Audit agent setup: what a repo's Claude Code configuration lacks
A generic recommender reads a codebase and proposes popular servers and
format hooks. On this machine the faults that cost time sit elsewhere: a
plugin switched off on purpose and proposed again, a .mcp.json URL
still holding a template placeholder that fails every session start, a
server at user scope that belongs to one tenant, an az tenant pin no
MCP server honours, a CLAUDE.md rule that nothing enforces. This skill
checks those first, then recommends from the code, reuse first. It
reports and stops: nothing is edited, written or connected.
Its neighbours own the rest. Where a line of guidance lives (root,
a rule, a nested file, a skill) is scope-instructions'. Install
health and unused extensions against their context cost are
/doctor's. Allowlists drawn from past sessions are
/fewer-permission-prompts'. Applying a recommendation is
/update-config, /permissions, /plugin or /mcp, on the user's
request.
Past references/, the live page: code.claude.com by the
claude-code-docs MCP server or
curl -s https://code.claude.com/docs/en/<page>.md, and another tool's
line, as husky's, by that tool's own markdown source, else WebFetch,
saying it read a summary. A live page that contradicts a line here wins:
say so, act on the page, and route the fix by capture-learning, bar
a dated measured line that names the page it overrules and the call that
measured it, which a page changed since its date sends back to re-measure.
An observed claim is cited or marked pending a doc check, and stands, marked
observed, only where no page covers it.
0. Whose setup
git rev-parse --show-toplevel
git remote get-url origin
printenv CLAUDE_CONFIG_DIR
printenv CLAUDE_CODE_PLUGIN_CACHE_DIR
- The account is a config directory.
CLAUDE_CONFIG_DIR, else
~/.claude. Its .claude.json sits inside it when the variable is
set, else at ~/.claude.json (measured on 2.1.293, 2026-10-16). Its
plugins root is CLAUDE_CODE_PLUGIN_CACHE_DIR, else <config>/plugins
(plugins loading reference). A work-root session and a personal one
read different directories, so name the one read in the report.
- The config source is the repo the user's own config is authored
in, where they keep one: a repo holding
claude/CLAUDE.md, the
source of their user-scope instructions, at its root. There a payload
recommendation is a proposed edit in this tree; anywhere else it goes
through capture-learning (step 5) where it is installed, and with
neither the report names it for the user to place.
- Not a git repo: audit the account alone, and say so.
- An argument narrows the run to that category's findings and
recommendations, three to five of them instead of one or two.
1. Read the repo's setup, with Read
Read each that exists, one Read per file:
.claude/settings.json, .claude/settings.local.json, .mcp.json
.claude/launch.json, the desktop app's Browser pane config (3.6)
- root
CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md
.pre-commit-config.yaml, .gitignore
package.json, for its scripts and a lint-staged key, and each
hook file in .husky/
.gitignore names what no Glob finds. A credentials file kept
outside the checkout, or ignored in it, may show only as a pattern
there: one run's strongest finding, a vendor credentials file holding a
client secret, came from its extension in .gitignore alone (observed
once, 2026-10-21). Count each credential-shaped pattern as the
catalog's credentials-file signal.
A Read of .mcp.json or a settings file loads the user-scope rule
claude-config-scoping.md by its glob, and a Read of a CLAUDE.md
loads agent-instructions-scoping.md; this skill applies their tables
and does not restate them. A pipe or a cat of several files loads
neither.
List .claude/agents/, .claude/skills/ and .claude/rules/ by Glob;
Read one only when a finding turns on it. Then whether each config file
is shared:
git ls-files -- .mcp.json .claude/settings.json .claude/settings.local.json CLAUDE.local.md .claude/launch.json
git check-ignore -v -- .mcp.json .claude/settings.local.json CLAUDE.local.md .claude/launch.json
A commit gate is configured, armed and passing: three facts, read
apart. Configured is a file: .pre-commit-config.yaml, a .husky/
folder, a lint-staged key, or a CI step running the tests or a linter.
Armed is a hook where git runs them:
git rev-parse --git-path hooks
It prints .git/hooks with core.hooksPath unset, and .husky/_ once
husky has set it (git 2.55, 2026-10-28). Read pre-commit in that
folder, from the repo root: missing, or only pre-commit.sample, the
gate is configured and not armed. Passing, this skill never measures
(3.5). Configured and not armed is a finding to explain, its likeliest
cause a baseline that fails, not a one-command fix.
Codebase signals by Glob, not a chain of ls: manifests
(pyproject.toml, package.json, *.csproj, *.sqlproj), CI
(.github/workflows/*, azure-pipelines.yml), infrastructure
(**/*.bicep, azure.yaml), deploy scripts (**/deploy*.ps1,
**/deploy*.sh), Fabric items (**/.platform), PBIP (**/*.pbip),
a Salesforce DX project (sfdx-project.json, .forceignore,
force-app/), .env* files, an MCPB manifest.json, and format or
lint configs.
references/catalog.md says which signal leads where.
2. Read the account's user scope
<config>/settings.json: enabledPlugins, hooks, permissions.
Names only in <config>/agents/ and <config>/skills/: what is
deployed, for reuse.
User-scope servers, one key of the account's .claude.json:
jq -b -r '.mcpServers // {} | keys[]' <file>. Never Read the whole
file: it is Claude Code's runtime state, the OAuth account and every
project's history beside that one key.
Plugins, by the bundled script, never by hand:
uv run --no-project ${CLAUDE_SKILL_DIR}/scripts/plugin-inventory.py
It reads enabledPlugins from the user, project and local settings,
and under the plugins root the install records, cache/, each
synced/<bucket>/ folder (the claude.ai account's, loaded as
<name>@synced with no install record) and every marketplace clone's
list. Per plugin it prints which settings file wins, the stage, the
<plugin>:<skill> names, and each server's url or command with its
header and env keys, never their values. Every run before it wrote
that read by hand, and three got it wrong (2026-10-21): a synced
folder is named by its own .claude-plugin/plugin.json, the name
enabledPlugins uses, not by the bucket's manifest.json, whose
name is the folder's, often a UUID. Exit 2 says a file's shape
changed: read that one by hand and call the inventory partial.
