Knowledge gaps

Knowledge Gap Detection

Analyze the current project's dependencies against Basic Memory coverage to identify packages that should be documented but aren't. Supports npm, Rust crates, Go modules, PHP Composer, PyPI, and RubyGems.

When invoked with --stale [<ecosystem>], the skill instead runs an alternative version-drift workflow (Mode A below) that flags documented brew/npm/cask/crate/vscode/plugin notes whose recorded version has fallen behind upstream — not a coverage audit.

Edge Cases

• No manifest files found — if none of the Step 0 globs match, report "No package manifest files detected" and skip Steps 1–5. Still proceed to Steps 6–9 to check for tool manifests.
• No tool manifests found — if Steps 6 finds nothing, report "No tool manifests detected" and skip Steps 7–9. Still run Step 10 for dead wiki-links.
• Ecosystem directory missing in BM — if list_directory returns nothing for npm/, crates/, etc., treat coverage as 0 documented for that ecosystem. Do not error.
• Empty manifest — if package.json has no dependencies or devDependencies, or Cargo.toml has no [dependencies] tables, report "No dependencies found in <manifest>" and skip that ecosystem.
• Brewfile with only comments or taps — if no brew "...", cask "...", or vscode "..." lines exist after filtering, report "No tools found in Brewfile".
• Workflow file with no uses: lines — if a .github/workflows/*.yml file exists but has no uses: lines matching the pattern, skip it silently.
• RETRO-*.md not committed — retro files are gitignored; find still detects them locally. This is expected behavior.
• Readwise not available — if readwise_search_highlights or reader_search_documents fails or returns no results, skip the reading- signal analysis in Step 14b and report only graph-based hub gaps in Step 15.
• bm CLI not available — if the bm project info command in Step 10 fails, skip the quick-exit gate and proceed directly with the relation index queries.
• brew CLI not available — if brew leaves in Step 7b fails (non-zero exit, command not found, or empty output on a non-macOS host), skip the Brewfile-vs-installed reconciliation silently and report brew coverage in Brewfile-only mode. This is the expected fallback for auditing a remote machine's declared deps from a different host.
• --global manifest absent — if ~/.claude/plugins/installed_plugins.json (plugins) or ~/.agents/.skill-lock.json (skills) is missing/unreadable in Step 7c, skip that population silently and note it; the other still runs. These are user-global, so a CI host without ~/.claude simply yields nothing.
• --global + --stale together — reject; they are separate modes (coverage vs drift). Run one at a time.
• --full with a scope flag — reject --full combined with any of --limit/--since/--sample ("--full can't be combined with --limit/--since/--sample — pick one"). --full is the opt-out from the scope preflight, not a narrowing flag, so pairing it with one is contradictory.
• Scope preflight is top-level-only — the --stale scope preflight (AskUserQuestion) fires only in the Mode A skill body, never inside a wave subagent. It must never migrate into references/staleness-detection.md (loaded verbatim into subagents), where AskUserQuestion auto-approves and returns empty. A sub-40 largest cohort, or any typed scope flag / --full, skips the prompt.
• recent_activity unavailable or empty (Step 10 recency-scoped sweep) — if the call errors, report "Recency sweep unavailable — could not confirm recent-note coverage" in the gap report rather than treating an empty or failed result as "zero recently active notes." If it returns zero results (no MCP error, empty list), that IS a valid clean state — distinguish the two explicitly in the report; do not conflate them.
• read_note fails mid-sweep (Step 10 recency-scoped sweep, step 3) — skip that note, name it in the report ("N of M recently-active notes could not be read — sweep is partial"), and continue with the remaining notes. Do not silently report the partial result as if it were the full recency sweep.

Workflow

This skill has two mutually exclusive modes. Pick exactly one based on the user's invocation, then run only that mode's steps. Never combine them.

Mode A — Stale mode (alternative entry point)

Triggered when: the user invoked the skill with the --stale flag (e.g., /knowledge-gaps --stale, /knowledge-gaps --stale npm), or used trigger phrases like "stale notes", "outdated formulae", "drifted notes", "which tools need updating".

