# Audit

> When the bundled update-config skill resolves in this session, prefer it to make a requested settings change; this skill to audit what is configured. Audit settings.json, settings.local.json, .mcp.json, hooks, plugins, permissions and env vars for correctness, security, and drift against current official docs. Use when: 'audit settings', 'check config', 'check for config drift', after a Claude Code update; pass --fix to apply auto-correctable findings with confirmation.

- Skill: `melodic-software/audit-19` (Agent Skill, multi-file: 28 files)
- Install (CLI): `npx get-skill add melodic-software/audit-19`
- Raw SKILL.md: https://api.skillmd.com/api/skills/melodic-software/audit-19/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: melodic-software (https://skillmd.com/u/melodic-software)
- Updated: 2026-11-12
- Page: https://skillmd.com/skills/melodic-software/audit-19

---


**Arguments.** `[--fix] [permissions|mcp|hooks|plugins|issues|effort-pins|all]`. default: all

## Pre-computed context

Claude Code version: !`claude --version 2>/dev/null || echo "unknown"`

## Purpose

Periodic audit of Claude Code configuration files against current official documentation and the
consuming project's own conventions: are our settings correct, secure, complete, and up to date?

## Adapting to your environment (graceful degrade)

Where a phase names an adjacent capability, such as a Claude Code issue-tracking skill or a hook
plugin's coverage manifest, treat it as optional: use it when your setup provides it, otherwise
follow the inline guidance here. Project-specific conventions (required permission patterns beyond
the baseline, documented reasons for disabled servers, launcher-script wrappers) come from the
consuming repo's own `CLAUDE.md` (or the `AGENTS.md` a session reads natively, whether in place of a
`CLAUDE.md` or alongside one) and `.claude/rules/`. Read them when present; never assume them.

Neighboring questions belong elsewhere: the sibling `audit-automation-gaps` skill asks whether the
configured automation SET is the right set (landscape gaps); the `audit` skill in the
`harness-memory` plugin audits the instruction layer (CLAUDE.md / a natively read AGENTS.md / rules
/ auto-memory). This skill asks whether the configuration FILES are correct against upstream truth.

## Arguments

Parse `$ARGUMENTS` for:

- **`--fix`**: Apply corrections automatically (with user confirmation per fix). Without this flag, report-only mode
- **Scope filter**: Limit audit to a single category. If omitted, run all categories
  - `permissions`: deny/ask/allow rules, security gaps
  - `mcp`: MCP server definitions, commands, env vars, connectivity
  - `hooks`: hook scripts exist, timeouts, matchers
  - `plugins`: enabled/disabled status, marketplace availability
  - `issues`: recheck known GitHub issues only (Category J and Phase 3.2)
  - `effort-pins`: run only the effort-pin drift check (Phase 3) and report its lines verbatim; no
    engine run and no fixes
  - `all`: run everything (default)

## Division of labor: the engine decides, the model judges

`scripts/audit-engine.sh` settles every row a script can decide without a reading and emits one
JSON document plus a findings file. The model never re-derives a row the engine settled; it reads
the document and does only what needs judgment:

