AGENT SKILL

setup

Verify harness-config's readiness for this repository: the external CLI prerequisites its audit scripts need, jq (the JSON-parsing audit scripts) and curl (the plugin-drift check), and the tracked suppression record in docs/conventions/harness-config.yaml, so the audit skills run instead of failing. Use when: 'set up harness-config', 'configure harness-config', 'is harness-config working', 'set up audit-pass suppressions', or an audit skill reported a missing prerequisite. Actions: check (read-only verification, default) | apply (resolve what check found) | wsl-profile (print an opt-in WSL hardening settings block, writes nothing). Re-runnable and safe.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 30
Forks
7

01 前言 Front Matter

shell
bash
description
Verify harness-config's readiness for this repository: the external CLI prerequisites its audit scripts need, jq (the JSON-parsing audit scripts) and curl (the plugin-drift check), and the tracked suppression record in docs/conventions/harness-config.yaml, so the audit skills run instead of failing. Use when: 'set up harness-config', 'configure harness-config', 'is harness-config working', 'set up audit-pass suppressions', or an audit skill reported a missing prerequisite. Actions: check (read-only verification, default) | apply (resolve what check found) | wsl-profile (print an opt-in WSL hardening settings block, writes nothing). Re-runnable and safe.
allowed-tools
[ "Bash(command -v jq*)", "Bash(command -v curl*)", "Bash(command -v awk*)", "Bash(command -v sort*)", "Bash(bash ${CLAUDE_PLUGIN_ROOT}/skills/setup/scripts/wsl-profile.sh *)" ]
argument-hint
[check|apply|wsl-profile]
user-invocable
true
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Pre-computed context

check's tool probes ran at load time. Read these rows instead of re-issuing them; each shows the tool's path when present, or absent when missing:

  • jq: !command -v jq 2>/dev/null || echo absent
  • curl: !command -v curl 2>/dev/null || echo absent
  • awk: !command -v awk 2>/dev/null || echo absent
  • sort: !command -v sort 2>/dev/null || echo absent

A row reading [shell command execution disabled by policy] carries no result: run that tool's command -v probe via Bash instead.

  • Pointer: for the placeholder a skipped load-time command leaves, fetch skills § Inject dynamic context live.
  • As of: 2026-11-11
  • Recheck trigger: the placeholder text changes.

Purpose

Setup per the uniform setup contract (docs/plugin-philosophy.md "Setup is explicit and repeatable" in the marketplace repository): check inspects and reports, apply resolves. This plugin declares no userConfig, and has two setup concerns:

  • the external command-line tools its bundled scripts require, where apply is guidance-and-verify with no write path: it points at platform install instructions and never installs a system package;
  • the tracked consumer-project configuration audit and audit-pass read, the suppressions mapping of <home>/harness-config.yaml, the one file apply may write. <home> is the convention home the repository binds in the pointer region of its AGENTS.md or CLAUDE.md (bash ${CLAUDE_PLUGIN_ROOT}/lib/resolve-convention-home.sh prints it), else docs/conventions.

Action routing: no argument or check runs the check; apply runs the check first, then remediation; wsl-profile runs only its own section. All are non-interactive, so never prompt when the action is given.

check (read-only)

The bundled scripts are the single source of truth for what this plugin requires.

Read it first. Probe what it actually does, don't recite this file. Then read the pre-computed tool rows, run the remaining probes via Bash, and report a PASS/FAIL/INFO table with one remediation line per FAIL. Do not modify anything.

