# Inventory

> List of every Claude Code surface this machine can invoke: built-in commands, subagents, tools, bundled skills and workflows, and installed plugin components. Use when: 'what slash commands do I have', 'what agents are available', 'is /foo a real command', 'what changed after the update', 'what does marketplace X give me'. On-disk audit: /harness-ops:audit-install-state.

- Skill: `melodic-software/inventory-2` (Agent Skill, multi-file: 17 files)
- Install (CLI): `npx get-skill add melodic-software/inventory-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/melodic-software/inventory-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: melodic-software (https://skillmd.com/u/melodic-software)
- Updated: 2026-11-12
- Page: https://skillmd.com/skills/melodic-software/inventory-2

---


**Arguments.** `[--builtin|--plugins|--bundled|--agents|--tools|--hooks|--docs] [--marketplace <m>] [--diff <f>]`. Full form: [--builtin|--plugins|--bundled|--agents|--tools|--hooks|--docs] [--marketplace <name>] [--diff <file>], or just ask in words

## Purpose

Answers one question completely: **what can this machine actually invoke, and where did each thing come from?**

That question has no single documented answer. The commands page publishes a table of built-in
commands, but the shipped executable registers names the table omits, so the executable is the
only complete source for the built-in surface and the page is a cross-check (`--docs`). Plugin
components, by contrast, are on disk and enumerable directly. This skill reads both and keeps them
clearly separated, because they are evidence of different quality.

The report is an inventory, not a judgment. Nothing here says a component is stale, misconfigured,
or wrong; the neighbors below own those verdicts.

## Scope boundary

| Question | Owner |
|---|---|
| What can I invoke, and where did it come from? | **this skill** |
| Is my install directory healthy. What is stale, what does the product manage? | `/harness-ops:audit-install-state` |
| Is the plugin fleet current, and at what scope? | `/harness-ops:plugins-audit` |
| Are settings, hooks, permissions, and MCP config correct? | `/harness-config:audit` |
| Which permission scopes hold which rules? | `/harness-config:audit-permission-state` |
| Are the configured MCP servers safe to run? | `/mcp-tools:audit-posture` (composed below when present) |

The distinction that matters most: `audit-install-state` inventories **files on disk**, this skill
inventories **capabilities that resolve**. A plugin can be present on disk and contribute nothing
because it is disabled, and a built-in command can be fully live while existing nowhere on disk.

## Run it

```bash
python3 "${CLAUDE_PLUGIN_ROOT}/skills/inventory/scripts/inventory.py" --deps-dir "${CLAUDE_PLUGIN_DATA}" --out ./claude-inventory.json
```

Python 3.11+ is the only requirement of the default bundle reader (`--reader regex`). No
`strings`, no `jq`, no PowerShell. The run takes about twenty seconds, dominated by reading and
tokenizing the executable once. `--reader parser` resolves bindings with a JavaScript parser and
reads a spread list as partial where code may change it; it needs Node.js with npm, installs its
pinned packages into the plugin data directory (`--deps-dir`) on first use, and without them
reports the binary source broken with the command that repairs it. `--reader compare` runs both.

Useful flags: `--binary <path>` (else the session's own CLI, then `PATH` and the usual install roots) · `--config-dir <path>` (else
`$CLAUDE_CONFIG_DIR`, else `~/.claude`) · `--binary-only` / `--disk-only` to skip a source ·
`--reader regex|parser|compare` (default `regex`) · `--docs` to add the `docs_crosscheck` block,
fetching the commands page, the changelog, and the tools reference · `--docs-file <path>` /
`--changelog-file <path>` / `--tools-docs-file <path>` to read one from a file instead (each
implies `--docs`). A failed fetch marks that block `unavailable` or `degraded` and never fails the
run.

**Extract once, filter at presentation.** The script always emits the whole inventory; the user's
filter selects what you *show*. Filtering in the script would mean a different extraction per
question and a cache that disagrees with itself. One JSON, many views.