| Engine (deterministic, emitted once) | Model (judgment, on the engine's output) |
| --- | --- |
| A: `$schema` presence and URL, misplaced `mcpServers`, personal `hooks` in the local file, whether each top-level key and each key inside an object the page documents children of (such as `permissions`, `sandbox`, `worktree` or `statusLine`) is documented or deprecated on the fetched `settings-reference` (the installed binary settles an undocumented one), and the `A/consent-receipt` label for an undocumented top-level key `reference/consent-receipts.json` records (gates in validation-categories.md Category A) | A: nothing |
| B: presence of each baseline pattern, deny rules in the local file, blanket `Bash(git *)`, the allow-completeness rows at `info`, narrowing 3 where a hook plugin ships a coverage manifest, suppression-record matching | B: narrowings 1 and 2 (a documented exemption, a documented hook convention), narrowing 3 for hooks with no manifest, the consuming repo's extra required patterns |
| C: command resolution, `${VAR}` syntax, URL shape, `enableAllProjectMcpServers`, enabled/disabled coverage and name validity | C: documented reasons for disabled servers, launcher-wrapper conventions |
| D: path resolution and readability, event names against the fetched `hooks` page, millisecond-shaped timeouts, matcher class and anchoring, placeholder quoting in shell form, duplicates, lever state, cache-versus-loaded divergence | D: whether a timeout is reasonable for its tool, exec-form resolution on a Windows-targeting repo |
| E: marketplace membership, every `false` key as an inventory row (a finding only when an enabled plugin depends on it), ORPHAN / RENAME / REMOVED drift, catalog plugins with no entry in any scope as one inventory row per marketplace, keys the drift check did not diff, `strict` versus `plugin.json` | E: the fix for a disabled dependency, orphan-`true` review, rename confirmation |
| F: token-shaped values, documentation status against the fetched `env-vars` page | F: whether an undocumented custom variable is justified |
| G: the measurement, read from an existing debug log; `skillOverrides` keys that name a known plugin (inert, `warning`), colon keys whose prefix names no plugin (`skip`), and entries in the user dir's `settings.local.json` (`info`) | G: the levers, scoped to the roster's composition, and what an undecided colon key names |
| H and I: every value check; the accepted `effortLevel` and `disableDeepLinkRegistration` values, the version `enforceAvailableModels` requires and the values of every other string key whose Type bullet lists them come from the fetched `settings-reference`, and the `fallbackModel` cap from the fetched `model-config` "Fallback model chains" section; a `deniedModels` or `availableModelsMatch` key outside managed settings is flagged from a rule kept in our words | H and I: nothing, once the Phase 3 fetch confirms the behavior the row rests on |
| J: each `Fixed in vX.Y.Z` a `reference/known-issues.md` row records, compared with the installed Claude Code version (the form is that file's "Recording a fix version"). K: the age of each suppression entry against `exceptions.max_days` (default 90; the team YAML may set another), a `warning` past it that names the entry, its reason and its age, never a removal | J: the live status of each issue (Phase 3.2), and whether a workaround an `info` row flags is still needed. K: nothing; report each aged entry for the operator to re-date or remove |

A row the engine marks `skip` or `not-inspectable` is exactly that in the report: never clean.

## Config Files

| File | How to read | Notes |
| --- | --- | --- |
| `.claude/settings.json` | the engine, or `jq` | Project-level, checked in |
| `.claude/settings.local.json` | the engine (structure and the four model and effort keys only) | Treated as secret-bearing; a project may deny-list it for the Read tool, though the baseline does not require that. **Never echo secret values** |
| `.mcp.json` | the engine, or `jq` | Project-level MCP server definitions |
| `~/.claude/settings.json` | the engine, or the Read tool | User-level defaults (optional; checked when it exists) |
| start-directory `.claude/settings.local.json` | `check-structure.sh` (structure only) | Only when the session's start directory is not the repository root AND a copy is there. We treat it as live beside the repository-root copy; how the two combine, and the version boundary behind the copy, are the dated record in `audit-permission-state/reference/criteria.md` §Scopes |
| `managed-settings.json` + `managed-settings.d/` | `check-structure.sh` (structure only) | Machine-scope managed policy, the highest-precedence layer. OS-specific path resolved by the script (Pointer: for each OS's location, fetch [managed settings: where each mechanism stores the policy](https://code.claude.com/docs/en/managed-settings#where-each-mechanism-stores-the-policy) live. As of: 2026-11-10. Recheck trigger: that section moves a file location). Findings on it are report-only routing; managed policy is the administrator's, never edited by `--fix` |
| managed policy outside the filesystem | not read | The OS-level policy stores (Windows registry policy keys, the macOS managed-preferences domain) listed at the same pointer. `check-structure.sh` names them so an absent `managed-settings.json` is never read as "no managed policy deployed", but it does not read them |
| `docs/conventions/harness-config.yaml` (`suppressions`, the team layer of three cascade layers) | the engine | The consumer's suppression record, keyed by `finding_id`, in the convention home the repository binds (default `docs/conventions`). A finding it carries is reported as suppressed with its reason and the layer and file that supplied it; only the team layer suppresses, a personal-only entry is reported as not applied, and a malformed entry is reported and never suppresses. `suppressions.layers` in the engine's output names each file read |

