doctor

omr:doctor

Diagnose omr installation issues. Read-only — emits CRITICAL / WARN / OK per check, then prints a summary that points at /omr:setup for any needed fix.

When this skill is invoked, immediately execute the workflow below. Do not just restate or summarize these instructions back to the user.

Note: paths under ~/.claude/... respect CLAUDE_CONFIG_DIR when set.

Invariants

1. Read-only. Never writes a file, never runs setup, never touches .env, CLAUDE.md, .omr-config.json, or HPC configs. If a fix is needed, point at /omr:setup with the appropriate flag.
2. Never echoes secrets. Token probes report presence/source only — same rules as /omr:setup Phase 2.
3. Cross-platform commands. Prefer constructs that work in bash 3.2 (macOS default) — no associative arrays, no mapfile. Use POSIX utilities where possible.
4. Aggregate at the end. Track each step's verdict so the final summary lists exactly what failed and how to fix it.

Diagnostic steps

Run each step in order. After each step, record one of OK / WARN / CRITICAL plus a one-line reason. Don't print the verdicts inline — collect them, then emit the summary in the last section.

Step 1 — Plugin integrity

Verify the plugin's own manifest and required files are present.

# Plugin manifest
jq -r '.name + " v" + .version' "${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json" 2>/dev/null \
  || echo "MISSING: ${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json"

# Required files
for f in .mcp.json bin/load-env-and-exec.sh skills/setup/SKILL.md; do
  if [ -e "${CLAUDE_PLUGIN_ROOT}/$f" ]; then
    echo "FOUND: $f"
  else
    echo "MISSING: $f"
  fi
done

# Shim is executable
if [ -x "${CLAUDE_PLUGIN_ROOT}/bin/load-env-and-exec.sh" ]; then
  echo "SHIM: executable"
else
  echo "SHIM: not executable"
fi

Diagnosis:

• Any MISSING: line → CRITICAL (install broken; recommend reinstalling the plugin).
• Plugin name from manifest ≠ omr → CRITICAL (wrong plugin or tampered manifest).
• Shim not executable → WARN (stdio MCPs will fail to launch; user can chmod +x to fix manually).
• Else → OK.

Step 2 — MCP server declarations

Confirm .mcp.json declares all expected servers.

jq -r '.mcpServers | keys[]' "${CLAUDE_PLUGIN_ROOT}/.mcp.json" 2>/dev/null | sort

Diagnosis:

• Output missing any of brave-search, exa, github, huggingface, tavily → WARN (one or more servers no longer declared; reinstall the plugin).
• All five present → OK.

Step 3 — CLAUDE.md install state

Check whether the omr marker block is installed in either the global or project CLAUDE.md, and whether its version matches the plugin.

PLUGIN_V=$(jq -r '.version' "${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json")
echo "plugin version: $PLUGIN_V"

for path in "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/CLAUDE.md" "./.claude/CLAUDE.md"; do
  if [ -f "$path" ]; then
    INSTALLED_V=$(grep -oE 'BEGIN omr version="[^"]+"' "$path" 2>/dev/null \
      | head -1 | sed -E 's/.*version="([^"]+)"/\1/')
    if [ -n "$INSTALLED_V" ]; then
      echo "$path: omr v$INSTALLED_V"
    else
      echo "$path: no omr block"
    fi
  else
    echo "$path: (file does not exist)"
  fi
done

Diagnosis:

• At least one target has omr v<X> matching PLUGIN_V → OK.
• Any target has omr v<X> ≠ PLUGIN_V → WARN per target (CLAUDE.md drift; run /omr:setup --global or --local to refresh).
• Both targets are no omr block or (file does not exist) → WARN (plugin known but never registered; run /omr:setup).