Argument parsing: --stale takes an optional ecosystem token — --stale [brew|npm|cask|crate|vscode|plugin]:

• bare --stale → check all six supported cohorts.
• --stale <token> where the token ∈ brew|npm|cask|crate|vscode|plugin → check only that cohort.
• any other token (e.g. action, gh, go, docker, skill, pypi, gem, composer) → reject with: "--stale supports brew, npm, cask, crate, vscode, plugin. action/gh/go/docker have no single canonical comparable version; skill is unsupported (no comparable version); pypi/gem/composer are deferred." Do NOT silently fall back to all.

Scope modifiers (additive): after the ecosystem token, --stale also accepts three optional scoping flags, in any order — --stale [<eco>] [--limit N] [--since <date>] [--sample N]. These exist for large cohorts (e.g. a 388-note npm directory) where the default full sweep is impractical in one turn. Applied per cohort — a bare --stale --limit 50 run against all six cohorts checks up to 50 notes in EACH of brew/npm/cask/crate/vscode/plugin, not 50 total.

• --since <date> — the highest-leverage flag: it restricts the cohort to notes NOT touched since <date> (ISO YYYY-MM-DD), and does so BEFORE the upstream fetch (S3) and before S2's per-note read storm — it shrinks N, it doesn't just filter the rendered report. Resolved via one recent_activity(timeframe="<date>") call, not N read_note calls. See the staleness-detection reference's S1 for the exact mechanics.
• --limit N — caps the (optionally --since-narrowed) cohort to the first N notes in listing order. Deterministic: the same input always yields the same slice — which is exactly why it is not safe to tile a large cohort into waves on its own (a second --limit N call re-fetches the identical first N notes rather than advancing). Only safe to layer on top of an already date-disjoint --since slice; partitioning itself uses successive --since cutoffs, not --limit (see "Large-cohort strategy" in the reference file for the actual mechanism).
• --sample N — takes a random N-note sample instead of the first N — an unbiased bounded slice for a cohort too large to sweep in full, its N notes checked the same way as any other scope (it estimates no drift rate and extrapolates nothing). Not partition-safe (overlapping/uncovered draws across separate runs) — use it for a single bounded run, never to tile a full sweep.

N must be a positive integer and <date> a valid ISO date; reject a malformed value by name rather than silently ignoring it. --limit and --sample both set a cohort size, so they're mutually exclusive — if both appear, reject with: "--limit and --sample can't both be set — pick one."

• --full — skips the interactive scope preflight below and sweeps the full cohort unscoped. It is the power-user opt-out from the preflight prompt, not a scope narrower — so it is mutually exclusive with the three scope flags. If --full appears alongside any of --limit/--since/--sample, reject with: "--full can't be combined with --limit/--since/--sample — pick one." (--full already implies the widest scope; a narrowing flag contradicts it.)

Scope preflight (interactive, top-level only). Before delegating to the staleness-detection workflow, and only in this top-level skill body (a human is present at every /knowledge-gaps invocation), run a one-question scope preflight for large unscoped sweeps. This mitigates a filed, still-open upstream limitation — Basic Memory has no bulk metadata/version projection, so the drift sweep must read_note every entity in a cohort (UPSTREAM-basic-memory.md, No bulk metadata/version extraction); the preflight lets a human shrink N before that O(N) read storm rather than after. It is a UX layer over the open gap, not a reopening of automated wave orchestration (spike vp-claude-4g53, DEFER verdict) — it automates nothing, it only helps a human pick an existing scope flag.

Never place this preflight in staleness-detection.md. That file is loaded verbatim into wave subagents, and AskUserQuestion auto-approves / returns empty answers inside a subagent — so an interactive gate there silently no-ops. The gate is safe by construction only here, in Mode A's top-level body.

Run the preflight when both conditions hold:

1. No scope modifier was typed — none of --limit/--since/--sample, and --full was not passed (either flag skips the preflight: a typed scope flag already answers the question; --full is the explicit opt-out).
2. The largest selected cohort exceeds 40 notes. Learn each selected cohort's size from one cheap list_directory(dir_name="<cohort-dir>", depth=1) per selected cohort (the same listing S1 needs anyway) — a count only, no read_note. For bare --stale that is one listing per cohort; for --stale <eco> it is a single listing. Compare the largest count to 40.

The threshold 40 is an honest heuristic (≈5 concurrent reads/turn; a serialized 1s/call crate cohort; well below the 122-note brew incident this preflight was built for) with no config surface by design — if it fires too eagerly or too late in practice, it is a one-literal edit here.

