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
- 参考文档agents/openai.yaml246 B
- 脚本scripts/evaluate-goal.mjs7.0 KB
- 脚本scripts/select-qa.mjs2.2 KB
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
- Put tickets sharing a product surface, file surface, or dependency chain in one sequential lane.
- Put only disjoint lanes in separate registered Git worktrees.
- Use no more writers than disjoint lanes.
- Start disjoint lanes together. Run tickets inside a lane sequentially.
- Keep the parent responsible for conflicts and one integrated candidate.
- 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.
Route for speed
- Keep trivial mapping and integration on the parent.
- Use
fast_implementeron Spark Low for exact bounded implementation. - If Spark is unavailable or fails the same bounded packet twice, use the fastest currently validated implementation route; do not wait for Luna or escalate to Sol merely because it is familiar.
- 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
inheritis not execution evidence.
Run the fixed implementation loop
For each ticket:
- implement the smallest complete behavior;
- run the affected linter or typecheck when it is locally available and fast;
- run only focused local tests that prove the changed contract;
- run React Doctor or the applicable Impeccable CLI only when React, interaction, or visual implementation changed and the repo exposes it;
- 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
working-backwards
Define a software feature through PRFAQ and staged documents when the user requests working backwards before coding.
flow-implement
Implement a requested software change through verification, corrections, and its authorized delivery endpoint.
development-steward
Run the scheduled, read-only weekly maintenance review for the six explicitly opted-in repositories and prepare one concise private report for Check-in.
grill-with-docs
Run one recoverable interview for functional, visual, or mixed software grills; reuse equivalent answers and offer the shared HTML questionnaire.
global-agent-guardrails
Install, audit, test, or update the Development System's global destructive-command guard for Codex, Claude Code and T3 Code. Use when configuring agent safety hooks or investigating whether dangerous shell operations are blocked.
impeccable
Design or refine frontend UI using project context and a focused design playbook; also handles visual critique and polish.