Ops Doctor

Runtime Context

Before diagnosing, load:

1. Preferences: cat ${CLAUDE_PLUGIN_DATA_DIR:-$HOME/.claude/plugins/data/ops-ops-marketplace}/preferences.json — check all configured channels and services
2. Daemon health: cat ${CLAUDE_PLUGIN_DATA_DIR}/daemon-health.json — primary diagnostic input
3. Secrets: Verify secret resolution chain works: Doppler MCP → env → Doppler CLI → password manager

OPS ► DOCTOR

Load ops-rules before acting. Public repo (no personal data). Outbound: one draft → one approval → one send. If AskUserQuestion / Workflow are missing, follow Rule 10 in ops-rules (Hermes: numbered options / two-turn Telegram card; delegate_task).

CLI/API Reference

ops-doctor bin script

Key files read by diagnostics

Phase 1 — Run diagnostics

Run the diagnostic script to get a full health report:

${CLAUDE_PLUGIN_ROOT}/bin/ops-doctor 2>/dev/null || echo '{"errors":["diagnostic_script_failed"],"warnings":[]}'

Parse the JSON output. Display a summary:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 OPS ► DOCTOR — [timestamp]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

 Plugin:     [version] at [plugin_root]
 Skills:     [count] defined
 Agents:     [count] defined
 Bin scripts:[count] available

 ERRORS      [count]
 [list each error with description]

 WARNINGS    [count]
 [list each warning with description]

 TOOLS
 [table of CLI tool availability]

 ENV VARS
 [table of env var status]

 Registry:   [status] ([project_count] projects)
 Preferences:[status]
 Cache:      [versions list]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Agent Teams support

If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is set, use Agent Teams when multiple independent fix categories are identified (e.g., manifest issues + permission issues + registry issues). This enables:

• Fix agents work in parallel on different issue categories without stepping on each other
• You can prioritize: "fix manifest errors first, then permissions"
• Agents share context so a manifest fix can inform the registry repair

Team setup (only when flag is enabled, multiple issue categories):

TeamCreate("doctor-fixers")
Agent(team_name="doctor-fixers", name="fix-manifest", subagent_type="ops:doctor-agent", ...)
Agent(team_name="doctor-fixers", name="fix-permissions", subagent_type="ops:doctor-agent", ...)
Agent(team_name="doctor-fixers", name="fix-registry", subagent_type="ops:doctor-agent", ...)

If the flag is NOT set or only one issue category exists, use a single doctor-agent subagent.

Phase 2 — Decision

If $ARGUMENTS contains --check-only: stop here, display results only.

If there are errors or warnings:

Display: "Found [N] issues. Spawning doctor agent to auto-fix..."

Then spawn the doctor agent (or Agent Team — see above):

Agent({
  subagent_type: "ops:doctor-agent",
  prompt: "Fix the following ops plugin issues.\n\nDIAGNOSTIC_JSON: [paste full JSON]\nPLUGIN_ROOT: ${CLAUDE_PLUGIN_ROOT}\nCACHE_DIR: ~/.claude/plugins/cache/ops-marketplace/ops\n\nFix all errors and warnings. Re-run diagnostics after to verify.",
  description: "Fix ops plugin issues"
})

If there are no errors and no warnings:

Display: "All checks passed. Plugin is healthy."

MCP watchdog health probe

After Phase 1 diagnostics, parse the 5f block warnings from bin/ops-doctor. The bin script emits these MCP-specific keys:

• claude_json_invalid — ~/.claude.json is not valid JSON (hard error — all MCP tooling breaks)
• claude_json_missing — no ~/.claude.json (warning — no servers configured)
• mcp_stdio_cmd_missing_<name> — stdio server command binary not on PATH
• mcp_watchdog_stale — watchdog health not updated in >1h (cron not running)
• mcp_watchdog_no_health — watchdog has never run (never registered)
• mcp_servers_degraded_long — one or more HTTP MCP servers degraded >1h

For each warning, surface a one-line fix hint:

Include these warnings in the doctor-agent prompt for auto-fix where applicable (e.g., crontab registration is safe to do automatically).

Pocket health probe

After Phase 1 diagnostics, run the pocket probe if the pocket section is configured in preferences or if --verbose was passed. The bin script already includes these checks (section 5e); this section describes what the SKILL layer adds on top.

Parse the pocket block from the bin/ops-doctor JSON output. Surface any pocket-specific warnings:

• pocket_health_stale_* — health file exists but mtime > 5 minutes
• pocket_health_missing_* — health file does not exist
• pocket_health_error_* — health file has "status": "error"
• pocket_tmux_missing — pocket-exec tmux session not found
• pocket_config_invalid_* — whatsapp-config.json or email-config.json not valid JSON
• pocket_email_auth — gog gmail status not authenticated
• pocket_bridge_port — Baileys bridge port 8080 not listening

For each warning, surface a one-line fix hint:

Include pocket warnings in the doctor-agent prompt so it can auto-fix where possible.

Phase 3 — Post-fix verification

After the agent completes, re-run diagnostics:

${CLAUDE_PLUGIN_ROOT}/bin/ops-doctor 2>/dev/null

Display updated results. If errors remain, report them to the user with manual fix instructions.

Native tool usage

WebSearch — known issue lookup

When diagnostics find errors, use WebSearch to check if the issue is a known Claude Code plugin bug, MCP server issue, or configuration problem. Include links to relevant GitHub issues or docs.

WebFetch — MCP health check

For MCP servers that appear disconnected, use WebFetch to test their underlying APIs directly (e.g., https://api.linear.app/graphql with a simple query) to distinguish between "MCP broken" and "API down".