dispatch-kit

Dispatch Kit — autonomous worker orchestration

Turn "here are issues" into "here are running workers." From inside a session you file an issue and dispatch a separate, durable Claude Code worker that implements it in an isolated git worktree, opens one PR, and stops. You review and merge. This is the execution-dispatch layer the SDD pipeline stops short of — to-prd/to-issues produce the spec; dispatch-kit runs it.

Model

• Issues + PRs both live in $ISSUES_REPO (the current repo by default).
• A slot is a git worktree of this repo at ../<repo>-wt-<id>, parked on wt/<id>. All slots share one .git object store, so a single fetch updates origin/<default> for every slot.
• Dispatch files/adopts an issue, resets an idle slot to task/<N> off the default branch, renders the worker prompt, and spawns a worker via the configured dispatcher.
• A worker implements the issue test-first, runs the green-light check, opens one PR, comments on the issue, and stops. It never merges.
• The human reviews + merges the PR (which closes the issue), then runs /reap to park the freed slot.

It's opt-in via the commands; normal direct work in the repo is unaffected.

Commands

Want to…	Use
Start a task from a one-liner	/dispatch "<text>"
Start a worker on an existing issue	/adopt <issue#>
See pool + PRs + issues	/status
Free slots for closed issues	/reap
Add/remove slots	/pool list | add <N> | remove <id>

Bootstrap (one-time, per repo)

# Optional config (else sensible defaults; ISSUES_REPO falls back to the gh origin):
mkdir -p .claude/brain && cp "${CLAUDE_PLUGIN_ROOT}/templates/config.sh.example" .claude/brain/config.sh

"${CLAUDE_PLUGIN_ROOT}/scripts/init-pool.sh" 3   # create 3 worktree slots
"${CLAUDE_PLUGIN_ROOT}/scripts/status.sh"        # inspect

state/ under .claude/brain is auto-gitignored; config.sh is committable.

Dispatchers (pluggable)

DISPATCHER (env or config.sh) selects what spawns the durable worker workspace:

• cmux (default) — a watchable cmux workspace per worker.
• tmux — a detached tmux session (tmux attach -t <name>). Portable fallback.
• plain — a backgrounded claude via nohup, PID + log in the slot's .claude/.

Each is a small file under scripts/dispatchers/ defining dispatcher_require, dispatcher_spawn, and dispatcher_close — add your own backend by dropping in another file.

Worker effort knob

WORKER_EFFORT (default max) encodes a real CLI trade-off:

• max — top raw reasoning, but multi-agent Workflow orchestration is OFF (the CLI gates orchestration on resolved-effort xhigh, and --effort wins the resolver).
• ultracode — expands to --effort xhigh --settings '{"ultracode":true}': orchestration ON, slightly lower raw effort.
• any other value passes through as --effort <value>.

Config knobs (.claude/brain/config.sh, all env-overridable)

ISSUES_REPO, WT_PREFIX, WT_BASE, WORKER_MODEL, WORKER_EFFORT, GREEN_LIGHT_CMD, DISPATCHER. Use : "${VAR:=value}" form so env still wins.

Dispatch flow you MUST follow

1. Take the one-liner. For non-trivial work, run to-prd → to-issues; for trivial work, draft inline from templates/issue.md.
2. Show the draft and WAIT for approval. This is the review gate — don't file until the user okays it.
3. On approval, write the body to a temp file and run scripts/dispatch.sh --create-issue --title "<t>" --body-file <f> (or --adopt <issue#>).
4. Report: issue #, slot id, branch, dispatcher, workspace name.

If dispatch reports "no idle slots", offer /pool add <N>.

Source of truth

• state/pool.json (gitignored) — slot allocation.
• GitHub — issue/PR state. /reap reconciles the two by closing-issue.