autonomous-fleet on Orca is a supervised fleet-mission layer on top of Orca's native
orchestration. This reference maps every case in Orca's orchestration skill to either this
adapter (fleet path) or a companion skill (orchestration, orca-cli).
| User intent | Skill / path | Lifecycle (taskId/dispatchId)? |
|---|---|---|
Run a fleet mission (doc-sync, adversarial-review-and-fix, fleet-program, …) |
This adapter + core + mission | Yes — supervised coordinator loop |
Supervise workers, wait for worker_done, coordinate a DAG |
Orca orchestration skill (this adapter implements it for fleet) |
Yes |
| "Hand off", "give to another agent", one-shot ownership transfer | orca-cli — worktree create --prompt or terminal send |
No — do not task-create / dispatch --inject / check --wait |
| Shell, worktree mgmt, read/wait terminals, embedded browser | orca-cli |
No |
| Desktop / external browser UI | Computer Use (outside Orca embedded browser) | No |
Fleet coordinators always classify before spawning: if the user did not ask to supervise,
monitor, wait for results, or coordinate a DAG, route to orca-cli — not this adapter.
When ownership transfers without supervision:
# New top-level worktree + agent + prompt (no orchestration lifecycle)
orca worktree create --name <task> --no-parent --agent codex --prompt "<brief>" --json
# Existing terminal
orca terminal send --terminal <handle> --text "<brief>" --enter --jsonCustom Codex model/effort (Orca cannot pass model flags on worktree create --agent):
orca worktree create --name <task> --no-parent --json
orca terminal create --worktree id:<id> --title <task> \
--command 'codex --model <model> -c model_reasoning_effort="<effort>"' --json
orca terminal wait --terminal <handle> --for tui-idle --timeout-ms 60000 --json
orca terminal send --terminal <handle> --text "<brief>" --enter --jsonAfter prompt delivery on a full handoff: stop monitoring unless the user explicitly asked for supervision.
--no-parentcontrols Orca lineage (top-level worktree in the Orca tree), not the Git base.- For independent fleet workers: pass
--repo REPO_ROOT; omit--base-branchto use the repo default (origin/main/orca repo show), or pass it explicitly. Never base on the current feature branch unless the mission requests stacked work. - Fleet isolated branches still use
BRANCH_PREFIX+<slug>-<run_short>per engine namespacing.
Include in check --wait --types:
worker_done, escalation, decision_gate, merge_ready
merge_ready: integrator may proceed toMERGE_PRwhen review flags are true in the ledger.decision_gatefrom workerask: answer withreply, notgate-create.gate-create/gate-resolve: coordinator DAG decisions only.
When the worker role is reviewer-only, worker_done reports findings — it does not authorize
the coordinator to edit files. After review-only completion: synthesize findings, use a decision
gate if ownership is unclear, then dispatch or orca-cli hand off fixes to the named builder role.
If the target terminal is a bare shell (not a recognized agent CLI): omit --inject on dispatch,
then deliver the task with orca terminal send --terminal <handle> --text "<spec>" --enter --json.
On Orca, fleet builders run as interactive agent terminals, not headless codex exec:
| Role | Terminal command | Notes |
|---|---|---|
| Builder | codex or grok |
Mission role pipeline wins; apply auto/yolo per CLI at spawn |
| Fresh build-blind reviewer | claude |
New terminal; never the builder session |
| Integrator | claude |
Opens/merges PRs; does not author fixes |
| Design missions | grok |
When mission specifies |
Log spawn flags + model/effort in DECISIONS.md. codex exec is for headless drivers on
other adapters — not the primary Orca supervised path.