--plugin <name> looks a plugin up in the marketplaces, for step 4.
claude.ai connectors are in no file. Read this session's own tool
list for connector servers and report their tool families; their
endpoints cannot be read from files. A desktop session's
mcp__ccd_connectors__session_connectors_status lists each server's
name, kind, status and tool count, and a connector's id, which is its
tool prefix, and connects nothing (2.1.304, more than one session,
2026-10-21). Its tool count is the server's own, which a
permissions.deny leaves as it was: read a deny's effect from the
session's tool list (2026-10-22). There they run in-process,
type: "sdk" servers that no
MCP setting reaches (MCP docs § "How connectors reach Claude Code",
read 2026-10-21).
The desktop app's own config is a third user-scope source. Its
%APPDATA%\Claude\claude_desktop_config.json mcpServers load in
every local Code-tab session beside ~/.claude.json and .mcp.json,
and the CLI reads none of them (Desktop docs, read 2026-11-14). The
status tool above lists each as kind: "desktop", and one named
github-mcp was the only github-mcp a session listed in a repo
whose .mcp.json enables its own (observed 2026-11-14, 2.1.321).
Read each entry's key, command or url, and env key names by
jq -b, never a value.
Where Read refuses a path outside the working directory, under
permissions.blockReadsOutsideWorkingDirectories, read the one key with
jq -b, which prompts rather than refuses (settings reference).
Connect nothing. No claude mcp list or get, /mcp,
claude plugin list, az, headersHelper run or call to a project
server's tool. claude mcp list health-checks every server it lists,
which runs each project helper, and an az-based helper takes its token
from whichever login the shared ~/.azure holds: from a client repo,
that is a call into another tenant, the hazard of check 4. The files
answer every question here.
The harness may have connected already. In a repo whose .mcp.json
servers are approved, Claude Code connects them at session start, so
their helpers ran before step 0 (observed once, 2026-10-21). Nothing
here can stop that: where check 4 fires, say a cross-tenant call at
launch is the launch's, not this audit's.
3. Findings
3.1 Plugins declared off
For each enabledPlugins entry set false in user, project or local
settings (the highest source naming the id wins: managed, then
--settings, local, project, user), read from step 2's inventory, the
report carries one line, though none of it is a fault (step 6):
- Its stage: installed, cached and not installed, synced, or not
fetched. Then whether this session's own skill listing carries the
plugin's
<name>: skills, which the inventory names: a false is
what the CLI reads, not what a running session loaded. With
code-review@synced set false, a desktop session three hours after
the deploy still listed its code-review: skill, while
claude plugin list on the same config directory printed it disabled
(2.1.307, 2026-10-21).
- Whose decision: an id the payload's
claude/settings.json sets
false was switched off on purpose, for its listing cost (2026-08-31),
since skillOverrides cannot reach a plugin's skills. Never
recommend installing a plugin that is installed and off. If it earns
its place in this repo, recommend enabling it here only, in
.claude/settings.local.json or .claude/settings.json, which
outrank the user false, against that reason.
- A plugin suggestion that never appeared has several documented
causes: the plugin is installed; a relevance suggestion comes only
from a marketplace an administrator lists in
pluginSuggestionMarketplaces; a CLI hint comes once ever per plugin,
and never with analytics off. Name the ones that fit; assert none.
3.2 .mcp.json hygiene
For each server:
- A placeholder still in place: a
<...> segment in url, args
or headers, as the project template ships them (<WorkspaceId>,
<GITHUB_PAT_VAR>; the user's MCP README, <config>/mcp/README.md
where their config holds one, lists them). It fails at
every session start. An empty url is not this: Claude Code lists it
not configured and never dials it.
- A header holding a literal value rather than a
${VAR} or a
headersHelper: a credential in a file. Name the key, never the
value. Tracked and pushed, it is recreate-repo's case; otherwise
move it to a variable named per account, with no generic fallback.
- A
${VAR} this session lacks: printenv NAME >/dev/null && echo set || echo unset, never printing the value. Unset with no
:-default, Claude Code warns and sends the text unexpanded. A
credential name such as ANTHROPIC_API_KEY or NPM_TOKEN reads as
empty in a remote url or headers whether set or not (MCP docs).
VS Code reads .mcp.json too and expands no ${VAR} at all
(1.140.0, 2026-10-03).
- Tracked or not: tracked, every clone gets the servers and is asked
to approve each; untracked or ignored, clones get none. Either can be
right: say which, and whether a server carries per-person values.
- A
timeout on the wrong transport. A hosted Fabric http
server with none fails any call past about 60 seconds with
The operation timed out.; on a stdio server the value is a hard
wall-clock cap on each call. The project template sets 120000 on the
measured hosted entries and none on stdio, and the user's MCP README,
where one exists, has why in § "Per-server timeout". One
.mcp.json copied from another project had it the other way round
(2026-10-21).
- Overlap: the same name at local or user scope with another
endpoint (Claude Code warns; sign-ins are per endpoint), or the same
endpoint as a plugin's server, a synced plugin's included, or a
claude.ai connector. Plugins and connectors are matched by endpoint
and the project server hides them (MCP docs, scope precedence).
Recommend keeping one. A connector is hidden so only where Claude Code
fetches connectors itself, in terminal, IDE and SDK sessions; a
desktop session's run in-process (step 2).
- Permission rules for a server's tools go in
.claude/settings.json,
not beside the server.
3.3 Scope
User scope holds what is bound to nothing: on this machine, the two
docs servers microsoft-learn-mcp and claude-code-docs. A server
that needs a workspace, a database, a connection string, a desktop
app, a GitHub account or a tenant belongs in this repo's .mcp.json.
A tool absent from user scope is not a gap; a recommendation to put
one there is the finding.
The same test places a hook: one every repo needs is payload, one for
this repo's tools is the repo's. A claude.ai connector (step 2) reaches
every repo its account opens, as user scope does, so a tenant-bound one
meets the same test; its fix is on claude.ai, in no file. So does a
server in claude_desktop_config.json (step 2), whose fix is deleting
its entry by hand, since no deploy reconciles that file.
3.4 Tenant pins that MCP servers ignore
When the repo pins az to a tenant (a script or hook setting
AZURE_CONFIG_DIR, a guard reading az account show, a CLAUDE.md
line naming the tenant) and also declares a tenant-bound server
(azure-mcp, a Fabric or Power BI headersHelper server,
microsoft-fabric-mcp, fabric-rti-mcp), the pin does not cover the
server. Servers and their helpers are spawned by Claude Code and
inherit its environment, so an az-based one reads the shared
~/.azure: the user-scope CLAUDE.md says MCP servers ignore
AZURE_CONFIG_DIR, and the user's MCP README, where one exists,
§ "The helper's login is the harness's, not the folder's", holds the
one relayed measurement (2026-09-22). Report it with that section's
three candidate remedies, all unrun, the third a stdio server's own
env, or with no such file that one alone: measure one before relying
on it. Never confirm the finding by calling the server.
