菜单
Audit the agentic configuration for consistency, coherence, and conciseness. Load when modifying agent definitions, skills, or pipeline structure, or to verify cross-tool parity.
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。
TDD cycle process and design-check decision tree for feature implementation. Load when implementing features using test-driven development.
Install or update the claude-pod tooling — the claude-pod command, the pod Dockerfile, the default config, the IDE-oracle preflight, and the egress filter pair — from this repo's tools/claude-pod/ into ~/.local/bin and ~/.config/claude-pod. Thin front-end for tools/claude-pod/install.sh: run its check mode to show drift, apply only on the user's approval. Use when the user asks to install claude-pod, set up the container pod, or update the pod script or image definition. Examples: "install claude-pod", "update the pod tooling", "set up the claude container".
Install or update the harness-stats tooling — harness-statusline.sh, cache-report.sh, and the cache-report skill — from this repo's tools/harness-stats/ into the user's ~/.claude/ directory. Thin front-end for tools/harness-stats/install.sh: run its check mode to show drift, apply only on the user's approval. Use when the user asks to install the cache tooling, set up the statusline, update the harness-stats scripts, or reconfigure cache reporting. Examples: "set up the statusline", "install harness-stats", "update the cache tooling".
Generate a per-agent cache-efficiency report for the current Claude Code session. Use when the user asks about cache stats, cache efficiency, prompt-cache performance, token savings, or whether subagents are using cache effectively. Examples: "show cache report", "how's the cache doing", "cache stats", "are subagents efficient", "which agents are wasting cache", "session token breakdown".
Periodic multi-angle improvement review of the reference. Always fans out all five read-only research agents in parallel — maintainer tooling, documentation, runtime cost, source duplication, consumer surface — so cross-angle tension informs the synthesis. Every agent carries the challenge charter: question the design against the README's stated philosophy, recommend removals that pay for themselves. Judges findings by the resilience-first doctrine; a settled ADR decision is rebuttable with a new argument, never silently suppressed. "/review-harness challenge" drops the ADR anchoring from the agent prompts for a zero-based outside look. Adversarially verifies structural findings and delivers one prioritized report. Assessment only — never edits. Complements /audit-harness, which gates defects against the current bar. Root-only (Claude Code).
Build, test, format, and lint requirements that must pass before code review. Load when checking implementation completeness or running the quality gate.
基于 SOC 职业分类
| name | audit-agents |
| description | Audit the agentic configuration for consistency, coherence, and conciseness. Load when modifying agent definitions, skills, or pipeline structure, or to verify cross-tool parity. |
| compatibility | ["claude-code","github-copilot","opencode","junie-cli"] |
| metadata | {"version":"2.0","author":"team"} |
This skill audits a committed harness runtime and applies to the copy channel only. On the marketplace channel the runtime ships inside the plugin, so install hygiene is the plugin repository's CI concern, not the consumer project's. Doc-form validation is owned elsewhere: the doctor skill (blocking, deterministic), audit-docs (advisory judgment), and document-writing (review-time form checks).
Run this audit after any change to:
.claude/agents/, .github/agents/, .opencode/agents/, .junie/agents/).claude/skills/).claude/templates/).claude/agents/README.md.claude/skills/ — no missing, no extras..claude/agents/README.md skills table matches CLAUDE.md.Load the X skill) resolves to an existing .claude/skills/X/SKILL.md.An agent file is the agent's job description; a skill is reusable mechanics. They overlap by design. The audit must distinguish real duplication (drift hazard) from parallel description (different audience, different level of detail) before flagging anything.
Positive shape of a thin agent file. A well-shaped agent body contains, in roughly this order: persona statement, skill/doc references, write scope (which files the agent may modify), a short process overview (≤7 lines naming what the agent does, not how the skill works), and any agent-specific conduct rules. Anything beyond this is a candidate for the drift test.
The drift test. For each candidate finding, run this test before tagging it a violation:
If the skill changes, must the agent file also change to keep the system correct? If yes, it's duplication. If no, it's parallel description — leave it.
