Wire cost accounting through to persistence and CLI #1

Summary

Token usage is captured during agent strategy execution and persisted to SQLite, but USD cost is calculated in-memory by LLMClient and lost when the process exits. CLI inspect recalculates cost at display time from stored tokens + pricing config — meaning cost is never a first-class persisted artifact.

Current Data Flow (with 6 break points)

Provider API → Adapter extracts usage dict → LLMResponse.usage
  → LLMClient._record_usage_and_enforce_post_call_limits()
    → Calculates cost_usd via _estimate_cost_usd() (lines 308-330)
    → Stores in self._run_usage[run_id]["cost_usd"]  ← IN-MEMORY ONLY (BREAK 1)
  → Strategy._aggregate_usage(turns)  ← aggregates TOKENS only, ignores cost (BREAK 5)
  → AgentResult.token_usage = aggregated tokens
  → Executor: execution.token_usage = agent_result.token_usage  ← NO cost_usd field (BREAK 2-3)
  → Storage: steps.token_usage_json column  ← EXISTS, persisted correctly
  → Storage: steps.cost_usd column  ← DOES NOT EXIST (BREAK 4)
  → CLI inspect: recalculates cost from tokens + pricing config at display time

What's Persisted Today

• token_usage_json (per step) — in SQLite steps table, line 234 of storage/sqlite.py
• model_name (per step) — in SQLite steps table

What's Lost

• LLMClient._run_usage[run_id]["cost_usd"] — calculated per-request with _estimate_cost_usd() using pricing_usd_per_1k_tokens config
• Per-step cost breakdown
• Run-level total cost

Key Code Locations

Implementation Path

1. Add cost_usd: Optional[float] = None field to StepExecution dataclass (core.py:211)
2. Add cost_usd column to SQLite steps table (storage/sqlite.py:234)
3. Update storage.append_step() to persist cost
4. Pass cost from LLMClient response through AgentResult to Executor._dispatch_agent()
5. Add run-level cost aggregation field to Run model
6. Update CLI inspect to display persisted cost instead of recalculating

Priority

Tier 1 — High Impact, Low Effort. The data is already captured; it needs ~6 touch points to wire through to persistence.

🤖 Generated with Claude Code