analyst

Cost tracking and learning agent that extracts patterns, maintains institutional memory, and proposes improvements.

Overview

The analyst is the terminal agent in the pipeline — it runs last (Phase 6) and passes to no one. It tracks token usage, generates cost reports, extracts recurring patterns into memory/patterns.json, and maintains lessons learned in memory/lessons.md. Every 10 sessions, it proposes updates to CLAUDE.md based on accumulated patterns (never auto-applies — always proposes for human review).

Quick Reference

Documentation & Management

Capability	Details
Cost tracking	Token usage per task in memory/cost-log.json
Cost reports	/meow:budget — spend by task, agent, model tier, over time
Pattern extraction	Recurring issues and solutions in memory/patterns.json
Lessons learned	Human-readable learnings in memory/lessons.md
CLAUDE.md proposals	After 10 sessions, proposes improvements (human-approved)
Cost optimization	Identifies tasks consistently over-classified to expensive tiers

How to Use

/meow:budget    # view cost report

The analyst also runs automatically at session end (Phase 6).

Under the Hood

Handoff Example

Analyst (session end):
  Token usage this session: 45,000 (Sonnet)
  Patterns found: "auth tests frequently need retry" → saved to patterns.json
  Lesson: "JWT refresh logic requires integration tests, not unit tests"
  Cost note: "3 trivial tasks were classified STANDARD — could be TRIVIAL"

  Sessions since last CLAUDE.md proposal: 8 (2 more until next)

  → No handoff (terminal agent)

Troubleshooting

Issue	Cause	Fix
Memory files corrupted	Interrupted session or manual edit	Analyst creates fresh files with empty structure
Cost data seems wrong	Token counting is approximate	Analyst logs what's available — never fabricates
CLAUDE.md proposal unwanted	Happens every 10 sessions	Review and accept/reject — never auto-applied