### Reading settings.local.json safely

Treat `settings.local.json` as secret-bearing regardless of deny rules: the engine and
`check-structure.sh` report key counts, validity, and the four model and effort values by value,
and never dump its contents. Supplemental jq recipes: [context/procedures.md](context/procedures.md)
"Reading settings.local.json safely".

The counts are not guaranteed. Where the project's configuration blocks the read, whether a sandbox
`denyRead` merged from a project's `Read` deny on the file or filesystem permissions, the engine reports the
scope as `unreadable` and its rows as `not-inspectable` instead of failing. Record the file as not
inspectable, carry that into the report, and do not reach for another reader to get the counts anyway.

---

## Track progress

For any full audit run (Phases 1-5), keep a phase checklist and tick each phase as it completes,
either in-response or by copying `${CLAUDE_PLUGIN_ROOT}/skills/audit/templates/checklist.md`
into wherever the consuming repo keeps working task notes. Phase 5 is SKIPPED in default report-only
mode.

---

## Phase 1: Run the engine

Run the engine once, with the findings file in the topic's memory slice (the `memory_dir` the
consuming repo binds, default `.work/`):

```bash
mkdir -p .work/harness-config-audit
bash "${CLAUDE_PLUGIN_ROOT}/skills/audit/scripts/audit-engine.sh" --table \
  --out .work/harness-config-audit/findings.json \
  [--docs-dir <dir of pages already fetched>] [--debug-log <file>]
```

The engine records the installed Claude Code version itself (`claude --version`, carried as
`claude_version`); version-gated rows are evaluated against it, and Phase 3.2 compares issue-fix
versions against it. An unreadable version turns those rows into `skip`, never clean.

It also reads the upstream pages its rows rest on, every run, so a default run needs the network (`--max-age 0` asks the server every time, so a failed fetch is `unread`, never a stale cached copy).
It hands the docs index (`llms.txt`) and each page it needs to the plugin's shared fetcher,
`${CLAUDE_PLUGIN_ROOT}/scripts/fetch-docs.sh --cache --max-age 0`, which resolves the page from a link there and reads it verbatim to a file.
The `docs` object in the document is the coverage record, built from the fetcher's manifest: the
index and each page with its URL or path, byte count, line count, `sha256`, content type, read time,
and one `state`: `read`; `unread`, with a `reason` such as `fetch-failed`, `http-404`,
`unexpected-content-type`, `not-in-index`, `off-origin`, or `redirected-off-origin`; or `unparsed`, for a
settings-reference that downloaded but has no heading for `permissions` or `enabledPlugins`, or a
hooks page with no Event table (a soft 404, a reshaped page). Only `read` means the engine decided
anything from the page; every row resting on an `unread` or `unparsed` page is `not-inspectable`.
The pages this covers today are `settings-reference`, `env-vars`, and `hooks`, which the engine
requests only when `settings-reference` links it; every other page is Phase 3's. `--docs-dir` is optional reuse: a page already fetched there is read
instead of fetched again, and a page missing from it is fetched as usual.

Run it with `--json` instead when you want the whole document; `--table` prints the version and
the docs coverage, the findings, each with its paste-ready `suppress:` line, then the suppressed,
skipped, and not-inspectable rows, then the skill-listing measurement. Exit `0` means no
error-severity finding, `1` at least one, `2` a fatal condition (no project settings, `jq`
missing). Read the document, not the exit code alone.

