一键导入
hd-review
Reviews harness health (full) or targets one layer/file (targeted). Writes dated report; emits summary with health bars + priorities in chat.
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
Reviews harness health (full) or targets one layer/file (targeted). Writes dated report; emits summary with health bars + priorities in chat.
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。
Answers questions about the five-layer design harness framework. Use when learning concepts, asking about a layer, or orienting before setup.
Walks the five-layer design harness layer by layer. Detects existing harnesses + tools, offers per-layer scaffold / review / create / skip. Use to set up or revisit any repo.
Captures design lessons and proposes rule adoptions from narrative to rule. Use when capturing a decision or promoting a recurring pattern to AGENTS.md.
基于 SOC 职业分类
| name | hd:review |
| description | Reviews harness health (full) or targets one layer/file (targeted). Writes dated report; emits summary with health bars + priorities in chat. |
| argument-hint | [full | <file-path-or-url> [--rubric <name>] | snapshot] |
Output destinations. Every /hd:review run writes the full report to a dated markdown file and emits a rich summary in chat. Chat summary carries the visualization (ASCII health bars, priorities table, cross-layer signals). Deep findings live in the file.
Host-agnostic by construction. File write works on every host (Claude Code, Codex CLI, Cursor IDE/CLI, Windsurf, any terminal). Chat summary works everywhere the transcript does. No Task dispatch is required — it's an optional speed-up for the evaluation phase on hosts that support parallel sub-agents.
Narrate rationale inline at each major step (preflight, per-layer evaluation, synthesis, write, summary). Users shouldn't have to ask "what's going on?"
Review harness health — either in full (every layer + cross-layer + budgets) or targeted (one layer, one file, or one rubric). One skill, two scopes of the IMPROVE family.
Declares protected paths so any external review/cleanup tool leaves our artifacts alone:
<protected_artifacts>
- docs/design-solutions/**
- docs/knowledge/**
- docs/context/**
- docs/plans/**
- docs/rubrics/**
- AGENTS.md
- hd-config.md
- loading-order.md
- skills/**
- agents/**
</protected_artifacts>
Multiple review tools can run on the same repo without modifying each other's outputs.
/hd:review or /hd:review full → full review (default)/hd:review snapshot → snapshot (bars-only chat output, no file write)/hd:review targeted <path> [--rubric <name>] → targeted review/hd:review <path> (bare path, no verb) → ask "Run a full review across all layers, or a targeted review of <path>?"Default is full review. Ambiguous input asks scope (full vs targeted), never legacy "audit vs critique" vocabulary.
hd:review full Progress:
- [ ] Step 1: Preflight — run budget-check.sh + detect.py; diff vs hd-config.md
- [ ] Step 2: Per-layer evaluation (inline serial baseline; parallel when host supports it)
- [ ] Step 3: Cross-layer consistency check (inline)
- [ ] Step 4: Synthesize findings + cross-check against <protected_artifacts>
- [ ] Step 5: Render full report; atomic write to docs/knowledge/reviews/<date>-harness-review.md
- [ ] Step 6: Emit rich chat summary (bars + priorities + cross-layer signals)
- [ ] Step 7: Suggest next
Per-layer evaluation: for each layer 1–5, load review-criteria-l<N>.md, gather evidence, emit YAML findings. Baseline is inline serial (any host). Parallel dispatch (Claude Task, Codex /agent, Cursor subagents) runs the same evaluation concurrently when available.
→ See references/review-procedure.md for full step detail.
Snapshot mode: /hd:review snapshot — ~30s preflight-only pass. Emits bars-only table + overall score; no file write, no deep layer reads.
hd:review targeted Progress:
- [ ] Step 1: Parse target + optional --rubric
- [ ] Step 2: Resolve rubric path (starter / user-defined)
- [ ] Step 3: Evaluate — SKILL.md target → skill-quality-auditor; otherwise → rubric-applier (inline or parallel)
- [ ] Step 4: Aggregate findings per targeted-review-format.md
- [ ] Step 5: Emit inline summary; write targeted report only if user confirms
→ See references/targeted-review-procedure.md for full step detail.
Path: docs/knowledge/reviews/<date>-harness-review.md
Template: assets/review-report.md.template. Carries per-layer findings with evidence, recommendations, consistency findings, hd-config drift, agent list, meta.
Rich summary with Unicode box-drawing tables. Mandatory sections:
Review complete · Full report: docs/knowledge/reviews/<date>-harness-review.md
═══════════════════════════════════════════════════════════════════
Harness health — <score> / 10 (<state>)
Layer Bar Score State
───────────────── ─────────── ────── ───────────────────────────
L1 Context ████████░░ 8.0 <one-line summary>
L2 Skill Curation ██░░░░░░░░ 2.0 <one-line summary>
L3 Orchestration ██████░░░░ 6.0 <one-line summary>
L4 Evaluation ████░░░░░░ 4.0 <one-line summary>
L5 Knowledge █████░░░░░ 5.0 <one-line summary>
═══════════════════════════════════════════════════════════════════
Top priorities
Sev # Layer Finding Effort
─── ─── ─────── ─────────────────────────────────────────── ──────
P1 1 L2 <one-line> S
P1 2 L1 <one-line> S
P2 3 L4 <one-line> M
═══════════════════════════════════════════════════════════════════
Cross-layer signals
Signal Status Evidence
─────────────── ──────── ─────────────────────────────────────
hd-config.md <status> <evidence>
Consistency <n> <evidence>
═══════════════════════════════════════════════════════════════════
Next · <one-line next-step suggestion>
Rules: ASCII bars always shown. Tables preferred over bullets for structured data. Box-drawing ═ as section dividers, ─ inside tables. No emoji, no color codes.
/hd:setup/hd:learn/hd:maintain capture (suggestion, not invocation)docs/solutions/ (reserved for other tools)docs/knowledge/reviews/<date>-harness-review.md (full review)<protected_artifacts> block declares our outputs as read-only for external review/cleanup toolsWhen context budget is tight: full review collapses to snapshot mode (bars-only, no file write). Targeted review applies fewer rubrics. No plan-hash mechanism (hd-review is read-mostly).
When the host supports sub-agent dispatch (Claude Task, Codex /agent, Cursor subagents API), per-layer evaluation fans out in parallel (≤5 agents per batch). Otherwise runs inline serial. Output file + chat summary are identical regardless of mode.
review-criteria-l5-knowledge.md + review-criteria-budget.md + review-criteria-consistency.md — per-scope health criteriarubric-template.md, all with ## Scope & Grounding sectionsscripts/budget-check.sh — deterministic always-loaded + SKILL.md budget enforcement; emits JSONAll Task calls use fully-qualified names in the harness-designing: namespace. Never dispatches into other plug-ins' namespaces.
Full review — Batch 1 (parallel, 5): harness-designing:analysis:harness-auditor × 5 (one per layer). Batch 2 (parallel, 2–3): harness-designing:analysis:rubric-recommender + harness-designing:research:lesson-retriever + conditional harness-designing:analysis:coexistence-analyzer (when other_tool_harnesses_detected[] non-empty).
Targeted review — harness-designing:review:skill-quality-auditor for SKILL.md targets; harness-designing:review:rubric-applier for all other targets (batch-parallel ≤5). Optional external agents from hd-config.md:review_agents (empty default).