fix(core): size background turns for hidden reasoning

[justinhelmer]

Intake and memory reflection now reserve room for hidden reasoning without misclassifying capped replies. Wrapped intake cuts retain their structured-attempt history, and production composition tests prove both background paths receive their model card’s cap field.

Why: #2267 blocks first-party OpenAI defaults because Responses counts hidden reasoning inside the output cap. The original fix missed a wrapped retry failure, production wiring proof, and the behavior specs.

Where to look

1. Wrapped cap classification preserves structured attempts while classifying the underlying cap cut before generic provider failure. ⚠ A mistake here silently changes outage handling.
2. The regression test reproduces a malformed first answer followed by a capped retry.
3. Intake production composition carries the resolved card’s cap field into the live decision dependencies.
4. Reflection production composition proof exercises scheduling through model-card resolution and observes the expanded completion request.
5. Memory contract defines the dynamic reasoning-aware cap and rejects every capped response before parsing.
6. Intake validation contract binds cap sizing, wrapped-cut behavior, and production wiring to named tests.

Feedback wanted: Check that unwrapping only StructuredAskError’s direct preserved cause is the right boundary, and that the two composition tests cover the live bridges without over-coupling startup.

Risk: Incorrect classification could turn an output-cap cut into a permanent provider failure or omit the reasoning allowance in production. Revert this single commit to restore the prior behavior.

Verified: 97 focused tests, four exact regression/wiring tests, root TypeScript, specs/docs/hygiene/coverage/test-guard, Prettier, and npm run fix passed; CI remains the full-suite gate.

Decisions (3)

• Unwrap only the structured seam wrapper. askStructured is the boundary that wraps a later model throw after collecting violations, so intake examines that known wrapper instead of recursively interpreting arbitrary error causes.
• Test the production bridge, not only each endpoint. intakeDecisionDeps is used by src/index.ts, while reflection is exercised through scheduleReflection; both tests observe the completion request after model-card resolution.
• Reject every capped reflection. A syntactically complete object can still be semantically incomplete when the provider reports the cap, so no max_tokens response reaches parsing or storage.

Validation (7 criteria)

Criterion	Proof
A capped retry after malformed intake output remains an output-cap error and keeps the first violation	npx vitest run src/core/intake.test.ts -t "an output-cap cut after a malformed answer stays an output-cap error and keeps the structured attempt" — passed
Intake production composition carries the resolved cap field into the verdict request	npx vitest run src/intakeModel.test.ts -t "production intake deps carry the completion's cap field into the verdict call" — passed
Reflection production composition carries the resolved cap field into the completion request	npx vitest run src/core/memory/index.test.ts -t "carries the reflection model card's cap field into the completion request" — passed
Complete capped reflection output is rejected without a write	npx vitest run src/core/memory/reflection.test.ts -t "every capped response is rejected even when the object is complete; another unusual stop is named" — passed
All focused intake/reflection behavior remains green	Four exact test files — 97 tests passed
The behavioral specs resolve and cover every changed source path	npm run specs:check; npm run specs:coverage -- --changed origin/main...HEAD --test-guard; npm run docs:check — passed
Changed code is typed, formatted, hygienic, and generated artifacts are current	root tsc --noEmit; changed-file Prettier; npm run hygiene:check; npm run fix — passed

For agents

Rebased onto current origin/main, resolving the intake test conflict by retaining both the typed provider-failure cases from the base and this PR’s cap tests. The branch is one coherent commit above main.

Red step: the new malformed-then-capped intake test failed because the decision reason was The model provider refused the call… rather than containing answer cut at the output cap; it passed after unwrapping the preserved OutputCapError.

The first npm run fix attempt used a symlinked dependency tree and rewrote generated Nuxt import paths; those unrelated changes were discarded. A local copy-on-write dependency tree then let npm run fix complete successfully without generated path drift.

[coreplane-switchboard]

Changes requested: Two minor issues remain: wrapped output-cap errors are misclassified, and the memory spec contradicts the new cap behavior.

Warning

Changes requested · head 6f0b663 · 2 findings: 2 minor

Severity	Finding	Where
minor	F1 Wrapped output-cap errors become provider failures	src/core/intake.ts:274
minor	F2 Spec contradiction — docs/reference/specs/memory.md item 12: reflection no longer uses or reports a fixed 1024-token cap	docs/reference/specs/memory.md:27

Full review

F1: If intake first receives a malformed answer and the re-ask hits the output cap, askStructured wraps OutputCapError in StructuredAskError. The instanceof OutputCapError check then misses it and records a generic permanent provider failure, potentially emitting the wrong failure sentence. Inspect the preserved cause when classifying truncation and retain the structured attempts.

F2: Memory spec item 12 and validation rows still require a fixed 1,024-token cap, report truncation at 1,024, and allow a complete object that survived truncation to be written. The implementation now uses a reasoning-aware dynamic cap and rejects every max_tokens response without writing. Update the behavior and validation rows to describe the new contract.

[justinhelmer]

Re-review requested at ad5cba5 — rebased onto current main; fixed wrapped output-cap classification, aligned memory and intake specs, added production cap-propagation proofs, and preserved the typed provider-failure work. Independent OpenCode review is in progress.

[justinhelmer]

OpenCode re-review at ad5cba5: ready to merge. All four prior findings are resolved, the #2232 typed-provider behavior remains intact, no additional findings were found, and all CI checks are green.

[coreplane-switchboard]

LGTM: The prior findings are resolved and the reasoning-aware cap behavior is now wired, specified, and covered end to end.

Note

Approved · head ad5cba5 · no findings

Full review

The prior findings are fully resolved; no remaining issues found.

[github-actions]

Auto-approved: coreplane-switchboard[bot] reviewed this PR and posted an LGTM verdict (see its review). This repository opted in through its REVIEW_BOT_LOGIN and REVIEW_BOT_ID variables.

[justinhelmer]

Production inclusion receipt — 1.260.6. v1.260.6 is full commit 903ae2855833637fbcc0c0a22d54e7268be75965; GitHub compare reports this PR's merge commit as the merge base (release ahead 2, behind 0). The release workflow succeeded, bot/memory/resident /healthz report the full tag SHA, and the authenticated sandbox gate passed on the tag. This is deployment evidence only; it does not claim intake or reflection hit an output-cap cut in production.