AGENT SKILL

work-multiple

Execute several authorized software tickets through conflict-aware parallel worktrees and one integrated candidate. Use only when the user explicitly invokes $work-multiple or clearly says to work multiple tickets in parallel; never infer activation from ticket count, project size, or urgency.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
work-multiple
description
Execute several authorized software tickets through conflict-aware parallel worktrees and one integrated candidate. Use only when the user explicitly invokes $work-multiple or clearly says to work multiple tickets in parallel; never infer activation from ticket count, project size, or urgency.

02 资源文件 3

03 SKILL.md 正文

Work Multiple

Optimize for one constant goal:

Deliver correct functionality to the authorized state as fast as possible, with useful quality, the lowest measured cost, and the least complexity.

This skill is an explicit execution mode, not a router. If the current request does not clearly authorize multiple-ticket execution, stop without creating worktrees or agents.

Freeze the run

Require at least two authorized tickets. For each ticket, record:

  • its acceptance criterion;
  • owned product and code surfaces;
  • declared blockers;
  • whether it changes behavior or only mechanical text/style;
  • the publication and production boundary.

Do not create, expand, reorder, merge, release, or deploy tickets unless that operation is separately authorized.

Build the smallest lane graph

  1. Put tickets sharing a product surface, file surface, or dependency chain in one sequential lane.
  2. Put only disjoint lanes in separate registered Git worktrees.
  3. Use no more writers than disjoint lanes.
  4. Start disjoint lanes together. Run tickets inside a lane sequentially.
  5. Keep the parent responsible for conflicts and one integrated candidate.
  6. Create one review and one pull request, when publication is authorized.

Before starting writers, capture each lane's absolute worktree path, branch, HEAD, and git worktree list --porcelain evidence. Include the expected absolute path and branch in every writer packet. The writer must verify both before its first read or edit and reject every patch target outside that Git root. A mismatch interrupts the writer before any relocation, restoration, or integration begins.

Route for speed

  • Keep trivial mapping and integration on the parent.
  • Use the active harness roster's explicit fast_implementer route for exact bounded implementation. Codex resolves that route to Luna High; Factory uses its separately versioned explicit droid mapping.
  • If the fast route fails the same bounded packet twice, use the active harness's judgment-capable implementer only when the packet needs judgment; otherwise keep the bounded work on the parent. Browser-QA models remain reserved for browser QA.
  • Use discovery or QA agents only when their parallel work saves more time than their startup and context cost.
  • Record actual harness runtime worker, model, reasoning, session, and event metadata. A planned model or inherit is not execution evidence.

Run the fixed implementation loop

For each ticket:

  1. implement the smallest complete behavior;
  2. run the affected linter or typecheck when it is locally available and fast;
  3. run only focused local tests that prove the changed contract;
  4. run React Doctor or the applicable Impeccable CLI only when React, interaction, or visual implementation changed and the repo exposes it;
  5. never run a full repository suite unless the user requests it for this run.

After integration, run two independent code reviews concurrently. Give both reviewers only the fixed diff, acceptance criteria, and their own review brief. Do not give either reviewer the implementation narrative or the other report. Resolve blocker/high findings before continuing.

Select manual QA by value

Run node scripts/select-qa.mjs --input <change.json> from this skill.

  • No UI change: manual UI QA is not applicable.
  • Label, copy, icon, or obvious style-only change with no behavioral risk: skip manual UI QA.
  • Interactive behavior, navigation, mutation, responsive behavior, a meaningful regression, or a critical user flow: run one direct manual acceptance flow.
  • If the useful flow no longer exists, verify the closest observable acceptance criterion instead of recreating obsolete ceremony.

Record the decision and one-sentence reason. Do not cite route, destination, or test totals as quality.

Measure the goal

Start timing when execution begins and stop functionalEvidenceMs when every acceptance criterion has observable evidence. Review, PR, CI, deployment, and promotion are separate timings.

Finalize a private run record, then execute:

node scripts/evaluate-goal.mjs --input /private/path/run.json

The evaluator reports correctness, quality, speed, cost, and simplicity as separate axes:

  • tiny work meets the speed goal only at five minutes or less;
  • five to ten minutes is an exception, not success;
  • above ten minutes stops the run for process audit;
  • multiple work needs a comparable sequential baseline to prove speedup;
  • missing comparable authoritative price evidence makes the cost goal unproven; token counts remain resource evidence, not price;
  • any missing acceptance proof, failed relevant check, non-independent review, or unresolved blocker/high finding fails quality.

Do not create a composite score. Return met, missed, or unproven with the exact failed or unavailable axes.

04 同仓库更多 Skills