Plugin Doctor

/plugin-doctor — Plugin Self-Diagnostic

Check that the plugin is installed correctly and operating within healthy bounds. Default mode is fast (linter-only, no LLM cost) and always passes on a clean install.

Invocation

Shipped commands:

/plugin-doctor                          # default smoke test (Tier 1 linters)
/plugin-doctor --calibrate-judge        # opt-in: judge calibration vs samples
/plugin-doctor --show-recent-errors     # top-5 hooks by ERROR/WARN count (audit WP-4.2)

Planned (not yet shipped):

/plugin-doctor --runs --last N          # summarize last N workflow runs
/plugin-doctor --health-trends          # context-health metrics summary (G8)

Both --runs and --health-trends will land once .ai-skills-memory/runs.jsonl accumulates enough entries to make the summary meaningful. Track progress in the plugin issue tracker.

Two-step boot model

Default mode (/plugin-doctor)

Wraps plugin/eval/runner.py --tier 1 — runs and reports on:

• Skill frontmatter — name + description present, kebab-case name matches parent dir, recognizable Use when trigger phrase (H5 regex from runner.py).
• Hook references — every entry in plugin/hooks/hooks.json resolves to an existing script under ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/. Both bare-path and the v0.3.3 python3 ${CLAUDE_PLUGIN_ROOT}/... wrapper form accepted.
• Hook scripts — Python syntax valid (py_compile in the runner).
• Run logs — .ai-skills-memory/runs.jsonl parses (JSON-syntax check; richer summaries are planned via --runs --last N, see Planned section above).
• JSON validity — plugin/schemas/*.json + eval/config.json + monitors/monitors.json + .claude-plugin/plugin.json parse cleanly.

Note: schema files are checked for valid JSON syntax, not full JSON Schema draft 2020-12 validation; the $schema URI declares draft 2020-12 but the runner does not currently re-validate against it.

Reports calibration as not yet run; use /plugin-doctor --calibrate-judge to run.

--calibrate-judge (explicit opt-in)

Delegates to the Tier 2 smoke runner — plugin/eval/runner.py --tier 2 — which performs a score-band tolerance check against plugin/eval/calibration/<rubric>/ samples. For each sampled rubric, the Haiku judge scores known good + known bad samples and the result must land within ±0.5 of the expected score encoded in the filename (<scenario>.score-X.X.md).

Catches three drift classes: rubric edits that break scoring, judge-model version changes, and calibration-sample edits that move them out of band.

Errors clearly when samples are missing or count too low (< 4 per rubric):

ERROR: Insufficient calibration samples at plugin/eval/calibration/<rubric>/.
       Expected ≥ 4 (good + bad pairs). Found N.
       See plugin/docs/concepts/eval.md for guidance.

Requires ANTHROPIC_API_KEY and the anthropic Python SDK; without them the run reports DRY-RUN ONLY and skips real judge calls. Default sample plan (seed 42): 10 rubrics × 2 samples = 20 judge calls. Override with --seed, --sample-rubrics, --samples-per-rubric on runner.py.

--show-recent-errors (recent-errors dashboard)

Surfaces the top hooks ranked by ERROR + WARNING count in .ai-skills-memory/errors.log over the last 7 days (default window). Ranking is ERROR * 10 + WARNING so a hook with a few ERRORs ranks above a hook with many INFO-level WARNINGs. For each hook the dashboard shows the top 3 issue strings and counts so the root cause of a noisy hook is one glance away.

Invocation:

python3 ${CLAUDE_PLUGIN_ROOT}/dev/recent-errors.py [--days N] [--top K] [--json]

Run from the project root (the script defaults to <cwd>/.ai-skills-memory/errors.log). No LLM cost; fail-open on missing log or parse errors. Typical output (real f4ai data, 7-day window):

Recent errors — last 7 day(s) (since 2026-05-06T13:40:33Z)
  Total events: 5918
  Distinct hooks: 7

Top 5 hooks by severity (ERROR weighted 10x):
  1. tool-output-normalize             ERR=0    WARN=5356 INFO=0
       └ wrap_marker_missing_before_normalize     x 5356
  2. subagent-stop-learnings           ERR=351  WARN=0    INFO=0
       └ return_contract_validation_failed        x 351
  …

Use --json for machine-readable output (e.g. CI dashboards). Use --days 1 for a same-day triage view.

Other modes (planned, not yet shipped)

--runs --last N and --health-trends are documented for design continuity. The runner does not currently expose these flags; invoking them will fall through to the default Tier 1 path (or error on the unknown flag depending on shell). Track shipping in the plugin issue tracker.

Output

• stdout: human-readable report
• .ai-skills-memory/plugin-doctor.log — append-only audit log

Failure modes

• No .ai-skills-memory/: report as "first run — no history yet"; suggest /ai-skills-init
• Hook script missing executable bit: expected since v0.3.3. Hooks are invoked via the wrapper form python3 ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/<name>.py (commit 4ea3347), which does not need the exec bit. Reverting to the bare ${CLAUDE_PLUGIN_ROOT}/... form would re-introduce the v0.3.2 "Permission denied" install bug. Report this as INFO only, not as a warning.
• hooks.json command starts with python3 ${CLAUDE_PLUGIN_ROOT}/... rather than ${CLAUDE_PLUGIN_ROOT}/...: expected since v0.3.3 — same reason as above. The Tier 1 linter (plugin/eval/runner.py) accepts both forms by searching for the placeholder substring rather than requiring it at position 0. If a third-party validator still flags the wrapper form, treat it as a linter-bug, not a plugin-bug.
• marketplace.json missing at the cache root (~/.claude/plugins/cache/<owner>/<plugin>/.claude-plugin/marketplace.json): expected in claude --plugin-dir install mode. The repo-level /<repo>/.claude-plugin/marketplace.json is the source-of-truth and is consumed only by /plugin marketplace add. --plugin-dir is the supported install mode for local development of this repo (per repo README.md and CLAUDE.md); the cache-root warning can be ignored unless distributing via marketplace.
• Run log too large to parse fully: stream-parse last 1000 lines; report partial stats with "(last 1000 events)" note

Integration

• Reads: every file under plugin/ for lints; .ai-skills-memory/runs.jsonl, plugin-doctor.log for summaries
• Calls: eval/runner.py --tier 1 (Tier 1 linters); eval/runner.py --tier 2 if --calibrate-judge
• Memory writes: .ai-skills-memory/plugin-doctor.log (audit trail)
• Used by: developers post-install, CI smoke before merge