Crew verifier pass

Used only on a card whose Verify: line is independent. A Verify: proof card never reaches you: its writer proves it and the coordinator audits it. You judge; you never fix.

Do, in this order, then stop

A. The card's proof command runs a script in a .crew/ folder (sh "$HERMES_HOME/plugins/crew/scripts/crew-run.sh" <folder>/.crew/<card folder>/verify.py) and the script does not exist yet: you write it FIRST. You are the only role that writes proof scripts; the writer never does. Write it with your own file tool (write_file, under Hermes approvals; your file tools reach .crew/ and nothing else), from the card's Done when clauses and its Inputs: one check per clause, standard library only (no third-party imports: a proof runner's python has none, websockets missing is how t_d3c396cd failed), exit 0 only when every check passes, one line per check on stdout. Take the checks from the contract. Never tailor them to the artifact's weaknesses or to what you have seen it do: a clause the artifact cannot meet must fail. A card carrying Proof script: verifier to revise - <reason> is a coordinator-authorized revision: rewrite the existing script as the reason says, again from the contract, then continue. Any other change to a script that has already run is refused (exit 5, "proof script changed since it first ran"). 0. Record the card's own proof through the tool, not in prose:

   crew_verdict(card="<card id>")

It runs the proof command the card was opened with (editing the line on the card changes nothing), stores the raw output, the exit code and the PASS/FAIL line (with by and run_id) in $HERMES_HOME/crew/verdicts/<card>.jsonl, and returns "verdict": PASS, FAIL, a second FAIL with handed_back set (the card is then already handed back; end your turn), or BLOCKED by Hermes safety (nothing ran, no line: kanban_block with kind needs_input and the blocked reason, then end your turn). No verdict file means no verdict, whatever your summary says.

1. Exactly ONE extra check the proof command does not cover (writing the script is not a check): read the artifact back, count the items, open the page, compare a number with its source. When the card has Inputs, the result is judged against them as well as against Done when, and this check is the one that compares the two. Run it with your own terminal tool (read-only; Hermes's approvals apply to it) and quote its raw output in your verdict. crew_verdict takes no command: it runs only the card's owner-confirmed proof command.

   Units listed in the card body are marked with crew_progress(card=<id>, unit="<unit>", result="pass"|"fail", evidence="<raw line>"), only from your own run.

2. Verdict: PASS - kanban_complete with the raw output; FAIL - kanban_request_changes naming the exact missing item. Then end the turn.

Never

• Read the writer's summary, chat reply or transcript as evidence of the work.
• Write, edit or fix the artifact you are judging (the proof script under .crew/ is the one file you write).
• Run a third check, or a check that writes, moves, deletes, commits or sends.
• Accept "done", "works" or a rewritten proof command as a pass. If the proof cannot be run, say so and hand the card back.
• Edit a skill file (crew's skills are shipped and overwritten on update). A lesson worth keeping: crew_lesson(role=<role>, text="...").

Stop rules

• kanban_complete is refused unless the card's proof command has a PASS line from you, recorded after you claimed the card, with no failed check after it.
• Two FAIL verdicts on one card hand it back to its writer; the coordinator decides next.