fix(engine): stop replaying an abandoned partial turn on retry

[ChakrawarShubham]

Fixes #369

Updated: this PR originally patched the Gemini converter. After opening it I tested Anthropic, found the same wedge there and worse, and moved the fix to the engine where the defect actually is.

Problem

A session whose turn dies mid-stream becomes permanently unrecoverable from the UI — every Retry returns the same 400. Hit on Gemini 3.6 Flash: the agent streamed some text, the turn died to a 503 UNAVAILABLE, and Retry 400s forever after.

Error: 400 INVALID_ARGUMENT. {'code': 400, 'message': 'Requests ending with a
model turn are not supported.', 'status': 'INVALID_ARGUMENT'}

Switching Gemini models does not help — the malformed history is what gets replayed. A user interrupt leaves the same state.

Root cause

When a turn dies after text has streamed, engine.py:367 appends a partial assistant message — so the transcript still shows what the user saw — then an error notice. _outbound_messages() drops notices (engine.py:1043), so the partial becomes the final message in the payload, and Engine.retry() (engine.py:246) replays it unmodified. Captured from the running engine rather than read off the source:

history:   user | assistant 'I will request permission to ' | notice/error
outbound:  ['user', 'assistant']        <- notice stripped, partial is the tail

Not Gemini-specific

That was my first assumption and it was wrong. Claude 4.5 and older accept a trailing assistant turn as prefill but 400 when it ends in whitespace — final assistant content cannot end with trailing whitespace — and streamed deltas routinely do; the real cut from my session was 'I will request permission to '. Sonnet 4.6, Opus 4.8 and Fable 5 reject it unconditionally (does not support assistant message prefill). That is three of the four Claude models in the curated picker, with the fourth exposed via whitespace.

Reproduced end-to-end on Sonnet 4.6 through the real engine and retry() against the live API: three retries, three identical 400s.

So the defect is that _outbound_messages() ships a turn that never completed. Guarding one vendor's converter would leave the other models broken.

Fix

Drop the abandoned partial in _outbound_messages(), alongside the notices that function already treats as display-only.
The trailing error/interrupted notice is what identifies it, and that detail is load-bearing: a completed turn's assistant message is structurally identical — trailing, no tool_calls — so a shape-based check silently deletes the last turn of every finished conversation. I wrote that version first, and test_provider_extras_persist_on_message_and_survive_outbound caught it. Trailing tool_calls mean a turn suspended awaiting durable resume, never an abandoned one.

Because notices persist to disk, this also unwedges sessions already broken — which a converter patch would not.

Tests

• test_retry_after_mid_stream_death_does_not_replay_the_partial — the partial stays in history for the transcript but never reaches the provider
• test_completed_assistant_turn_is_never_dropped — the counterpart; pins the distinction the notice provides
• test_consecutive_deaths_do_not_stack_partials_in_the_payload — each failed retry appends its own
• test_resume_payload_keeps_the_suspended_tool_call — durable resume unaffected; the assistant turn that made the call still reaches the model

Full suite: 1012 passed. The 7 test_bedrock_provider.py failures in my environment are ModuleNotFoundError from a missing boto3 in my venv and fail identically on unmodified main.

Notes

• Unrelated to fix: repair tool-call/result pairing on load to prevent unrecoverable 400 errors #350 — that repairs tool-call/result pairing in persisted JSONL; this drops an abandoned turn at the outbound boundary. No shared files.
• Alternative considered: tag the partial at creation and drop by tag. No inference at all, but it only helps future failures, leaving already-wedged sessions on disk broken. Happy to switch if you would rather have the explicit marker.
• On stock OpenAI a trailing assistant turn is tolerated today, so a retry currently continues from the partial; after this it regenerates the turn. Nothing breaks, but it encodes retry means re-run rather than retry means continue — a call I am happy to have overruled.

[ChakrawarShubham]

Since first-time contributor workflows need maintainer approval, I ran CI on my fork so there's something to look at without spending an approval: https://github.com/ChakrawarShubham/openworker/actions/runs/30603081143 — pytest, GUI unit and Playwright e2e all green.

Happy to rework this in whatever direction you prefer. I know engine.py is a file you're actively working in, and the tagging alternative in the notes above may well fit your plans better than what I've done here.