The launch may already carry a pin, so read the harness's own value
before wording the finding: $env:AZURE_CONFIG_DIR in the PowerShell
tool, which runs -NoProfile, or printenv AZURE_CONFIG_DIR in Bash
only when /proc/$$/cmdline shows a snapshot (bash -c source …). A
login-mode Bash reruns the profile, which sets it from the folder
whatever the harness has. Set, report the servers covered by the
launch, not by the repo, so a session started another way has nothing
pinning them; unset, report as above. Whether a project helper keeps
it past Claude Code's credential stripping is unprobed. The same README
section has the measurements (2026-10-21).
3.5 CLAUDE.md rules nothing enforces
For each line in the files read in step 1 that says something must or
must not happen, look for what enforces it in the settings, hooks,
pre-commit config and tests. CLAUDE.md is "context, not enforced
configuration", while "permission rules are enforced by Claude Code,
not by the model" (memory and permissions docs). Where nothing does:
| The line says |
Enforce with |
Say also |
| ask before X |
permissions.ask with X's rule |
prompts even in acceptEdits and bypassPermissions; dontAsk denies instead |
| never read or edit F |
permissions.deny with Read(...) or Edit(...) |
covers the file tools and Bash file commands Claude Code recognizes, not grep -r or a script opening files; the sandbox is the OS-level answer |
| never run C, never print secrets |
a PreToolUse hook that exits 2 |
a failing hook lets the call through unless it sets "onFailure": "block" (2.1.295+); a Bash rule matches command text only |
| only the user starts X |
permissions.ask naming X's script, a rule per shell |
disable-model-invocation: true stops Claude loading a skill, not running its script or the command it wraps (skills docs) |
| run X before committing |
a git hook in .pre-commit-config.yaml |
binds every committer, not only Claude |
| change A with B and C, keep these in step |
a test comparing them, run in CI |
binds every change, not only Claude's; a value that may differ on purpose goes in the test's exception list, with its reason (2026-10-22) |
A gate counts only once it passes. The last two rows, and the
catalog's Hooks rows, count a gate step 1 found, or propose one, only
over a baseline that passes: failing, it blocks every commit touching a
failing file, or rewrites each whole by a --write or --fix step. One
client run recommended npm install to arm a husky gate configured and
not armed; once installed, its ESLint, Prettier and Jest steps failed
across most of the repo (2026-10-28). This skill never runs a gate,
which would run the repo's code and its formatters:
- Unmeasured, say so, and make running the gate in check mode the
first Recommended action, ahead of any arm:
prettier --check,
eslint without --fix, the test script. Install by npm ci, not
npm install, with HUSKY=0 set: npm ci runs the prepare script
too, which arms husky (npm scripts docs; HUSKY=0 is husky's own
switch for an install that must not, 2026-10-28).
- Measured failing, by the user or a CI log, report its counts; the
fix is a baseline cleanup, or scoping the gate to changed files, never
arming it as it is. So too a CI step proposed to run a script the repo
already has.
A rule on command text has twin routes. Check a "never run" or
"never print" line against each, and answer each with a matcher or with
permissions.ask or deny:
- Another command runner. The desktop app's
mcp__terminal__run_in_terminal runs commands no Bash|PowerShell
matcher sees and no permission rule reaches: Claude Code skips an
mcp__ rule with parentheses, and a bare one asks on every command
(permissions docs § "Match by input parameter", read 2026-11-14). So
answer a Bash(...) or PowerShell(...) rule there with a
PreToolUse hook on the tool that returns permissionDecision ask
or deny; a matcher naming the tool fired in a live Desktop session
(observed 2026-11-14, 2.1.321).
- A declared server's tools. Azure MCP's
appservice webapp settings get-appsettings is annotated Secret, its
values able to include connection strings (Learn, 2026-10-21): a rule
on mcp__<server>__<tool>. In azure-mcp's default namespace mode
that is mcp__azure-mcp__appservice, one tool for every App Service
operation, so the rule denies its writes too (read from a session's
tool list 2026-10-21).
- A second CLI for the same API. Az PowerShell's
Get-AzStorageAccountKey and Get-AzFunctionAppSetting pass a guard
matching az as a word, and ignore AZURE_CONFIG_DIR, so a
tenant-pin rule has the same hole (a guard fed synthetic payloads,
2026-10-21): widen the pattern to the second CLI.
- The same CLI's REST verb.
az rest, gh api and fab api take
the operation as a URL, so its name arrives as a path segment:
az rest against a site's config/appsettings/list or
config/connectionstrings/list, or a storage account's listKeys,
passes a guard matching appsettings list or keys list (a client
guard fed synthetic payloads, 2026-10-21; its patterns re-run on the
URLs 2026-10-22). Match the path segments too, or deny the REST verb
where the repo has no use for it.
Whether the line then stays as the reason, moves or goes is
scope-instructions' question, as is a side-effecting workflow
written as prose: name it, do not answer it.
3.6 .claude/launch.json
The desktop app's Browser pane config: one entry in configurations
per dev server it starts, or per URL it opens (Desktop § "Configure
preview servers", read 2026-10-29). It takes JSON with comments, so
Read it rather than jq it. Each check here is from the docs' text,
and none has yet met a broken file. With no file, the line says so, and
the catalog's dev-server row weighs proposing one.
- Tracked or ignored, by step 1's two
git lines. The docs treat
the file as committed: env reads "Don't put secrets here since this
file is committed to your repo". So ignored is the finding: a clone
has no entries until the app detects a dev server on its first
preview. An allowlisting .gitignore (/.claude/*, then a ! line
per file) ignores a new one silently, exit 0, as the config source's
did until it gained a line (2026-10-29). Which copy a desktop-app
worktree session reads is undocumented; one relayed note found a
worktree's own copy unread (2026-10-20).
- A literal value in an entry's
env that reads as a credential:
name the key, never the value, as 3.2 does for a header. Tracked and
pushed, it is recreate-repo's case; otherwise it moves to a user
environment variable on Windows, or the local environment editor,
then a Desktop restart (§ "Set environment variables for dev
servers"). A dev server gets neither ~/.claude/settings.json's env
nor a saved environment's.
