orchestrate

name

orchestrate

description

Multi-goal harness — runs N `/cmax` pipelines in parallel for DIFFERENT goals at the same time. Each goal gets its own deepresearch → multispec → /goal → blind /verify, independent state, independent verdict. Rollup status at end. Effectiveness-driven defaults (--tdd --confidence 0.85). Use when you have several distinct things to ship at once and don't want to babysit them sequentially. Distinct from /parallel (which fans out one goal across N sub-Specs) and /hive (which gives N agents the SAME problem).

/orchestrate — N /cmax pipelines, one command

Ask for several things at once. Each gets its own full pipeline; they run in parallel.

CLI equivalent: cmax orchestrate "<goal A>" "<goal B>" "<goal C>" — or cmax multi ...

Pipeline (per goal, in parallel)

For each goal you provide, the orchestrator spawns a cmax ask "<goal>" --tdd --confidence 0.85 subprocess. Each subprocess runs the standard /cmax pipeline independently:

1. /deepresearch
2. multispec decompose
3. /specqa + /introspect gates
4. Parallel /goal per DAG leaf
5. Per-sub-Spec /verify (blind Opus)
6. Rollup /verify

Each goal writes to its own memory dir (.claudemax/state/orchestrator-<ts>/<goal-id>.memory.sqlite) so they don't collide.

Live status

A status table updates every 5 seconds (configurable via --tick-ms):

─── status @ 03:14:22 ───
  running    142s  add-health-endpoint-with-tests        add a /health endpoint that returns build sha…
  running    138s  migrate-user-model-to-drizzle         migrate the user model to drizzle, preserve a…
  finished    97s  add-otel-to-the-worker                add OpenTelemetry instrumentation to the wor…

Rollup verdict

all-finished (every goal shipped clean) / partial (some did, some didn't) / all-failed. Exit 0 only on all-finished.

Flags

Flag	Default	Effect
--goals-file <path>	(none)	Read newline-separated goals from a file (in addition to positional args)
--max-parallel <n>	all	Cap concurrent subprocesses. Default = run all goals in parallel
--variant opussonnet|opusolo	opussonnet	Per-goal exec tier — opusolo is ~3× cost, max effectiveness
--mode auto|solo|teams	auto	Per-goal parallelism mode for the inner multispec
--no-tdd	(TDD on)	Skip the TDD enforcement inside each goal
--confidence <n>	0.85	Verifier confidence threshold per goal
--tick-ms <n>	5000	Live status refresh interval

When to invoke /orchestrate

• You have several distinct features / fixes / migrations to ship and don't want to do them sequentially.
• Overnight: queue 5–10 goals before bed, wake up to a rollup verdict.
• You're evaluating which approach works — give 2-3 phrasings of the same goal, see which ones the verifier accepts.

When NOT to invoke /orchestrate

• One goal: just use /cmax or cmax ask.
• Goals that touch the same files heavily (writeSet overlap across goals → merge conflicts). Run them sequentially instead.
• You haven't yet exercised cmax ask against your billable pool once. Validate the single-goal path first.

Pre-split era warning (today, until 2026-06-15)

Per code.claude.com/docs/en/agent-sdk/overview (accessed 2026-05-21): until 2026-06-15, every cmax ask consumes your shared 5-hour rolling subscription pool. N parallel goals = N× faster envelope burn. Use a small N first (2–3) to feel the cost before queuing 10. After 2026-06-15 the harness auto-switches to the dedicated monthly Agent SDK credit pool.

Differentiation from sibling skills

• /orchestrate vs /parallel: /orchestrate runs N DIFFERENT goals (different deepresearch, different multispec each). /parallel fans out ONE goal into distinct packets.
• /orchestrate vs /hive: /orchestrate gives N distinct problems to N pipelines. /hive gives the SAME problem to N agents and merges their proposals.
• /orchestrate vs /cmax: /cmax is one pipeline for one goal. /orchestrate is N parallel /cmax invocations.