## Resolving what the user asked for

Both spellings reach the same place. Treat a flag and its sentence as identical requests.

| Flag | Sentences that mean it | Show |
|---|---|---|
| `--builtin` | "built-in commands", "what ships with Claude Code", "is /foo real" | `builtin_commands` |
| `--bundled` | "bundled skills", "Anthropic's skills", "bundled workflows", "built-in plugins", "what `cc-plugin-*` ships" | `bundled_skills`, `bundled_workflows`, `builtin_plugins` |
| `--docs` | "is /foo documented", "what's undocumented", "was /foo removed" | `docs_crosscheck` |
| `--plugins` | "only the plugin ones", "what did my plugins add" | `disk.marketplaces` |
| `--marketplace <name>` | "what does melodic-software give me" | that marketplace only |
| `--agents` | "what agents do I have", "what subagent types ship built-in" | `builtin_agents`, then agents across every other source |
| `--tools` | "what tools does Claude Code have", "is Monitor a real tool" | `builtin_tools` |
| `--hooks` | "what hooks are wired", "what's hooked" | hooks across every source |
| `--diff <file>` | "what changed after the update", "does the cloud session match my machine" | this run against a saved one |
| none | "which CLAUDE.md files load", "what instruction files do I have" | `instruction_files` |

With no argument, report every section at summary depth and offer to expand one. A full unfiltered
listing runs to several hundred rows, which buries the answer to whatever prompted the question.

Two habits worth keeping. When a name the user asked about does not appear, say which sources were
searched. Absence from a filtered view is not absence from the machine. And when they ask about one
name, answer it directly first, then offer the surrounding list.

## Report structure

Lead with the shape of the answer, then the rows. Each `<n>` below takes its lane's integrity
status ([table](#staying-current-as-claude-code-ships)): a total when `ok`, `at least <n>` when
`degraded`. A `broken` lane gets no number anywhere, the Summary and its heading included: write
`withheld (lane broken: <cause>)`, where `<cause>` quotes the lane's `problems` text verbatim,
and list none of its rows.

```
# Claude Code inventory — <host>, <date>

## Summary
<n> built-in commands · <n> bundled skills · <n> plugins across <n> marketplaces · <n> enabled

## Built-in CLI commands (<n>)
One per line, alphabetical: name, argument hint, description, aliases, hidden/gated markers, and
the Invocable-by marker.

## Bundled skills (<n>)
One per line, alphabetical, in the same shape.

## Bundled workflows (<n>)
One per line: name, description, phases, Invocable-by marker.

## Built-in plugins (<n>)
One per plugin: `id`, description, load (`unconditional`, or `conditional` with its guards),
default enabled, gate flags, then its skills, agents and commands one per line with descriptions,
and its hook events on one line. Beside each gate flag's default, its `cached` value from
`builtin_plugin_state` (`absent` when `cached_present` is false), and any `enabledPlugins` setting
for the id with its scope. Open the section with the mods rollout flag
(`builtin_plugin_state.mods_flag`) and the pinned-gate caveat. Name any plugin whose `partial` is
non-empty.

## Built-in subagents (<n>)
One per line: name, description (`whenToUse`), roster (`default`, `conditional`, `absent`), model,
tools or disallowed tools with their `_source`, Invocable-by marker.

## Built-in tools (<n>, plus <n> factory-built)
One per line: name, description or search hint, `deferred` / `always_load` (null means decided
per session), gated marker, aliases. Name the factory count from `builtin_tool_notes`.

## Docs cross-check
Only with --docs. Counts per status, then the names in every status except `documented` and
`alias`, each with its docs text; kind mismatches and alias disagreements after. Say that the
changelog columns are a heuristic. Then the nested `tools` block the same way, with its own status.

## Plugin components
Installed plugins grouped by marketplace, then plugin, with a component-type breakdown.
Catalog-only entries — offered by a cached marketplace but not installed — are listed separately.

## Project scope
Skills, agents, and wired hook events from the current project's `.claude` tree, when present.

## Instruction files (<n>)
One per line, in the array's order: path, scope, kind, loads, bytes.

## MCP coverage
The posture report's inventory and findings with its saved path, or the one line saying mcp-tools
is not enabled. Then the connectors sentence from "MCP servers" below.

## Provenance
Which source produced which section, and anything the run could not resolve.
```

One per line, alphabetical, is the default for every list. The point of this report is scanning for
a name; a prose paragraph or a multi-column grid defeats that.

Never merge the sections into one list. A built-in command, a bundled skill, and a plugin skill
behave differently, different namespacing, different update path, different removal story, and a
merged list is unusable for deciding what to do about any of them.

