AGENT SKILL

audit

Coordinate Git/GitHub hygiene across a cross-repository fleet: discover canonical repositories, collect and roll up cross-repository evidence (including merged remote-tracking heads still on origin), and hand an action plan to repo-hygiene/source-control, which own per-repository cleanup. The current collector is read-only and emits detailed exact handoffs; it never deletes, prunes, repairs, fetches, checks out, or rewrites. Use when: 'audit repositories across a fleet', 'stale branches across repos', 'orphaned worktrees across repos', 'merged remote branches still on origin', 'moved or renamed GitHub repos'.

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

01 前言 Front Matter

metadata
{ "cadence": "weekly", "summary": "Discover a repository fleet and coordinate read-only evidence handoffs", "workflow-stage": "operator" }
description
Coordinate Git/GitHub hygiene across a cross-repository fleet: discover canonical repositories, collect and roll up cross-repository evidence (including merged remote-tracking heads still on origin), and hand an action plan to repo-hygiene/source-control, which own per-repository cleanup. The current collector is read-only and emits detailed exact handoffs; it never deletes, prunes, repairs, fetches, checks out, or rewrites. Use when: 'audit repositories across a fleet', 'stale branches across repos', 'orphaned worktrees across repos', 'merged remote branches still on origin', 'moved or renamed GitHub repos'.
allowed-tools
[ "Bash(${CLAUDE_SKILL_DIR}/scripts/audit-fleet.sh:*)" ]
argument-hint
[[<dir>]... [--root <dir>]... [--repo <dir>]... [--config <file>] [--detail]|--apply-plan <path>]
user-invocable
true
disable-model-invocation
false

02 资源文件 6

03 SKILL.md 正文

Arguments. [[<dir>]... [--root <dir>]... [--repo <dir>]... [--config <file>] [--detail]|--apply-plan <path>]. Full form: [[<dir>]... [--root <dir>]... [--repo <dir>]... [--config <file>] [--canonical <github.com/owner/repo=path>]... [--skip <name>]... [--extend-skip <name>]... [--max-depth <1..12>] [--detail] [--plan-file <path>]|--apply-plan <path>]

Purpose

Coordinate cross-repository hygiene. This skill owns bounded fleet discovery, canonical checkout resolution, fleet-scale evidence collection, rollup, and action-plan routing. It does not own per-repository cleanup decisions or execution; those belong to repo-hygiene and source-control.

The collector produces the detailed read-only report described below, including the compact machine-readable rollup and action-plan artifact. Tell the user to run /repo-fleet-hygiene:apply --plan-file <path> (dry-run by default; --apply plus confirmation or --yes to mutate; --remote-branches additionally opts in to per-branch remote deletion, which --yes never answers). Do not add an execute flag to this audit script.

Non-negotiable boundary

Never run or suggest running inline from this skill: git fetch, git worktree prune, git worktree repair, git worktree remove, git branch -d/-D, git remote set-url, or any filesystem deletion. The bundled script has no mutation mode. --apply-plan is a read-only dry-run approval artifact over a prior plan file. A report may name a command/tool as a future handoff; the fleet action plan lists those invocations once per repository behind one confirmation gate. Actual fleet mutation belongs to /repo-fleet-hygiene:apply, not this skill.

Input resolution

Parse $ARGUMENTS as opaque arguments for the bundled script. Supported flags:

  • <dir>: a bare positional path, treated as --root. A drive root (D:, D:/) is a legitimate discovery root and normalizes to D:/. This is the form /repo-fleet-hygiene:audit D: uses.
  • --root <dir>: bounded recursive repository discovery (repeatable).
  • --repo <dir>: exact repository/worktree target (repeatable).
  • --config <file>: explicit Git-format config (at most one).
  • --canonical <github.com/owner/repo=path>: invocation-specific canonical checkout override (repeatable; explicit wins over config).
  • --skip <name>: discovery directory-name skip (repeatable). Explicit --skip / fleet.skip entries replace the default skip list rather than appending. Otherwise shrinking is impossible. Default (neither CLI nor config): vendor plus the package-manager cache trees node_modules, .venv, .pnpm-store, .yarn, .npm, .cargo, .rustup, .gradle, .m2, .nuget, __pycache__, and .tox. To shrink (e.g. reach a repo under vendor/), list only the names you still want skipped. CLI and config compose additively with each other like other scope inputs. Values must be bare directory names (no empty value, no path separator). ., .., and .git stay skipped unconditionally even when an explicit list omits them.
  • --extend-skip <name>: discovery directory-name skip that adds to whichever list is in effect (the defaults, or an explicit --skip / fleet.skip list) instead of replacing it (repeatable). Config equivalent: repeatable fleet.skipAppend. Same bare-name validation.
  • --max-depth <1..12>: discovery bound; explicit wins over config/default 5.
  • --project-dir <dir>: the session's project directory, used for the project-scoped config rung. It is not a scope fallback; see the no-scope ladder below.
  • --detail: emit collapsed per-target evidence after the rollup (default is rollup + action plan only).
  • --plan-file <path>: write the machine-readable action-plan JSON to this path (otherwise a temp file is created and named in the report).
  • --apply-plan <path>: standalone read-only mode. Render the ordered dry-run approval artifact for a previously written plan (cannot combine with discovery flags).

