forge

1. 1. Click run sample investigation below. The input is a hardcoded Sentry-style payload (a TypeError stack trace, a deployed commit SHA, a timestamp) — see show input ↓.
2. 2. Four lanes fire in parallel. Watch them transition queued → running → done. Each lane calls Claude Sonnet 4.6 with its own focused system prompt and a small tool set (fixture-backed in this demo; real GitHub/Sentry adapters are a follow-up).
3. 3. When all lanes finish, the coordinator merges into ranked hypotheses with calibration-weighted confidences. You'll see panels for calibration (Brier scores per lane) and cost (per-call token usage in USD, with cache hit rates).
4. 4. Try refreshing the page mid-run — the session is durable, the UI rehydrates from server state via the resumable-stream layer.

what's real vs fixtures

The LLM calls, structured outputs, parallel fan-out, abort plumbing, calibration log, prompt caching, and cost math are all real. The data sources the subagents call into are fixture-backed for this public demo:

• fixture src/auth/session.ts — a small TypeScript snippet shaped like a real auth handler with a null-deref bug. Returned by fetch_file when source-reader navigates to it.
• fixture the candidate commits ( a3f1c92, 1b8e44d) and their diffs. Returned by git_log and git_diff when blame-correlator queries the history.
• fixture the error metrics (1,247 occurrences, 184 affected users, hourly spike). Returned by the error-tracking query tools when frequency-analyzer asks for blast radius.
• real everything else — the LLM's reasoning, the structured output coercion, the confidence values, the merge logic, the per-lane cost in USD, the speculator hit rate. Swap the fixtures for real GitHub + Sentry adapters and the rest of the system works identically.