Concretely, this means:
Mandatory false-positive examples — these have been wrongly flagged as duplication; do NOT flag them:
pipeline-coordinator.md Step 5 enumerating which schema gates which transition (product-requirements-expert→system-design-expert, system-design-expert→implementer, etc.). The list is the coordinator's surface area, not the skill's content.system-design-expert.md Responsibilities listing architectural validation, security/reliability, understandability, defense in depth, integration analysis. These are the system-design-expert's judgement criteria; design-validation § Design Principles details how to apply them.*-review skill carries a parallel section. The agent overview signals what the reviewer does first; the skill section is the full mechanics.Real-duplication checklist:
Grep patterns — candidates only, run the drift test before flagging:
| Pattern | Where to Search | Likely violation when… |
|---|---|---|
- \[ \] (3+ rows) | Agent body | The checkbox list mirrors a skill's checklist body. Single-row reminders are fine. |
\*\*Red\*\*.*failing test | Agent body | Restates TDD cycle mechanics from tdd-workflow. |
## Review Focus with criteria | Agent body | Restates a review skill's checklist. |
## PRD Boundary with rules | Agent body | Restates prd-authoring validation rules. |
## Output Format with full template | Agent body | Restates a template that exists in .claude/templates/ or a skill's output contract. |
| Numbered process 5+ steps mirroring a skill | Agent body | The agent's process is the skill's process verbatim. |
Patterns that look like duplication but routinely pass the drift test (do not flag without confirming):
## Responsibilities or ## Process list naming what the agent does.<skill> § X" followed by a 1-2 sentence summary of why the agent invokes that section.For each agent, compare all four tool versions (.claude/, .github/, .opencode/, .junie/):
Same persona text (first paragraph after frontmatter).
Same skill references (identical skill names in body).
Same document references (same files and sections).
Same write scope (if defined in any version, must be in all).
Same review process steps (same numbered list).
Client-specific tools are expected, not parity gaps: an MCP server or skill wired to only one runtime (declared in that client's tools: or skill set) is correct by design. Its absence from .opencode/, .github/, or .junie/ is not a finding.
Correct model mapping. Each tier maps across tools as follows; flag only deviations from this table:
| Tier | Claude Code | GitHub Copilot | OpenCode | Junie |
|---|---|---|---|---|
| Sonnet | claude-sonnet-4-6 | Claude Sonnet 4.6 (copilot) | openrouter/anthropic/claude-sonnet-4.6 | sonnet |
| Opus | claude-opus-4-8 | Claude Opus 4.7 (copilot) | openrouter/anthropic/claude-opus-4.8 | opus |
The Opus tier is asymmetric across tools: Anthropic's current latest is 4.8, available on Claude Code and OpenRouter today; GitHub Copilot's catalog tops out at 4.7. Junie still uses the alias form (opus/sonnet) because its docs do not document a pinned-ID format. Run deps-upgrade to bump pins when upstream catalogs advance.
Tool permissions match intent (reviewers need write for output file).
The rule is uniform: every path-shaped string in agent and skill files must resolve to an existing file or directory. Path-shaped means a token containing / and ending in a known extension (.md, .yaml, .yml, .json, .jsonl, .sh, or a source-file extension) or referring to a known directory (docs/, .claude/, .github/, .opencode/, .scratch/, schemas/, or a source root declared in scripts/layout.toml).
.claude/agents/, .claude/skills/, .claude/templates/, .github/agents/, .opencode/agents/, .junie/agents/, CLAUDE.md, and docs/ resolves to a real file or directory. The check includes — but is not limited to — docs/X.md, docs/X.md#anchor, .claude/templates/X.md, .scratch/*, source files, schemas/scratch/X.schema.json.docs/X.md#anchor reference points to an existing heading or <a id="..."> anchor..claude/skills/audit-agents/SKILL.md). Stale references in the audit skill itself propagate into every audit run.Use grep to find candidates. The alternation below covers the universal doc and config extensions; append your stack's source-file extension(s) — a Go project adds go, a Java project adds java — so source references are caught too:
grep -rohE '[A-Za-z0-9_./-]+\.(md|ya?ml|json|jsonl|sh)' \
.claude/ .github/ .opencode/ .junie/ CLAUDE.md docs/ | sort -u
Then check each against the filesystem. Same for directory references.
Verify the floor author values match across all locations:
author value)review-workflow skill reviewer table.claude/agents/README.md agent tableThe roster is the floor plus any extra_reviewers declared in scripts/layout.toml [harness]. schemas/scratch/review-feedback.schema.json validates author by the *-reviewer shape pattern, not a closed enum — the roster's authority is the floor (harness-owned) plus the declared extras (doctor-enforced), not the schema.