Always pass --project-dir "${CLAUDE_PROJECT_DIR}" on audit runs (not on --apply-plan). The script does not read the project directory from its environment, so passing it in is what makes the project rung below reachable at all.

If no explicit scope and no config-supplied fleet.root/fleet.repo resolve, the run uses the shared no-scope ladder: --named paths, then ghq root --all when ghq is installed, then the nearest of the 4 parents above the working directory's checkout (above the working directory itself outside a checkout) that directly holds 2 or more Git repositories, so a run inside one checkout covers its sibling fleet, then the checkout holding the working directory, else exit 3 naming every rung. A Git checkout in the working directory is a rung; the session project directory is not one on its own. Claim: ghq root --all prints every configured root. Basis: the ghq README (ghq root [--all]; "Without '--all' option, the primary one is shown") and ghq help root on ghq 1.10.1 ("--all Show all roots"). As of: 2026-11-10. Recheck: the README's root entry drops or renames --all. Pass that guidance through rather than re-deriving a root yourself. Config resolution is the script's own ladder. Do not pre-resolve or pass a probed path yourself: explicit --config wins, else the script probes <project-dir>/.claude/repo-fleet-hygiene.conf (project-scoped), else ~/.claude/repo-fleet-hygiene.conf (user-global, a machine-scoped fleet config placed there is recorded user intent, not a guessed root). The report header names the consumed config and its source, or states that none was consumed. Never guess a broader machine root from the current path beyond that ladder.

Config-supplied scope is additive to CLI-supplied scope: a --repo X run still walks every configured root. The header's Scope: line names each contributing rung and its entry count, so report that line rather than assuming the arguments were the whole scope.

Before execution, reject any arguments outside this grammar, noting that a bare positional path is in the grammar, so /repo-fleet-hygiene:audit D: and /repo-fleet-hygiene:audit /path/to/tree are valid invocations to pass through, not arguments to refuse. What stays rejected is an unrecognized flag: anything beginning with - that is not listed above. Pass every path/override as a quoted argument; never assemble a shell fragment from config, repository, remote, or branch text.

Run exactly once:

${CLAUDE_SKILL_DIR}/scripts/audit-fleet.sh --project-dir "${CLAUDE_PROJECT_DIR}" <validated-and-quoted-arguments>

The script validates config with git config --file; it never sources or executes it.

Evidence rules

The bundled collector is authoritative for classifications. Preserve its evidence in the report:

  1. Canonical checkout: explicit remote-keyed override → configured remote-keyed override → the repository's main worktree, read from git worktree list --porcelain, never from the checkout discovery reached first. Which record is the main worktree: git-worktree's list entry and List output format. When a supplied or discovered path resolves to a different main worktree, the header states the substitution on one Resolved to main worktree: line per repository, naming every path that resolved into it. Relay it, because the operator named one path and the report is about another. The report always shows discovered and canonical paths. An override target with a missing/non-GitHub remote, or a different identity that cannot be proven to resolve to the same GitHub repository, stops that repository's local audit before evidence combines.

  2. GitHub identity: read the selected fetch remote with git remote get-url; accept only github.com/owner/repo; query GET /repos/{owner}/{repo}. If returned full_name differs, report HIGH transfer/rename evidence and continue branch/worktree analysis against that resolved identity. A moved remote is not a reason to skip local classification or merge evidence. A 404/403/network error is UNKNOWN, never "deleted" or "moved". A 404/403 on an identity listed in fleet.ackUnavailable is demoted to ACKNOWLEDGED, still reported, never suppressed; acks never touch non-404/403 failures or successful-response evidence.

