fix(ui): don't evict a user message sent during local-model warm-up (#11670)

[lalalune]

Closes #11670

Root cause

useChatSend's post-turn reconcile (loadConversationMessages) full-replaces the thread with server truth. When the server never persisted the user's turn — a send during local-model warm-up where the runtime-ready hold expired (503 Agent is not running from conversation-routes.ts; the user message is only persisted after runtime resolution), or a stream that completed empty with nothing persisted — the reload returned a thread without the user's message and silently evicted the optimistic bubble. No failed state, no notice on the empty-success variant, text unrecoverable. The suite never caught it because its loadConversationMessages mock was a no-op { ok: true } instead of the real full-replace.

Fix (root cause, no fallback sludge)

• restoreEvictedUserTurn() — after every post-send reconcile, detect that the just-sent user turn is missing from server truth (text match among user turns no older than the send minus 60s clock-skew slack, so an identical message from an earlier exchange can't mask an eviction) and re-attach the user's bubble plus a retryable failed assistant turn (failureKind: "provider_issue" → the overlay's existing Retry chip). The send fails loudly; one tap re-delivers once the model is ready — mirroring the existing reattachInterruptedPartial pattern for assistant partials.
• Wired into both reload sites of runQueuedChatSend (success + non-404 error path) and both sites of sendActionMessage. Validation-4xx rejects excluded (their draft goes back to the composer — re-attaching would duplicate). No-op when the server persisted the turn or the reload never replaced local state.
• handleChatRetry fallback now also drops the optimistic (temp-) user turn it retries, so the resend's fresh optimistic bubble doesn't transiently duplicate the message.

Evidence

Row	Artifact
Before (develop) — message evicted
After — optimistic bubble in flight
After — bubble survives 503 + reconcile, Retry chip attached
After — Retry delivers exactly once when model ready

• Regression tests (red on develop): 8 new tests in useChatSend.test.tsx that emulate the REAL full-replace reload. 6 red against develop's useChatSend.ts (verified by swapping it in: 6 failed | 2 passed), all green with the fix. 2 guard the no-false-positive side (persisted silent turn → no duplicate/no spurious failed turn; validation reject → composer restore only). Full file: 40/40 green. Two pre-existing fixtures updated from toy timestamp: 1 to realistic epoch-ms (a persisted turn's real timestamp) — intent unchanged.
• Real-pixel e2e: new bun run --cwd packages/ui test:warmup-eviction-e2e — bundles the REAL useChatSend pipeline + real ContinuousChatOverlay, warm-up 503 simulated at the client-API boundary, real typing/Enter/click in headless chromium. 8/8 assertions green (screenshots + console log committed under src/components/shell/__e2e__/output-warmup-eviction/). Same harness against develop's useChatSend: reds exactly at "user bubble SURVIVES the warm-up 503 + reconcile reload".
• Frontend logs: output-warmup-eviction/console.log shows the real path: send → 503 notice → reload conv-1: server holds 0 messages (the eviction moment) → retry → server holds 2 messages.
• Suites: bunx vitest run src/state/ src/components/shell/ → 95 files, 1166 passed. bun run --cwd packages/ui typecheck clean. Biome clean on touched files. Re-verified after rebase onto origin/develop (cdd87af).
• Backend logs / real-LLM trajectory: N/A — client-side state-management fix; no agent/prompt/model behavior change. The server's warm-up 503 contract (resolveRuntimeForChatTurn) is unchanged and is what the boundary simulation reproduces.
• Video: N/A — the three sequential full-page frames above capture the entire flow (send → survive → retry-deliver); the e2e re-generates them deterministically on every run.

🤖 Generated with Claude Code

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 4cc8ab4c-4682-4e68-bf05-916a5a620592

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/11670-warmup-message-eviction

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.