| name | orchestrator-parallelization |
| description | When the dispatcher has N tasks to execute -- use this before fanning out to decide whether to run in parallel or sequentially. Default is sequential; parallelize only when all three conditions hold (different file surface, independent semantics, predictable pattern). Different repos is the canonical parallel case; same-repo parallel needs worktrees. |
Parallelization heuristics
The default is sequential. Parallelize when ALL of these hold:
- Different file surface. Tasks A and B touch disjoint files.
Same-file parallel = merge-conflict hell.
- Independent semantics. Task B's prompt doesn't reference the
merged result of A.
- Predictable per-task pattern. First-run of a new dispatch
shape goes serial; parallel obscures which dispatch produced
which lesson.
Different repos is the canonical parallel case -- zero file
overlap by construction.
How to parallelize
- Each dispatch runs in its own branch + draft PR + watch loop. The
lifecycle is already parallel-safe.
- For same-repo parallelism: use
isolation: "worktree" on each
Task dispatch. Each runner gets its own git checkout; the
dispatcher's working tree is unaffected. Fan out N dispatches;
wait for notifications; push each worktree's changes as
notifications arrive, then remove the worktree. For Bash-based
dispatch, use claude -p with a separate worktree checkout
(or roba's -w if roba is installed).
- Multiple Task dispatches (or
run_in_background=true Bash calls)
fire simultaneously.
- Cap concurrency at 3-5. Beyond that, cognitive load and token
cost outpace wall-clock savings.
- Wait for ANY notification, then handle that one PR. The harness
notifies you per completed job.
- Merge in any order CI lands. The last few may need rebase if main
moved -- that's the runner's problem, not yours.
When NOT to parallelize
- Same files (sequential is faster end-to-end than rebase-conflict
resolution).
- Hard dependency (#X blocks on #Y -- finish Y first).
- Soft dependency (B's prompt references "the result of A" -- you'd
be writing B's prompt against a stale assumption).
- New pattern dogfooding (first run of a new dispatch shape, lean
serial so the lesson is clean).
- Review bandwidth (if YOU can't review N PRs concurrently, fanning
out costs you more than it saves the user).
Cost awareness
Parallel = N× tokens per round. Honest tradeoff. Worth it when
wall-clock matters (multi-repo work where the human's blocked); not
worth it for casual backlog grinding.
Anti-patterns
- Parallelizing tasks that touch the same files -- results in merge-conflict hell and costs more than sequential execution.
- Fanning out beyond 5 concurrent dispatches -- cognitive load and token cost outpace wall-clock savings above that cap.
- Fanning out before verifying task independence -- if B references A's result, parallelizing produces wrong output.
Related
dispatch-options -- the
dispatch mechanism each parallel run uses.
dispatch-wait-react -- how to
coordinate the multi-dispatch wait + notification fan-in.
orchestration-patterns --
which pattern you're parallelizing within (P1 same-repo
parallel; P2 multi-repo parallel is the natural case).