feat: add persisted whole-chat summary with background generation - #26657

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0bc176d8c4

Deep review of PR #26657

Well-structured feature addition. The optimistic concurrency control, cost attribution, and background generation patterns are solid. Three findings worth addressing before merge.

🔴 Migration version collision

000530_chat_summary collides with 000530_relay_host_nats_port which already exists on main. golang-migrate/iofs returns source.ErrDuplicateMigration for duplicate version numbers, blocking all deployments. Renumber to 000531 (or whatever the next free slot is at merge time).

🟡 Stale chat struct in cadence gate

generateAndStoreChatSummary receives the chat struct snapshotted when the turn finished, but messages are re-read from DB. If two consecutive turns complete quickly, both background goroutines see the same stale chat.Summary.Valid / chat.SummaryGeneratedAt and both pass the cadence gate. Both generate a summary and call updateChatSummary. Since UpdateChatSummary checks history_version but does not update it, both writes can succeed; the second silently overwrites the first. The summaries are likely identical (same input), so data is not corrupted, but one LLM call is wasted.

A cheap fix: re-read the chat row at the top of generateAndStoreChatSummary:

chat, err = p.db.GetChatByID(authCtx, chat.ID)

🟡 p.inflight.Go bypasses shutdown admission gate

maybeGenerateChatSummaryAsync calls p.inflight.Go(func(){...}) directly instead of p.goInflight(func(){...}). The comment explains why (deadlock avoidance with drainInflight). The goroutine IS tracked in p.inflight so it cannot outlive the server. But combined with context.WithoutCancel(ctx) and chatSummaryWorkTimeout = 120s, a summary launched right before shutdown could block Close() for up to 2 minutes. This creates an asymmetry: finalizeSuccessfulTurnStatusLabelAndPush uses goInflight and is silently dropped during shutdown, but summary generation runs regardless.

If the 120s shutdown delay is acceptable, this is fine. If fast shutdown matters, consider adding an early select on p.ctx.Done() inside generateAndStoreChatSummary before the model call.

Solid feature addition. The optimistic concurrency via history_version, cost attribution model, and background generation pipeline are well-designed. The test suite covers the critical paths: cadence gate, transcript rendering, validation, override resolution, SQL staleness race, cost source tagging, and frontend event merging. Six reviewers independently confirmed the event isolation pattern works correctly.

1 P0, 2 P2, 7 P3, 3 P4, 5 Nit.

The P0 (migration collision) blocks all CI. The two P2s (frontend cross-clobber between event types, shutdown admission bypass) are the substantive findings requiring fixes.

"The goroutine IS tracked in p.inflight so it can't outlive the server, but combined with context.WithoutCancel(ctx) and chatSummaryWorkTimeout = 120s, a summary launched right before shutdown could block Close() for up to 2 minutes." (Hisoka)

🤖 This review was automatically generated with Coder Agents.

All P0/P2/P3 findings from round 1 are addressed. The fixes are well-executed: the deduplication extractions (CRF-1, CRF-6) are genuine generalizations, the frontend cross-clobber fix (CRF-3) correctly scopes each field to its own event type with regression tests, the shutdown admission guard (CRF-4) avoids the deadlock concern while preventing 120s drain delays, and the cadence gate re-read (CRF-5) eliminates redundant LLM calls.

Contested findings closed by panel:

• CRF-14 (P3, no orchestration test): 5/6 accept. Components are well-tested; orchestration is a linear pipeline with no conditional branching the unit tests miss.
• CRF-18 (P4, first-turn threshold): 6/6 accept. Product decision.
• CRF-21 (P3, premature full load): 6/6 accept. Messages are needed for both the cadence gate and the transcript renderer.

1 new P4 below.

"The code held up. The recordHiddenUsageMessageTx refactor is clean, the UpdateChatMessageCostSource two-step insert-then-tag avoids threading cost_source through the batch insert, and the history_version staleness guard mirrors the proven UpdateChatLastTurnSummary pattern." (Hisoka)

🤖 This review was automatically generated with Coder Agents.

All 23 findings resolved. CRF-23 (P4, test pin for dotted-identifier sentence counting) addressed in 5d0a775 with TestCountSentenceTerminators. The test asserts dotted identifiers (auth.rbac.Policy, file.go) are not counted as sentence boundaries and that a dotted-identifier-dense summary passes validation. Removing the unicode.IsSpace boundary guard now fails the test.

CI: 25 passed, 2 failing (UI Review, UI Tests). The UI failures appear unrelated to the changes in this PR.

No further findings.

🤖 This review was automatically generated with Coder Agents.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 877bbf0ffb

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7a57ac09c5

Suggestions for the remaining mechanical fixes. The complete suggestion set was applied together in an isolated worktree and the affected package tests passed.

🤖 This review was automatically generated with Coder Agents.