Universal diagnostic & repair toolkit for AI coding agents. Exposes agent-fix engine as native MCP tools for checking and fixing agent issues, managing configurations, and providing diagnostics.
14 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and error guidance. Tool names are action-oriented (doctor, check, apply, audit, backup, restore, provider, hooks, self_heal) which is good. Descriptions range 50-250 chars and explain WHAT and WHEN. However, many parameters lack descriptions (e.g., 'timeout' in net tool has description but 'depth' in audit is bare; 'confirm' params lack guidance on consequences). No output schemas documented. Error handling is minimal, tools return diagnostics but don't guide recovery. Composition is reasonable (single responsibility per tool), but some tools combine concerns (e.g., 'provider' generates AND writes config). Risk annotations present (READ_ONLY, WRITE, REVERSIBLE) but not formalized as toolAnnotations in schema.
List the agent registry and which agents are installed on this machine. Use first when diagnosing any agent problem.
Apply the fixes for one issue id, then verify. Default is a DRY RUN listing the fix commands; pass confirm=true to actually execute. Use after `check` shows a FAIL, or directly when the user reports a known symptom.
Scan agent config files for JSON parse errors and leaked API keys (masked). Use before committing configs or when an agent ignores its config.
Snapshot every detected agent's config dir into ~/.agent-fix-backups/<timestamp>.zip (excludes node_modules/sessions/logs). Use before any repair or upgrade.
Run diagnostics for one issue id. Issue ids (see also `info`): agent-broken-generic, npm-postinstall-skipped, gui-path-blind, node-version-too-old, npm-registry-mirror, agent-auth-broken, provider-config, net-connectivity, opencode-mcp-schema, deepseek-harness-broken.
Run every catalog check, including per-agent binary checks for all detected agents. Returns a health report. Use to answer 'is anything broken?'.
No output schemas documented. Tools return diagnostics/results but LLMs cannot predict structure (fields, types, pagination). Downstream tool chaining is blind.
Parameter descriptions missing or incomplete. 'depth' in audit has no description; 'timeout' in net lacks units/range; 'confirm' params don't explain consequences (dry-run vs destructive). LLMs cannot infer parameter semantics.
Error handling is minimal. Tools return diagnostics but do not guide recovery. E.g., 'check' may report FAIL but does not suggest 'call apply' or explain why. No actionable error messages.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Manage agent-fix self-heal startup hooks. action: status (default) | install | uninstall. agent_id optional (omit = all installed hook-capable agents). install registers the startup hook so the agent auto-checks+repairs on every launch; uninstall removes it; status shows what is registered.
Print the knowledge-base doc for an issue id (symptoms, root cause, manual fix, verification).
Scan agent log locations for recent ERROR/WARN/Traceback lines. Use when an agent fails without a clear message.
Check TCP connectivity + latency to every agent's API endpoint (anthropic/openai/deepseek/moonshot/google/zhipu/alibaba/github/npm) and show proxy env. Use when an agent 'suddenly stopped working' or for network diagnosis.
Generate per-agent config snippets for ANY provider (deepseek|openai|anthropic|google|moonshot|zhipu|qwen|openrouter|ollama|custom). Pass provider + api_key (optional base_url/model overrides). Key is MASKED in output by default; pass show_key=true to reveal it, or apply=true to write ~/.claude/settings.json.
List config backups, or restore one. Pass backup='latest' or a filename AND confirm=true to actually restore (zip-slip guarded; restores only into currently-known agent config dirs).
Run the full check pipeline once (same engine as the startup hooks). Default apply=false = diagnose-only (reports broken issues, changes nothing); pass apply=true to auto-fix anything broken. Use when the user reports any agent symptom, or as a periodic health pass.
Compare installed vs latest version for every detected agent (npm agents query the registry; GUI desktop apps are never probed and get update hints). Use before/after upgrades.
Tool 'provider' combines two concerns: generate config snippets AND write to settings.json (apply=true). Should split into 'generate_provider_config' + 'apply_provider_config' for composability.
Risk annotations (READ_ONLY, WRITE, REVERSIBLE) are metadata but not formalized as toolAnnotations in MCP schema. readOnlyHint/destructiveHint/idempotentHint should be in tool definitions for spec compliance.