Toolshed Doctor

Toolshed Doctor

Run the updater in check mode first. It only reads the installed-plugin registry and runs the Codex gateway health check when that plugin is installed:

node "${CLAUDE_PLUGIN_ROOT}/bin/update-toolshed.js" --check

Then run the same read-only session health audit used at startup:

node -e "const { audit } = require(process.env.CLAUDE_PLUGIN_ROOT + '/hooks/session-start-freshness.js'); const result = audit({ currentProject: process.cwd() }); console.log(JSON.stringify({ problems: result.problems, boards: result.mappings, staleProcesses: result.staleProcesses }, null, 2));"

For Model Gateway, report the audit finding exactly. It distinguishes a missing checker, a spawn error, a three-second startup timeout, a nonzero doctor result, and empty output. A nonzero doctor result includes one bounded, secret-filtered diagnostic line and is never a missing-checker finding. The SessionStart warning uses the same cause in compact form. These checks only report observed health; do not call the gateway healthy because a later command happened to pass.

An enabledPlugins entry only selects an installed plugin. If the audit says a plugin's hooks are not running, that settings file enables a plugin with no matching install row. Offer to install it at the reported project or user scope, or remove the dead enabledPlugins entry from the named file. Do not call that plugin configured or healthy until one of those actions is complete.

When staleProcesses has rows, report each PID, start time, and stale worktree path. They are read-only findings: do not kill any process from the doctor.

Observability, when that plugin is installed

Telemetry lives in the separate Observability plugin. Its files are under a different plugin root, so resolve that root first and skip this whole section when the plugin is absent:

node -e "const fs=require('node:fs'),os=require('node:os'),path=require('node:path');const reg=JSON.parse(fs.readFileSync(path.join(os.homedir(),'.claude','plugins','installed_plugins.json'),'utf8'));const installs=reg?.plugins?.['observability@eigenwise-toolshed']||[];const hit=installs.map(i=>i?.installPath).filter(Boolean).find(p=>fs.existsSync(path.join(p,'bin','setup-observability.js')));console.log(hit||'');"

An empty result means observability is not installed. Say so in one line and move on. Otherwise use that path as <OBS_ROOT> in the commands below.

Read the consent/config record and report the managed observability health. This command is read-only and uses the configured ports/container:

node "<OBS_ROOT>/lib/observability/ensure.js" --health

If it reports configured: false, observability was never consented to and needs no repair. If it reports enabled: false, say it is deliberately disabled. For an enabled record, report observer health, Collector listening state, selected sink, configured ports, dashboard/Docker state, and the storage block. Treat a listening observer with a failed /health response as unhealthy, and a configured dashboard without Docker as optional/unavailable rather than a pipeline failure.

Report storage.overDatabaseLimit: true or storage.overWalLimit: true as a real finding with the two byte counts, not as a note. Over the database limit means retention pruning alone cannot get back under the cap, so the file can keep growing; the managed prune drops the oldest days and uses incremental vacuum. Do not describe this as a blocking full VACUUM, and do not predict its duration from the current file size.

When the observer is healthy, audit project attribution. Hook events reach the observer from any directory, but the claude_code_* metrics only exist where Claude Code found the telemetry env in the settings of the directory the session started in. A project with observer events and no native samples in the same window is half-wired, and its dashboard reads empty rather than broken. This command is read-only:

node "<OBS_ROOT>/bin/verify-project-telemetry.js" --audit --project "<absolute-current-project-dir>"

Report every UNWIRED session directory by name, every half-wired: line, and the printed fix: command verbatim, adding that each affected session has to restart before its metrics appear. native-samples=unknown means Grafana was unreachable or unconfigured, so say the check could not run instead of calling it healthy. A not opted in: line is a hint, not a fault: those project names never opted in.

Then run the local report too:

node "<OBS_ROOT>/bin/token-usage-report.js" --format json

From the JSON report, call out outbox queue depth/capacity, drops, schema drops, telemetry conflicts, sessions missing SessionEnd, and the newest event/source. The SessionStart ensure hook repairs stopped managed processes on the next startup/resume; if immediate repair is requested, rerun /observability:enable-project-telemetry and keep the current observability choices.

Agent teams

Check whether a project-level env block masks the global agent-teams setting. This is read-only and prints nothing when the project has no env block or already enables teams:

node -e "const { agentTeamsWarning } = require(process.env.CLAUDE_PLUGIN_ROOT + '/lib/project-settings.js'); const warning = agentTeamsWarning(process.cwd()); if (warning) console.log(warning);"

If it prints a warning, report it with the one-line remedy verbatim. Do not fix it from the doctor.

Codex auto-compact window

Check the effective autoCompactWindow from user and project settings. This is read-only; the project setting wins when both files set the key:

node -e "const { compactionWindowFinding } = require(process.env.CLAUDE_PLUGIN_ROOT + '/lib/project-settings.js'); console.log(compactionWindowFinding(process.cwd()));"

Report the one-line finding verbatim. When the key is unset, it advises the optional setting "autoCompactWindow": 325000 in the named user settings file for a consistent 292000-token trigger instead of a model-dependent compaction point. Do not write or offer to fix it from the doctor.

Report all results together. Explain each concrete problem and give the smallest next step. This skill does not update, install, uninstall, reload, or edit anything. If freshness is proven stale, tell the user to run /quartermaster:update-toolshed, then /reload-plugins or restart before retrying the blocked work.