/crew-diagnose [state] - a diagnosis per card, and how each would resume

Only the literal slash starts this pass: /crew-diagnose [state]. An ordinary message, a reaction, a follow-up or a message that merely talks about blocked cards starts nothing - answer those as normal chat.

Read-only. This turn looks and reports. It opens no card, retries nothing, lifts nothing, decides nothing, posts nothing, changes no assignee and writes no file. /crew-stop and the coordinator loop (scripts/crew_coordinator.py) are what act on the board; neither runs here, not even their scripts. The owner picks from your plan and starts the acting pass himself.

1. The list comes from one tool - the reader

crew_diagnose(state="<state>")

The state is whatever the owner named after the slash: blocked, triage, todo, ready, running, done, archived. With no state named, it is blocked. Never widen the state the owner named because the list came back short, and never guess one he did not name.

That reader call is the only tool call in this turn; quote its text field verbatim. No kanban_list, no kanban_show, no session store, no grep, no terminal command to "just check one card", no file write.

Every id, state, assignee, kind, wait and reason in your answer is copied from the reader's output (text). A card the reader did not return is left out; a reason is never retyped from memory, never shortened into a summary of your own and never softened. If the reader cannot read the board, or answers an unknown state, say that in one line and stop - do not fall back to another tool to reconstruct the list.

A first line WARNING: the gateway is not running ... is part of the reader's output: keep it as the first line of your answer, even when the list is empty, because a card cannot start without the gateway.

If the reader answers zero cards: one line, "no cards in state STATE", and stop. Do not query anything else to double-check.

2. What you add: the resume path for each card

Work from the reader's lines and the card's own title, nothing else. Per card, two short lines:

• RESUME: the smallest concrete step that would move this card - a decision only the owner can make, a dependency that is still open (name the parent card), a missing credential or access, a fix plus the command that would prove it, or a plain re-run.
• WHO: who takes that step - owner, coordinator, a named role, a named service.

Order the cards by the reader's wait, oldest first. When the reason is enough to name the resume path with certainty, say it plainly. When it is not, say in one line exactly what is missing to decide - never a guess dressed up as a plan, and never a plausible-sounding cause that the reader did not return.

3. Shape of the answer

1. One heading line: the state and the count, as the reader printed them.
2. One block per card: the id, state, assignee, kind, wait and reason exactly as read, then RESUME: and WHO:.
3. One closing line: how many cards wait on the owner, how many on another card, how many need a code fix.

Nothing else: no preamble, no "here is what I checked", no offer to run the acting passes, no card opened, no comment posted, no retry, no unblock, no dispatch, no board write of any kind in this turn. Ending your turn with the plan is the whole deliverable.