**Invocable by** comes from `user_invocable` and `model_invocable`, present on every command,
bundled-skill, and bundled-workflow record:

| `user_invocable` | `model_invocable` | Marker |
|---|---|---|
| true | true | model+user |
| true | false | user-only |
| false | true | model-only |
| any | null | runtime (say which flag decides it, from `flag_driven`) |

`model_invocable` is true only for what the Skill tool admits: a `type:"prompt"` command declaring
`source:"builtin"`, or a bundled skill, in each case without `disableModelInvocation`. `local` and
`local-jsx` commands are false. A per-machine skill override in settings can still turn one off;
that is config, and the inventory does not read it. A built-in subagent in the default or
conditional roster is model+user; one outside the roster (`absent`, spawned only by a feature) is
runtime. A built-in tool is model-only.

## Reading the evidence honestly

**Extraction proves a command exists in the build, not that you can type it.** `hidden` and `gated`
mean a runtime predicate decides visibility from plan, platform, session type, or config. Report the
marker; never promote "present in the binary" to "available to you". `/skills` and `/help` are the
authority on what resolves right now.

**`enabledPlugins: false` does not settle enablement.** It spans several scopes, and a plugin's
hooks live in its own manifest. Report the map as read and tell the user to run
`/harness-ops:plugins-audit` for the verdict.

**Counts that disagree are data, not noise.** `bundled_skill_notes` carries `registrations_seen`
alongside `resolved`, and `dynamic_roster` when a registration builds its names at runtime (a loop
over a table, or a template literal). When seen exceeds resolved, say so rather than reporting the
smaller number as complete.

**A description or hint can depend on the session.** `description` and `argument_hint` each
carry a `_source` (`literal`, `template`, `constant`, `call`, `getter`, `arrow`, `roster`,
`unresolved`, `absent`). A getter or ternary yields `_variants`; the value shown is the
fallthrough branch, which is what a default session displays. An ellipsis marks a runtime
substitution. `unresolved` means the text is computed at runtime: say so rather than leave the
line blank. `integrity.undetermined` lists every name with a null invocability field or an
unresolved description or hint.

**The docs classify; the binary decides what exists.** `docs_crosscheck.names.<name>.status` is one
of `documented`, `undocumented`, `alias`, `docs_alias_but_registered`, `removed_in_docs`,
`removed_in_docs_but_registered`, `docs_only`. An `undocumented` name can be internal plumbing or a
gated feature, and a `docs_only` name can be one this build dropped; neither is a defect by
itself. `kind_mismatch` means the page and the binary disagree on command versus skill versus
workflow. The `changelog` field is a heuristic: the earliest version mentioning `/name`, and lines
using an add, rename, remove, deprecate, or alias word.

**One name can be two registrations.** When two distinct bundled registrations share a name, the
extraction keeps both as a list under that name with `collision: true`, and
`bundled_skill_notes.collisions` names them. Report each registration on its own line with its
description and invocation fields; never pick one and present it as the name's single meaning.

**A roster status is not availability, and a tool definition is not a loaded tool.** An agent's
`roster` says whether the default roster function registers it unconditionally (`default`), under a
runtime condition (`conditional`), or not at all (`absent`: a feature such as fork, coordinator
mode, or a workflow spawns it). A tool's `deferred` says whether it loads behind tool search; null
means a getter decides per session. Tools built by a factory at runtime (connector listings,
artifact family members) have no static name: `builtin_tool_notes.factory_definitions` counts them,
so say the named list excludes them rather than reporting it as every tool.

**A built-in plugin has three switches, and the inventory reads each separately.** `load` says
whether this session type requires the plugin at all (`load_guards` is the binary's own condition
text); `gated` with `gate_flags` says whether a runtime check, often a feature flag with the default
shown, decides that it is offered; `default_enabled` is what `enabledPlugins` falls back to when no
setting names `<id>`. A plugin can load, be gated off by its flag, and still read
`default_enabled: true`. Report all three; never collapse them into "on" or "off".

