Implement observability: token tracking, error logging, and action audit trail

[martinjms]

Summary

Before optimizing token usage or debugging unconscious processes, we need visibility into what's happening. Track every API call, every error, and every action so we can reconstruct what happened and find issues.

Requirements

Token tracking

• Log every Anthropic API call: model, input tokens, output tokens, cache hits, cost
• Track per-component: brain, worker, heartbeat, cron, (future) unconscious processes
• Daily/hourly summaries: total spend, breakdown by component
• Expose via tool so the agent can check its own costs

Error logging

• Structured error log with timestamps, component, error type, stack trace
• Tool call failures with the tool name, args, and error
• Hot-reload failures and rollbacks
• Worker task failures

Action audit trail

• Every tool call the brain makes (name, args, result summary, tokens)
• Every channel message (direction, channel, user, timestamp)
• Every file change detected by the kernel
• Every heartbeat/cron cycle and what it did

Access

• Log files in .state/logs/ (rotated daily)
• Tool for the agent to query its own logs (log_search, log_summary)
• Dashboard endpoint /api/logs for the UI (future)

Implementation

• server/modules/logger.js — structured logger used by all components
• server/modules/tools/logs.js — tools for the agent to query logs
• All existing components updated to log through the structured logger

Why first

Every unconscious process we build after this will automatically have observability. Build the foundation once, benefit forever.