plan-hardening
Team-lead drives plan hardening after the current plan state already exists in repo docs.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Team-lead drives plan hardening after the current plan state already exists in repo docs.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
Reusable QA orchestration skill for GitHub PRs. Use for multi-pass QA, CI monitoring, and template-driven findings and final quality reports.
Orchestrate Wyvern sprint/phase work inside a single Cursor session. Parent spawns reviewers and rust-developer; Task(cursor-quality-mgr) enforces spawn proof, aggregates fenced JSON evidence, triages findings, publishes PR reports, and owns the QA gate. Use when the user asks for /cursor-orchestration, Cursor-session phase/sprint orchestration, or same-session QA via cursor-quality-mgr. Never use ATM quality-mgr or codex-orchestration while this skill governs the session.
Orchestrate sprint work where team-lead coordinates, cwy is the sole developer, and quality-mgr enforces the QA gate.
Orchestrate pre-dispatch QA finding triage as team-lead. Launch one qa-triage agent per finding, collect phase-scoped Turtle records, aggregate by promoted branch, and only then dispatch branch-scoped fix assignments to arch-ctm.
Orchestrate multi-sprint phase execution as team-lead. Manages sprint waves, scrum-master lifecycle, PR merges, cwy reviews, and integration branch strategy. This skill is for team-lead only, not for scrum-masters.
Repair Claude teammate routing after same-session compaction or resume when wyvern still exists on disk and the saved leadSessionId still matches the current SESSION_ID, but SendMessage or teammate reachability is broken.
| name | plan-hardening |
| version | 1.5.0 |
| description | Team-lead drives plan hardening after the current plan state already exists in repo docs. |
| depends_on | {"codex-orchestration":"0.x"} |
Audience: team-lead only.
Use this only for phase-plan hardening before implementation starts or resumes.
team-lead routes the process but is not the authority for rewriting the
plandevelop before
startingteam-lead is a coordinator only and must not redirect hardening flow,
offer accept-and-proceed shortcuts, or stop to ask the user what to do when
an in-scope hardening loop is still mechanically actionableSprint plan approved by:
plan-scope-reviewercritical-plan-reviewerquality-mgrAlways use:
.claude/skills/plan-hardening/sprint-planning-guidelines.md| # | Route to | Input required | Output expected | Read before executing |
|---|---|---|---|---|
| 1 | arch-ctm | vars file | step-1 fenced JSON | steps/step-1.md |
| 2 | plan-scope-reviewer (background) | context + step-1 JSON | step-2 fenced JSON | steps/step-2.md |
| 3 | arch-ctm | step-2 JSON | step-3 fenced JSON | steps/step-3.md |
| 4 | critical-plan-reviewer (background) | context + step-3 JSON | step-4 fenced JSON | steps/step-4.md |
| 5 | arch-ctm | step-4 JSON | step-5 fenced JSON | steps/step-5.md |
| 6 | quality-mgr | step-5 JSON + QA vars file | codex-orchestration plan-QA handoff | steps/step-6.md |
team-lead must keep a round table for every /plan-hardening run.
Minimum columns:
| Round | Step | Reviewer | reviewed_commit | status | blocking | important | minor | findings_hash | supersedes | Note |
|---|
Use the example in:
.claude/skills/plan-hardening/examples/plan-hardening-rounds.example.mdplan-scope-reviewer and critical-plan-reviewer both default to a
3-cycle capplan_scope_review_cycle_limitcritical_review_cycle_limitreview_cycle_limitreview_cycle_indexteam-lead must default them to 3Cycle-cap behavior:
FAIL from plan-scope-reviewer or critical-plan-reviewer must be
routed to arch-ctm immediately through the matching plan-editing stepFAIL on the final allowed reviewer cycle, team-lead
must still send those findings to arch-ctm for one final correction passcap-exhausted / not converged and report status plainlyteam-lead only checks the top-level status and expected mode fields on
each fenced JSON response before advancingreviewed_commit changed or
findings_hash changedreviewed_commit and the same
findings_hash again, treat it as a stale replay and do not open a new
hardening roundarch-ctm and report cap-exhausted / not converged; do not continue launching background reviewers and do not ask
the user for a decision mid-loop.claude/skills/plan-hardening/01-plan-scope-review.xml.j2.claude/skills/plan-hardening/02-sprint-scope-hardening.xml.j2.claude/skills/plan-hardening/03-consistency-hardening.xml.j2.claude/skills/plan-hardening/steps/step-1.md.claude/skills/plan-hardening/steps/step-2.md.claude/skills/plan-hardening/steps/step-3.md.claude/skills/plan-hardening/steps/step-4.md.claude/skills/plan-hardening/steps/step-5.md.claude/skills/plan-hardening/steps/step-6.md.claude/skills/plan-hardening/examples/plan-hardening-vars.example.json.claude/skills/plan-hardening/examples/plan-hardening-rounds.example.md.claude/skills/plan-hardening/examples/plan-hardening-qa-vars.example.json.claude/skills/plan-hardening/sprint-planning-guidelines.md