  3. Merged branch: one aliased gh api graphql query per repository page of local branches (up to MERGED_PR_GRAPHQL_ALIAS_PAGE headRefName aliases per call, first:1, states:[MERGED]). GraphQL's headRefName argument is an exact match, never the search API's prefix-matching head: qualifier, so feature/auth and feature/auth-v2 never conflate. Measured rate cost stays 1 per call (nodeCount equals the alias count), well under GitHub's limits: Primary rate limit and Node limit, as of 2026-11-03; recheck when either section changes its limit. There is no merged-PR result window and no per-branch fallback: every non-default local branch the operator asked about is queried by exact name, including heads GitHub auto-deleted and a later fetch pruned. Fail closed when gh/GraphQL is unavailable. Emit github-pr-evidence-unavailable and never infer unmerged from a missing row after a failed page. Identical branch names in another repository are unrelated. HIGH requires the PR headRefOid to equal the current local tip. Tip drift is MEDIUM manual review. Git ancestry without GitHub evidence is LOW and never called merged-by-PR, and under squash merges that ancestry predicate is near-inert, so on a squash-merging fleet GitHub evidence is effectively the only merge evidence.

  4. Merged remote branch: after local classification, the same merged-PR rows are matched against each remote-tracking tip under the selected remote. When headRefOid equals that tip and the branch is not the default, probe live existence with git ls-remote --heads <remote> refs/heads/<branch>. A matching tip → HIGH merged-remote-branch (remote head still present after merge, unset or blocked delete_branch_on_merge). ls-remote failure → MEDIUM cached observation (may be stale after a prune-less fetch). Empty ls-remote → no finding (head already gone upstream). Unlike every other probe, ls-remote keeps global and system Git config, where the transport lives (core.sshCommand, credential.helper, proxies). It runs only when the remote URL names the same github.com repository with and without that config; otherwise it counts as a failure. When every attempted probe in the run fails (non-zero status; empty success does not count), emit one fleet-level UNKNOWN ls-remote-fleet-unavailable finding and withhold the per-repository MEDIUM rows, because no live probe succeeded. Remote-only heads (local already deleted) are included. The handoff is an optional git push --delete --dry-run preview naming the remote and branch; this skill never runs it and never calls org-admin APIs to flip repository settings. Enabling delete_branch_on_merge is complementary (it stops the class accruing) and is not a substitute for this fleet visibility. When the merged PR's headRefOid differs from the remote-tracking tip, the tip is still merged content if it is an ancestor of headRefOid and that commit is in the clone (git merge-base --is-ancestor <tip> <headRefOid> exits 0). That case is MEDIUM merged-remote-branch with the reason tip is an ancestor of the merged head. It runs no ls-remote probe, so it is never HIGH. A head commit absent from the clone, or a tip that is not an ancestor, emits nothing. Remote branch families: each remote-tracking branch under the selected remote whose name belongs to a family below, and that no merged-remote-branch finding covers, gets one LOW remote-branch-family row: the family, the tip, its age in days from the committer date, and whether the tip is an ancestor of <remote>/<default> (yes, no, or unknown). The row reads the last-fetched refs/remotes/<remote>/ inventory and adds no remote probe. It is report-only: no handoff, no deletion, no deletion preview. Whether and when such branches are deleted is the owner's decision.

    Family Branch name Source
    agent agent- followed by hex digits only Claude Code subagent worktrees
    claude claude/* Claude Code cloud sessions (on the web, routine runs)
    plan plan/* named by hand
    stranded stranded/* named by hand
    pre-wipe pre-wipe/* ad hoc safety pushes made before a reimage
    Claim Basis As of Recheck
    agent-<hex> branches come from Claude Code subagent worktrees. plugins/source-control/scripts/worktree-create.sh uses the harness-supplied name verbatim as the branch. 2026-09-29 The harness stops supplying agent-<hex> names, or that script changes how it names the branch.
    claude/* branches come from Claude Code cloud sessions. Observed on fleet remotes (no probe recorded). Pointer: the routine-run branch prefix is in https://code.claude.com/docs/en/routines#repositories-and-branch-permissions; https://code.claude.com/docs/en/claude-code-on-the-web names none. 2026-10-11 That routines section changes the prefix or is renamed, the web page names a prefix, or a claude/* branch turns up with another origin.

