P2: Structured worker result in the lead's wake notice, rendered as a card #1427

Activity

1. currentbits

   added

   plan:todoPlanboard statusPlanboard status

   on Oct 4, 2026
2. currentbits commented on Oct 4, 2026

   

   currentbits

   on Oct 4, 2026

   OwnerAuthor

   More actions

   A code-level pass produced a design for this issue. Full detail: reports/T3 Code orchestrator deep dive.md on branch coder/i-want-you-to-research-t3-codes-orchestr-fae77a. Revised size: M, about 350–450 lines including tests (2–3 days), up from S+S.

   Corrections to the issue body

   • T3 has nothing to port here. In T3's code a child's result is just its last assistant message, and the parent is woken with a fixed pointer sentence. The "decisions, files, tests, open issues" delta exists only in T3's design docs. This is original design work.
   • The diff can miss the last turn. The checkpoint on the non-verify path is fire-and-forget (runner.js:1526-1532), so the branch can lag behind the worker's final edits. Workers must await it first. This is race R3 in the new race issue.
   • The itinerary has no open items. It has risks (reviewItinerary.js:43-53), so use those.
   • Appending text breaks the transcript fold. src/workerActivity.ts:51-62 exact-matches the canned notice text. Carry the data on the message row instead.

   Design

   interface WorkerResult {
     workerId: string; title: string; status: "done" | "failed";
     branch: string | null; baseSha: string | null;
     diffstat: string | null;   // git diff --stat=100 --summary <base> HEAD, ≤30 lines / 1,500 chars
     flags: string[];           // --summary lines matching /mode 120000|delete mode/ (the core/dist symlink case)
     uncommitted: number;       // git status --porcelain count
     verify: { ok: boolean; command: string; attempt: number; timedOut: boolean } | null;
     ruledOut: string[];        // invalidated hypotheses, ≤5
     risks: string[];           // itinerary risks, ≤5
   }
   // ChatMessage gains workerResults?: WorkerResult[] on fromNotice rows

   1. buildWorkerResult(worker) never throws and uses async git (gitTryAsync). The base comes from snapshotSha in crewIntegration.js:379-385. It returns null diffstat for worktree:false workers and remote projects.

   2. queueOrchNotice builds the result, then enqueues the P1: Durable crew notices and honest restart recovery with consent-gated resume #1426 record with result plus a text block of at most ~2,000 chars, then flushes.

   3. flushOrchNotices passes workerResults through startRun into the appendMessage extras (runner.js:8461-8474). "Not delivered" events carry them too.

   4. Renderer: a WorkerResultCard shows

      • status and verify chips;
      • the diffstat, collapsed after 8 lines;
      • a warning row for flags;
      • "Ruled out" and "Risks" lists;
      • a Review & integrate button wired to the existing onOpenCrewIntegration (ThreadView.tsx:804).

      It renders in UserMessageBlock before the worker-activity fold (ThreadView.tsx:1403-1466), so it's visible the moment it arrives. That's the lesson from T3's #15063.

   Tests

   • A worker that commits a symlink: the record and the wake row both carry create mode 120000.
   • On the non-verify path, the diff includes the last turn's files.
   • Caps and truncation hold.
   • A failed worker reports uncommitted > 0.
   • The renderer shows [data-worker-result], and rows without it still fold.

   Depends on #1426's record shape and the R3 fix.
3. currentbits

   added

   plan:doingPlanboard statusPlanboard status

   and removed

   plan:todoPlanboard statusPlanboard status

   on Oct 12, 2026
4. currentbits commented on Oct 12, 2026

   

   currentbits

   on Oct 12, 2026

   OwnerAuthor

   More actions

   Implemented in f0c1b70 on branch coder/p2-structured-worker-result-in-the-lead--7e14e2. Not landed yet.

   • Step 1: workerResultBlock in electron/runner-orch-notices.js adds a [worker result <id>] … [/worker result] block to the done/failed notice. The block holds the diff stat and summary of leadSnapshotSha..branch, the verify verdict (only while a verifyCommand is set), invalidated hypotheses, and review-itinerary risks. The body is capped at 2000 chars, and mode 120000 lines are moved to the top so the cap can't cut them.
   • Step 2: src/workerResult.ts parses the block. WorkerResultCard renders it in the lead's transcript, and its "Integrate into lead" button calls git:integrateWorker directly, which is a human action through IPC. Notice lines outside the block stay visible next to the card.
   • Tests: electron/test/worker-result-block.test.js covers the symlink line, the cap and itinerary risks. test/threadView.test.tsx covers rendering the card, clicking the button, and parsing the runner's real block output. Electron suite: 4672 pass. The 2 failures (automation-create-mcp trigger fields, store prLinks) also fail on clean HEAD.
5. currentbits

   added a commit that references this issue on Oct 12, 2026

   Structured worker result in the lead's wake notice, rendered as a card (

   ...

   8ff5207
6. currentbits

   closed this as completedin #1606on Oct 12, 2026
7. currentbits

   added 2 commits that reference this issue on Oct 12, 2026

   Structured worker result in the lead's wake notice, rendered as a card (

   ...

   3b650a5

   Structured worker result in the lead's wake notice, rendered as a card (

   ...

   ebf4391