Concept 6 of 10

Every move the model made, on the record.

The agent tried to deploy on a red build. Did it actually get stopped? 'The model did something' is not an answer — but there's a line in the log that is. Flowgate emits one structured event per step, for free, because it's already in the loop on every move.

You didn't add logging. It's already there.

That second line is the receipt: your agent reached for deploy before tests had run, and because deploy isn't a legal move yet, flowgate refused it — recorded as transition.rejected. You ran Claude Code or Cursor through the gateway and didn't write a line of logging; every workflow start, transition, and executor attempt — including the retries and the moves the agent tried and couldn't make — emits a structured JSON event on its own. Route it to stdout, a file, or wherever you keep logs. A correlationId threads a whole run together, so you can reconstruct exactly what the agent did.

audit.jsonl

{ "type": "workflow.started",     "definitionId": "fix_failing_test", "correlationId": "c_91a3" }
{ "type": "transition.rejected",  "transition": "deploy", "reason": "INVALID_TRANSITION", "correlationId": "c_91a3" }
{ "type": "executor.retrying",    "attempt": 2, "backoffMs": 400, "correlationId": "c_91a3" }
{ "type": "workflow.transitioned","transition": "run_tests", "to": "tests_green", "correlationId": "c_91a3" }

A log you can reconstruct from, not just read

Because the record is structured and correlated, it answers the questions you actually ask after a run: did tests really pass before it merged? How many retries did that refactor cost you? Which move did it try first? Filter by correlationId to pull a whole run in order; every executor attempt carries the hash of the exact script it ran — a fingerprint of the inputs — so you can reproduce the precise command that fired, not a paraphrase of it. (Those same structured events are also what feed dashboards and cost attribution later, if you grow into wanting them.)

replay.jsonl

# every event for one logical run, in order
$ grep '"correlationId":"c_91a3"' audit.jsonl

{ "type": "workflow.started",    "seq": 1, "correlationId": "c_91a3" }
{ "type": "executor.succeeded",  "seq": 2, "scriptHash": "sha256:9f2c…", "correlationId": "c_91a3" }
{ "type": "workflow.transitioned","seq": 3, "to": "tests_green", "correlationId": "c_91a3" }