    Unmerged remote branch: a remote-tracking head with no MERGED PR row (not the default, current, or worktree-attached branch) gets one more read-only GraphQL query for pull requests whose headRefName is exactly that branch, in any state, plus totalCount. With a live ls-remote tip: no PR at all and a complete list → HIGH unmerged-remote-branch, class never-pr; only CLOSED PRs with one headRefOid equal to the live tip → HIGH, class closed-unmerged. An OPEN or MERGED PR emits nothing. ls-remote failing → MEDIUM unmerged-remote-branch-unverified (no tip, no deletion candidate). Only CLOSED PRs and none at the live tip (commits pushed after close), or a totalCount above the returned page → MEDIUM unmerged-remote-branch-review. Only the two HIGH classes reach the action plan, as delete-remote-branches rows carrying the live tip as expected_oid, the remote's remote_key, and the github_repo that apply rechecks. A branch whose only PR is merged stays merged-remote-branch and is never a deletion candidate here.

  5. Local inventories: parse only git worktree list --porcelain -z registrations and NUL-delimited git for-each-ref branch/tip records. Directory naming is never worktree evidence. Compare each existing registered path's actual --git-common-dir with the canonical checkout's expected common dir. A mismatch is HIGH evidence of an administrative linkage problem but manual review only. Missing/prunable registrations never trigger pruning. Linked, unlocked registrations with reliable admin emit one MEDIUM worktree-status-handoff per repository naming those paths. Disposability (stranded / unknown / safe) is owned by /source-control:worktree status, and this collector emits no git status-based substitute verdict. Separately, every linked worktree that passes existence and root-verifiability checks is classified against its own repository's configured worktree root (worktreeroot.path read with git -C <canonical>, which honors includeIf, else the project directory's key, then source-control worktree_root): conforming, outside/wrong-layout (expected <root>/<owner>-<repo>-<slug> or <root>/<repo>-<slug> without origin, matching /source-control:worktree create; create-shaped basenames stay conforming after branch rename/detach; comparisons use physical paths so symlink aliases of the configured root do not false-positive), or tool-owned (Codex/Cursor). Missing, prunable, non-root, and root-unverifiable registrations keep their own finding kinds and are excluded from conformance denominators. When no root is configured, placement is reported without asserting a convention. A repository with no worktreeroot.path of its own falls back to the project directory's key, then pluginConfigs; it never borrows another target's root. The header names the first resolvable root. The fleet summary lists each distinct root with its repository count when repositories resolve different roots, and lists each repository whose own root differed, with the root it used. Advice for a repository's worktrees names that repository's root. If pluginConfigs cannot be read because jq is missing, emit worktree-root-pluginconfigs-unreadable rather than pretending the key is unset. Per-repository and fleet rollups always state the classifiable counts. If either inventory command fails or emits malformed/partial output, discard it, emit UNKNOWN, stop local branch/worktree classification, and do not count that repository as successfully audited; an empty/failed inventory never means no branches are attached.

  6. Protection: current/default/worktree-attached branches are never emitted as standalone branch cleanup candidates. A merged worktree is routed to worktree dry-run first. merged-remote-branch is independent of local attachment. It describes the remote ref.

Every emitted finding kind, both confidence axes, and the merge-strategy and gc.worktreePruneExpire dependencies the tiers rest on: reference/confidence-model.md. The official Git/GitHub behaviors this collector relies on: reference/official-sources.md. The read-only enforcement model and its threat assumptions: reference/security-review.md.

Presentation

Default output is screen-scale:

  1. Fleet header (config, scope, discovery counts).
  2. Repository rollup. One row per repository with CLEAN / N candidates / BLOCKED (evidence gap), plus counts by finding kind. Every UNKNOWN kind except the disclosure-only discovery-skip and discovery-symlink-skip makes a repository, and so the fleet, BLOCKED. Fleet-level findings (stale config, duplicate checkouts) get their own row. A fleet verdict summarizes blocked vs candidate vs clean.
  3. Fleet action plan. Recommended skill invocations once per repository (not once per finding), ordered so branch cleanups precede worktree cleanups, behind one confirmation gate.
  4. Path to the machine-readable action-plan JSON (and the --apply-plan dry-run invocation).

Pass --detail when the operator needs evidence: targets are collapsed (one entry per path/branch carrying every applicable finding), never duplicated across confidence groups. Never collapse same-named branches across repositories.

