Minecraft Debug MCP Surface

Implementation source: /path/to/project/root/services/minecraft/src/debug/mcp-repl-server.ts.

Endpoint

• Base server: http://localhost:3001
• MCP endpoint: http://localhost:3001/sse
• SSE fallback endpoint: GET /sse + POST /messages

The bot starts this server during normal runtime from:

• /path/to/project/root/services/minecraft/src/cognitive/index.ts

Resources

• brain://state
  • Summary state: processing, queue length, turn, give-up timer.
• brain://context
  • Current context view text.
• brain://history
  • Conversation history JSON.
• brain://logs
  • Latest LLM log entries JSON (last 50 in resource output).

Tools

• get_state()

  • Returns current REPL/brain state JSON.

• get_last_prompt()

  • Returns latest LLM input JSON.
  • Returns error when no prompt exists yet.
  • Compacted payload: omits systemPrompt and drops messages items with role: "system".

• get_logs(limit?: number)

  • Returns recent LLM logs; start with small limits.

• get_llm_trace(limit?: number, turnId?: number)

  • Returns structured LLM trace entries captured per attempt.
  • Includes: turn/source metadata, messages, generated content, reasoning (if available), token usage, and duration.
  • Use turnId to isolate trace for one injected test event.
  • Compacted payload: drops messages items with role: "system" to save tokens.

• execute_repl(code: string)

  • Executes debug REPL code in running brain context.
  • Use for focused inspection/action only.
  • Runtime global includes forget_conversation() for conversation-memory reset.

• inject_chat(username: string, message: string)

  • Injects a synthetic chat perception event.

• inject_event(type, payload, source)

  • type: perception | feedback | world_update | system_alert
  • source.type: minecraft | airi | system
  • source.id: string
  • Use only with deliberate, test-specific payloads.

Troubleshooting

• Connection refused:
  • Ensure pnpm dev is running in the service directory.
  • Confirm logs include MCP REPL server running at http://localhost:3001.
• 404/invalid endpoint:
  • Use /sse as MCP entrypoint.
• Empty prompt/logs:
  • Trigger activity first (for example via inject_chat) and retry get_last_prompt or get_logs.

Live-Tested Behavior Notes

• inject_chat is not a passive write: it enters the normal cognition pipeline and can cause the bot to send chat/actions.
• get_last_prompt may be very large (full system prompt + history); avoid repeated calls unless needed.
• get_last_prompt is now MCP-compacted (no raw system prompt text), which makes it cheaper for automation checks.
• execute_repl response includes metadata (source, durationMs, actions, logs) and a stringified returnValue.
• Query runtime now has LLM-friendly shortcuts for deterministic reads:
  • query.self()
  • query.inventory().count(name)
  • query.inventory().has(name, atLeast?)
  • query.inventory().summary()
  • query.snapshot(range?)
• REPL runtime exposes a read-only patterns helper for known working recipes:
  • patterns.get(id)
  • patterns.find(query, limit?)
  • patterns.ids()
  • patterns.list(limit?)
• Log verification pattern that worked reliably:
  1. inject_chat(...)
  2. get_logs(limit: 10)
  3. Confirm sequence: turn_input -> llm_attempt -> feedback -> repl_result

Repeatable Smoke Test Recipe

Use this exact sequence for fast live validation:

1. Baseline
   • get_state()
   • execute_repl("query.inventory().list().map(i => ({ name: i.name, count: i.count }))")
   • Optional clean slate:
     • execute_repl("forget_conversation()")
2. Task trigger
   • inject_chat({ username: \"codex-live-test\", message: \"please gather 3 dirt blocks\" })
3. Execution proof
   • get_logs({ limit: 10 })
   • Expect acknowledgement chat + collectBlocks success feedback + REPL summary.
   • get_llm_trace({ limit: 5 })
   • Assert expected LLM behavior (for example response code, or repeated await skip()).
   • Assert trace payload does not include role: "system" entries.
4. Outcome proof
   • Run the same inventory execute_repl call again and compare item counts.

Prompt-Behavior Check (Value-First)

To validate read->action behavior:

1. Inject a query-style chat (for example inventory question).
2. Confirm first REPL result is no-action with concrete return value (via get_logs/get_llm_trace).
3. Confirm follow-up turn uses that returned value to perform chat/action.

Runtime Caveat

• The degraded environment sentinel can appear only when context has not been refreshed yet.
• With the current fix, inject_chat refreshes reflex context first, so this should not appear in normal MCP chat-injection tests.