A large configuration takes a minute or two (106 seconds for 83 plugins and 259 hooks on a busy Windows host), so give the call a 5-minute timeout.
Each category writes an `audit-engine: <category>` line to stderr as it starts; `SETTINGS_AUDIT_ENGINE_PROGRESS=0` turns them off.

Pass `--debug-log` when you know where this session's debug log is; otherwise the engine looks in the documented locations
(the debug-log variable first, then the newest file under `<user dir>/debug/`; Pointer: for that
variable and what it names, fetch
[environment variables: variables](https://code.claude.com/docs/en/env-vars#variables) live. As
of: 2026-11-10. Recheck trigger: the debug-log variable is renamed or changes what it names). A log it found rather than was given settles the row only when the log
names this project.

### 1.0 Hook inventory

The engine runs `scripts/check-hook-coverage.sh` itself and carries its result in
`hook_inventory`. That script enumerates settings-declared hooks **and** the hooks shipped by every
enabled plugin, read from the directory the session actually loads (a `directory` marketplace's
checkout first, the installed-plugin registry otherwise), plus the levers (`disableAllHooks`,
`allowManagedHooksOnly`, `strictPluginOnlyCustomization` in the forms that cover hooks; Pointer:
for those forms, fetch
[settings reference: strictPluginOnlyCustomization](https://code.claude.com/docs/en/settings-reference#strictpluginonlycustomization)
live. As of: 2026-11-10. Recheck trigger: that section changes which values lock hooks) that
switch hooks off, the mod-plane keys reported apart from them (`D/mod-plane`), and any
divergence between the loaded directory and the registry's cache snapshot.

**Read the inventory state, not just the rows.** `complete` means every enabled plugin resolved.
`partial` means some plugin did not resolve or a hook config did not parse, and `unreadable` names
which; for anything a partial inventory could not see, keep the conditional posture in
[required-permissions.md](reference/required-permissions.md) "Fail open where the inventory is
incomplete". `none` means no inventory was taken; treat it as partial for every family.

The inventory never runs a hook. Whether a hook *covers* a family is decided by the engine only
where the hook's plugin ships a coverage manifest (`hooks/coverage.json`); every other hook stays
the model's judgment, against the three preconditions in "Narrowing the baseline".

### 1.1 JSON validity

A scope in state `invalid` blocks every other row about that file; the engine reports the invalid
file as an error and nothing else about it.

### 1.2 Structure inventory

Fill the summary table from the engine's `scopes` and `rows`, or from `check-structure.sh` when a
managed layer or a start-directory copy is in play (the engine does not read those):

| File | Valid JSON | Keys | Deny | Ask | Allow | Hooks | MCP Servers | Plugins |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |

### 1.3 Baseline snapshot

Record the engine's `summary` for before/after comparison if `--fix` is used: findings by severity,
permission rule counts, MCP server count, hook count, plugin count.

---

## Phase 2: Validate

Load the audit checklist: [audit-checklist.md](reference/audit-checklist.md)

The engine's rows are settled. Work the judgment column of the division-of-labor table, category by
category; **full per-check criteria in
[context/validation-categories.md](context/validation-categories.md)** (read it when running Phase 2).