ACKNOWLEDGED is a prominence demotion, not a fifth confidence tier: the evidence stays exactly as weak as the UNKNOWN it came from. A rollup CLEAN verdict means no actionable cleanup-plan candidates (the kinds that produce skill invocations) and no UNKNOWN evidence gap for that repository (discovery-skip and discovery-symlink-skip are disclosure-only and do not count), not "GitHub was unreachable so nothing was wrong." Manual-review HIGH/MEDIUM findings (for example locked-worktree or merged-pr-tip-drift) remain in kind counts but do not inflate N candidates when the action plan correctly lists Actions: none.

When the user wants to act on a fleet report, tell them to run /repo-fleet-hygiene:apply --plan-file <path> (dry-run, then --apply) rather than driving per-repository skills by hand.

Fleet cleanup plan

This section defines how audit relates to the execute verb; it does not add a mutation command to the audit argument grammar.

The cleanup-plan consumer is /repo-fleet-hygiene:apply. It takes only the machine-readable rollup artifact. Never parse this skill's human report into executable operations. The apply verb:

  1. rejects an incomplete, invalid, or non-audit artifact and preserves every repository-qualified target, confidence, evidence gap, and disposition;
  2. owns batched merged-local-branch deletion (with fail-closed OID refresh) and worktree cleanup in plan order rather than widening this audit script;
  3. presents one fleet action plan and obtains one explicit confirmation (or --yes) before any mutation; and
  4. re-derives mutable facts, including relevant branch/worktree OIDs, at execution time. An old artifact is evidence, not authorization.

The rollup, the --apply-plan dry-run, and /repo-fleet-hygiene:apply all ship. Do not invent --cleanup-plan, --execute, or report-and-execute behavior on audit-fleet.sh. Return the report, rollup, and plan path; tell the user to run /repo-fleet-hygiene:apply for execution. A HIGH evidence tier is never itself permission to delete a branch or worktree.

Related fleet contracts that remain separate:

  • merged remote branches carry a distinct safety gate: this skill reports them, and their deletion is not part of /repo-fleet-hygiene:apply. Only never-PR and closed-unmerged remote branches can be deleted, and only through /repo-fleet-hygiene:apply --remote-branches (per-branch prompt, never bypassed by --yes, tip recorded first).

Graceful degradation

  • Git missing or too old: stop before scanning and give the prerequisite error.
  • Invalid config SYNTAX, invalid override, or an invalid CLI-supplied --repo/--root path: report the exact invalid input and stop; never silently fall back.
  • No scope given and no rung of the no-scope ladder resolves (exit 3): stop, and relay the script's remedy block verbatim. The operator did not choose that path, so the rejection alone is not actionable.
  • A config-sourced fleet.repo/fleet.root path that is missing or not a Git working tree degrades per-entry, not per-run: the entry becomes an UNKNOWN stale-config-entry finding and the rest of the fleet is still audited (deleting repositories right after an audit must not abort every subsequent run until the config is edited).
  • A path discovered under --root that is unreadable or not a Git working tree (despite a .git marker) degrades the same way: an UNKNOWN discovery-skip finding, header skip counts, and the rest of the fleet is still audited. The finding is disclosure-only: it does not make the fleet BLOCKED. An explicitly named --repo that is not a working tree still hard-fails.
  • A directory that itself carries a .git marker (directory or file) is treated as a nested repository: discovery add_targets it and returns without descending into its children. A repository buried inside another repository's working tree therefore never appears as its own audit target unless named explicitly via --repo / fleet.repo.
  • A symlinked or junctioned intermediate directory under --root is not followed, but is disclosed as an UNKNOWN discovery-symlink-skip finding and counted on the discovery-skips header line. Like discovery-skip, it is disclosure-only and does not make the fleet BLOCKED. What a Windows directory junction is: Creating directory junctions. Symlinked discovery roots remain a hard refusal (CLI) or stale-config-entry (configured).
  • gh missing/unauthenticated or API/timeout failure: continue Git/worktree checks, report GitHub evidence as UNKNOWN, and make no merged/migration claim. Compatible timeout/gtimeout is preferred; otherwise use the collector's finite TERM-to-KILL Bash watchdog.
  • Non-GitHub or ambiguous remote: continue local checks; GitHub identity/PR evidence is UNKNOWN.
  • Canonical override is missing a GitHub remote or does not resolve to the same normalized GitHub identity as the discovered repo: surface UNKNOWN, stop that repository, and do not merge evidence.
  • Worktree porcelain fails: surface UNKNOWN and stop local branch/worktree classification for that repository so no branch can be mislabeled unattached.
  • Branch enumeration fails or is malformed: discard every partial record, surface UNKNOWN, stop branch classification, and exclude the repository from the successful-audit count.