**A flag default is not this account's value, and the cached value is not the running session's.**
A full run (not `--binary-only`) adds `builtin_plugin_state`. Per gate flag, `cached` is the value
in the global config's `cachedGrowthBookFeatures` (`~/.claude.json`, or `.claude.json` inside a
custom config dir); `cached_present: false` means the key is absent and the binary default applies;
`null` means no cache was read. `enabled_overrides` lists each user, project and local
`enabledPlugins` entry for the plugin's id, and `enabled_setting` is the winning one (local, then
project, then user), or null when none names it and `default_enabled` decides. The scope order is
read live from
[settings: settings precedence](https://code.claude.com/docs/en/settings#settings-precedence)
(as of 2026-11-10; recheck when that section reorders the scopes). A scope listed in
`enabled_plugins_rejected` holds a non-Boolean value, so the script lets that file's whole map
contribute no override, per the `audit-skill-visibility` probe record for that rule; name the
offending keys. `mods_flag` carries
`tengu_plugin_hooks_modules`, the rollout flag for mods in installed plugins. A gate read through
the per-process pin is fixed at session start, so a cached value read later can differ from what a
running session uses: say "cached", never "live".

**A marketplace checkout is not an installation, and neither is enablement.** Three different sets:
a cached marketplace is a catalog of what is *available*, `disk.installed_plugins` is what is
*present locally*, and `enabledPlugins` governs what *loads*. They routinely disagree. Report the
one the question is actually about, and say which you used.

**A hook script on disk is not a wired hook.** Project scope reports hook *events* declared in
settings, not files sitting in a `.claude/hooks/` directory. A script nothing references is dead
weight, and listing it as a hook repeats the same present-versus-active error.

## Instruction files

A disk run adds `instruction_files`: every managed, user and project file Claude Code reads as
instructions. The project root is the git toplevel of the working directory (or of
`--project-dir`), so a run from a subdirectory lists the same files. Gitignored trees such as
`node_modules/` are skipped, since a fresh clone lacks them; a gitignored `CLAUDE.local.md` is
still listed. Each entry carries:

- `path`: `$MANAGED_POLICY_DIR/CLAUDE.md` for the managed-policy file in this OS's managed
  location, `~/`-relative for user files (`$CLAUDE_CONFIG_DIR/`-relative when the config dir is
  outside the home directory), repo-relative for project files, never absolute;
- `scope`: `managed`, `user`, `project`, or `local` (a `CLAUDE.local.md`);
- `kind`: `claude-md`, `claude-local-md`, `agents-md`, or `rule`;
- `loads`: `launch`, `path-scoped` (a rule with `paths` frontmatter), `on-demand` (a file in a
  subdirectory, read when Claude works there), or `not-by-default` (an `AGENTS.md` with a
  `CLAUDE.md`, `.claude/CLAUDE.md` or `CLAUDE.local.md` in its folder or any folder above it);
- `bytes` and `sha256` of the file.

Which file kinds exist and when each loads is the memory page's to say:
<https://code.claude.com/docs/en/memory>, the file table and "When Claude Code reads AGENTS.md".
As of 2026-10-03. Recheck when the memory page's file table changes. The script encodes the default
loading only: it does not expand `@path` imports, apply `claudeMdExcludes` or the Project
instructions setting, or read directories above the git toplevel.

Because no entry holds an absolute path, two runs over the same repository from different home
directories (a cloud session and this machine) produce identical `instruction_files` when the
files match. Save each run with `--out`, then compare them with
`python3 "${CLAUDE_PLUGIN_ROOT}/skills/inventory/scripts/compare_reports.py" <first> <second>`: a
changed `sha256` or an added or removed entry is a real difference between the hosts. Other
blocks (`disk.config_dir`, `project.root`) carry absolute paths and differ by design.

## Skills index

`--skills-index` prints only a JSON array, one record per skill, read from disk; the binary is
not read. It is the index a skill router shortlists from. `--config-dir`, `--project-dir` and
`--out` apply; `--exclude disabled,not-installed` drops records in those states.

```bash
python3 "${CLAUDE_PLUGIN_ROOT}/skills/inventory/scripts/inventory.py" --skills-index --exclude not-installed
```

