fix(web-ui): keep failed agent runs in review instead of auto-completing them

[Atman36]

Fixes #524

Problem

When an agent run fails mid-task (LLM/provider error), the session lands in the Review column with reviewReason: "error" — but the auto-review automation never looks at that signal. If the failed run left partial changes on disk, useReviewAutoActions auto-commits them and auto-moves the card to Done. The card face shows a green "success" dot with the error text. As reported in #524: you return in the morning to a Done card that is actually broken, with no visual trace it failed.

Root cause

useReviewAutoActions receives only board (which carries no session state) and keys its decisions solely on autoReviewEnabled + changedFiles. RuntimeTaskSessionSummary.reviewReason already reaches the web-ui (sessions state in App.tsx, threaded to BoardCard as sessionSummary), and is even in scope at the hook's call site in use-board-interactions.ts — it just was never passed in. Similarly, getCardSessionActivity in board-card.tsx branches on state === "awaiting_review" without checking reviewReason, so failed runs get the green success dot.

Change

• Thread sessions into useReviewAutoActions (ref pattern, same as the existing boardRef), and re-run evaluateAutoReview when sessions changes — so a failure arriving over the state stream is acted on immediately, not only on the next board change.
• In evaluateAutoReview, skip ALL auto actions (arming, auto-commit/PR, move-to-done) for a card whose session has reviewReason: "error" or state: "failed" — mirroring the existing disabled-path cleanup. The same failure gate is re-checked inside both delayed callbacks at timer-fire time (next to the existing column/enabled/mode re-checks), closing the race where a session flips to error after an action was scheduled but before its timer fires. Cards with reviewReason of exit/attention/hook/null (incl. manually dragged cards with no session) behave exactly as before.
• In getCardSessionActivity, failure now dominates success-style final messages: awaiting_review + reviewReason: "error" and state: "failed" render a red error dot with the failure message (falls back to warningMessage, then a static label for PTY exits with no hook activity), keeping the credit-limit branch's precedence. The previously unreachable late failed branches are removed (the compiler flags them as dead after the new early branch).

No server-side or contract changes — the signal already flows to the client.

Tests

• use-review-auto-actions.test.tsx: bug repro (error session + changed files → no git action, no auto-done); happy-path guard (clean exit review still auto-commits/auto-dones); race guard (session fails after the action is scheduled, before timers advance → nothing fires).
• board-card.test.tsx: error dot + message for failed reviews (incl. state: "failed" carrying a success-style final assistant message); green dot unchanged for clean reviews.
• npm run check, npm --prefix web-ui run test (70 files / 503 tests), npm run build all green locally (Node 22).

Prior art

A similar fix was made independently in a fork (VictorGambarini#28) but never upstreamed; this implementation was written from the root cause against current main and additionally covers the PTY exit path (no hook activity) via the warningMessage fallback.

[greptile-apps]

PR author is not in the allowed authors list.