When both conditions hold, ask once with a single AskUserQuestion — one question object regardless of how many cohorts qualify (fold the per-cohort counts into the question body), neutral framing. The menu is three options, or four when the single-ecosystem sample option below applies.

Scope question (always asked):

1. Full sweep (all N) — check every documented note in the selected cohort(s).
2. Untouched in the last 90 days (recommended) — resolves to --since <date> where <date> is today minus 90 days in ISO YYYY-MM-DD. This is the broader narrowing: it keeps every note not refreshed in roughly the last three months. Recommended because it gives the widest coverage short of a full sweep, and because --since's result set nests cleanly across repeat runs (per this skill's own S1 mechanics).
3. Untouched in the last 180 days — resolves to --since <date> where <date> is today minus 180 days, ISO. This is a tighter subset of option 2, not a wider one: a note untouched for 180 days is necessarily also untouched for 90, so option 3 ⊆ option 2 ⊆ full. Pick it to check only the most-neglected notes (not refreshed in roughly six months) — a smaller cohort than option 2, not a larger one. When framing the options to the user (labels and question body), state this direction explicitly so a longer window is never mistaken for a bigger cohort.
4. Random sample of 30, excluding the last 7 days (recency-floored, unbiased slice) — offered only for a single-ecosystem --stale <eco> run whose one cohort is very large (> 150 notes); never for a bare --stale (see below for why). Resolves to --since <today−7d> --sample 30 — compute the 7-day cutoff at runtime, the same way options 2/3 compute theirs — a random 30-note draw from the notes NOT touched in the last week, checked in full, the same per-note enumeration as any other scope. It computes no drift rate and extrapolates nothing; it is simply the only narrowing that stays representative when a cohort is both huge and actively touched, so neither --since window is tractable on its own: the 90-day narrowing still leaves more than one turn can process (dogfood: ~70 on a ~500-note npm cohort) while the 180-day option resolves to ≈0 (nothing sits untouched that long on a maintained graph), and --limit's alphabetical slice (brew starts at aarch64…) would be a biased subset. The 7-day floor is a narrow guard, not a staleness-narrowing strategy: it excludes only notes an upgrade-haul (or any refresh) may have stamped with today's date moments earlier, so the draw stays representative of the wider cohort — unlike options 2/3, whose 90/180-day floors deliberately bias toward the most-neglected notes as their whole strategy. Its non-partition-safety (overlapping draws across runs) does not matter here — the preflight is a single-shot pick, not a tiled sweep. Below the 150 bar this option is not shown (a 90-day narrowing suffices — brew's 122-note cohort collapsed to 2), keeping the menu at three options for the common case. A different cutoff or sample size stays available as a typed --since <date> --sample N combo (already valid — only --limit+--sample are mutually exclusive, not --since+--sample; a typed combo bypasses this preflight).

Compute all cutoff dates at runtime — never hardcode them. --limit is never a preflight option (it stays a typed flag): its alphabetical slice is the very bias the sample option above is chosen to avoid, so it belongs layered on a typed --since, not offered here.

The sample option is single-ecosystem-only. A bare --stale --sample 30 would draw 30 notes from each selected cohort and silently truncate the smaller ones (a 45-note cask cohort → a random 30), so the > 150 gate is scoped to a single --stale <eco> run by construction — never offered for a bare --stale. The 150-note bar (like the 40-note trigger above) is a fixed heuristic with no config surface — a one-literal edit if it fires too eagerly or too rarely.

On no response (timeout): default to option 2 (--since <today−90d>) — never the full sweep, and never the sample even when it was on the menu. The full sweep is the costliest option, so defaulting to it on silence would reintroduce the exact blow-up this preflight exists to prevent. The sample, though cheaper than the 90-day narrowing, is a random draw — a silent automatic default must be deterministic and reproducible, which a stochastic sample is not — so the deterministic 90-day --since is always the fallback, regardless of whether the sample was offered. How long to wait before defaulting is the executing session's own timeout config, not a number hardcoded here.

One caveat where the sample option was on the menu (a single >150 cohort whose 90-day narrowing is itself known to leave more than one turn can process): the 90-day timeout fallback keeps determinism but does not by itself make that cohort one-turn-tractable. Flag the resulting default sweep as possibly partial and prefer the Large-cohort wave-tiling strategy (see staleness-detection.md) over silently attempting one oversized turn — a truncated single-turn sweep would under-report drift with no signal that it stopped early.