Each record carries `plugin` (`<plugin>@<marketplace>`, or `user` / `project` for a skill in
`~/.claude/skills/` or the project's `.claude/skills/`), `skill` (the skill's directory name),
`name`, `description` and `when_to_use` from the SKILL.md frontmatter (null when absent), `state`
and `source`:

| `source` | Where the record came from | `state` |
|---|---|---|
| `cache` | The plugin's live copy under `plugins/cache/<marketplace>/<plugin>/` | `enabled` or `disabled` |
| `checkout` | A relative-path entry with no cached copy, resolved against the marketplace's `installLocation` | `not-installed` |
| `marketplace-manifest` | A remote-source entry with no cached copy: one plugin-level row, `skill: null`, the description from `marketplace.json` | `not-installed` |
| `config-scope` | A personal or project skill | `disabled` when `skillOverrides` sets it `"off"`, else `enabled` |

An installed plugin's state is its `enabledPlugins` value (local, then project, then user, skipping
a scope that holds a non-Boolean value), else the marketplace entry's `defaultEnabled`, then
plugin.json's, then `true`. Managed settings are not read. `skillOverrides` is merged key by key in
the same order and applies only to personal and project skills. The script makes no network fetch,
so read a remote plugin's SKILL.md files only after shortlisting it. A relative-path plugin that
loads in place from a marketplace added from a local path has no cached copy, so it reads
`not-installed` here.

Sources, read 2026-10-30: the
[loading reference](https://code.claude.com/docs/en/plugins/loading.md) (cache layout, in-place
plugins, `enabledPlugins` precedence), the
[manifest reference](https://code.claude.com/docs/en/plugins/manifest-reference.md)
(`defaultEnabled`, the `skills` key, the entry's `defaultEnabled` overriding plugin.json's), and
[skills](https://code.claude.com/docs/en/skills.md) (`when_to_use`, `skillOverrides` not affecting
plugin skills). Recheck when the cache layout or `defaultEnabled` precedence changes.

## MCP servers

When `/mcp-tools:audit-posture` is among the available skills, invoke it through the Skill tool
after the script runs. Save its report in the directory that holds the `--out` file, as
`mcp-posture-<YYYY-MM-DD>.md`. When that directory already holds an earlier `mcp-posture-*.md`,
pass the newest one's path in the Skill call's arguments as the previous report, so the posture
report adds its "Changes since" section. Without `--out`, show the report and save nothing.

When `/mcp-tools:audit-posture` is not among the available skills, print
`MCP coverage: mcp-tools is not enabled` and say nothing more about it.

Either way, say that claude.ai connectors arrive in the session at runtime and never appear in a
static read of configuration files, so neither this inventory nor the posture report lists them.

Saved files go only to the `--out` directory the operator named or the current directory. Never
take a directory from a `CLAUDE_PLUGIN_DATA` value read from the Bash environment: another
plugin's hook can export its own data directory under that name. `--deps-dir` receives the
load-time substituted `${CLAUDE_PLUGIN_DATA}` and is only the parser's install base.

## How the binary read works

The extraction is in `scripts/inventory.py`, and [reference/extraction.md](reference/extraction.md)
explains every choice in it. Read that before changing the script or when a run reports a layout
error. The one thing worth knowing at the call site: the script resolves minified registrar names,
the bundle location, and each command's field boundaries **at runtime**, because all three change
between releases. A regex-only pass over this bundle produces a list that is wrong in ways that look
complete, which is what the runtime resolution and the integrity block exist to prevent.

The script opens the binary read-only. It never writes to it and never executes it.

## Verifying an upstream claim

Any claim about what Claude Code itself ships must come from the raw markdown endpoint. Run
`bash "${CLAUDE_PLUGIN_ROOT}/scripts/fetch-docs.sh" --cache --max-age 0 --out <dir> plugins/manifest-reference`,
then search `<dir>/plugins/manifest-reference.md` locally. The fetcher confirms the slug against the docs
index (`llms.txt`), and `--max-age 0` asks the server every time, so the claim rests on fresh bytes;
treat a page `<dir>/manifest.json` reports `unread` as no evidence. Never rest an absence claim on
a summarizing fetch; what one returns is read live from
[tools reference: WebFetch tool behavior](https://code.claude.com/docs/en/tools-reference#webfetch-tool-behavior)
(as of 2026-11-10; recheck when that section stops describing WebFetch output as an extraction).