- **A, Schema & Structure**: engine-decided
- **B, Permissions**: for each baseline row the engine left at full severity, check narrowing 1 (a documented exemption in the consuming repo's rules) and narrowing 2 (a documented project hook convention); for a hook with no coverage manifest, take narrowing 3 by hand against the three preconditions in [reference/required-permissions.md](reference/required-permissions.md); report a missing `Read(./…)` row as present when `permissions.deny` carries the standards floor's bare spelling of it (same file, "Relation to the standards permission floor"); add any patterns the consuming repo's own rules declare as required
- **C, MCP Servers**: documented reasons for disabled servers; launcher conventions
- **D, Hooks**: timeout reasonableness, exec-form resolution on Windows-targeting repos (the four-part record is the Category D checklist row). Event names are the engine's; for a `hook-event-page-not-read` row it left `not-inspectable`, read the live hooks page
- **E, Plugins**: for each `dependency-disabled` finding, whether to enable the dependency or disable the plugins that need it; orphan-`true` and rename review. The engine merges `enabledPlugins` from the user, project and local files only (managed settings are not merged) and checks direct dependencies only
- **F, Environment Variables**: whether a variable the engine reports as not on the env-vars page is documented elsewhere or justified by the repo
- **G, Skill-listing budget**: the levers, scoped to the roster's composition (`skillOverrides` reaches project and user skills; plugin skills are managed through `/plugin`). The engine already reports `skillOverrides` entries that cannot take effect: a key naming a known plugin (`G/skill-override-plugin`, `warning`) and entries in the user dir's `settings.local.json` (`G/skill-override-home-local`, `info`); a colon key it left as `skip` is yours to name or leave undecided. When the engine reports `not measured`, name the routes (`/doctor` interactively, a `--debug` relaunch headless) and never report clean
- **H, Model and effort settings** and **I, Deep-link registration**: engine-decided; Phase 3 confirms the behavior each finding rests on before it is reported
- **J, Known-issues fix versions** and **K, Suppression age**: engine-decided (`known-issue-fixed` and `suppression-age` rows); for a J `info` finding, Phase 3.2 confirms the workaround is retired only if the live issue agrees; report each aged K entry, which keeps suppressing until the operator re-dates or removes it

---

## Phase 3: Research & Recheck

The engine already read `settings-reference` and `env-vars` in Phase 1 and decided the rows that
rest on them. This phase covers the pages and questions it does not: the pages the citation check
names, model configuration, permission syntax, and known issues.

**Read every page in this phase verbatim, not through a summarizer.** A cut read can make a row
that is present look absent. So for each page,
run `bash "${CLAUDE_PLUGIN_ROOT}/scripts/fetch-docs.sh" --cache --max-age 0 --out <dir> <page>...` to fetch
fresh bytes into one directory and grep the files, per the
[fetch route](https://github.com/melodic-software/claude-code-plugins/blob/main/docs/contracts/upstream-drift/README.md#reading-the-basis-the-fetch-route).
**A truncated read supports NO finding.** Say so and move on, in either direction: neither "the key is
gone" nor "the key is unchanged" is reportable from a read that may have been cut.

The fetcher writes `<dir>/manifest.json` beside the pages: per page `url`, `retrieved`, `sha256`,
`status`, `content_type`, `bytes`, `lines`, `state`, and `reason`, plus `claude_version` for the run.
A page it reports `unread` supports no finding.

Then run the citation check over that directory (it reads pages through the same fetcher):

```bash
bash "${CLAUDE_PLUGIN_ROOT}/skills/audit/scripts/check-doc-citations.sh" --docs-dir <dir>
```

It greps every key and sentence this skill's references cite
([reference/doc-citations.tsv](reference/doc-citations.tsv)) in the fetched pages. A `MISS` means a
checklist row cites text the page no longer carries, and that row is re-derived before it is used;
a `SKIP` means the page could not be read and its rows stay unverified this run.

Then run the effort-pin drift check, which the `effort-pins` scope runs on its own:

```bash
bash "${CLAUDE_PLUGIN_ROOT}/skills/audit/scripts/check-effort-pins.sh" --docs-dir <dir>
```

It hashes three parts of the model-config page (the Levels column table, the level rows of the
level-guidance table, and the resolution-list item that states each model's default level),
compares the hash with [reference/effort-table.baseline](reference/effort-table.baseline), and
lists every effort pin: `effort:` frontmatter in agents and skills, lane configs (JSON `"effort"`
keys in skill `context/` files), the lane launch lines, and Workflow literals in skill `context/` files and plugin `workflows/` scripts.
Report its lines verbatim. It never edits a pin. Exit 0: unchanged, and every pin names a listed level. Exit 1: the page changed
(`reason=table-changed` on every pin) or a pin names a level the page does not list
(`reason=level-not-in-table`). Exit 3: the page was unread or reshaped, which supports no claim
about any pin. Exit 2: fatal.