Integration

Finding Handoff (not executed here)
merged-local-branch Run /repo-hygiene:clean git in the named canonical repository
merged-remote-branch Optional preview only: git push --delete --dry-run <remote> <branch> in the canonical repository (never executed here). Enabling GitHub delete_branch_on_merge is complementary and owned by the repository's settings automation. This audit does not change it
unmerged-remote-branch Gated delete: apply-plan.sh --plan-file <path> --apply --remote-branches in the operator's own terminal, one prompt per branch, deleted only at the recorded tip; /repo-fleet-hygiene:apply --remote-branches from a session without a terminal only previews. Preview by hand: git push --delete --dry-run <remote> <branch> in the canonical repository
unmerged-remote-branch-unverified Confirm git ls-remote --heads works by hand, then rerun. No deletion is planned
unmerged-remote-branch-review Manual review of the branch and its closed PRs; never a deletion candidate
remote-branch-family None. Report only: no deletion and no deletion preview. Read the family, age and on-default fields and decide by hand
merged-worktree, prunable-worktree, missing-worktree Run /source-control:worktree cleanup --dry-run in the canonical repository
worktree-status-handoff Run /source-control:worktree status in the canonical repository (stranded-work axis); use cleanup --dry-run only after Work is safe. If source-control is not installed, name the listed worktree targets and the missing collaborator. Emit no porcelain-based substitute verdict
worktree-admin-mismatch Manual inspection; git worktree repair is an option only after validating which administrative directory is authoritative
worktree-not-a-root Manual inspection of the registered path; a git -C probe of it describes the CONTAINING repository, so no cleanup handoff is safe until the path is resolved
worktree-root-unverifiable Manual inspection of the registered path. Root-ness is unproven here rather than disproven. The probe itself failed, so infer nothing about the path in either direction
worktree-nested-in-repository Recreate at an external root with /source-control:worktree create, then remove the nested one
worktree-outside-configured-root Recreate at the expected location named in evidence with /source-control:worktree create, then remove the misplaced one
worktree-wrong-layout Recreate at the expected <owner>-<repo>-<slug> path under the configured root, then remove the wrong-layout one
worktree-tool-owned Leave to Codex/Cursor lifecycle, or migrate deliberately to the configured root
worktree-root-conformance Read the per-worktree outside/wrong-layout findings for expected paths; migrate toward the configured root
worktree-root-conformance-summary Same as per-repository conformance; fleet-scale migration toward the configured root
worktree-root-unconfigured Set worktreeroot.path (git config) or source-control worktree_root, then rerun
ls-remote-fleet-unavailable Confirm git ls-remote --heads works by hand with the operator's usual Git transport, then rerun
worktree-root-pluginconfigs-unreadable Install jq, or set worktreeroot.path; do not treat the fleet as unconfigured
worktree-placement-unverifiable Inspect the canonical checkout; placement was not checked for any of its worktrees, so their placement is unknown rather than confirmed
bare-repo-with-working-tree Manual review: core.bare=true coincides with working-tree content or registered linked worktrees. What the setting does: the core.bare entry
github-remote-moved Human-reviewed git remote set-url; this plugin never changes remotes

This plugin remains useful if those optional collaborators are absent: the report names the local Git/GitHub evidence and target so another tool or human can act. For worktree-status-handoff, when source-control is missing, name the listed worktree paths and the missing collaborator and do not invent a porcelain-based disposability substitute.

Gotchas

  • Which config is consumed depends on where the audit runs. Config resolution follows the ladder in the Input resolution section: explicit --config, else the project-scoped .claude/repo-fleet-hygiene.conf, else the user-global one. A project-scoped config is invisible when the audit runs from a different project, so confirm the consumed config named in the report header before trusting a run's scope.
  • Config paths resolve relative to the config file's directory. A relative root/repo/canonical path is anchored at the config directory, not the audit's working directory. Absolute paths work but are what a consumer's write-time path guard flags, so author config via /repo-fleet-hygiene:setup apply, which prefers the portable relative form.

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