[dadachi]

Closes #167.

Each run now reports what it cost, instead of leaving the user to filter tmp/trace/ by timestamp.

What changed

• src/usage.ts: a per-run ledger (AsyncLocalStorage around dispatch(), so concurrent MCP calls and the smoke tests' many dispatches don't mix). The planner and each judge call record their usage; the repair agent records the Agent SDK's total_cost_usd. Priced exactly like scripts/cost.sh (list rates from OPUS_PRICE_PER_MTOK; calls served by a refusal-fallback model are counted in fallbackCalls and priced at the same rates).
• report.json gets a cost block (estimatedUsd, calls, byAgent.{planner,judge,repair} with tokens and USD, pricing). Absent when the run made no API call (stub mode).
• HTML report: an "API cost (estimate)" table and the rounded total in the header.
• CLI: cost: ≈ $0.2197 (25 API calls: planner $0.0273, judge $0.1924) after the overall: line; the same line goes to tmp/trace/dispatch.log.
• MCP: comes with report in structuredContent, no extra field.
• Docs: README Runtime/Cost (per-run cost, the 2026-10-18 Console reconciliation, --tests adds ~2.5 min and no API cost since the suites run in parallel), docs/validation-report.md (section + data model), docs/PLUGIN-GUIDE.md, and the generate-app skill (report shape; quotes the cost in its headline).

Tests (npm test: 206 pass; build + typecheck:tests clean)

• Pricing uses the Orange Baton run that matched the Console export token-for-token: planner $0.0273 + judge $0.1924 = $0.2197, with the hand calculation in the test.
• Fallback counting and repair SDK cost taken as-is; concurrent ledgers stay separate and nothing is recorded outside one; stub dispatch has no cost; report/HTML/CLI line carry it.
• Seen red: zeroing the cache-read rate, sharing one ledger across runs, and dropping the fallback count each fail the new tests.

Run for real (local build, --visual=0, barbershop spec): report.json cost = planner 2,722 in / 849 out = $0.0279, identical to scripts/cost.sh on that run's trace; the CLI line and the HTML section show it. The judge path wasn't exercised live (no simulator booted at the time); it's the same recordMessageUsage call next to the existing usage: trace line.

🤖 Generated with Claude Code

https://claude.ai/code/session_016UPsTupJqJ5X1NK2gZEN5c