codastre

Decide WHICH repo or document set to open before searching inside one. Uses Codastre's CORPUS_SEARCH tool (CLI: `codastre corpora`) to rank whole corpora for ticket-, incident- or feature-shaped prose on aggregated evidence plus a per-repo identity card, then hands off to a QUERY scoped to the winner. Use when the input is prose rather than code vocabulary and the owning repo is unknown — a pasted Jira/GitHub ticket, an incident summary, a bug report, a feature description, or a service/repo name you can't place. Triggers: "which repo owns", "which service handles", "where should I start on this ticket", "customers report …", "investigate this incident", "find the code for this bug", "what repo is X in", a pasted ticket body.

Get the actual source behind a Codastre QUERY result when the response carries no `snippet` — resolve the hit's repo, find or create a local checkout, verify it is the exact indexed blob, then read the line range. Results are `(repo, path, line range, blob_sha)`, so on deployments where snippets aren't hydrated QUERY locates code without showing it. Use whenever a Codastre hit needs to be read and the repo may not be checked out locally, or whenever you are about to Read a `real_path`/`path_token` returned by QUERY. Triggers: "no snippet in the result", "read this codastre hit", "repo isn't checked out", "open the file codastre found", "clone the repo codastre pointed at", "which repo is this hit in".

This skill should be used for structural code questions — "what calls X", "what would break if I change X", "what consumes this Kafka topic", "which services talk to each other", impact analysis before a rename/refactor/delete, tracing a request across services, or fleet-wide boundary questions — "which topics have no consumer", "which routes does nothing call". It teaches the Codastre GRAPH tool (directions, edge kinds, depth, confidence/resolution) and the CONTRACTS tool (cross-repo route/topic orphans).

This skill should be used when searching or exploring a codebase — finding where something is implemented, locating definitions or usages, answering "where is X handled", "code that does Y", or "which repo owns Z". It teaches when and how to use the Codastre QUERY tool instead of Grep/Glob/rg/find, and when text search is still the right choice.

This skill should be used when the user asks to measure, demonstrate, or compare the cost/benefit of Codastre versus plain text search — "how many tokens did search use", "show the savings", "is codastre actually worth it", "compare codastre vs grep", or when running /codastre:tokens or /codastre:compare. It defines the measurement methodology so numbers are honest and reproducible.