Upstream facts this skill depends on, each with the trigger that obliges re-deriving it:

| Claim | Basis | Recheck trigger | Verified |
|---|---|---|---|
| The commands page publishes a partial built-in table, so it is a cross-check and the binary stays the source: its "All commands" table carries rows for commands, aliases, removed commands, and **Skill**/**Workflow** markers, and the binary registers names it omits | `docs/en/commands.md` parsed by `--docs` (115 rows) against a `--binary-only --docs` run on this machine, which classed 39 registered names `undocumented` | The commands page drops or restructures its table (the docs block goes `broken`), or the undocumented count reaches zero | 2026-09-29, Claude Code 2.1.284 |
| The plugin component set is skills, commands, agents, workflows, output-styles, themes, monitors, hooks, bin, settings.json, .mcp.json, .lsp.json, dependencies; the manifest also declares `channels`, each bound to one of the plugin's MCP servers, `types` (a mod's `.d.ts` declaration file) and `experimental.evals` (the eval-case directory), none of which this skill scans | `docs/en/plugins/manifest-reference.md` manifest schema and standard plugin layout table | The manifest schema gains or drops a component key | 2026-10-07 |
| A user reaches a built-in or custom subagent by @-mention or `--agent`, and the model by the Agent tool; the page documents Explore, Plan, general-purpose, claude, statusline-setup, and claude-code-guide, and the binary also defines fork, web-fetch, worker, workflow-subagent, and comment-thread-analyst | `docs/en/sub-agents.md` ("Built-in subagents", "Invoke subagents explicitly") against a `--binary-only` run on this machine | The page changes its built-in list or invocation patterns, or a run's `builtin_agents` names change | 2026-09-29, Claude Code 2.1.285 |
| Claude Code ships built-in plugins as `<name>@builtin`, registered in the binary rather than installed on disk, so the binary read stays the source; the overview's roster of built-in mods is a cross-check, as the commands page is for commands, because the binary registers built-ins the roster omits | Pointer: [Mods built into Claude Code](https://code.claude.com/docs/en/plugins/mods/overview#mods-built-into-claude-code), against an inventory run on this machine whose `builtin_plugins` lane resolved 11 names, every roster row among them | The roster section moves, its rows come to match the `builtin_plugins` names, or the `builtin_plugins` lane goes `broken` | 2026-10-02, Claude Code 2.1.288 |
| A built-in plugin's gate reads a GrowthBook flag cached in the global config's `cachedGrowthBookFeatures`; `cc-plugin-diff`'s `isAvailable` is `ft()&&lo(Ce(),at)`, where `lo` stores the first value per flag in `pinnedFeatureValues` for the process, so availability is fixed at session start; installed mods load under `tengu_plugin_hooks_modules` (default true) | The 2.1.287 bundle on this machine (`var vr=()=>ft()&&lo(Ce(),at)`; `function lo(e,n){let r=Co().pinnedFeatureValues??=new Map;if(!r.has(e))r.set(e,C(e,n));...}`; `var HXe="tengu_plugin_hooks_modules";var kjt=!0;var dQ=()=>C(HXe,kjt)`) and `jq .cachedGrowthBookFeatures ~/.claude.json` | A release renames `cachedGrowthBookFeatures` or `pinnedFeatureValues`, drops the rollout flag string (`builtin_plugin_notes.mods_flag_in_bundle` goes false), or docs publish per-plugin availability | 2026-10-01, Claude Code 2.1.287 |
| The tools reference table lists tools by exact name and is partial: 45 of 80 statically named tools are in it, and it keeps `TaskOutput`, which the binary only names in a retired-names list | `docs/en/tools-reference.md` parsed by `--docs` (46 rows) against a `--binary-only` run on this machine | The tools table restructures (the nested `tools` block goes `broken`) or the `undocumented` count reaches zero | 2026-09-29, Claude Code 2.1.285 |

The changelog at `https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md` is the
fastest way to explain a diff between two runs; `--docs` attaches its per-name history. Commands
appear and disappear between releases. Cite the changelog line that introduced or removed a
command rather than asserting it changed.

## Staying current as Claude Code ships

This skill reads Claude Code internals that change every release, so **drift must never be
silent**.