**Rebaseline.** After `table status=changed`, a person reads the changed rows and defaults and
re-decides each flagged pin and lane level. Then, from this skill's directory in the plugin's source
repository, they write the new baseline and commit it with those edits:

```bash
bash scripts/check-effort-pins.sh --print-baseline > reference/effort-table.baseline
```

The script never rewrites the baseline itself.

- **Pointer**: for the levels, per-model defaults and level guidance the check reads, see
  [Adjust effort level](https://code.claude.com/docs/en/model-config#adjust-effort-level) and
  [Choose an effort level](https://code.claude.com/docs/en/model-config#choose-an-effort-level).
- **As of**: 2026-10-10
- **Recheck trigger**: the check reports `table status=unparsed`, or either section is renamed or
  moved.

### 3.1 Official docs check

Keys live on [settings-reference](https://code.claude.com/docs/en/settings-reference); the
[settings](https://code.claude.com/docs/en/settings) page carries file locations, precedence, and
the `$schema` guidance. Compare against both:

- Are there new settings the project should consider adopting?
- Have any settings in use been deprecated or renamed?
- Has permission rule syntax changed?

### 3.2 GitHub issue recheck

For each issue in [known-issues.md](reference/known-issues.md), check live status by the first
route that works, and never let a route that does not work block the run:

1. The REST endpoint, `gh api repos/anthropics/claude-code/issues/<number>`, which some sessions
   serve where GraphQL is refused.
2. The consuming environment's own issue-tracking skill when one is installed
   (`/harness-ops:known-issues` in this marketplace).
3. Neither reachable: report the issue as **unverified since its last-verified date**, taken from
   that column of the table, so the report states how old the recorded state is instead of an
   unqualified "unverified".

The engine has already compared each fix version `known-issues.md` records with the installed Claude
Code version (Category J, `known-issue-fixed` rows). For each `finding` row, confirm the
settings-specific workaround is still needed and recommend retiring it if not. Compare versions by
hand only for an issue whose live thread names a fix release the file does not record, or when the
engine's row is a `skip` because the installed version was unreadable.

### 3.3 Model configuration verification

Any category H finding requires a fetch of
[code.claude.com/docs/en/model-config](https://code.claude.com/docs/en/model-config) confirming the
behavior the finding rests on. Do NOT report a category H finding from this checklist's wording
alone: the fallback-chain cap and the allowlist wildcard rule are upstream-owned and move with the
harness. The accepted `effortLevel` values the engine already reads from `settings-reference`.

### 3.4 Permission syntax verification

Fetch [code.claude.com/docs/en/permissions](https://code.claude.com/docs/en/permissions) and verify:

- The project's permission patterns match current documented syntax
- No new permission types added (e.g., new tool types)
- Wildcard behavior unchanged

---

## Phase 4: Report

Present all findings as a severity-rated GFM table:

| # | Category | Severity | Finding | Current | Recommended |
| --- | --- | --- | --- | --- | --- |
| 1 | B, Permissions | error | Deny rule placed in `settings.local.json`, where bug #8961 leaves it inert | `settings.local.json` → `permissions.deny: ["Bash(rm -rf:*)"]` | Move the rule into `settings.json`; keep `settings.local.json` deny empty |

Then three short sections the engine's document already carries:

- **Suppressed**: each finding the consumer's record retired, with its reason, layer and date, plus
  every `personal-only, not applied` and `malformed` entry by name.
- **Not decided**: every `skip` and `not-inspectable` row, with the reason. An unmeasured skill
  listing, an unreadable local file, and an unfetched page all land here, never in the clean count.
- **Findings artifact**: the path passed to `--out`. Its rows carry audit-pass's identity tuple, so
  `audit-pass` appends them unchanged and a later run diffs against them. When the model adds a
  judgment finding of its own, derive its identity with the engine (`audit-engine.sh anchor` and
  `audit-engine.sh finding-id`) and append the row in the same shape with `"tier": "judged"`.

