What it does
Impact Analyser answers the question a diff never shows you: what else breaks. You supply the affected-module code-graph data for a pull request or commit, and it scores every module and service by breaking-change risk and consumer fan-in — how many call sites actually depend on what moved. A finding reads "modified, fan-in 9 consumer call sites, BREAKING contract change, POST /v1/checkout" — enough to judge severity without opening the code. Alongside it you get the risk matrix, the specific breaking call chains a reviewer must check, the consumer services that need notifying, and a regression test set scoped to the blast radius rather than the whole suite. The design decision that matters most is how it treats missing data. A module whose fan-in is absent is treated as unmeasured, not safe. It is scored at least warning, printed as fan-in UNKNOWN, and always kept in the regression set. That is exactly the failure mode of an un-indexed cross-repo target, and the agent refuses to report it green — which is the difference between a tool you can trust before a merge and one that quietly reassures you. `score_pr_risk` rolls everything into a single merge decision: a 0-100 score with its band, the critical, warning and breaking counts, the consumer call sites at risk, and a recommendation. `list_impacted_consumers` answers "who do I have to tell" on its own. `list_breaking_points` and `suggest_regression_tests` each stand alone too, so you can ask one question without running the whole analysis. Scoring is deterministic and driven by real inputs — whether the change is contract-breaking, and how many consumers sit downstream. The language model only narrates the likely root cause and what to verify before merge; it never invents a module, a service or a number, and a template fallback keeps the agent useful with no key. One precondition it states rather than hides: cross-repo analysis assumes the data for every target repo has been provided or indexed. It does not verify that for you. No connectors and no credentials. Built for tech leads and staff engineers reviewing changes whose blast radius is wider than the diff suggests.
Example prompts
- Score the merge risk of this PR and give me a recommendation
- Which consumer services do I need to notify about this change?
- What are the breaking call chains a reviewer must check here?