- A
url the app refuses. One on localhost, a *.localhost
subdomain, 127.0.0.1 or ::1 must be the server's origin alone, its
port the entry's port, 3000 when unset: a path, a query or another
port is reported as a configuration error. Any url must be http
or https, with no username or password. An external one may carry a
path, and asks permission the first time the preview opens it.
autoVerify and autoPort are no finding on their own.
4. Recommendations
Reuse first. Before anything new, match the need against what step
2 found deployed: the user-scope agents and skills,
microsoft-learn-mcp, the payload's hooks, and the built-in skills. A
need one of them meets is a "use X" row, or no row.
Then Read references/catalog.md and match step
1's signals. Keep one or two per category, the strongest signal first,
or three to five in the category an argument names; skip a category
with nothing. Every item names the file or signal behind it.
Name only what exists here: a plugin the inventory's --plugin
lookup finds in a marketplace's list, a server from the project
template the MCP README describes, or one Learn documents through
microsoft-learn-mcp. Found nowhere, it is not recommended; say what
was looked for.
A copied skill goes only if its group can lose it. Its description
saying it does not apply here is not enough. copy-skills.ps1 selects
whole groups, so a skill in .claude/skills/.managed-skills.json whose
group under the payload's skills/ holds one the repo keeps cannot be
dropped by that script, and a hand delete is planned as a copy on the
next refresh (read from the script, not run; 2026-11-01). Grep the kept
skills for its name too, and read each hit as a dependency or only a
boundary. Where either stops it, or the payload cannot be read, recommend
no removal and say why.
5. Place each item
| Lands in |
When |
How it gets there |
| the repo |
bound to its tools, files or tenant |
.claude/settings.json or .mcp.json for everyone, .claude/settings.local.json for one person, .claude/agents/, .claude/skills/, .pre-commit-config.yaml |
| the payload |
every repo needs it, or a payload file is wrong |
in the config source, a proposed edit; elsewhere capture-learning where installed, which writes the inbox note and cites a client's estate by kind; with neither, named for the user |
| nowhere |
reuse covers it, the code makes it derivable, or a payload decision rules it out |
say which |
6. Report
Plain markdown, no emoji headings, about a screen:
## Profile the account read, the commit read (git rev-parse --short HEAD), the stack, the signals that drove the report
## Findings 3.1 to 3.6, a line each: the file, what is wrong, the fix, where it lands
## Recommendations by category, one or two each: what, the signal, where it lands
## Next steps in other skills what scope-instructions, /doctor or /fewer-permission-prompts owns here, for the user to run
## Recommended actions numbered: each fix or recommendation above that calls for an edit, and where it lands
Findings gives every check a line, a clean one included, since a
missing line reads the same as a check not run: one run gave 3.4 none,
having read no AZURE_CONFIG_DIR (2026-10-21). It also gives each
plugin set false its line from 3.1: stage, whether the session
lists its skills, and whose decision. One run read every stage, then
reported only that the ids matched the payload (2026-10-21).
Recommended actions ends the report, numbered 1., 2., … across
Findings and Recommendations: one line per finding fix or
recommendation that calls for an edit, naming the finding or category
it comes from and where it lands, as 1. 3.5.1, the inbox-delete rule: a PowerShell twin of its ask rule, in the payload's claude/settings.json. A clean check, a 3.1 line, a finding the report
says to leave, a category with nothing new and an item under Next steps
in other skills get none. With nothing to edit, it reads _(none)_. It is the list a
handoff transcribes: in the config source, file-audit-findings, where
it has one, briefs each item without choosing among the findings, and
elsewhere nothing reads it.
A gate's check-mode run comes first, where 3.5 calls for one: it is
1., though it edits nothing, ahead of any fix that would arm the gate.
Leave out any other empty section. Then stop: applying a row is a
separate request.
Constraints
- Edits nothing and writes nothing, an inbox note included: that is
capture-learning's, run by the user.
- Connects nothing: no
claude mcp, claude plugin or az
command, and no tool of a project server.
- Runs no gate: no test, linter or formatter of the repo's, in check
mode or not, and no install that would arm one (3.5).
- Prints no secret: a literal header or a variable's value is named
by its key, never shown.
- Recommends nothing at user scope that is bound to a workspace, an
account or a tenant.
- A plugin the payload switched off stays a decision: weigh its
reason; never recommend installing it again.
- A client's evidence stays in its repo: a payload row from a client
repo reaches the config source only through
capture-learning or the
user, cited by kind.
- Not this skill: where guidance lives (
scope-instructions),
install health and context cost (/doctor), allowlists from past
sessions (/fewer-permission-prompts).
