evidence-investigatelisted

# Evidence Investigate ## Purpose Answer code-state questions through a dedicated read-only sub-agent so every factual claim is backed by a cited command output or file:line reference. This skill exists because main-thread Claude has repeatedly claimed things about running code from memory (grepping for a symbol name remembered from a commit message, for example) and been wrong. ## When this skill fires The skill description auto-triggers on questions like: - "Is function X defined in file Y?" - "What branch of the plugin clone is currently checked out?" - "Does the running task-board MCP import `debounce.ts`?" - "Is commit SHA X reachable from origin/main?" - "What files does branch A have that branch B does not?" - "When was the last modification to `server.ts` on main?" ## When this skill does NOT fire - Trivial file existence checks answerable by a single `ls` — just run the ls. - Implementation work (edits, writes, builds, deploys) — different skills. - Questions about mutable external state (APIs, databases, live processes, Telegram message history) — these are non-deterministic and not a good fit for an "evidence replayability" contract. - "Should we do X?" — that's a decision, not a fact. The sub-agent refuses these. ## Red Flags — STOP and Spawn the Sub-Agent These thoughts mean you are about to violate this skill. Stop. | Thought | Reality | |---|---| | "I remember this function is called X" | Symbol names recalled from memory are the #1 source of hallucina