 152 passed, 8 deselected
FAILED test_finish_reason_mapping[OTHER-OTHER]
FAILED test_finish_reason_mapping[TOO_MANY_TOOL_CALLS-TOO_MANY_TOOL_CALLS]
FAILED test_finish_reason_mapping[NO_IMAGE-NO_IMAGE]
FAILED test_finish_reason_mapping[IMAGE_OTHER-IMAGE_OTHER]
FAILED test_unmapped_finish_reason_still_attaches_usage_on_streamed_final_chunk

Restored the fix, reran, back to GREEN (157 passed).

Also ran pyright and ruff check/ruff format --check against the changed files only; all clean. Did not run the .NET suite, the full Python monorepo test suite, or integration tests (no Gemini/Vertex credentials in this environment).

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]

Pull request overview

This PR fixes a bug in the Gemini Python chat client where unmapped provider finish reasons were silently dropped to None. #7105 added a raw-string fallback to the other chat clients (ag, bedrock, claude, core, github_copilot, ollama, openai) but missed gemini. Because streamed-chunk usage is attached only when finish_reason is truthy, an unmapped reason (e.g. TOO_MANY_TOOL_CALLS) dropped both the finish reason and the turn's token/billing accounting. The fix mirrors the established ollama/openai fallback pattern, wrapping the raw reason as FinishReason(reason) so it type-checks under this package's strict Pyright config.

Changes:

• _map_finish_reason now falls back to FinishReason(reason) for values absent from _FINISH_REASON_MAP, while FINISH_REASON_UNSPECIFIED and absent reasons still map to None; return type widened to FinishReasonLiteral | FinishReason | None.
• Added the FinishReason import and updated the docstring to explain the fallback behavior.
• Expanded finish-reason mapping tests and added a regression test confirming usage is still attached to the final streamed chunk under an unmapped reason.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
python/packages/gemini/agent_framework_gemini/_chat_client.py	Adds raw-string fallback in _map_finish_reason, imports FinishReason, widens the return type, and updates the docstring.
python/packages/gemini/tests/test_gemini_client.py	Expands parametrized mapping cases (mapped, unmapped-passthrough, None/UNSPECIFIED) and adds a streamed-chunk usage-cascade regression test.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/gemini/agent_framework_gemini
_chat_client.py	494	17	96%	415, 748, 788–789, 798–799, 802–803, 835, 842, 967, 978, 1170–1171, 1175, 1186–1187
TOTAL	48158	4496	90%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
9740	36 💤	0 ❌	0 🔥	2m 42s ⏱️