The runtime scripts and their tools:

  • ${CLAUDE_PLUGIN_ROOT}/skills/audit/scripts/check-plugin-drift.sh: jq and curl
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit/scripts/check-structure.sh, check-hook-coverage.sh and fix-plugin-drift.sh: jq
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit/scripts/audit-engine.sh: jq plus awk and sort; check-effort-pins.sh: jq plus awk
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-automation-gaps/scripts/inventory.sh: jq plus awk
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-permission-grants/scripts/permission-rule-check.sh: jq plus awk and sort
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-permission-state/scripts/*.sh: jq (required; permission-state.sh exits 2 without it), awk and sort; automode-block-lint.sh and automode-entry-diff.sh add python3 for an optional lane
  • ${CLAUDE_PLUGIN_ROOT}/skills/draft-auto-mode-rules/scripts/draft-automode-block.sh: jq
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-pass/scripts/run-state.sh: jq (falls back to python3, then to a scan that announces itself)
  • ${CLAUDE_PLUGIN_ROOT}/lib/check-retirements.sh: awk (no jq)
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-pass/scripts/state-digest.sh and finding-identity.sh: sort
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-instructions/scripts/instruction-scan.sh: grep, tr and awk (no jq)
  • ${CLAUDE_PLUGIN_ROOT}/skills/audit-instructions/scripts/conflict-scan.sh: awk and sort (no jq)

Most scripts that need jq probe for it, but only conflict-scan.sh and instruction-scan.sh probe for awk and sort; the rest call those with no guard, so read each script's actual calls rather than trusting a single script's prerequisite block to speak for the plugin.

  1. jq, from the pre-computed jq row. FAIL if absent: the JSON-parsing scripts need it (inventory.sh degrades to an empty inventory; the others exit 2 with an install remediation). Missing jq blocks every skill whose scripts parse JSON (audit, audit-automation-gaps, audit-permission-grants, audit-permission-state, draft-auto-mode-rules, and audit-pass's run-state helper); audit-instructions scans markdown and is unaffected.
  2. curl, from the pre-computed curl row. FAIL if absent, but scoped: only the plugin-drift check (check-plugin-drift.sh) uses it and exit 2s without it. The rest of audit and every other skill still run; say so in the remediation line.
  3. awk and sort, from the pre-computed awk and sort rows. FAIL if either is absent, and not scoped to one skill: conflict-scan.sh and instruction-scan.sh probe and exit 2 naming the missing one, while the other scripts in the list above that use them have no prerequisite check at all, so their skills fail mid-run on a bare command not found rather than on a named prerequisite. Say in the remediation line that this FAIL reaches six skills (audit, audit-automation-gaps, audit-pass, audit-permission-grants, audit-permission-state and audit-instructions), not just audit-instructions. Report awk and sort by name rather than as one row, since a minimal shell can carry one and not the other.
  4. Bash shell, INFO: the scripts are bash (arrays, [[ ]], process substitution, BASH_SOURCE), run through Claude Code's Bash tool. Report the resolved interpreter; FAIL only if no bash is resolvable. Pointer: for which shell a platform's sessions get, native Windows included, fetch tools reference: PowerShell tool live.
  5. Network reachability, INFO only: audit's drift/freshness fetches read raw.githubusercontent.com, and a failed fetch degrades to SKIP rather than a setup failure. Do not fetch here; check performs no network call.

The audit-pass suppression record

audit-pass reads a tracked suppression record layered per the marketplace's config-cascade convention. Read the operative shape from ${CLAUDE_PLUGIN_ROOT}/skills/audit-pass/reference/suppression.md rather than inferring it. That reference ships inside this plugin, so it resolves in an installed cache where a path out to the marketplace's own docs does not. The cross-consumer key contract is the marketplace's published finding-suppression convention, which is not a runtime dependency of this plugin. All layers absent is a valid state (no suppressions), so report INFO, never FAIL, when none exists.

Anchor at the repo root (${CLAUDE_PROJECT_DIR}, else git rev-parse --show-toplevel), never a CWD-relative read, which resolves a different (or missing) file depending on the subdirectory or nested worktree the skill was invoked from, then report one row per layer. The same tracked/ignored question has opposite correct answers per layer, so verify each on its own terms:

  • user-global ~/.claude/audit-pass.md: outside the worktree; no git command applies. INFO only.
  • team <home>/harness-config.yaml: must be tracked, probed as the pair: git check-ignore -v reports no match (a match is FAIL with the pattern) AND git ls-files --error-unmatch exits 0. Untracked while present is a hard STOP: teammates never receive the shared suppressions. Parse it with awk -v STRICT=1 -f ${CLAUDE_PLUGIN_ROOT}/lib/yaml-subset.awk <file>, the parser the audit engine uses: a non-zero exit is FAIL with the line it prints, because the engine then drops the whole team layer. The common cause is an unquoted anchor/v1 key; the parser takes a plain key only of letters, digits, _ and -, so the key is written "anchor/v1":.
  • personal override .claude/audit-pass.local.md: must be gitignored and never staged. Staged or tracked is a FAIL: a personal deviation can reach team history.