For a finding the operator keeps, print its `suppress:` line as a stanza to paste under
`suppressions:` in `docs/conventions/harness-config.yaml` (team layer, in the bound convention home):
`check`, `claim`, `sites` with `surface` and a quoted `"anchor/v1":` key, a non-empty `reason`, and
today's `date`, keyed by the engine's `finding_id`. A suppression entry is the declared route for
session posture; the skill never infers posture from the environment.

### Severity guide

| Severity | Criteria |
| --- | --- |
| error | Security gap, broken config, or enforcement bypass |
| warning | Deprecated syntax, missing best practice, stale issue status |
| info | Enhancement opportunity, new feature available |

### Interactive checkpoint

**If `--fix` flag is set:** Present findings table, then ask user which items to fix. Do NOT proceed without their response.

**If report-only (no `--fix`):** Present findings table and summary. Note which items could be auto-fixed.

> "Which findings would you like me to fix? Reply with numbers, 'all', or 'skip'."

---

## Phase 5: Fix (only with `--fix` flag)

For each user-approved fix:

1. Make the edit. Route each approved edit to `settings.json` or `settings.local.json`
   through the built-in `update-config` skill when it resolves in this session, and write directly
   when it does not (`.mcp.json` is outside its scope; edit it directly); the one exception is orphan-`false` plugin removal, which goes through `scripts/fix-plugin-drift.sh --yes` so its lower-precedence-scope
   check still runs. In auto mode a settings edit needs the `[Self-Modification]` handshake: the
   classifier asks, and the user's explicit approval of that fix is the consent. Done when the
   target file carries the change and nothing else in it moved.
2. Validate with `jq . <file> >/dev/null` after each edit. Done when jq exits 0; on a parse error,
   revert that edit before touching the next one.
3. Report what changed, as the file, the key, and the before and after values. Done when every
   applied fix has a line in the report.

After all fixes:

- Re-run the engine and present the before/after `summary` (findings by severity, rule counts,
  server counts)
- Verify all config files are still valid JSON

### Refusals in auto mode

Two operations are refused in auto mode. Writing the team-layer suppression record is refused as
`[Instruction Poisoning]`; re-running `scripts/audit-engine.sh` for the after-fix summary, as
`[Self-Modification]`. Never retry around a refusal. Hand the operator the fallback: they apply the
suppression edit themselves, or run the engine and paste its output back. Report the before/after comparison from what they return.

Claim: auto mode refuses those two operations under those two category names. Basis: an empirical `harness-config:audit@0.48.2`
`--fix` auto-mode run on Claude Code 2.1.283, recorded in [melodic-software/.github PR #153](https://github.com/melodic-software/.github/pull/153)
when the team record lived under `.claude/`. As of 2026-09-27. Recheck when a release changes auto-mode classifier
categories, either refusal stops firing, or an auto-mode run first writes `harness-config.yaml`.

### Fixes the skill can apply

Auto-fixable (add `$schema`, **move** deny rules from local to project, plugin orphan-`false`
removal via `scripts/fix-plugin-drift.sh --yes`, which never adds a key) vs judgment-required (**adding** a baseline deny
rule, new settings from docs, permission restructure, MCP config, orphan-`true` removal, heuristic
rename), with the full matrix in [context/procedures.md](context/procedures.md) "Phase 5, fixes the skill can
apply". Adding and moving a deny rule are graded apart on purpose: moving one is #8961 placement, while
adding one has to be checked against hook coverage that may already hold and against an ask-gate the
addition would suppress.

---

## Required permission patterns

Category B (Phase 2) checks that the project's `.claude/settings.json` contains a security baseline of
deny/ask patterns: secret-file Read denies, destructive-git Bash denies, and a `git push` ask-gate. The
concrete list is in [reference/required-permissions.md](reference/required-permissions.md), which the
engine reads directly, so the list is never transcribed anywhere else. Projects with a stricter
posture declare additional required patterns in their own rules files; include any such list in the
Category B check.

