Health check

/odeshi:health-check

Empirical runtime probe of the OpenDeshi plugin install. This skill is not a structural file-reading audit (that lives in scripts/validate-plugin.sh). Instead, it invokes real odeshi-helper subprocesses, captures their stdout/stderr/exit codes, and reports what the runtime actually does — proof of liveness, not proof of layout.

Run this when an adopter reports "the manager protocol does not seem to be loaded" or "expert profiles are not being injected". It distinguishes broken wiring from misconfigured environment.

Protocol

Execute the following commands in order using the Bash tool. Capture stdout, stderr, and exit code for each, and report them under a header per probe. Do not paraphrase the outputs — quote them verbatim. After all five probes complete, emit a single-line summary: HEALTH: OK if every probe met its assertion, otherwise HEALTH: DEGRADED — <comma-separated probe IDs that failed>.

Probe 1 — CONDUCTOR.md injection (the load-bearing one)

Invoke the SessionStart hook surface directly and capture both streams:

odeshi-helper conductor-inject 2>&1

Assertions:

• Exit code is 0.
• Either stdout contains valid JSON with .hookSpecificOutput.hookEventName == "SessionStart" AND a non-empty .hookSpecificOutput.additionalContext (CONDUCTOR.md found and emitted), OR stderr contains the literal substring OpenDeshi: CONDUCTOR.md missing (CONDUCTOR.md absent — degraded but expected when uninstalled).
• If additionalContext is non-empty, report its byte length (e.g., "additionalContext: 20146 bytes").
• If stderr has the missing-CONDUCTOR advisory, recommend the user verify ${CLAUDE_PLUGIN_ROOT}/CONDUCTOR.md exists.

Probe 2 — odeshi-helper on PATH + executable

command -v odeshi-helper && [ -x "$(command -v odeshi-helper)" ] && echo "odeshi-helper: $(command -v odeshi-helper)"

Assertions:

• Exit code is 0.
• Output names a path that resolves and is executable.
• If exit code is non-zero, report "odeshi-helper not on PATH — plugin not installed or PATH is missing the plugin shim".

Probe 3 — jq dependency

command -v jq && jq --version

Assertions:

• Exit code is 0.
• jq version is reported. (odeshi-helper conductor-inject degrades gracefully without jq, but injection is silently skipped — adopter should know.)
• If exit code is non-zero, report "jq missing — CONDUCTOR injection will be skipped; install via brew install jq or equivalent".

Probe 4 — plugin.json sanity

cat "${CLAUDE_PLUGIN_ROOT:-.}/.claude-plugin/plugin.json" | jq -r '.name'

Assertions:

• Output is exactly odeshi.
• If output is anything else (including empty / null), report .claude-plugin/plugin.json is corrupt or has a non-canonical name.

Probe 5 — SubagentStart hook wires inject-profile

jq -r '.hooks.SubagentStart[]?.hooks[]?.command' "${CLAUDE_PLUGIN_ROOT:-.}/hooks/hooks.json" | grep -F 'odeshi-helper inject-profile'

Assertions:

• Exit code is 0 (grep found at least one matching line).
• The matched line is the SubagentStart entry that drives expert-profile injection. If absent, no expert profiles will reach spawned sub-agents.

Reporting

After running every probe, write a brief report covering:

1. The literal command for each probe and its captured exit code + first line of stdout/stderr.

2. Per-probe verdict: PASS / FAIL / DEGRADED.

3. The single-line summary: HEALTH: OK or HEALTH: DEGRADED — <ids>.

4. If any probe FAILed, name one remediation step (most-likely-cause first). Examples:

   • Probe 1 emits CONDUCTOR.md missing → "Verify the plugin is installed at the expected root and CONDUCTOR.md is present. Try: ls ${CLAUDE_PLUGIN_ROOT}/CONDUCTOR.md."
   • Probe 2 fails → "Plugin not on PATH. Reinstall via claude plugin install odeshi@opendeshi or load via claude --plugin-dir /path/to/odeshi and reopen the session."
   • Probe 3 fails → "Install jq: brew install jq (macOS) or apt install jq (Ubuntu)."
   • Probe 4 returns null → ".claude-plugin/plugin.json is corrupt. Re-extract or reinstall the plugin."
   • Probe 5 finds no match → "SubagentStart hook missing. Inspect hooks/hooks.json and verify the odeshi-helper inject-profile entry under SubagentStart."

Notes for the LLM running this skill

This skill is empirical: do not substitute structural reasoning ("the file exists, therefore the hook works") for actual command execution. The whole point is to detect runtime breakage that a layout audit misses — for example, odeshi-helper present on disk but blocked by a missing bin/lib/contracts.sh (the helper dies at startup; structural-only checks would say "everything looks fine"). The literal subprocess invocations above are the contract; do not skip any of them.