Expected floor author values for review-feedback records:
code-quality-reviewertest-reviewersecurity-reviewerdoc-reviewerEach reviewer appends one review-feedback record per dispatch to .scratch/handoff.jsonl. There is no per-reviewer markdown file.
design-validation, code-quality-review, …) copies a tactical rule the brief owns as adaptable (Pattern Catalog realizations, persistence/mapping/ACL strategy, naming/suffix conventions). Enforcers carry only the closed-kernel checks, may state the stack's realization of them, and defer the rest to the brief with one "as written" line. Bounded-module structure and the inward-dependency check are kernel, not findings. Language idioms the brief does not own — line length, package-name hygiene, framework wiring — stay with the reviewer. A brief-owned adaptable rule copied verbatim into an enforcer is a finding. The closed kernel is:
Verify state file references match across:
handoff-routing skill state files table.claude/agents/README.md scratch directory structure.claude/templates/ directory (markdown helpers only)schemas/scratch/*.json (record schemas)Expected state files:
.scratch/handoff.jsonl (append-only; record types: prd-entry, design-block, consultation-request, consultation-response, dispatch-start, build-failure, build-pass, review-feedback, review-plan, design-doc-autofix, prd-autofix, grader-features, grader-verdict).scratch/implementation-plan.md (feature-implementer self-tracking).scratch/escalations.md (feature-implementer on escalate-tag findings; root on the router's blocked decision for prerequisite-missing aborts, reviewer stalls, and escalate findings on an approved verdict; never the coordinator itself)The change-grader writes no separate state files (both records live in .scratch/handoff.jsonl).
Expected schema files (one per record type):
schemas/scratch/prd-entry.schema.jsonschemas/scratch/design-block.schema.jsonschemas/scratch/consultation-request.schema.jsonschemas/scratch/consultation-response.schema.jsonschemas/scratch/dispatch-start.schema.jsonschemas/scratch/review-feedback.schema.jsonschemas/scratch/review-plan.schema.jsonschemas/scratch/build-failure.schema.jsonschemas/scratch/build-pass.schema.jsonschemas/scratch/design-doc-autofix.schema.jsonschemas/scratch/prd-autofix.schema.jsonschemas/scratch/grader-features.schema.jsonschemas/scratch/grader-verdict.schema.jsonExpected design-block.verdict enum: covered, minor, new, refactor-first, foundational, conflicting. Flag any occurrence of the old enum values (approved, needs_changes, blocked, revised, escalated) in this project's docs, skills, agents, or schemas — they are stale and must not leak.
Expected review-feedback.verdict enum (distinct from design-block): approved, changes_requested, blocked. Do not confuse the two enums when auditing. Other project domains may reuse some of these tokens (e.g. as work-unit outcome values in their PRD or system-design) — those are unrelated to the design-block verdict.
Verify the quality gate matches across all locations:
.claude/skills/code-quality-gate/SKILL.md required checks table matches CLAUDE.md.build-pass record for the rest of the gate; they do not re-run build/test..claude/settings.json registers the hooks; format-command permissions belong in the uncommitted .claude/settings.local.json or the tool's allowlist, never a committed file.Verify the handoff-routing skill contains:
Verify agents do NOT contain:
For each reviewer agent in all four tool directories — the four-reviewer floor (code-quality, test, security, doc) plus any extra_reviewers declared in scripts/layout.toml [harness]:
/tmp prohibition: "Never use system /tmp; use .scratch/tmp/"..scratch/ as the only write surface and the single review-feedback record as the deliverable. Every path in a skill's reads: frontmatter exists and appears in the doctor manifest (scripts/doctor-expectations.toml) roster — skills may only bind to documents the harness-project API guarantees.
doc-reviewer agent references document-writing skill (for validation categories and review process).
doc-reviewer agent references prd-authoring skill (for PRD boundary enforcement).
doc-reviewer agent references review-workflow skill (for output format).
feature-implementer agent references tdd-workflow skill.
feature-implementer agent references code-quality-gate skill.
feature-implementer agent references review-workflow skill.
pipeline-coordinator agent references handoff-routing skill.
Every record-writing agent — all project-defined agents except pipeline-coordinator — references the handoff-append skill.
change-grader agent references change-grading skill.
system-design-expert agent references design-validation skill (for triage modes, verdicts, and consultation handling).