1---2name: audit-agent-setup3model: inherit4effort: xhigh5disable-model-invocation: true6argument-hint: "[hooks|permissions|skills|agents|plugins|mcp]"7allowed-tools: Read Grep Glob Bash(git rev-parse *) Bash(git ls-files *) Bash(git check-ignore *) Bash(git remote *) Bash(uv run --no-project ${CLAUDE_SKILL_DIR}/scripts/plugin-inventory.py *)8description: "Audits a repo's Claude Code setup on this machine and reports what to fix and what to add: hooks, permission rules, skills, subagents, plugins and MCP servers. Edits, writes and connects nothing. Runs the checks a generic recommender cannot: plugins declared off and the payload's reason for it, .mcp.json placeholders, literal header values and tracking, servers at the wrong scope, tenant pins that MCP servers ignore, and CLAUDE.md rules nothing enforces. Reuses the deployed agents and skills and microsoft-learn-mcp before proposing anything new, names scope-instructions, /doctor and /fewer-permission-prompts where one owns the question, and places each item in the repo, the repo the user's own config is authored in (through capture-learning from another repo) or nowhere. Typed only: audit-agent-setup [category]."9metadata:10 source: https://skillmd.com/skills/wardawgmalvicious/audit-agent-setup11---1213# Audit agent setup: what a repo's Claude Code configuration lacks1415A generic recommender reads a codebase and proposes popular servers and16format hooks. On this machine the faults that cost time sit elsewhere: a17plugin switched off on purpose and proposed again, a `.mcp.json` URL18still holding a template placeholder that fails every session start, a19server at user scope that belongs to one tenant, an `az` tenant pin no20MCP server honours, a `CLAUDE.md` rule that nothing enforces. This skill21checks those first, then recommends from the code, reuse first. It22**reports and stops**: nothing is edited, written or connected.2324Its neighbours own the rest. **Where a line of guidance lives** (root,25a rule, a nested file, a skill) is `scope-instructions`'. **Install26health and unused extensions against their context cost** are27`/doctor`'s. **Allowlists drawn from past sessions** are28`/fewer-permission-prompts`'. Applying a recommendation is29`/update-config`, `/permissions`, `/plugin` or `/mcp`, on the user's30request.3132Past `references/`, the live page: code.claude.com by the33`claude-code-docs` MCP server or34`curl -s https://code.claude.com/docs/en/<page>.md`, and another tool's35line, as husky's, by that tool's own markdown source, else `WebFetch`,36saying it read a summary. A live page that contradicts a line here wins:37say so, act on the page, and route the fix by `capture-learning`, bar38a dated measured line that names the page it overrules and the call that39measured it, which a page changed since its date sends back to re-measure.40An observed claim is cited or marked pending a doc check, and stands, marked41observed, only where no page covers it.4243## 0. Whose setup4445```bash46git rev-parse --show-toplevel47git remote get-url origin48printenv CLAUDE_CONFIG_DIR49printenv CLAUDE_CODE_PLUGIN_CACHE_DIR50```5152- **The account is a config directory.** `CLAUDE_CONFIG_DIR`, else53 `~/.claude`. Its `.claude.json` sits inside it when the variable is54 set, else at `~/.claude.json` (measured on 2.1.293, 2026-10-16). Its55 plugins root is `CLAUDE_CODE_PLUGIN_CACHE_DIR`, else `<config>/plugins`56 (plugins loading reference). A work-root session and a personal one57 read different directories, so name the one read in the report.58- **The config source is the repo the user's own config is authored59 in**, where they keep one: a repo holding `claude/CLAUDE.md`, the60 source of their user-scope instructions, at its root. There a payload61 recommendation is a proposed edit in this tree; anywhere else it goes62 through `capture-learning` (step 5) where it is installed, and with63 neither the report names it for the user to place.64- **Not a git repo**: audit the account alone, and say so.65- **An argument narrows the run** to that category's findings and66 recommendations, three to five of them instead of one or two.6768## 1. Read the repo's setup, with Read6970Read each that exists, **one Read per file**:7172- `.claude/settings.json`, `.claude/settings.local.json`, `.mcp.json`73- `.claude/launch.json`, the desktop app's Browser pane config (3.6)74- root `CLAUDE.md`, `.claude/CLAUDE.md`, `CLAUDE.local.md`75- `.pre-commit-config.yaml`, `.gitignore`76- `package.json`, for its `scripts` and a `lint-staged` key, and each77 hook file in `.husky/`7879**`.gitignore` names what no Glob finds.** A credentials file kept80outside the checkout, or ignored in it, may show only as a pattern81there: one run's strongest finding, a vendor credentials file holding a82client secret, came from its extension in `.gitignore` alone (observed83once, 2026-10-21). Count each credential-shaped pattern as the84catalog's credentials-file signal.8586A Read of `.mcp.json` or a settings file loads the user-scope rule87`claude-config-scoping.md` by its glob, and a Read of a `CLAUDE.md`88loads `agent-instructions-scoping.md`; this skill applies their tables89and does not restate them. A pipe or a `cat` of several files loads90neither.9192List `.claude/agents/`, `.claude/skills/` and `.claude/rules/` by Glob;93Read one only when a finding turns on it. Then whether each config file94is shared:9596```bash97git ls-files -- .mcp.json .claude/settings.json .claude/settings.local.json CLAUDE.local.md .claude/launch.json98git check-ignore -v -- .mcp.json .claude/settings.local.json CLAUDE.local.md .claude/launch.json99```100101**A commit gate is configured, armed and passing: three facts, read102apart.** Configured is a file: `.pre-commit-config.yaml`, a `.husky/`103folder, a `lint-staged` key, or a CI step running the tests or a linter.104Armed is a hook where git runs them:105106```bash107git rev-parse --git-path hooks108```109110It prints `.git/hooks` with `core.hooksPath` unset, and `.husky/_` once111husky has set it (git 2.55, 2026-10-28). Read `pre-commit` in that112folder, from the repo root: missing, or only `pre-commit.sample`, the113gate is configured and not armed. Passing, this skill never measures114(3.5). Configured and not armed is a finding to explain, its likeliest115cause a baseline that fails, not a one-command fix.116117**Codebase signals by Glob**, not a chain of `ls`: manifests118(`pyproject.toml`, `package.json`, `*.csproj`, `*.sqlproj`), CI119(`.github/workflows/*`, `azure-pipelines.yml`), infrastructure120(`**/*.bicep`, `azure.yaml`), deploy scripts (`**/deploy*.ps1`,121`**/deploy*.sh`), Fabric items (`**/.platform`), PBIP (`**/*.pbip`),122a Salesforce DX project (`sfdx-project.json`, `.forceignore`,123`force-app/`), `.env*` files, an MCPB `manifest.json`, and format or124lint configs.125`references/catalog.md` says which signal leads where.126127## 2. Read the account's user scope128129- `<config>/settings.json`: `enabledPlugins`, `hooks`, `permissions`.130- Names only in `<config>/agents/` and `<config>/skills/`: what is131 deployed, for reuse.132- User-scope servers, one key of the account's `.claude.json`:133 `jq -b -r '.mcpServers // {} | keys[]' <file>`. **Never Read the whole134 file**: it is Claude Code's runtime state, the OAuth account and every135 project's history beside that one key.136- Plugins, by the bundled script, never by hand:137138 ```bash139 uv run --no-project ${CLAUDE_SKILL_DIR}/scripts/plugin-inventory.py140 ```141142 It reads `enabledPlugins` from the user, project and local settings,143 and under the plugins root the install records, `cache/`, each144 `synced/<bucket>/` folder (the claude.ai account's, loaded as145 `<name>@synced` with no install record) and every marketplace clone's146 list. Per plugin it prints which settings file wins, the stage, the147 `<plugin>:<skill>` names, and each server's url or command with its148 header and env keys, never their values. Every run before it wrote149 that read by hand, and three got it wrong (2026-10-21): **a synced150 folder is named by its own `.claude-plugin/plugin.json`**, the `name`151 `enabledPlugins` uses, not by the bucket's `manifest.json`, whose152 `name` is the folder's, often a UUID. Exit 2 says a file's shape153 changed: read that one by hand and call the inventory partial.154 `--plugin <name>` looks a plugin up in the marketplaces, for step 4.155- **claude.ai connectors are in no file.