Step 4 — MCP token presence

Probe each server's expected env var. Presence-only — never echo values.

for pair in \
    "EXA_API_KEY:exa:stdio" \
    "TAVILY_API_KEY:tavily:stdio" \
    "BRAVE_API_KEY:brave-search:stdio" \
    "GITHUB_PERSONAL_ACCESS_TOKEN:github:http" \
    "HF_TOKEN:huggingface:http"; do
  v="${pair%%:*}"
  rest="${pair#*:}"
  server="${rest%%:*}"
  transport="${rest#*:}"
  if [ -n "$(printenv "$v")" ]; then
    echo "$server: $v set in shell"
  elif [ -f "./.env" ] && grep -E "^${v}=." ./.env >/dev/null 2>&1; then
    if [ "$transport" = "http" ]; then
      echo "$server: $v in .env (HTTP — not read at runtime; export in shell)"
    else
      echo "$server: $v set in .env"
    fi
  else
    echo "$server: $v missing"
  fi
done

Diagnosis:

• All servers set in shell (or stdio servers set in .env) → OK.
• Any stdio server missing → WARN (run /omr:setup --audit for a full status report, then /omr:setup to scaffold .env if needed).
• Any HTTP server missing or only-in-.env → WARN (HTTP MCPs need shell-exported tokens; .env is invisible to them).

Step 5 — HPC config drift (optional)

If ./.omr/hpc/ contains any *.yaml files, check each against the current plugin version. (HPC configs are always project-local — there is no ~/.claude/hpc/.)

PLUGIN_V=$(jq -r '.version' "${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json")
found_any=0
for dir in "./.omr/hpc"; do
  [ -d "$dir" ] || continue
  for f in "$dir"/*.yaml; do
    [ -f "$f" ] || continue
    found_any=1
    INSTALLED_V=$(awk '/^template_version:/{print $2; exit}' "$f")
    if [ "$INSTALLED_V" = "$PLUGIN_V" ]; then
      echo "$f: template_version $INSTALLED_V (aligned)"
    else
      echo "$f: template_version $INSTALLED_V (plugin is $PLUGIN_V — drift)"
    fi
  done
done
if [ "$found_any" = "0" ]; then
  echo "no HPC configs installed (skip)"
fi

Diagnosis:

• no HPC configs installed → OK (nothing to check).
• All aligned → OK.
• Any drift line → WARN per file (rerun /omr:setup to refresh).

Step 6 — Config marker sanity

Read .omr-config.json and validate it.

PLUGIN_V=$(jq -r '.version' "${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json")
CFG="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/.omr-config.json"
if [ -f "$CFG" ]; then
  if jq -e '.setupCompleted, .setupVersion' "$CFG" >/dev/null 2>&1; then
    INSTALLED_V=$(jq -r '.setupVersion' "$CFG")
    if [ "$INSTALLED_V" = "$PLUGIN_V" ]; then
      echo "config: ok (setupVersion $INSTALLED_V, aligned)"
    else
      echo "config: ok (setupVersion $INSTALLED_V, plugin is $PLUGIN_V — drift)"
    fi
  else
    echo "config: malformed (missing setupCompleted or setupVersion)"
  fi
else
  echo "config: not present (setup never completed?)"
fi

Diagnosis:

• aligned → OK.
• drift → WARN (run /omr:setup to refresh the marker).
• malformed → WARN (manual edit broke it; run /omr:setup --force).
• not present → WARN if Step 3 found an omr block (setup was partially done and the marker got removed) or if any other signal suggests setup ran; else OK (fresh install, expected before first setup run).

Step 7 — External dependencies

for cmd in jq node npx; do
  if command -v "$cmd" >/dev/null 2>&1; then
    ver=$("$cmd" --version 2>&1 | head -1)
    echo "$cmd: $(command -v "$cmd") ($ver)"
  else
    echo "$cmd: MISSING"
  fi
done

Diagnosis:

• All three present → OK.
• jq missing → WARN (config-marker writes degrade to overwrite; Phase 3's CLAUDE.md install loses some features).
• node or npx missing → CRITICAL (stdio MCPs cannot launch at all).

Summary

After all steps, print exactly one block:

omr:doctor — health report

  ✓ OK        <count>
  ⚠ WARN      <count>
  ✗ CRITICAL  <count>

Findings:
  ⚠ <one-line description> — fix: <command or action>
  ✗ <one-line description> — fix: <command or action>
  ...

Next moves: <single-line suggestion based on highest-severity finding>

If every step is OK:

omr:doctor — all checks passed.

End there. Do not run any fix on the user's behalf. Do not ask follow-up questions. The user decides whether to run /omr:setup.