Once the scope is resolved (a typed flag, an interactive pick, or the timeout default), carry it forward as if it had been typed and proceed to What to do below. A natural-language trigger with no flags is treated exactly like a bare --stale — the same threshold check applies (the gate lives in these shared Mode A instructions, not in slash-command parsing, so it fires identically on both routes).

What to do: load and follow the staleness-detection reference file in full — do NOT execute any of the Mode B steps below in the same session:

${CLAUDE_PLUGIN_ROOT}/skills/knowledge-gaps/references/staleness-detection.md

Pass the parsed ecosystem scope (one cohort, or all six) AND the resolved scope value — any typed --limit/--since/--sample, or what the scope preflight resolved to: a --since <date> (90d/180d pick or timeout default) or a --since <today−7d> --sample 30 (the large single-cohort recency-floored sample option) — to that workflow; it runs the per-cohort drift check and renders one ### Version Drift — <eco> section per checked cohort. Also pass how the scope was chosen (typed flag / interactive pick / timeout default) so S8 can stamp its provenance footnote.

Mode B — Standard mode (manifest-driven coverage)

Triggered when: the user invoked the skill without the --stale flag.

The remaining steps below (numbered 0 through 15) describe this mode only. Skip them entirely when --stale is present — Mode A is a complete alternative workflow.

The --global flag is a Mode B addition: when present, it activates Step 7c (user-global installed-plugin / skill coverage) alongside the project-manifest steps. --global and --stale are mutually exclusive — if both appear, reject with "--global (coverage) and --stale (drift) are separate modes; run one."

0. Detect project ecosystems

Before parsing dependencies, scan for manifest files using the Read tool (not Bash). Check for the following in the current working directory:

Manifest file	Ecosystem	BM directory
package.json	npm	npm/
Cargo.toml	Rust / crates	crates/
go.mod	Go modules	go/
composer.json	PHP / Composer	composer/
requirements.txt or pyproject.toml	Python / PyPI	pypi/
Gemfile or Gemfile.lock	Ruby / RubyGems	gems/

A project may have multiple manifest files (e.g., a monorepo with both package.json and Cargo.toml). Process each detected ecosystem separately and combine results in the final report.

Check for root-level manifest files using Read — do not use Glob for root manifests, as it recurses into node_modules/ and similar directories:

Read("./package.json")
Read("./Cargo.toml")
Read("./go.mod")
Read("./composer.json")
Read("./pyproject.toml")
Read("./requirements.txt")
Read("./Gemfile")

If Read succeeds, the ecosystem is present and the content is already loaded for Step 1 (no re-reading needed). If Read returns "file not found", skip that ecosystem.

1. Parse dependencies

For each detected ecosystem, read the manifest and extract dependencies:

npm (package.json): Read package.json and extract all dependencies and devDependencies keys. Exclude workspace packages (check workspaces field and skip matching entries).

Rust (Cargo.toml): Read Cargo.toml and extract [dependencies], [dev-dependencies], and [build-dependencies] table keys. Exclude workspace members.

Go (go.mod): Read go.mod and extract all require directives. Ignore indirect dependencies (// indirect comments) unless explicitly requested.

PHP Composer (composer.json): Read composer.json and extract require and require-dev keys. Skip php and ext-* entries (platform requirements, not packages).

Python PyPI (pyproject.toml or requirements.txt):

• pyproject.toml: extract [project].dependencies array and [project.optional-dependencies] entries
• requirements.txt: extract package names (strip version specifiers)

Ruby (Gemfile): Read Gemfile and extract all gem '<name>' lines. Group by Bundler groups (group :development, etc.).

2. Check Basic Memory coverage

For each ecosystem, get all documented packages in one lightweight call:

list_directory(dir_name="<ecosystem-dir>", depth=1)

This returns all <prefix>-* note titles without loading content. Cross-reference against the dependency list to classify each package:

• Documented — a <prefix>-<package-name> note exists
• Undocumented — no dedicated note

For undocumented packages that land in Tier 1 (after step 3), check if they're mentioned in engineering notes:

search_notes(search_type="text", query="<package-name>", page_size=3)

Classify matches as:

• Mentioned — appears in a note but isn't the primary subject

Limit this fallback to Tier 1 candidates to avoid excessive API calls.