Parse each present layer and report as malformed any entry missing any of its five required keys: check, claim, sites, reason, date. Checking only reason and date would pass an entry that audit-pass itself rejects, so readiness would report green on configuration that cannot suppress anything. Report as malformed too an entry whose stored constituents do not hash to its own key: the constituents are authoritative and the key is derived from them. A malformed entry does not suppress, and a silent partial parse would turn a formatting slip into a lost check.

Report a user-global or personal override entry for an id the team layer does not carry as INFO personal-only, not applied, naming promotion to the team layer as what makes it take effect. The team layer is the only one that enacts a suppression, so a personal-only entry that looks live is a finding the operator believes is accepted and is not.

apply (idempotent)

Run check, then for each FAIL give the platform install instructions from the README Requirements. This skill never installs system packages:

  • missing jq: the platform's jq install (for example winget install jqlang.jq, brew install jq, or apt-get install jq); rerun check after.
  • missing curl: the platform's curl install. Modern Windows and Git Bash already ship curl. Only the plugin-drift check needs it, so the rest of the plugin works meanwhile.
  • missing awk or sort: both ship with every POSIX userland, so absence means the shell environment is minimal rather than that one package is missing. Git Bash and Windows busybox shims are where this shows up. Remediate by installing a full userland rather than the single tool: Git for Windows, which bundles both; the distribution's gawk/mawk and coreutils on Linux; brew install gawk coreutils on macOS. Report the two separately, since a minimal shell can carry one and not the other. The skills check step 3 names depend on them, and only conflict-scan.sh and instruction-scan.sh exit 2 cleanly, so do not offer the others as still working meanwhile.
  • no resolvable bash: also not remediable by one package. The scripts use arrays, [[ ]], process substitution, and BASH_SOURCE, so they need a real bash on PATH: Git for Windows on native Windows, the distribution's bash elsewhere. Nothing bundled runs until it resolves, so say that this FAIL blocks the plugin rather than offering a partial workaround.

The network row stays INFO and has no apply entry on purpose: a failed fetch degrades to SKIP by design, so there is nothing to remediate.

After any install, re-run the relevant probe via Bash (command -v <tool>; the pre-computed rows predate the install) and report its actual result. Never claim resolved on the install command's exit code alone. Re-running apply once every probe passes changes nothing and reports "already configured".

Then converge the one surface this plugin owns, conservatively:

  • No suppression record anywhere: leave it that way and say so. Absent is valid; an empty scaffold is noise. Scaffold <home>/harness-config.yaml with the documented shape only on an explicit request.
  • Team layer present but untracked: report the STOP and the exact git add the operator should run. Never stage on their behalf.
  • Personal override present but not ignored: recommend the recursive .claude/**/*.local.* line and leave the .gitignore edit to the consumer.
  • Malformed or unrecognized entries: report them and stop. Never rewrite, reorder, or drop an operator's suppression: an entry this skill cannot reconcile is a question for the operator, and a silent rewrite would hide the very finding the entry was suppressing.

Every write names the file and the exact change before making it, and preserves unrelated content.

wsl-profile (print only)

An opt-in settings profile for Claude Code inside WSL. Read reference/wsl-profile.md for what it holds and its limits.

