[Bindy-lbb]

First, isolated #411 fix: indeterminate timeout/cancellation copy now receives whether model.delta output exists. It only says partial output was preserved when the turn record actually has partial output; otherwise it says no output was received. This intentionally does not change adapters or timeout semantics.

Validation: npm run build; node --test apps/server/dist/test/turns.test.js apps/server/dist/test/turns-cancel.test.js (29 passed).

[waterbro-8]

Code review APPROVE (not product 验收), head 0c09b382a9c9a34d73df4827778699edd6f7c964.

roleweave #411 Split A only：indeterminateMessage 吃到 partialOutput !== undefined，timeout/cancel 只在 turn record 真有 model.delta 时才说 partial output was preserved；否则 “before any output was received”。适配器和超时模型未改，不应 close #411。

覆盖：AC-003（有 delta）、新用例（无 delta）、cancel 无输出路径的文案已改对。CI 11/11 green。

非产品验收。不自动合。胡奕舟 can merge. Gemini stream-json / idle timeout 仍走后续 PR。

[PeterGuy326]

Review (roleweave #419, first slice of #411)

Re-review of head 0c09b38. Base verified: single commit on top of origin/main.

Verified by execution, in a clean worktree: npm run build exit 0; turns.test.js + turns-cancel.test.js = 29 tests / 29 pass / 0 fail, matching your body exactly. CI green on Node 24 macOS + Ubuntu.

The claim holds, and the way you did it is the reason it holds:

• extractPartialOutput filters strictly on e.type === "model.delta", so "partial output was preserved" is backed by an actual model delta and nothing else.
• The message argument and the persisted field come from the same partialOutput !== undefined expression (apps/server/src/routes/turns.ts:455 and :458). Copy and record cannot diverge, which is the failure mode this fix is about — deriving hasPartialOutput separately would have left room for exactly the drift #411 reported.
• Streaming and timeout semantics untouched: the diff is message branches plus tests; retryable logic is unchanged, and the other indeterminate codes still resolve through INDETERMINATE_MESSAGES.
• Moving turn_timeout/turn_cancelled out of the table into the function is fine, though it does mean those two codes no longer sit with their siblings — worth a comment if the table grows.

Non-blocking

• turn_cancelled positive branch has no test. After your change, turns-cancel.test.ts:64 asserts only "before any output was received". The hasPartialOutput === true path for cancellation is untested, while turn_timeout has both (AC-003 plus your new case). Since this PR's whole point is that the two claims must track reality independently per code, add the symmetric cancel-with-delta case.
• Empty deltas count as preserved. A turn that emitted only zero-length model.delta events yields partialOutput === "", so partialOutput !== undefined is true and the message still says "partial output was preserved" while record.output is the empty string. Strictly the field is present, so this is not a lie, but decide whether an empty join should read as "no output received" — if so, the predicate becomes partialOutput !== undefined && partialOutput !== "".
• Rebase note, not a defect here. #418 also edits apps/server/src/routes/turns.ts (different hunks: :69/:386 vs :455/:585/:592). Both are mergeable against main today, but whichever lands second will need a rebase.

Verdict: no correctness concerns in scope. I'd merge this once the cancel-branch test is added, or as-is if you'd rather keep the slice minimal and follow up.