3. Tier by import frequency

For undocumented packages, count imports using the Grep tool. Ripgrep automatically respects .gitignore, skipping node_modules, .git, etc.

npm:

Grep(pattern="from ['\"]<package-name>['\"/]", glob="**/*.{js,ts,mjs,cjs}", output_mode="count")
Grep(pattern="require\\(['\"]<package-name>['\"/]", glob="**/*.{js,ts,mjs,cjs}", output_mode="count")

Rust:

Grep(pattern="use <crate_name>::", glob="**/*.rs", output_mode="count")
Grep(pattern="extern crate <crate_name>", glob="**/*.rs", output_mode="count")

(Replace hyphens with underscores for the crate name in use statements.)

Go:

Grep(pattern="\"<module/path>\"", glob="**/*.go", output_mode="count")
Grep(pattern="\"<module/path>/", glob="**/*.go", output_mode="count")

PHP Composer:

Grep(pattern="use <Vendor>\\\\<Package>", glob="**/*.php", output_mode="count")

Python:

Grep(pattern="import <package_name>", glob="**/*.py", output_mode="count")
Grep(pattern="from <package_name>", glob="**/*.py", output_mode="count")

(Replace hyphens with underscores for import names.)

Ruby:

Grep(pattern="require ['\"]<gem_name>['\"]", glob="**/*.rb", output_mode="count")

For scoped npm packages (e.g., @fastify/postgres), match the full name — the @ and / don't need escaping in the pattern.

Classify:

• Tier 1 (3+ files import it): Must document — core dependency
• Tier 2 (1-2 files): Should document — used but limited scope
• Tier 3 (devDependencies/dev only, 0 runtime imports): Optional — tooling

4. Generate gap report

Present a structured report with a section per ecosystem. See ${CLAUDE_PLUGIN_ROOT}/skills/knowledge-gaps/references/report-templates.md ("Package coverage report") for the full format: a ## Knowledge Gap Report — <project> header, one ### <eco> Coverage: X/Y (Z%) section per ecosystem (each with Tier 1 / Tier 2 / Tier 3 / Already Documented subsections, packages ranked by import count), then an ### Overall Summary.

5. Offer enrichment

For the top 3-5 undocumented Tier 1 packages across all ecosystems, offer to run /intel with the appropriate prefixed invocation:

• npm: /intel fastify
• crate: /intel crate:serde
• go: /intel go:github.com/gin-gonic/gin
• composer: /intel composer:laravel/framework
• pypi: /intel pypi:requests
• gem: /intel gem:rails

Present packages ranked by import count.

---

6. Detect tool manifests

Glob for tool manifest files in the current working directory:

Read("./Brewfile")
Glob(pattern=".github/workflows/*.yml")
Glob(pattern=".github/workflows/*.yaml")
Read("./Dockerfile")
Glob(pattern="*.dockerfile")
Glob(pattern="Dockerfile.*")
Read("./.vscode/extensions.json")

Announce which manifests are found. If none are found AND --global was not passed, skip Steps 7–9 and note "No tool manifests detected" in the report. If --global was passed, still run Step 7c — it reads user-global manifests independent of any project tool manifest.

7. Parse tool manifests

For each detected manifest, read and extract tool identifiers:

Brewfile: Read the file and extract entries by line pattern:

• brew "<name>" or brew '<name>' → brew:<name>
• cask "<name>" or cask '<name>' → cask:<name>
• vscode "<publisher>.<ext>" or vscode '<publisher>.<ext>' → vscode:<publisher>.<ext>

Skip comment lines (#) and other directive types (tap, mas, whalebrew).

Keep the parsed brew: set as the Brewfile-declared set. Step 7b reconciles it against actual installed leaves before coverage is computed.

.github/workflows/*.yml / *.yaml: Read each workflow file. Grep for uses: lines:

Grep(pattern="uses:\\s+[^./]", glob=".github/workflows/*.{yml,yaml}", output_mode="content")

From each match, extract the action reference:

• uses: actions/checkout@v4 → action:actions/checkout
• uses: docker://alpine:3.18 → skip (docker:// protocol, not an action)
• uses: ./.github/actions/my-action → skip (local action, ./ prefix)

Strip @version suffix. Deduplicate across all workflow files.