A hook plugin can declare which baseline patterns its guards block in `hooks/coverage.json`. The
engine takes narrowing 3 from that manifest mechanically: the pattern drops to `info` with the plugin,
hook, and opt-out levers named, and only while the hook is live (no suppression lever set, plugin
resolved). Without a manifest the narrowing stays the model's reading of the guard's source.

Report the secret-file Read denies with their scope, not as protection: a `Read(...)` deny covers the
built-in file tools and the Bash file commands Claude Code recognizes, but not a subprocess that opens
the path itself. "Scope of a Read deny" in
[reference/required-permissions.md](reference/required-permissions.md) carries the covered /
not-covered split, the ranked remedies, and the platform limit. Carry it into the finding rather than
implying the file is unreachable.

CC settings schema, MCP server shape, hook event names, and permission glob syntax are upstream
invariants resolved against their own official pages when a check needs them, rather than asserted
as fixed patterns here.

## Boundary, the bundled `update-config` skill

Both skills handle `settings.json` files, so a request about settings can mean either.

- **`update-config` (bundled skill)**: edits the matching `settings.json` or `settings.local.json` for a described change: hooks for
  automated behaviors, permissions, environment variables, and hook troubleshooting. The model and
  the person can both invoke it.
- **This skill (marketplace plugin).** Audits the configuration that exists for correctness,
  security, and drift against current official docs, across settings, MCP, hooks, plugins, and
  permissions. It reports only, unless `--fix` is passed.

**Routing.** When the bundled `update-config` skill resolves in this session, prefer it for making
a settings change the person requested; prefer this skill for auditing what is configured. A
request such as "allow npm commands" or "add a hook that runs when Claude stops" is a change, not
an audit.

**Mutation gate.** `update-config` writes settings files as its job. This skill writes only in
Phase 5, under `--fix`, one confirmed fix at a time; Phase 5 routes each such settings edit through
`update-config`, and outside `--fix` this skill never chains into it on its own behalf.

**Availability is never assumed.** Settings, version and host decide whether the surface
resolves; this section states what to do when the
surface resolves, never that it is present. The four-part records live in
[reference/native-update-config.md](reference/native-update-config.md).

## Spoke paths

The `context/` files write this skill's directory as `<skill-dir>`, which is `${CLAUDE_SKILL_DIR}`.
Put that path in place of the placeholder before running a command or writing it into a brief. Those
files arrive through the Read tool as plain bytes, so a `${…}` token in them would reach the Bash
tool unsubstituted, and the Bash tool's environment has no `CLAUDE_SKILL_DIR` to expand it from.
Basis: the plugins reference,
<https://code.claude.com/docs/en/plugins/manifest-reference#where-each-variable-resolves>, verified
2026-10-07; recheck when that table adds supporting files to where a `${…}` reference resolves.

## Next

- A finding is a permission grant that auto mode drops or a hardcoded path:
  `/harness-config:audit-permission-grants`.
- The configuration is correct and the question is now what automation is missing:
  `/harness-config:audit-automation-gaps`.
- Several audit skills need one reconciled, resumable pass: `/harness-config:audit-pass`.

## Post-Audit

### Consumer workaround inventory

When a settings-affecting upstream issue drives a project-specific workaround (or one becomes
obsolete), record that in the consuming repo's own conventions/rules files. The plugin's bundled
[known-issues.md](reference/known-issues.md) tracks only broadly-applicable upstream issues and is
refreshed via plugin updates, not per-consumer edits.

### Suggest frequency

Based on findings count:

- 0 findings: "Config is clean. Recheck after next Claude Code major update."
- 1-3 findings: "Minor drift. Monthly recheck recommended."
- 4+ findings: "Significant drift detected. Consider running after every CC update."

