[Gastown] PR 17: Agent CVs & Performance Analytics

[jrf0110]

Parent: #204 | Phase 3: Multi-Rig + Scaling

No major architectural changes — Agent Identity DO accumulates data from DO RPCs regardless of execution model.

Goal

Build the structured work ledger for agent performance tracking.

Agent Identity DO

Each agent gets a persistent Durable Object that accumulates:

• Bead closures (type, time, quality signal from refinery)
• Molecule step completions
• Convoy participations
• Escalation history
• Session count/duration
• Model used per session

Dashboard Views

• Agent performance cards: beads closed, avg time to close, quality rate (refinery pass rate)
• Model comparison: same work type with different models → which performs better
• Cost per bead: LLM usage from gateway, attributed to agent

Agent Identity DO Schema

CREATE TABLE work_log (
  id TEXT PRIMARY KEY,
  bead_id TEXT NOT NULL,
  bead_type TEXT NOT NULL,
  rig_id TEXT NOT NULL,
  convoy_id TEXT,
  started_at TEXT NOT NULL,
  completed_at TEXT,
  quality_signal TEXT,        -- 'passed', 'rework_requested', 'failed'
  model TEXT NOT NULL,
  token_usage INTEGER,
  cost_cents INTEGER
);

CREATE TABLE escalation_log (
  id TEXT PRIMARY KEY,
  bead_id TEXT NOT NULL,
  severity TEXT NOT NULL,
  category TEXT,
  created_at TEXT NOT NULL
);

Dependencies

• PR 1 (Rig DO)
• PR 9 (Town DO)
• PR 12 (Refinery — quality signals)

Acceptance Criteria

• Agent Identity DO with work log and escalation log
• Automatic logging on bead close, escalation, session events
• Agent performance card component
• Model comparison view
• Cost attribution per agent/bead