feat(own-loop): final-summary turn after a stuck-loop exit

[jonnyparris]

What

When the own-loop exits because the model got stuck (doom-loop, no-text-loop, or text-loop), run one more bounded streamText call with no tools to elicit a real conclusion from the model. Without this, users saw only the harness's own [Stopped: ...] notice and no real answer — even though the prior tool calls had often produced useful information that just never got written up.

Why

Reproduced today on Gemma 4 26B after PR #78 landed. Two sessions hit the same-tool repetition hard-break at step 10/13:

session 9f37534e (auto-nudge): hard-break, tokenInput=262k, content="[Stopped: codemode called 10 times in a row...]"
session fd8c6595 (skills UI):  hard-break, tokenInput=339k, content="[Stopped: codemode called 10 times in a row...]"

Both ran 10 productive codemode calls each — the model had real findings — but never converted them into a final text response. PR #78 stops the loop cleanly; this PR closes the loop on the user-visible side.

How

After the loop exits and exitReason is tagged, check if a stuck signal fired and the model never wrote a real conclusion. If so, run a single streamText with:

• tools: {} — empty, so the model can ONLY write text
• A pure-text system injection: "write your final answer now, summarise what you were trying to do, what you found, what you'd do next"
• 30s timeout via a chained AbortController
• 800 max output tokens

Bounded and one-shot. Failure (timeout, network error, etc.) is non-fatal — the existing [Stopped: ...] notice is already in the response.

Skipped on

• Abort (user asked to stop)
• natural / step-limit / budget-limit exits (auto-continuation handles those, or the model already wrapped up naturally)
• When the model already wrote >=200 chars of non-harness-notice text (no point asking for more)

The <200 chars threshold deliberately scrubs the harness's own bracketed notices before measuring, so a 60-char [Stopped: ...] doesn't count as "model wrote text".

Tests

• New src/final-summary-policy.ts exports two pure functions:
  • stripHarnessNotices(text) — removes [Stopped: ...], [Compacting context ...], [Loop detected ...]
  • shouldRunFinalSummary(inputs) — the boolean decision
• 16 unit tests cover all branches: abort, natural exit, custom threshold, harness-notice scrubbing, stuck exit with empty/short/long text
• All 38 existing tests still pass; typecheck clean

Linked

• PR fix(compaction): force/truncated bypass the minMessages guard #73, fix(compaction): mid-loop trigger uses force=true; guard slice underflow #74, feat(own-loop): in-memory tool-result prune harness safety net #75, fix(own-loop): prune fires on cumulative tokens; compaction logs truthfully #76, fix(own-loop): budget thresholds use projected context, not cumulative cost #77, feat(own-loop): explicit stop guidance + same-tool repetition detector #78 (compaction + prune + projected-budget + same-tool detector chain)
• memory/learnings/dodo-orchestrator-model-matters.md

beep-boop-🤖