** Read this session's own tool156 list for connector servers and report their tool families; their157 endpoints cannot be read from files. A desktop session's158 `mcp__ccd_connectors__session_connectors_status` lists each server's159 name, kind, status and tool count, and a connector's id, which is its160 tool prefix, and connects nothing (2.1.304, more than one session,161 2026-10-21). Its tool count is the server's own, which a162 `permissions.deny` leaves as it was: read a deny's effect from the163 session's tool list (2026-10-22). There they run in-process,164 `type: "sdk"` servers that no165 MCP setting reaches (MCP docs § "How connectors reach Claude Code",166 read 2026-10-21).167- **The desktop app's own config is a third user-scope source.** Its168 `%APPDATA%\Claude\claude_desktop_config.json` `mcpServers` load in169 every local Code-tab session beside `~/.claude.json` and `.mcp.json`,170 and the CLI reads none of them (Desktop docs, read 2026-11-14). The171 status tool above lists each as `kind: "desktop"`, and one named172 `github-mcp` was the only `github-mcp` a session listed in a repo173 whose `.mcp.json` enables its own (observed 2026-11-14, 2.1.321).174 Read each entry's key, `command` or `url`, and `env` key names by175 `jq -b`, never a value.176177**Where Read refuses a path outside the working directory**, under178`permissions.blockReadsOutsideWorkingDirectories`, read the one key with179`jq -b`, which prompts rather than refuses (settings reference).180181**Connect nothing.** No `claude mcp list` or `get`, `/mcp`,182`claude plugin list`, `az`, `headersHelper` run or call to a project183server's tool. `claude mcp list` health-checks every server it lists,184which runs each project helper, and an `az`-based helper takes its token185from whichever login the shared `~/.azure` holds: from a client repo,186that is a call into another tenant, the hazard of check 4. The files187answer every question here.188189**The harness may have connected already.** In a repo whose `.mcp.json`190servers are approved, Claude Code connects them at session start, so191their helpers ran before step 0 (observed once, 2026-10-21). Nothing192here can stop that: where check 4 fires, say a cross-tenant call at193launch is the launch's, not this audit's.194195## 3. Findings196197### 3.1 Plugins declared off198199For each `enabledPlugins` entry set `false` in user, project or local200settings (the highest source naming the id wins: managed, then201`--settings`, local, project, user), read from step 2's inventory, the202report carries one line, though none of it is a fault (step 6):203204- **Its stage**: installed, cached and not installed, synced, or not205 fetched. **Then whether this session's own skill listing carries the206 plugin's `<name>:` skills**, which the inventory names: a `false` is207 what the CLI reads, not what a running session loaded. With208 `code-review@synced` set `false`, a desktop session three hours after209 the deploy still listed its `code-review:` skill, while210 `claude plugin list` on the same config directory printed it disabled211 (2.1.307, 2026-10-21).212- **Whose decision**: an id the payload's `claude/settings.json` sets213 `false` was switched off on purpose, for its listing cost (2026-08-31),214 since `skillOverrides` cannot reach a plugin's skills. **Never215 recommend installing a plugin that is installed and off.** If it earns216 its place in this repo, recommend enabling it here only, in217 `.claude/settings.local.json` or `.claude/settings.json`, which218 outrank the user `false`, against that reason.219- **A plugin suggestion that never appeared** has several documented220 causes: the plugin is installed; a relevance suggestion comes only221 from a marketplace an administrator lists in222 `pluginSuggestionMarketplaces`; a CLI hint comes once ever per plugin,223 and never with analytics off. Name the ones that fit; assert none.224225### 3.2 `.mcp.json` hygiene226227For each server:228229- **A placeholder still in place**: a `<...>` segment in `url`, `args`230 or `headers`, as the project template ships them (`<WorkspaceId>`,231 `<GITHUB_PAT_VAR>`; the user's MCP README, `<config>/mcp/README.md`232 where their config holds one, lists them). It fails at233 every session start. An empty `url` is not this: Claude Code lists it234 `not configured` and never dials it.235- **A header holding a literal value** rather than a `${VAR}` or a236 `headersHelper`: a credential in a file. **Name the key, never the237 value.** Tracked and pushed, it is `recreate-repo`'s case; otherwise238 move it to a variable named per account, with no generic fallback.239- **A `${VAR}` this session lacks**: `printenv NAME >/dev/null && echo240 set || echo unset`, never printing the value. Unset with no241 `:-default`, Claude Code warns and sends the text unexpanded. A242 credential name such as `ANTHROPIC_API_KEY` or `NPM_TOKEN` reads as243 empty in a remote `url` or `headers` whether set or not (MCP docs).244 VS Code reads `.mcp.json` too and expands no `${VAR}` at all245 (1.140.0, 2026-10-03).246- **Tracked or not**: tracked, every clone gets the servers and is asked247 to approve each; untracked or ignored, clones get none. Either can be248 right: say which, and whether a server carries per-person values.249- **A `timeout` on the wrong transport.** A hosted Fabric `http`250 server with none fails any call past about 60 seconds with251 `The operation timed out.