feat(pilot): workflow-driven AI agent for MCP Mesh

Pilot is an event-driven workflow executor that serves as the central
AI brain for the MCP ecosystem. Key features:

- Event-driven: Subscribes to user.message.received, publishes agent.response.*
- Workflow-based: Declarative JSON workflows define execution patterns
- Full mesh access: Uses all connected MCPs as tools
- MCP Tasks protocol: Persistent task tracking and status
- Conversation support: Long-running threads with memory

Built-in workflows:
- fast-router: Routes to direct response, tool call, or async workflow
- conversation: Long-running conversation with history
- direct-execution: Execute with all tools, no routing
- execute-multi-step: Complex multi-step tasks

MCP Tools:
- WORKFLOW_START, MESSAGE, CONVERSATION_START/END
- TASK_GET, TASK_RESULT, TASK_LIST, TASK_CANCEL
- WORKFLOW_LIST, WORKFLOW_GET, WORKFLOW_CREATE
- ON_EVENTS for mesh event subscription

Works with mesh-bridge for WhatsApp, CLI, and other interfaces.

Author: vibegui

bun.lock

@@ -30,6 +30,7 @@
     "object-storage",
     "openrouter",
     "perplexity",
+    "pilot",
     "pinecone",
     "readonly-sql",
     "registry",

package.json

@@ -0,0 +1,51 @@
+# Pilot Agent Guidelines
+
+## Debugging Complex AI Flows
+
+When debugging multi-step workflows, LLM calls, or async task execution where terminal logs may be truncated or lost:
+
+**Use temporary file-based logging** to capture the full picture:
+
+```typescript
+const fs = await import("fs");
+const logPath = "/tmp/pilot-debug.log";
+const log = (msg: string) => {
+  const line = `[${new Date().toISOString()}] ${msg}\n`;
+  fs.appendFileSync(logPath, line);
+  console.error(msg); // Also emit to stderr for STDIO capture
+};
+
+log(`🔍 LLM CALL: model=${modelId}, messages=${messages.length}`);
+log(`📝 PROMPT: ${prompt.slice(0, 300)}`);
+const result = await callLLM(...);
+log(`📤 RESULT: text=${!!result.text} (${result.text?.length || 0} chars)`);
+```
+
+This technique is essential when:
+- Pilot runs as STDIO subprocess (Mesh only captures stderr)
+- Terminal output is truncated or scrolling
+- Async callbacks (setTimeout) fire after parent logs
+- You need timestamps to trace execution order across concurrent flows
+
+**Always use `console.error` instead of `console.log`** in Pilot - Mesh's STDIO transport only pipes stderr to the main console.
+
+Remember to clean up debug logging before committing.
+
+## Common Gotchas
+
+### Model ID Resolution
+When spawning child workflows via `start_task`, always pass the resolved model IDs from the parent config:
+
+```typescript
+// ❌ Wrong - passes literal strings
+config: { fastModel: "fast", smartModel: "smart" }
+
+// ✅ Correct - passes actual model IDs
+config: { fastModel: config.fastModel, smartModel: config.smartModel }
+```
+
+### STDIO Logging
+Pilot runs as an STDIO process under Mesh. Only `stderr` is captured:
+- Use `console.error()` for debug output
+- `console.log()` output will be lost
+