Autonomous orchestrator

The autonomous orchestrator dispatches AI coding agents end-to-end against the dependency-ready frontier of work. Operators tell it which task families it's allowed to pick from; it decides which task to ship next, allocates an isolated worktree, runs the pipeline, opens the PR, and surfaces decisions when human judgment is required.

RFC: RFC-0015 — Autonomous Pipeline Orchestrator

Operational runbook: docs/operations/orchestrator-promotion.md

CLI docs: pipeline-cli/docs/orchestrator.md

Why it exists

Every dispatch the operator makes by hand is a context switch. Picking the next ready task, opening a worktree, dispatching a developer subagent, watching for review feedback, signing the attestation — all of that is mechanical when the dependency graph + the Definition-of-Ready gate have already done the judgment work upstream. The orchestrator wraps that mechanical loop.

How it works

The orchestrator is a tick loop:

1. Frontier: read the dispatchable frontier from cli-deps frontier (tasks whose dependencies are all complete and that pass the DoR gate).
2. Admit: filter out tasks marked dispatchable: false, in flight in another worktree, or operator-blocked.
3. Dispatch: allocate a worktree, write the active-task sentinel, fan out to a developer subagent.
4. Pipeline: the developer runs the two-tier pipeline — implement, lint, test, review fanout, attestation sign, push, open PR.
5. Re-tick: on completion (success / failure / human-attention), record the outcome to events.jsonl and tick again against the new frontier.

Activation

The orchestrator is off by default. Opt in with:

export AI_SDLC_AUTONOMOUS_ORCHESTRATOR=experimental
node pipeline-cli/bin/cli-orchestrator.mjs start

status, tick, and start are the three subcommands. See cli-orchestrator --help for flags (--max-ticks, --tick-interval-sec, --dry-run).

When to use it

• You have a populated backlog with declared dependencies.
• You've signed off the DoR config (operator stays in the loop only when the gate flags ambiguity).
• Your subscription / API budget can absorb the dispatch rate (set --max-concurrent accordingly).

When NOT to use it

• You're iterating on a single task in a tight loop — /ai-sdlc execute <task-id> is more direct.
• You haven't tuned the DoR gate yet — without it, the orchestrator dispatches half-baked tasks and burns budget.