fix(agents): tolerate output_text parts with None text (gemini-3.6-flash run crash)

[inf-quantavius]

Summary

The first fixed-analysis HALO run on gemini-3.6-flash (pinned by monorepo #4289, merged 2026-07-24 10:31Z) crashed within one turn:

TypeError: sequence item 0: expected str instance, NoneType found

(dev halo_run 99790c24-…, analysis run iar_vEmBDW05ukFmvwyEofnke, smoke run 30091185355; dev Loki shows subagent sub-8e9b236a failed at depth=1: TypeError: … → Emitting halo.failed.)

Root cause

Gemini via the LiteLLM Responses translation returns assistant messages whose output_text part carries no text field on tool-calls-only turns — dev ClickHouse shows both gateway calls returned HTTP 200 with content [{"type":"output_text","annotations":[]}] alongside function_call items (ids carrying Gemini __thought__ blobs). The SDK materializes these leniently as ResponseOutputText with text=None, and the mapper's

text = "".join(part.text for part in parts if isinstance(part, ResponseOutputText))

(engine/agents/openai_event_mapper.py:96) raises the exact TypeError, killing the run. gpt-5.5 and deepseek-v4-pro runs on the same project earlier the same day did not hit this — the model swap is the delta.

Change

• _map_assistant_message: join part.text or "" so a text-less part maps to empty content (content=None downstream, same as tool-calls-only turns from other providers).
• Regression test building the real SDK shape via model_construct (lenient parse), asserting the mapper emits a normal assistant item instead of raising.

Validation

• pytest tests/unit: 480 passed (12 in the mapper file, incl. the new regression test)
• ruff check + ruff format --check: clean
• basedpyright on both touched files: 0 errors

Related: #79 handles the finalization-protocol side of the model swap; this fixes the orthogonal crash that prevents gemini runs from getting past their first tool-calling turn at all.

Provenance

• Requested by: CI Notifications (Slack U0B5KMSTCKX)
• Source: https://inference-net.slack.com/archives/C0B57KU5JNT/p1784894278780589?thread_ts=1784894278.351109&cid=C0B57KU5JNT
• Quant turn: 3f49f4c0

---

Note

Low Risk
Single-line defensive join in assistant message mapping with a regression test; no auth, data, or API contract changes.

Overview
Fixes HALO run crashes on Gemini when assistant turns are tool-calls-only: LiteLLM/Gemini can return output_text parts with no text field, which the SDK parses as text=None. Joining those values in _map_assistant_message raised TypeError and aborted the run.

The mapper now joins part.text or "" so missing text becomes empty assistant content (content=None), matching other providers’ tool-only turns. A unit test covers the lenient ResponseOutputText(text=None) shape.

^{Reviewed by Cursor Bugbot for commit 5309501. Bugbot is set up for automated code reviews on this repo. Configure here.}