`; on a stdio server the value is a hard252 wall-clock cap on each call. The project template sets `120000` on the253 measured hosted entries and none on stdio, and the user's MCP README,254 where one exists, has why in § "Per-server timeout". One255 `.mcp.json` copied from another project had it the other way round256 (2026-10-21).257- **Overlap**: the same name at local or user scope with another258 endpoint (Claude Code warns; sign-ins are per endpoint), or the same259 endpoint as a plugin's server, a synced plugin's included, or a260 claude.ai connector. Plugins and connectors are matched by endpoint261 and the project server hides them (MCP docs, scope precedence).262 Recommend keeping one. A connector is hidden so only where Claude Code263 fetches connectors itself, in terminal, IDE and SDK sessions; a264 desktop session's run in-process (step 2).265- **Permission rules for a server's tools** go in `.claude/settings.json`,266 not beside the server.267268### 3.3 Scope269270User scope holds what is bound to nothing: on this machine, the two271docs servers `microsoft-learn-mcp` and `claude-code-docs`. A server272that needs a workspace, a database, a connection string, a desktop273app, a GitHub account or a tenant belongs in this repo's `.mcp.json`.274**A tool absent from user scope is not a gap**; a recommendation to put275one there is the finding.276The same test places a hook: one every repo needs is payload, one for277this repo's tools is the repo's. A claude.ai connector (step 2) reaches278every repo its account opens, as user scope does, so a tenant-bound one279meets the same test; its fix is on claude.ai, in no file. So does a280server in `claude_desktop_config.json` (step 2), whose fix is deleting281its entry by hand, since no deploy reconciles that file.282283### 3.4 Tenant pins that MCP servers ignore284285When the repo pins `az` to a tenant (a script or hook setting286`AZURE_CONFIG_DIR`, a guard reading `az account show`, a `CLAUDE.md`287line naming the tenant) and also declares a tenant-bound server288(`azure-mcp`, a Fabric or Power BI `headersHelper` server,289`microsoft-fabric-mcp`, `fabric-rti-mcp`), **the pin does not cover the290server**. Servers and their helpers are spawned by Claude Code and291inherit its environment, so an `az`-based one reads the shared292`~/.azure`: the user-scope `CLAUDE.md` says MCP servers ignore293`AZURE_CONFIG_DIR`, and the user's MCP README, where one exists,294§ "The helper's login is the harness's, not the folder's", holds the295one relayed measurement (2026-09-22). Report it with that section's296three candidate remedies, all unrun, the third a stdio server's own297`env`, or with no such file that one alone: measure one before relying298on it. Never confirm the finding by calling the server.299300**The launch may already carry a pin**, so read the harness's own value301before wording the finding: `$env:AZURE_CONFIG_DIR` in the PowerShell302tool, which runs `-NoProfile`, or `printenv AZURE_CONFIG_DIR` in Bash303only when `/proc/$$/cmdline` shows a snapshot (`bash -c source …`). A304login-mode Bash reruns the profile, which sets it from the folder305whatever the harness has. Set, report the servers covered by the306launch, not by the repo, so a session started another way has nothing307pinning them; unset, report as above. Whether a project helper keeps308it past Claude Code's credential stripping is unprobed. The same README309section has the measurements (2026-10-21).310311### 3.5 `CLAUDE.md` rules nothing enforces312313For each line in the files read in step 1 that says something must or314must not happen, look for what enforces it in the settings, hooks,315pre-commit config and tests. `CLAUDE.md` is "context, not enforced316configuration", while "permission rules are enforced by Claude Code,317not by the model" (memory and permissions docs). Where nothing does:318319| The line says | Enforce with | Say also |320| --- | --- | --- |321| ask before X | `permissions.ask` with X's rule | prompts even in `acceptEdits` and `bypassPermissions`; `dontAsk` denies instead |322| never read or edit F | `permissions.deny` with `Read(...)` or `Edit(...)` | covers the file tools and Bash file commands Claude Code recognizes, not `grep -r` or a script opening files; the sandbox is the OS-level answer |323| never run C, never print secrets | a `PreToolUse` hook that exits 2 | a failing hook lets the call through unless it sets `"onFailure": "block"` (2.1.295+); a Bash rule matches command text only |324| only the user starts X | `permissions.ask` naming X's script, a rule per shell | `disable-model-invocation: true` stops Claude loading a skill, not running its script or the command it wraps (skills docs) |325| run X before committing | a git hook in `.pre-commit-config.yaml` | binds every committer, not only Claude |326| change A with B and C, keep these in step | a test comparing them, run in CI | binds every change, not only Claude's; a value that may differ on purpose goes in the test's exception list, with its reason (2026-10-22) |327328**A gate counts only once it passes.** The last two rows, and the329catalog's Hooks rows, count a gate step 1 found, or propose one, only330over a baseline that passes: failing, it blocks every commit touching a331failing file, or rewrites each whole by a `--write` or `--fix` step. One332client run recommended `npm install` to arm a husky gate configured and333not armed; once installed, its ESLint, Prettier and Jest steps failed334across most of the repo (2026-10-28). **This skill never runs a gate**,335which would run the repo's code and its formatters:336337- **Unmeasured**, say so, and make running the gate in check mode the338 first Recommended action, ahead of any arm: `prettier --check`,339 `eslint` without `--fix`, the test script. Install by `npm ci`, not340 `npm install`, with `HUSKY=0` set: `npm ci` runs the `prepare` script341 too, which arms husky (npm scripts docs; `HUSKY=0` is husky's own342 switch for an install that must not, 2026-10-28).343- **Measured failing**, by the user or a CI log, report its counts; the344 fix is a baseline cleanup, or scoping the gate to changed files, never345 arming it as it is. So too a CI step proposed to run a script the repo346 already has.347348**A rule on command text has twin routes.** Check a "never run" or349"never print" line against each, and answer each with a matcher or with350`permissions.ask` or `deny`:351352- **Another command runner.** The desktop app's353 `mcp__terminal__run_in_terminal` runs commands no `Bash|PowerShell`354 matcher sees and no permission rule reaches: Claude Code skips an355 `mcp__` rule with parentheses, and a bare one asks on every command356 (permissions docs § "Match by input parameter", read 2026-11-14). So357 answer a `Bash(...)` or `PowerShell(...)` rule there with a358 `PreToolUse` hook on the tool that returns `permissionDecision` `ask`359 or `deny`; a matcher naming the tool fired in a live Desktop session360 (observed 2026-11-14, 2.1.321).361- **A declared server's tools.** Azure MCP's362 `appservice webapp settings get-appsettings` is annotated Secret, its363 values able to include connection strings (Learn, 2026-10-21): a rule364 on `mcp__<server>__<tool>`. In `azure-mcp`'s default namespace mode365 that is `mcp__azure-mcp__appservice`, one tool for every App Service366 operation, so the rule denies its writes too (read from a session's367 tool list 2026-10-21).368- **A second CLI for the same API.** Az PowerShell's369 `Get-AzStorageAccountKey` and `Get-AzFunctionAppSetting` pass a guard370 matching `az` as a word, and ignore `AZURE_CONFIG_DIR`, so a371 tenant-pin rule has the same hole (a guard fed synthetic payloads,372 2026-10-21): widen the pattern to the second CLI.373- **The same CLI's REST verb.