dispatch

Prompt file imported from uncleSoWise/gulp-khup (.github/prompts/dispatch.prompt.md). Fill in {{args}} before use. Copyright stays with the author.

/dispatch — fill orchestration slots

Operate in caveman mode (load the caveman skill) — orchestration is high-volume and tokens compound across rounds. Keep gh commands, labels, lock markers, and slot/PR numbers byte-exact; compress only the narration.

Arguments: {{args}}

Parse them as:

• A single number N is the total slot count. Default split: rescue = max(1, ⌊N/8⌋), flex = ⌊N/4⌋, dev = the remainder. /dispatch 8 → dev:5 rescue:1 flex:2.
• dev:<n>, rescue:<n>, flex:<n> tokens (anywhere in the args) set counts explicitly and always win over the derived split. Any subset is fine — unallocated remainder from a total goes to dev. /dispatch 8 rescue:2 → dev:4 rescue:2 flex:2.
• No arguments is an error: ask the operator for a slot count rather than guessing — slot capacity is operator policy (see orchestrating-slots).

Slot semantics (full protocol in the orchestrating-slots skill — read it before acting):

• A dev slot runs a sub-agent implementing a groomed issue from the queue. Dedicated: it refills with dev work when freed.
• A rescue slot is held for inbound incidents: CI failures on owned PRs, bot-review thread sweeps, DIRTY/stuck PRs, deployment or smoke-check failures. Dedicated: never backfill rescue capacity with dev work just because it's idle — rescue capacity exists so a queue incident never waits behind feature work.
• A flex slot works in both directions: fill it with dev work when the pipeline is healthy, but it is the first capacity reclaimed when incidents exceed rescue capacity — preempt its dev dispatch only if the work isn't yet near a push (otherwise let it finish, then convert). Flex slots are also the first to go idle under backpressure.
• A slot is held from claim-lock placement until the linked PR closes. Existing in-flight work you own counts against today's slots — reconcile before dispatching anything new.

Steps

1. Resolve identity and reconcile. ME=$(gh api user --jq .login). Survey open PRs by ownership: yours (each holds a dev or flex slot already), other operators' (their files become the conflict-zone list passed into every brief — never dispatch onto them), unassigned. Release any stale claims per the issue-locking skill (gh-issue-locking / jira-issue-locking).

2. Fill rescue slots first if rescue-type work already exists: owned PRs with failing required checks, unresolved bot-review threads, DIRTY state, or red deploys. These outrank new dev work — a blocked merge pipeline starves everything downstream. Overflow incidents claim flex slots next. Triage per the driving-prs-to-merge ladder before spending any slot (a flake rerun is free; a rescue dispatch is not).

3. Fill dev then flex slots from the issue queue: open issues sorted P0 → P1 → P2, excluding claimed, operator-held, and not-dispatchable labels (taxonomy in the issue control-field skill — gh-issue-labels for GitHub, jira-issue-fields for Jira — and the issue-filing skill — gh-issue-filing / jira-issue-filing). Bundle same-surface small issues into one dispatch — review cost is per-PR, not per-issue. For each dispatch: lock per the issue-locking skill (gh-issue-locking / jira-issue-locking), resolve model/effort per the control-field skill, write the eight-section brief per dispatching-subagents, and launch independent dispatches in the same message so they run in parallel.

4. Hold flex (then dev) slots when the queue is gated. If every open PR is blocked on one fix landing (a platform-wide advisory, a broken required check), dispatching more dev work only adds rebase thrash — fix the gate, hold the rest, resume filling after it clears.

5. Track and report. Keep one task-list entry per slot (so state survives a context reset). End with a slot table: slot #, type (dev/rescue/flex), assignment (issue/PR numbers), model, status. Then operate the event loop per orchestrating-slots — verify every agent completion (verify-then-trust) before refilling its slot, and rebalance flex slots toward whichever side (dev/rescue) is under pressure each round.

If the slot count the operator gave conflicts with what reconciliation finds (e.g. /dispatch 4 but you already own 5 open PRs), report the discrepancy and operate at the higher of the two rather than abandoning in-flight work — slots track reality, not aspiration.