agent-execution-receipts

# Agent Execution Receipts Use this skill when the user asks who or what ran a command, whether an agent action reached a provider, or which agent-originated changes lack provider attribution. ## Procedure 1. Open the Cerebro Agent Receipts app or read `receipts.ndjson` from the configured receipt directory. 2. Verify each receipt signature and the append-only digest chain before using it as evidence. 3. Import a CloudTrail `LookupEvents` response when provider observations are available. 4. Report one of three states without collapsing them: - `Local evidence only`: a signed local receipt exists. - `Candidate correlation`: one provider event has a one-to-one evidence match, but no trusted binding. - `Provider bound`: an authenticated provider event passes the configured account, dedicated role, action ID, action, and time checks. 5. Start from the provider-event population and identify every provider mutation without a completed one-to-one action match. Never treat a process name, user agent, `startedBy` value, local transcript, session-wide identifier, or user-imported JSON as provider-bound attribution.