token-data-sources

Analysis Page Token Data Sources

When to use: When debugging why the analysis page shows missing or zero token data, or when adding new cost/usage visualizations.

Two Token Data Paths

The analysis page has two independent data sources for token information. Understanding which is active and when is critical.

Path 1: Live Agent Objects (In-Memory)

• Source: AnalysisPage.tsx reads agent.inputTokens / agent.outputTokens from the app store
• Populated by: WebSocket agent:usage events → appStore updates agent objects in memory
• Works when: Session is active and agents are running
• Fails when: Session is inactive (agents gone from memory → all zeros → "Waiting for token data...")
• Used for: Time-series token usage chart, total token counts

Path 2: DB-Persisted Cost Records

• Source: CostBreakdown.tsx fetches /costs/by-agent and /costs/by-task API endpoints
• Populated by: CostTracker.recordUsage() → writes to task_cost_records DB table
• Works when: Data has been recorded to the DB (persists across sessions)
• Fails when: The dagTaskId gate blocks recording (see below)
• Used for: Per-agent and per-task cost attribution breakdown

The dagTaskId Gate Bug

In AgentManager.ts, token usage recording has a gate condition:

// packages/server/src/agents/AgentManager.ts ~line 725
if (dagTaskId && agent.parentId) {
  tracker.recordUsage(agent.id, dagTaskId, ...);
}

This requires BOTH:

1. dagTaskId — agent must have an active DAG task
2. agent.parentId — agent must have a parent (not be the lead)

Problem: The lead agent (which has no parentId) never records usage. And child agents only record when they have an active DAG task. This means task_cost_records is often empty → "No token attribution data yet."

Meanwhile, emit('agent:usage') at line ~730 fires unconditionally, so the time-series chart works fine.

Fix Plan

1. Remove the parentId requirement: The lead's token usage should be tracked too
2. Fall back to a synthetic task ID: When no dagTaskId exists, use agent.id or a session-level bucket
3. Add DB fallback for inactive sessions: When live agents aren't available, fetch historical data from /costs/by-agent instead of reading from in-memory agent objects

Key Files