  1. Run bash ${CLAUDE_PLUGIN_ROOT}/skills/setup/scripts/wsl-profile.sh detect. Exit 1 means this is not a WSL host: say the profile does not apply and stop.
  2. Run bash ${CLAUDE_PLUGIN_ROOT}/skills/setup/scripts/wsl-profile.sh print. It fills in the running user's paths and the Windows host's /mnt/<drive>/Users/<user> directories. Show the block, relay its stderr notes (several Windows users, or placeholders left to fill in), and tell the user to merge it into their own settings file.
  3. State the limit from the reference: deny rules do not stop a script that opens files itself; the sandbox is the OS-level control.
  4. Name appendWindowsPath=false in wsl.conf and the PowerShell symlink as host steps the human runs.

This action writes no file, edits no settings and never touches wsl.conf or creates the symlink. To check an edited profile, run ... wsl-profile.sh lint <file>.

Gotchas

  • Never run a git command against the user-global layer. ~/.claude/audit-pass.md is outside the worktree, so git check-ignore and git status return a meaningless verdict there, or a confidently wrong one when the home directory is itself a git repository.
  • Missing config is not a failure. All three suppression layers absent means no suppressions, which is the normal state for a repo that has never suppressed a finding. INFO, never FAIL.
  • Recommend the recursive gitignore line, and leave the edit to the consumer. .claude/**/*.local.* covers flat, folder-form, and profiled personal overrides alike; the narrower .claude/*.local.* silently misses any nested personal override. The consumer's ignore file is their artifact; this skill never writes it.
  • An install command's exit code is not verification. After any install, re-run the probe and report its actual result.

What this skill does NOT do

  • Run an audit; that is the plugin's audit and drafting skills (/harness-config:audit, audit-automation-gaps, audit-instructions, audit-pass, audit-permission-grants, audit-permission-state, audit-prompting-postures, draft-auto-mode-rules, unhobble).
  • Write the plugin cache, Claude Code user settings, or pluginConfigs; wsl-profile prints a block and the user merges it.
  • Install system packages.
  • Write the consumer's .gitignore, stage anything, or edit an operator's suppression entries.
  • Download anything. check makes no network call; the audit skills' own doc/marketplace fetches are theirs, not setup's.

04 同仓库更多 Skills

melodic-software
melodic-software/claude-code-plugins

track

Backlog reads through the bound tracker: search open and closed items, view one item with its comments and state history. Use when another skill searches or reads tracker items.

★ 30
melodic-software
melodic-software/claude-code-plugins

reviewer-media

Reviewer media for one pull request: decides whether it needs a narrated walkthrough, builds the storyboard from /review:explain-change's digest, has /explainer-video:produce render it and hands the manifest to /source-control:pull-request media. Use when asked for 'reviewer media', 'a walkthrough video for this PR', 'a produced video of this change', or when pull-request ready or monitor hands a PR over.

★ 30
melodic-software
melodic-software/claude-code-plugins

route

Model and effort per agent role (orchestrator, worker, verifier, retrieval) for a multi-agent run, resolved through the bundled defaults and the user, team and personal override config layers into fenced JSON with single-agent and fan-out variants and the layer behind each value. Use when asked 'which model for the subagents', 'route the roles' or 'model routing', or before a skill passes a role's model on a spawn.

★ 30
melodic-software
melodic-software/claude-code-plugins

route

Model and effort per agent role (orchestrator, worker, verifier, retrieval) for a multi-agent run, resolved through the bundled defaults and the user, team and personal override config layers into fenced JSON with single-agent and fan-out variants and the layer behind each value. Use when asked 'which model for the subagents', 'route the roles' or 'model routing', or before a skill passes a role's model on a spawn.

★ 30
melodic-software
melodic-software/claude-code-plugins

principles

Module-design vocabulary for architecture findings, and the checks built on it. Use when: 'what is a deep module', 'what is a seam', 'deep vs shallow module', 'what is the deletion test', 'leverage and locality', 'module design vocabulary', 'which word do we use for X in architecture findings'. Scanning a codebase for friction is /architecture:improve.

★ 30
melodic-software
melodic-software/claude-code-plugins

write

Structured five-field bug report (title, steps to reproduce, expected vs actual, justified severity, suggested fix location) from an informal description, read-only. Use when the user names a defect they observed, or asks for one written up ('report a bug', 'file a bug', 'write this up as a bug').

★ 30