blast-radius

Blast radius

Find breakage a caller list will miss. The deliverable is not a convincing risk essay; it is the one safety-critical fact either proved against real behavior or marked unproven.

Evidence ladder

For every fact the change's safety depends on, climb as far as practical and report where you stopped:

1. An unsupported assertion—never sufficient.
2. A concrete file:line, pinned dependency source, schema, or contract.
3. A traced bad case showing whether it can reach the changed behavior.
4. A script or test that calls the real code and fails loudly if the fact is false.
5. A reproduction through the running product.

Any safety fact below level 4 stays explicitly unproven.

Tool availability

Need	Preferred route	Fallback
Change and history	Local git diff/log plus an available source-control connector	Local repository only; name missing PR context
Downstream consumers	Code search, schemas, generated clients, migrations, pinned dependency source	Trace the reachable local chain and mark external consumers unknown
Runtime proof	Existing tests, verification harness, or a small deterministic script	Minimal isolated repro; if execution is impossible, stop at the evidence level reached
Independent review	Invoke interrogate when the user authorizes a multi-agent review	One lead-review pass; disclose the missing independent signal

Process

1. Read the change. Inspect the diff, changed symbols, deleted behavior, commits, and surrounding contracts. State what now behaves differently, including effects the diff does not spell out.
2. Name the safety fact. Find the one or two assumptions that collapse most scary cases if true. Spend effort here instead of producing a long speculative list.
3. Look past symbol search. Check pinned library behavior and local patches; event timing and teardown; serialized data, database columns, wire formats, other-language consumers, feature flags, jobs, caches, and code several hops downstream.
4. Trace real risks. For each candidate, show a reachable failure path, likelihood, cost, and a concrete check. A search that finds nothing is evidence when the search scope is stated.
5. Prove the safety fact. Prefer the smallest script or test importing the same code the product ships. Run it and retain output. Reproduce through the product when proportionate.
6. Widen only when useful. For a broad or high-risk change, suggest interrogate; do not launch agents unless current policy and user intent allow it.

Output contract

• What changed: the visible and implicit behavior change.
• Safety fact: the assumption, evidence level reached, and proof output; otherwise UNPROVEN.
• Risks: only reachable risks, each with location, likelihood, impact, and check.
• Cleared: candidates investigated and why they do not reach.
• Before merge: the cheapest test or repro that catches the real failure.
• Coverage gaps: unavailable repositories, services, connectors, or runtime paths.

Use plain language, cite real code, and remove private material before public sharing.

<!-- Adapted from poteto/pstack blast-radius, MIT License, Lauren Tan, cursor/plugins@51a96e0. -->