feat: /work skill — plan-to-implementation orchestrator

[shohu]

Summary

• Adds /work, a new skill that closes the workflow gap: plan → work → review → ship
• Reads plan artifacts from /plan-ceo-review, /plan-eng-review, and /plan-design-review
• Automatic mode selection: Standard (single-agent Ralph Loop) for small plans, Parallel (subagent + git worktrees) for larger plans with independent lanes
• Includes Code Standards enforcement in spawned agent prompts ("every line is a liability")
• Logs build results to reviews.jsonl so /ship's Review Dashboard tracks implementation

How it works

/plan-ceo-review  ─┐
/plan-eng-review  ─┼──▶  /work  ──▶  /review  ──▶  /qa  ──▶  /ship
/plan-design-review┘

1. Gather — reads CEO plan, eng review, test plan, design doc from ~/.gstack/projects/
2. Decompose — breaks plan into discrete tasks with dependency tracking and lane assignment
3. Execute — Standard mode (Ralph Loop: pick → implement → test → commit) or Parallel mode (subagent per lane in isolated git worktrees + dedicated reviewer agent)
4. Verify — full test suite, plan conformance check
5. Log — persists build result to reviews.jsonl
6. Handoff — suggests /review → /qa → /ship

Why a new PR (not #218 or #580)

	#218 /build	#580 /work	This PR
Parallel mechanism	TeamCreate (experimental)	Agent tool only	Subagent + worktree (stable API, file isolation guaranteed)
Build/verify separation	No	No	Yes — dedicated reviewer agent in parallel mode
Code Standards	Custom section	None	ETHOS.md reference + Liability Principle in agent prompts
Codex dependency	No	Required for key feature	None — works with Claude Code alone
Task decomposition guide	Wave-based (conceptual)	Undefined	Explicit rules (1 task = 1 file, max 15, same-dir = same-lane)
Template size	500 lines	303 lines	350 lines (balanced)

Key differentiators from industry best practices (Addy Osmani, Claude Code docs):

• 1-file-1-owner rule via worktree isolation (no merge conflicts between agents)
• Build/verify separation (reviewer agent has fresh context, not self-assessment)
• Task granularity guide (5-15 min per task, explicit dependency tracking)

Design decisions

• No Codex delegation — kept out of scope to minimize complexity. Can be added in a follow-up PR for users who have Codex installed
• No Agent Teams (TeamCreate) — experimental feature, disabled by default. Worktree isolation achieves the same file safety with stable APIs
• No gen-skill-docs.ts change needed — template auto-discovery in scripts/discover-skills.ts picks up new */SKILL.md.tmpl files
• "Trust the plan" — /work executes, it does not re-review. Scope debates belong in /plan-ceo-review and /plan-eng-review
• Tests are part of each task — no separate "write tests" phase. Each task includes tests for its code

Test plan

• bun run gen:skill-docs generates work/SKILL.md without errors
• bun run gen:skill-docs --dry-run shows FRESH for all 33 files
• bun test — 1 pre-existing failure in path-validation.test.ts (unrelated, fails on clean main)
• Manual: invoke /work after running /plan-eng-review on a feature branch (Standard mode)
• Manual: invoke /work on a plan with 5+ tasks across multiple directories (Parallel mode)
• Manual: verify reviews.jsonl entry appears after /work completes
• Manual: verify /ship Review Dashboard shows /work run

Files changed

File	Change
work/SKILL.md.tmpl	New skill template (~350 lines)
work/SKILL.md	Generated from template

Related

• Closes there's no /implement skill after /office-hours completes plan #242 (there's no /implement skill after /office-hours completes plan)
• Addresses 没有代码配套生成的 skill #182 (没有代码配套生成的 skill)
• Alternative to feat: /build skill — plan-to-implementation orchestrator #218 and feat: /work skill with Codex delegation mode #580 with different architectural choices

---

This contribution was developed with AI assistance (Claude Code).

🤖 Generated with Claude Code