**Read the integrity block before quoting any number, lane by lane.** Every run carries one, and it
states per lane (`builtin_commands`, `bundled_skills`, `plugin_backed`, `bundled_workflows`,
`builtin_agents`, `builtin_tools`, `builtin_plugins`) whether that lane's counts are verified or merely believed; the top-level status is the worst lane.
`docs_crosscheck` carries its own `status` (`ok`, `degraded`, `broken`, `unavailable`) and never
changes the integrity status or the self-check exit code:

| Status | Means | Do |
|---|---|---|
| `ok` | Build matches the last validated release, every check passed | Report that lane's counts as totals |
| `degraded` | Extraction worked, but something is unaccounted for | Report that lane's counts as **floors**, and say what is unaccounted for |
| `broken` | A canary is missing or nothing resolved in that lane | Do not report that lane's counts at all, not even caveated or as the number of entries the file lists; name the lane, call it `broken`, and quote its `problems` text verbatim; the other lanes' counts stand |

A single broken lane makes the top level `degraded`, not `broken`, so the other lanes' counts are
never withheld. Top-level `broken` means every lane is broken or the binary could not be read.

The dangerous failure is a clean, confident, short list from a new registration path, not a crash,
so the checks catch shortfall: canary commands, a canary workflow, canary subagents and tools, a
minimum ratio of resolved commands to registration tokens present, a sweep for registrar-shaped
exports the script does not know about, the resolved-versus-seen gap on bundled skills, and
unresolved subagent and tool names.

**Run the drift check on a schedule, not on incident:**

```bash
python3 "${CLAUDE_PLUGIN_ROOT}/skills/inventory/scripts/inventory.py" --deps-dir "${CLAUDE_PLUGIN_DATA}" --self-check
```

It prints one verdict line and exits `0` ok, `1` broken, `3` degraded; `2` is argparse's usage
error, never a degraded run. `/harness-ops:changelog apply` runs it after integrating a CLI
release, with a `--docs` extraction it diffs against the previous run's, and files a work item when
the verdict is not `ok`.

**What a maintainer updates.** Most releases need no change: registrar names are discovered and the
bundle is found by export name. When a run goes `degraded` or `broken`, each verdict maps to one
edit in `scripts/inventory.py`, per the verdict table in
[reference/extraction.md](reference/extraction.md). After revalidating against the new build, bump
`VALIDATED_AGAINST` to that version; it is the one constant that turns "believed" back into
"verified".

**For downstream consumers.** An older plugin version against a newer CLI reports `degraded` or
`broken` rather than a wrong answer; fixes ship as a new plugin version, which
`/harness-ops:plugins sync` carries. Never quietly widen a count to make a status look better.

## Next

/harness-ops:audit-native-overlap

It reads this inventory's invocability, argument-hint, and workflow fields to find where local
skills duplicate a native surface.

## Gotchas

- **A name in the build is not always a command.** Strings such as `alias` and `todos` match a
  naive `name:"…"` search and are not commands. The brace-depth reader plus the `type:` requirement
  is what excludes them. Verified 2026-09-29 against Claude Code 2.1.284, by running
  `inventory.py --binary-only` on this machine: neither name appears under `builtin_commands`.
  Recheck when the extractor's `type:` requirement changes or a release adds a command by either
  name.
- **An alias is not a separate command.** Count commands once and list aliases beside them, or
  your total will drift from `/help`. Pointer: when stating which commands are aliases, fetch
  [commands: all commands](https://code.claude.com/docs/en/commands#all-commands) live; our probe,
  a run of `inventory.py --binary-only` on Claude Code 2.1.284 (2026-09-29), reported `usage` with
  aliases `cost` and `stats`. As of: 2026-11-10. Recheck trigger: the commands page changes the
  alias rows for `/usage` or the extraction reports a different alias set for `usage`.
- **A command in the docs can be a skill in the binary.** The commands page lists `/schedule` as a
  command; the binary registers it as a bundled skill, so `docs_crosscheck` flags `kind_mismatch`.
  Report the binary's kind, since it decides namespacing and model invocability. Verified
  2026-09-29 against Claude Code 2.1.284 by a `--binary-only --docs` run on this machine. Recheck
  when the page adds a **Skill** marker to `/schedule` or the binary registers it as a command.
- **A bundled skill can also appear as a command object.** When a name registers as both, the skill
  registration wins; the script drops the duplicate so one capability is not counted twice.
- **An npm install has no embedded bundle.** The launcher script is small and loads its bundle
  elsewhere. The script reports this rather than parsing a shim, and the disk half still works.
- **The bundled-skills directory is a lazy cache.** Skills are extracted there on first use, so it
  under-reports. It is never the source for the bundled-skill list.