** `az rest`, `gh api` and `fab api` take374 the operation as a URL, so its name arrives as a path segment:375 `az rest` against a site's `config/appsettings/list` or376 `config/connectionstrings/list`, or a storage account's `listKeys`,377 passes a guard matching `appsettings list` or `keys list` (a client378 guard fed synthetic payloads, 2026-10-21; its patterns re-run on the379 URLs 2026-10-22). Match the path segments too, or deny the REST verb380 where the repo has no use for it.381382Whether the line then stays as the reason, moves or goes is383`scope-instructions`' question, as is a side-effecting workflow384written as prose: name it, do not answer it.385386### 3.6 `.claude/launch.json`387388**The desktop app's Browser pane config**: one entry in `configurations`389per dev server it starts, or per URL it opens (Desktop § "Configure390preview servers", read 2026-10-29). It takes JSON with comments, so391Read it rather than `jq` it. Each check here is from the docs' text,392and none has yet met a broken file. With no file, the line says so, and393the catalog's dev-server row weighs proposing one.394395- **Tracked or ignored**, by step 1's two `git` lines. The docs treat396 the file as committed: `env` reads "Don't put secrets here since this397 file is committed to your repo". So ignored is the finding: a clone398 has no entries until the app detects a dev server on its first399 preview. An allowlisting `.gitignore` (`/.claude/*`, then a `!` line400 per file) ignores a new one silently, exit 0, as the config source's401 did until it gained a line (2026-10-29). Which copy a desktop-app402 worktree session reads is undocumented; one relayed note found a403 worktree's own copy unread (2026-10-20).404- **A literal value in an entry's `env`** that reads as a credential:405 name the key, never the value, as 3.2 does for a header. Tracked and406 pushed, it is `recreate-repo`'s case; otherwise it moves to a user407 environment variable on Windows, or the local environment editor,408 then a Desktop restart (§ "Set environment variables for dev409 servers"). A dev server gets neither `~/.claude/settings.json`'s `env`410 nor a saved environment's.411- **A `url` the app refuses.** One on `localhost`, a `*.localhost`412 subdomain, `127.0.0.1` or `::1` must be the server's origin alone, its413 port the entry's `port`, 3000 when unset: a path, a query or another414 port is reported as a configuration error. Any `url` must be `http`415 or `https`, with no username or password. An external one may carry a416 path, and asks permission the first time the preview opens it.417- **`autoVerify` and `autoPort` are no finding** on their own.418419## 4. Recommendations420421**Reuse first.** Before anything new, match the need against what step4222 found deployed: the user-scope agents and skills,423`microsoft-learn-mcp`, the payload's hooks, and the built-in skills. A424need one of them meets is a "use X" row, or no row.425426Then Read [references/catalog.md](references/catalog.md) and match step4271's signals. Keep one or two per category, the strongest signal first,428or three to five in the category an argument names; skip a category429with nothing. Every item names the file or signal behind it.430431**Name only what exists here**: a plugin the inventory's `--plugin`432lookup finds in a marketplace's list, a server from the project433template the MCP README describes, or one Learn documents through434`microsoft-learn-mcp`. Found nowhere, it is not recommended; say what435was looked for.436437**A copied skill goes only if its group can lose it.** Its description438saying it does not apply here is not enough. `copy-skills.ps1` selects439whole groups, so a skill in `.claude/skills/.managed-skills.json` whose440group under the payload's `skills/` holds one the repo keeps cannot be441dropped by that script, and a hand delete is planned as a `copy` on the442next refresh (read from the script, not run; 2026-11-01). Grep the kept443skills for its name too, and read each hit as a dependency or only a444boundary. Where either stops it, or the payload cannot be read, recommend445no removal and say why.446447## 5. Place each item448449| Lands in | When | How it gets there |450| --- | --- | --- |451| the repo | bound to its tools, files or tenant | `.claude/settings.json` or `.mcp.json` for everyone, `.claude/settings.local.json` for one person, `.claude/agents/`, `.claude/skills/`, `.pre-commit-config.yaml` |452| the payload | every repo needs it, or a payload file is wrong | in the config source, a proposed edit; elsewhere `capture-learning` where installed, which writes the inbox note and cites a client's estate by kind; with neither, named for the user |453| nowhere | reuse covers it, the code makes it derivable, or a payload decision rules it out | say which |454455## 6. Report456457Plain markdown, no emoji headings, about a screen:458459```text460## Profile the account read, the commit read (git rev-parse --short HEAD), the stack, the signals that drove the report461## Findings 3.1 to 3.6, a line each: the file, what is wrong, the fix, where it lands462## Recommendations by category, one or two each: what, the signal, where it lands463## Next steps in other skills what scope-instructions, /doctor or /fewer-permission-prompts owns here, for the user to run464## Recommended actions numbered: each fix or recommendation above that calls for an edit, and where it lands465```466467**Findings gives every check a line**, a clean one included, since a468missing line reads the same as a check not run: one run gave 3.4 none,469having read no `AZURE_CONFIG_DIR` (2026-10-21). **It also gives each470plugin set `false` its line from 3.1**: stage, whether the session471lists its skills, and whose decision. One run read every stage, then472reported only that the ids matched the payload (2026-10-21).473474**Recommended actions ends the report**, numbered `1.`, `2.`, … across475Findings and Recommendations: one line per finding fix or476recommendation that calls for an edit, naming the finding or category477it comes from and where it lands, as `1. 3.5.1, the inbox-delete rule:478a PowerShell twin of its ask rule, in the payload's479claude/settings.json`. A clean check, a 3.1 line, a finding the report480says to leave, a category with nothing new and an item under Next steps481in other skills get none. With nothing to edit, it reads `_(none)_`. It is the list a482handoff transcribes: in the config source, `file-audit-findings`, where483it has one, briefs each item without choosing among the findings, and484elsewhere nothing reads it.485486**A gate's check-mode run comes first**, where 3.5 calls for one: it is487`1.`, though it edits nothing, ahead of any fix that would arm the gate.488489Leave out any other empty section. Then stop: applying a row is a490separate request.491492## Constraints493494- **Edits nothing and writes nothing**, an inbox note included: that is495 `capture-learning`'s, run by the user.496- **Connects nothing**: no `claude mcp`, `claude plugin` or `az`497 command, and no tool of a project server.498- **Runs no gate**: no test, linter or formatter of the repo's, in check499 mode or not, and no install that would arm one (3.5).500- **Prints no secret**: a literal header or a variable's value is named501 by its key, never shown.502- **Recommends nothing at user scope that is bound** to a workspace, an503 account or a tenant.504- **A plugin the payload switched off stays a decision**: weigh its505 reason; never recommend installing it again.506- **A client's evidence stays in its repo**: a payload row from a client507 repo reaches the config source only through `capture-learning` or the508 user, cited by kind.509- **Not this skill**: where guidance lives (`scope-instructions`),510 install health and context cost (`/doctor`), allowlists from past511 sessions (`/fewer-permission-prompts`).