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: