scout

What problem does it solve? Debugging often stops at a plausible guess: an agent reads code, picks a likely culprit, and ships a fix that misses the real cause. Scout enforces evidence-based diagnosis by investigating a defect with graphify and code references, then certifying the root cause with runtime proof before any fix is attempted. ## Core Features & Use Cases - Parallel investigation engine: Dispatches code-tracer, blast-radius-mapper, and repro-verifier agents that verify leads against real source and write durable findings files. - Adversarial certification gate: A hypothesis only becomes a diagnosis with reproduce-plus-state-match, a counterfactual, or an unbroken evidence chain; after 3 rounds it reports UNCERTIFIED instead of guessing. - Read-only discipline: Never modifies application code; uses SELECT queries, tinker one-shots, existing tests, and logs for runtime inspection. - Structured handoff: Writes diagnosis.md and report.md in a reserved .planning/scout/S-NNNN directory and routes the result to /quick, backlog, or /recon via an operator decision. - Use Case: Run /scout OrderService::total() returns the pre-discount total to get a certified root cause with a pasted runtime transcript, then hand the verified diagnosis to a fix workflow. ## Quick Start Ask the agent to investigate a bug by running /scout followed by the bug description, error message, or failing test name.