Dockerfile / *.dockerfile / Dockerfile.*: Read each Dockerfile. Extract FROM lines:

Grep(pattern="^FROM\\s+", glob="{Dockerfile,*.dockerfile,Dockerfile.*}", output_mode="content")

From each match, extract the image identifier:

• FROM node:22-alpine → docker:node (strip :tag)
• FROM node:22-alpine AS builder → docker:node (strip AS alias and tag)
• FROM gcr.io/distroless/node → skip (non-Docker-Hub registry)
• FROM ghcr.io/owner/image → skip (non-Docker-Hub registry)
• FROM quay.io/org/image → skip (non-Docker-Hub registry)

Skip registries with a . or / prefix that indicates non-Docker-Hub. Strip version tags (:tag) and AS aliases. Deduplicate across files.

.vscode/extensions.json: Read the file and extract the recommendations array. Each entry is a vscode:<publisher>.<extension-id> identifier. Example:

{ "recommendations": ["esbenp.prettier-vscode", "dbaeumer.vscode-eslint"] }

→ vscode:esbenp.prettier-vscode, vscode:dbaeumer.vscode-eslint

7b. Reconcile Brewfile against brew leaves (ground truth)

The Brewfile is aspirational (what the user declared they want installed) — brew leaves is actual (formulae installed and not pulled in as a dependency of something else). The two diverge in two directions worth surfacing:

• Installed but not Brewfile-declared — silent leaves that snuck in via brew install outside the Brewfile. These are real usage signals worth documenting, even though no manifest declares them.
• Brewfile-declared but not installed — dead declarations the user intends but never realised, or formulae they have since uninstalled.

When to run: only if Step 7 parsed at least one brew:<name> entry from a Brewfile (no Brewfile → nothing to reconcile, skip this step).

Detection: check whether the brew CLI is available and runnable:

Bash: command -v brew >/dev/null 2>&1 && brew leaves 2>/dev/null

If the command succeeds, treat its stdout (one formula name per line) as the installed-leaves set. If it fails (non-zero exit, command not found, or empty output on a non-macOS host), skip the reconciliation silently and proceed to Step 8 with the Brewfile-declared set unchanged — this is the expected fallback when auditing a remote machine's declared deps from a different host.

Why brew leaves and not brew list or the Homebrew MCP:

• brew list includes transitive dependencies — the audit would flood with formulae the user never asked for (e.g., openssl@3, libffi).
• The Homebrew MCP's list endpoint returns plain text without installed_on_request metadata, so leaves cannot be distinguished from transitive deps.
• Per-formula Homebrew MCP info calls would need ~190 invocations at 2-5s each — minutes of latency for one audit. brew leaves is the canonical leaf-finder and returns in ~200ms.

Compute the diff between brewfile_declared and installed_leaves:

• installed_unlisted = installed_leaves − brewfile_declared
• declared_uninstalled = brewfile_declared − installed_leaves
• installed_and_declared = brewfile_declared ∩ installed_leaves

Update the brew set for Step 8: the canonical installed-formulae set used for BM coverage classification becomes installed_leaves ∪ brewfile_declared (union — document anything the user either declared or actually installed as a leaf). The two diff buckets are carried forward for Step 9 to surface in the report.

If brew leaves was unavailable, the set used in Step 8 is just brewfile_declared, and Step 9 reports brew coverage in Brewfile-only mode without the diff buckets.

7c. Detect host-installed sources (--global; plugins + skills today)

When to run: only when invoked with --global. --global audits what is INSTALLED ON THIS MACHINE (vs. what the project declares) against coverage — it reads USER-GLOBAL manifests in $HOME (~/.claude/plugins/*, ~/.agents/.skill-lock.json), not the project. That is why it is opt-in (the result varies by machine and is not CI-reproducible) and why the Step 9 report section is labelled "user-global". Today it covers Claude Code plugins + skills.sh bundles; other host-installed sources (e.g. brew leaves, installed VSCode/gh extensions — bd vp-claude-1u1 et al.) are candidate sub-steps under the same flag. This inverts the default's gh: handling: gh: has no project manifest, but it does have an installed list (gh extension list) — detecting that is a future --global sub-step, not a permanent exclusion.