The consultation roundtrip is the mechanism by which an in-flight specialist (typically feature-implementer) gets a focused answer from another specialist (typically system-design-expert) or from the human without advancing the pipeline. Verify the semantics are consistently described:
handoff-routing skill: documents Gate 2b for consultation records; states that after a consultation-response the router returns control back to the requesting specialist, not forward to the next pipeline stage.pipeline-coordinator agent: defers consultation transitions to route (which executes both hops) and never re-decides them; any consultation prose it keeps follows the back-route semantics above.tdd-workflow skill: the design-check decision tree directs the implementer to append a consultation-request rather than block waiting; the inner loop resumes when the matching consultation-response arrives.design-validation skill: describes both triage mode (returns one of the six design-block verdicts) and consultation mode (returns a consultation-response); the agent reads the input record type and acts accordingly.system-design-expert agent: write scope includes appending consultation-response records to .scratch/handoff.jsonl; docs/ubiquitous-language.md is in scope only during the foundational triage path.feature-implementer agent: write scope includes appending consultation-request records; agent does not modify docs/ directly.handoff-routing skill and route-spec.md Gate 2b: a consultation-request with target: "human" yields blocked (rule human-consultation); root runs the conversation and appends the consultation-response with author: "human"; consultation-return resumes the requester.The system-design-expert operates in two demand-driven modes; verify each is documented consistently:
system-design-expert agent (all four tool versions) names triage + consultation as the two modes and lists the six verdicts.design-validation skill enumerates the six verdicts with content guidance per verdict..claude/skills/handoff-routing/agentic-harness.md § The system-design-expert role in depth lists the same six verdicts.design-block.schema.json enum exactly matches the six verdict names.foundational path covers both greenfield projects and adoption (extracting candidate vocabulary from existing docs and source); same description across the system-design-expert agent, design-validation, and agentic-harness.md.Per agentic-harness.md § Principles Over Rigid Rules, every instruction in an agent or skill is a hard contract or a judgment instruction, written differently. This check keeps the split from decaying into a flat rule list. As with the drift test, flag only a clear miss, not every terse line.
Hard contract — schema field, routing rule, write scope, dispatch step, record shape. Stays a bare imperative.
Judgment instruction — a classification, sizing test, verdict, or escalate-or-proceed call where no enumeration is complete. Carries one compact rationale clause: the why an agent generalizes from on an unlisted case.
Each canonical judgment surface states its why, not only its what. The surfaces: the six triage verdicts (design-validation), the design-check decision tree (tdd-workflow), the review-feedback tags (review-workflow), slice-sizing (prd-authoring), severity classification (security-review), and the consult-vs-escalate call (handoff-routing).
No hard contract is padded with rationale prose — a schema field, routing row, or write-scope line stays bare; the why belongs in an ADR.
Each agent persona states the spirit of the role (what it protects, the judgment it owns), not a restatement of its mechanical steps.
A newly added judgment surface ships with its clause; a newly added contract does not grow prose.
Truncation recovery fires on a deterministic signal read from .scratch/handoff.jsonl alone — a dispatch-start with no subsequent substantive record from the same (req_id, author). An earlier design gated recovery on an out-of-band signal from root; that trigger is superseded. Verify every description of the mechanism agrees:
handoff-routing skill § Dispatch Truncation Detection states the deterministic, state-only rule; its route-spec.md marks the old root-signal trigger as superseded.route executes the implementer's recovery rows; the pipeline-coordinator (all four tool versions) fires recovery for escalated states. The test is behavioral, not lexical. Flag any router or coordinator prose that makes recovery wait on, depend on, or defer to anything outside .scratch/handoff.jsonl — a root or parent signal, external confirmation, human notification. Also flag prose that calls the state-only signal insufficient, ambiguous, or unreliable. If recovery could stall while the truncation signal already sits in state, it is a finding regardless of wording..claude/skills/handoff-routing/agentic-harness.md § Dispatch-Event Contract and Recovery Paths describes the same deterministic, filesystem-only detection.SUBSTANTIVE constant in scripts/handoff.py (the executable source), the handoff-routing skill (SKILL.md and route-spec.md), and .claude/skills/handoff-routing/agentic-harness.md. The coordinator must reference the term, not restate the enum.The check is on the detection mechanism, not a single stale phrase: flag any file that describes truncation as undetectable from state or dependent on an out-of-band trigger.
Report each item as:
[OK] — checked and correct[ISSUE] file:line — description and fix[DUPLICATION] file:line — what is duplicated and where[TAXONOMY] file:line — judgment surface missing its rationale clause, or a hard contract padded with proseWhen applying fixes for the issues this audit surfaces, three anti-patterns recur and have to be resisted explicitly — otherwise the fix re-creates the same class of problem the audit caught.
Prefer pattern phrasing over instance enumeration. When a finding cites a stale list of paths, the replacement should name the pattern, not enumerate current instances:
config/foo/defaults, config/bar/defaults — silently rots when a third config module is added.the relevant defaults file in config/<config-domain>/ (currently foo, bar) — survives module additions.The rule: if the list might grow, name the shape and parenthesize the current members.
Spread-check every stale reference. A stale string almost never appears in one file. Before declaring a fix done, grep the entire harness corpus (.claude/, .github/, .opencode/, .junie/, docs/, CLAUDE.md) for the original token and fix every occurrence in one pass. Piecemeal fixing across audit runs is how zombie references accumulate.
Redundancy check on new content. When a finding's fix adds a checklist item, process step, or section, verify no sibling item already covers it. If step 1 says "read every Go file under internal/", a step 2 enumerating four files under internal/ is dead weight — fold or delete.