mcp-kymograph
recording schema 4 · rules v4
Verdict · read back from the recording, nowhere else

A run that is not a result is not “did not find”

$ kymograph verdict {{run.path}}/events.jsonl
state
{{run.state}}
exit {{run.exit}}
{{run.meaning}}

Causes — a fixed key beside the words

{{run.entry}}
  • {{c.key}} {{c.words}} {{c.tag}}

{{run.note}}

What the recording holds — and where it falls short

events in order · the last line is read back, never assumed
  1. {{e.type}} {{e.detail}}

verdict.CAUSES — can the server bring it about?

decides how a repeated series treats the run: worst case, or run again
yes — stays in the worst case
{{k}}
no — not counted, the repetition is run again (at most 5 per key)
{{k}}
sandbox_failed is one key for two phases: before the target exists it cannot be the server’s; after, collecting the account can fail under what the target sent. Until it is split, it counts as “yes”.

A recorded cap is complete

One message records at most 20 findings. Past that, findings_capped names every statement and signal it cut, with the true numbers. What was cut, and why, is known — the run is complete. If that event is lost, or its numbers do not add up, the run is incomplete again (cap_unaccounted).

The same idea for pagination: a listing cut at the driver’s page limit is capped, not complete — a truncated listing must not read as a full one. A comparison marks a gone finding whose pair the later run cut (cap_hint).

Whose exit code is it?

drivethe run’s state; the server’s own exit is recorded (server_exited) and said on the terminal, never handed through runthe server’s — a client started us in the server’s place. A server ending with 2, 3, 4, 5 or 70 then looks like one of our codes; the terminal says so botha server that never started, or wrote nothing before it ended, is server_failed and exits 4 — there is no observed outcome to hand through verdictJSON on stdout, exits with the state’s code — the one place a run’s state lives is its recording