The resolution (the <name>@<marketplace> → owner/repo join across installed_plugins.json + known_marketplaces.json + each marketplace's marketplace.json, the four source shapes, and skill grouping-by-source) is deterministic, so it lives in a script — not in this prose:

Bash: node ${CLAUDE_PLUGIN_ROOT}/scripts/list-installed-plugins.mjs

It reads no stdin and emits one NDJSON record per installed plugin / skill-bundle:

{"identifier":"plugin:voxpelli/vp-claude#vp-knowledge","title":"plugin-voxpelli-vp-claude-vp-knowledge","installedAt":"…","members":[],"sourceResolved":true}
{"identifier":"skill:basicmachines-co/basic-memory-skills","title":"skill-basicmachines-co-basic-memory-skills","installedAt":"…","members":["memory-notes","memory-research"],"sourceResolved":true}

• identifier is the /intel plugin:/skill: address; title is the pre-normalized BM-note title Step 8 matches on; members is the grouped skill-dir roster (the report's "Skills installed" column); installedAt drives the Step 9 recency cap.
• sourceResolved: false means owner/repo could not be determined (no marketplace.json / unknown source shape) — always Undocumented, shown as "resolve manually".
• A non-zero exit or empty output → skip that population silently and note it (a fresh machine / CI host with no ~/.claude is normal — mirrors the command -v brew fallback in Step 7b). Do NOT abort the audit.

Carry the parsed records forward to Step 8.

8. Check Basic Memory coverage for tools

For each tool type with detected entries, get all documented tools in one call:

list_directory(dir_name="brew", depth=1)
list_directory(dir_name="casks", depth=1)
list_directory(dir_name="actions", depth=1)
list_directory(dir_name="docker", depth=1)
list_directory(dir_name="vscode", depth=1)

Only query directories for tool types that had manifest entries detected. Cross-reference against the parsed identifiers to classify each tool:

• Documented — a <prefix>-<name> note exists
• Undocumented — no dedicated note

When Step 7c ran (--global), get the documented claude_plugin set — matched by NOTE TYPE, not directory, since legacy notes may live outside plugins/:

list_directory(dir_name="plugins", depth=1)
search_notes(query="plugin skill", note_types=["claude_plugin"], page_size=100)

Match each Step 7c record's title against a claude_plugin note. For a miss, fall back to search_notes(query="<bare-name>", note_types=["claude_plugin"], page_size=5) where <bare-name> is the last /- or #-segment of the identifier — this catches legacy-titled notes (e.g. a note titled "beads-marketplace" for an installed beads plugin). Classify Documented / Undocumented as above; sourceResolved:false records are always Undocumented.

9. Add tools section to gap report

Append a tools section to the gap report after the package sections. See ${CLAUDE_PLUGIN_ROOT}/skills/knowledge-gaps/references/report-templates.md ("Tool coverage report") for the full format: one ### <type>: X/Y documented section per tool type (no import-count tiering — group by type, documented vs undocumented), with a Brewfile ↔ installed reconciliation sub-section under Homebrew Formulae only when Step 7b ran (brew leaves available; otherwise the "Brewfile-only mode" note), then a ### Tool Summary.

No import-count tiering for tools — all manifest entries are equally "used". Group by type, show documented vs undocumented count per type.

For the top undocumented tools, offer /intel invocations:

• brew: /intel brew:ripgrep
• cask: /intel cask:warp
• action: /intel action:actions/checkout
• docker: /intel docker:node
• vscode: /intel vscode:esbenp.prettier-vscode

When Step 7c ran, append a Plugin/Skill Coverage section (template in report-templates.md, labelled user-global). The coverage TABLE lists ALL installed plugins/skills (the X/Y documented count needs the full denominator; first dedup the records by title — the same plugin installed from two marketplaces resolves to one title and would otherwise inflate Y), but the actionable /intel OFFER below it is capped to the top 5 undocumented by installedAt (most recent first), with "…and N more — re-run to see all" when truncated. When 0 claude_plugin notes match, lead the section with "No plugin/skill notes yet — here are the 5 most-recently-installed to start" rather than an all-red wall. Offer (use each record's identifier verbatim):

• plugin: /intel plugin:<owner>/<repo> (with #<name> when the record has it)
• skill: /intel skill:<owner>/<repo>

---

10. Detect dead wiki-links

Check the graph for wiki-links referencing non-existent notes — organic documentation debt surfaced by the graph itself.

