feat(firstmate): walk everything that needs you as one ranked queue

[thiagown1]

Stacked on #5 (CI fix) and #3 (status bugs). Merge #5, then #3; this PR then shows only its own commit.

What Changed

The FirstMate chat's Decisions panel is now a queue of everything in the project that needs the user, shown one item at a time:

• Sources: pending decision cards (turn review / Jev, supervisor questions, provider requests), a pending approval or question with no card yet, a delegated thread whose run failed or was interrupted, and a delegated PR with failing, action-required or conflicting checks.
• Enough context to decide in place: topic goal, the last round summary, the question (for a turn review, the tail of the agent's last message), judge verdict and confidence, PR links with head SHA, "Open thread", the card's options, and a reply box.
• Reply from the panel: the text is queued on the thread with after-current-turn (starts now on an idle thread, never interrupts a running one). The card it answered is closed (turn review → "I'll answer in the thread", supervisor question → cancelled). Provider requests have no reply box; they are answered through their options.
• The item being read never changes or jumps. It is pinned when shown; new or more urgent items only reorder "Up next" behind it. If it is resolved elsewhere it stays on screen with "Already handled elsewhere · Next". Answering, Later (to the back of the queue) or Dismiss moves to the highest-priority item.
• The turn-review card now stores the judge's verdict (source.verdict: outcome, confidence, in-scope) so the queue can rank by it. It is optional, so older cards still decode.

Priority formula (firstMateQueuePriority)

Item	Priority
Provider request (agent paused mid-turn)	100
Turn review, blocked	90 + 10 × confidence
Turn review, needs_user	80 + 10 × confidence
Supervisor question	70 blocking / 40 otherwise
Failed or interrupted run	65
PR checks failing	60
PR action required	55
PR conflicting	50
Turn review, done / continue (judge leaned towards acting alone)	30

Ties go to the newer item. A turn review stored without a verdict ranks as needs_user at 0 confidence when blocking, otherwise 30. Also in docs/internals/firstmate-domain.md.

Why

The user wants to read and decide everything in sequence without opening thread after thread. Before, the panel only listed cards, so a failed run or red checks never reached it (on the real data below it said "Nothing to decide" while six items needed attention), and a list that reorders while you read it makes you lose your place.

Clearing an answered item uses the same mechanism as #3: starting the thread's next turn cancels its turn-review card, and a running thread's items leave the queue.

UI Changes

Worktree dev server seeded with a VACUUM INTO copy of the real FirstMate database, 1440×950.

Before	After

After Later: the failed run moved to the back and the next item (checks failed on pingdotgg#2202) took its place:

Not exercised in the browser: Send. On the copied database it would start a real provider turn in a real worktree. The reply wiring was checked by typecheck and the queue/cursor tests; no Jev card was pending in the copied data, so the verdict line was only covered by tests.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes (no animation)

Tests

• FirstMateDecisionQueue.logic.test.ts (new): priority order and the no-verdict fallback; queue building from every source with context; turn review and PR items leave while the thread runs; no duplicate for a provider request that has a card, and no reply on it; newer first on ties; the item being read stays pinned while a more urgent one arrives; a resolved-elsewhere item stays until "Next"; Later/handled ordering.
• FirstMateTurnReviewReactor.test.ts: the opened card now carries the verdict.
• vp test run on the FirstMate web components, the turn-review and delivery reactors, and packages/contracts: 37 files, 550 tests, all passing.
• tsc --noEmit: apps/web, apps/server, packages/contracts, packages/shared clean.
• CI: stacked on fix(ci): drop the msgpackr-extract dependency upstream removed #5, which fixes the three jobs that were already red on main.

Done by Claude Opus 5.5 (Claude Code, via T3 Code).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	13.5 KiB	—	15.1 KiB	✅
Codex	Thread snapshot wire	—	7.1 KiB	—	7.3 KiB	✅
Codex	Live turn WebSocket wire	—	6.4 KiB	—	7.8 KiB	✅
Codex	Live turn WebSocket decoded	—	56.2 KiB	—	66.4 KiB	✅
Codex	Live turn messages	—	9	—	21	✅
Claude	Total thread wire	—	13.5 KiB	—	15.1 KiB	✅
Claude	Thread snapshot wire	—	7.1 KiB	—	7.3 KiB	✅
Claude	Live turn WebSocket wire	—	6.4 KiB	—	7.8 KiB	✅
Claude	Live turn WebSocket decoded	—	57.0 KiB	—	66.4 KiB	✅
Claude	Live turn messages	—	9	—	21	✅

Baseline: unavailable · PR result: 8860ba2 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.9 KiB
• Claude decoded thread snapshot: 114.6 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.