Quick-exit gate: Check the unresolved count first:

Bash: bm project info main --json | jq '.statistics.total_unresolved_relations'

If the count is 0, skip this step. If the CLI is unavailable, proceed with the search.

Query the relation index for each ecosystem detected in Steps 0–9. Use entity_types=["relation"] to search the relation index directly — this returns relation objects with from_entity and to_entity fields:

search_notes(query="npm-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="crate-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="go-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="composer-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="pypi-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="gem-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="brew-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="action-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="docker-", entity_types=["relation"], output_format="json", page_size=50)
search_notes(query="vscode-", entity_types=["relation"], output_format="json", page_size=50)

Only query prefixes for ecosystems detected in Steps 0–9.

Identify unresolved relations: For each result, check whether to_entity is present in the JSON response. Relations missing to_entity are unresolved — the wiki-link target has no corresponding note.

Extract the target name from the relation title (format: "source → target") or matched_chunk. Cross-reference against the list_directory results from Steps 2 and 8 to confirm.

Deduplicate: If a dead-linked package already appears in Tier 1/2/3 from manifest parsing, add "(also wiki-linked)" annotation rather than listing it twice.

Recency-scoped sweep (large-directory blind spot): the relevance-ranked query above is biased toward old, frequently-linked notes — in a large, active ecosystem directory (e.g. brew/, ~190 notes) the top 50 relevance-ranked rows can be entirely old resolved links, never reaching notes created that same session. Confirmed empirically: a brew- prefix relation query returned only resolved old links, while 8 brew-* notes created that day carried 11 dangling edges that were only found by reading those notes directly — the relevance sweep reported a false-clean for the largest, most-active directory in the graph. Run this sweep in addition to the relevance-ranked query above, not instead of it:

1. Get recently-active ecosystem/tool notes in one call:

   recent_activity(timeframe="7d", type="entity", output_format="json")
   

2. Filter the result to titles matching a detected ecosystem/tool prefix (npm-, crate-, go-, composer-, pypi-, gem-, brew-, action-, docker-, vscode-) — the same prefixes queried above.
3. For each surviving note, read_note it and extract every [[Target]] wiki-link in its ## Relations section.
4. For each target that itself matches an ecosystem/tool prefix, check it against the list_directory results already gathered in Steps 2 and 8 — a target absent from that title set is a dangling edge the relevance sweep may have missed.

This is O(recently-active notes), not O(corpus) — it stays cheap because it is bounded by the 7-day window rather than directory size, and it is recency-complete regardless of how large or old-link-dominated the directory is. Merge findings from both sweeps into one table, deduplicating by target.

Add dead-link findings to the gap report:

#### Referenced but not documented (dead wiki-links)
| Link | Referenced in | Detected via |
|------|--------------|--------------|
| [[npm-some-pkg]] | npm-fastify, engineering/patterns/http | relevance |
| [[brew-some-tool]] | brew-newly-added | recency |

Add dead-link counts to the Overall Summary:

- Dead wiki-links: Q (across R unique notes)

When offering enrichment (Steps 5, 9), include dead-link targets annotated with "(wiki-linked in N notes)" to show organic graph momentum.

Limitation: page_size=50 per prefix on the relevance-ranked query is a sample, not exhaustive — the graph may have more unresolved relations than one page returns, and in a large directory the sample skews toward old, frequently-linked notes rather than recent ones (see the recency-scoped sweep above, which exists specifically to cover that blind spot). Even with both sweeps combined this remains a bounded gap-detection pass, not a full audit — a note edited outside the 7-day window whose dangling edge also falls outside the top-50 relevance-ranked sample is still missed by design (the gardener handles comprehensive auditing). The highest-scored results surface the most commonly referenced dead links; the recency sweep surfaces the most recently introduced ones.

---

11–13. Domain standard detection

Read the standard detection reference file for Steps 11–13: ${CLAUDE_PLUGIN_ROOT}/skills/knowledge-gaps/references/standard-detection.md

This covers detecting domain standards in Basic Memory, classifying them by codebase reference count, and adding a standards section to the gap report.

---

14–15. Concept-level gap detection

Read the concept detection reference file for Steps 14–15: ${CLAUDE_PLUGIN_ROOT}/skills/knowledge-gaps/references/concept-detection.md

This covers mining the relation graph for implicit hub gaps, checking Readwise for reading signals, and adding a concept coverage section to the gap report.