Skip to content

SKILL.md

Latest commit

 

History

History
257 lines (211 loc) · 14.2 KB

SKILL.md

File metadata and controls

257 lines (211 loc) · 14.2 KB
Raw
name autonomous-fleet-adapter-orca
description First-class Orca adapter for autonomous-fleet-core — the reference runtime for supervised fleet missions. Maps engine primitives to Orca orchestration CLI commands (task-create, dispatch --inject, check --wait, worktree/terminal placement, worker_done/ask/reply, gh PR pipeline). Includes routing to orca-cli for full handoffs vs supervised fleet runs. Load with autonomous-fleet-core and one mission on Orca. Default roles: interactive @codex or @grok builds, fresh build-blind @claude reviews, @claude integrates. Trigger on Orca fleet runs, multi-agent PR pipelines on Orca, or when the host is Orca and autonomous-fleet is active.
license MIT
compatibility Requires Orca orchestration CLI, git, and gh CLI
metadata
author version fleet-component reference-runtime
ravidsrk
1.2.3
adapter
orca

Adapter: Orca (reference runtime)

Runtime: Orca — Settings → Experimental orchestration enabled; orca status --json must show a running runtime. Branch prefix: BRANCH_PREFIX from core self-orientation (default fleet/; recorded in DECISIONS.md).

Orca is the reference runtime for autonomous-fleet. It is the only host that gives structural build-blind review (separate terminals per role) and matches the production directives this framework was distilled from. Other adapters emulate Orca's loop via subagents and ledger polling; they are supported but secondary for distribution.

This adapter resolves the core's PRIMITIVES to Orca commands. Where Orca's CLI differs across versions, try X, fall back to Y — never hard-fail on one syntax.

Read references/orca-platform.md for the full routing table (fleet vs orca-cli full handoff vs companion orchestration skill).

ROUTING — fleet mission vs platform handoff

Before any primitive call, classify the user's intent:

Intent Path
Fleet mission or campaign (doc-sync, adversarial-review-and-fix, fleet-program, …) This adapter — supervised loop with task-create + dispatch --inject + check --wait
User says "hand off", "give to another agent/worktree" without asking to supervise/wait/DAG orca-cli full handoff — no task-create, no dispatch --inject, no check --wait
Lightweight terminal prompt, worktree ops, embedded browser orca-cli
User explicitly asks to supervise, monitor, wait for worker_done, coordinate a DAG Supervised orchestration (this adapter for fleet; raw orchestration skill otherwise)

Full-handoff examples (stop monitoring after prompt delivery):

orca worktree create --name <task> --no-parent --agent codex --prompt "<brief>" --json
# or: orca terminal send --terminal <handle> --text "<brief>" --enter --json

Custom Codex model/effort: create worktree, then terminal create --command 'codex --model …', wait tui-idle, terminal send — see references/orca-platform.md.

PRIMITIVE SUPPORT MATRIX (issue #93 — honest per-primitive status)

Primitive Status Mechanic
PLACE real orca worktree create / terminal --worktree active
SPAWN_WORKER real worktree create --agent + terminal wait --for tui-idle
DISPATCH real task-create + dispatch --inject
WAIT real native blocking check --wait --types …
INSPECT real task-list, inbox, dispatch-show
WORKER_DONE / ASK / REPLY real send --type worker_done, ask --to, reply --id
OPEN_PR / MERGE_PR / CLEANUP real gh + version-tolerant worktree remove/archive
SYNC_TASK_STATE real task-update --status (cross-session)
SET_GOAL family (9–12) absent by design no /goal API — file ledger + check --wait loop suffices
LOOP_POLL real check --wait loop
CONTINUE_WORKER aliased → SPAWN_WORKER (no documented session restore)

PRECONDITIONS (the core calls for these; here's the Orca form)

orca status --json (running runtime) · orchestration experimental flag on · gh auth status (else local merge-commits into BASE) · gitleaks availability · BASE exists (create off the default branch at current HEAD if absent).

Machine-readable preflight requires-block:

bins: [orca, git, gh]
env: []
auth:
  - check: "gh auth status"
    skip_if_intent: "no_scm"
intent_gated:
  scm: "willClaimExistingPR"

PRIMITIVE → ORCA COMMAND

PLACE(kind) → worker location

  • independent → NEW worktree on its own branch off BASE (parallel PR). Pass --repo REPO_ROOT. Git base: omit --base-branch for repo default (origin/main), or set explicitly — never the current feature branch unless stacked work is intended. --no-parent only affects Orca lineage.
  • dependent → ACTIVE worktree, fresh terminal session (same checkout; uncommitted state OK).
  • Fresh worker ≠ new git worktree. Review-fix cycles and same-branch validation use dependent placement only.

SPAWN_WORKER(role, placement) → handle, in auto/max mode

  • INDEPENDENT: orca worktree create --name <slug>-<run_short> --agent <cli> --repo REPO_ROOT --json → orca terminal list --worktree id:<newId> --json (read the handle) → orca terminal wait --terminal <handle> --for tui-idle --timeout-ms 60000 --json

  • DEPENDENT: orca terminal create --worktree active --title <slug> --command "<cli>" --json → orca terminal wait --terminal <handle> --for tui-idle --timeout-ms 60000 --json

  • Interactive agent CLIs (supervised Orca — primary path):

    Role --command / --agent Notes
    Builder codex or grok Mission role pipeline wins; @grok when mission says so
    Fresh build-blind reviewer claude Separate terminal — never the builder session
    Integrator claude PR open/merge only
    Design missions grok When mission specifies

    Apply each CLI's auto/yolo/skip-permissions + effort tier at spawn; log in DECISIONS.md. Do not use codex exec for supervised Orca fleet runs — that subcommand is for headless drivers on other adapters. For custom Codex --model / effort, use the terminal create --command 'codex …' pattern in references/orca-platform.md.

  • If an older CLI rejects worktree create --agent, create the worktree, then orca terminal create --worktree <sel> --command "<cli>" --json.

  • Ready = tui-idle. NEVER DISPATCH before tui-idle (inject on a non-idle terminal is lost).

  • Reuse an idle agent in the required worktree only when the mission allows; otherwise spawn fresh.

DISPATCH(task, handle)

Build the inject payload: (1) mission ## Worker skills for this role → Worker skills: "Activate and follow: <names>" per core engine.md; (2) task spec + completion contract (worker_done once, with taskId + dispatchId). Create task first: orca orchestration task-create --spec "<spec>" [--deps <json>] [--parent <id>] --json. Then: orca orchestration dispatch --task <taskId> --to <handle> --inject --json.

  • Bare shell target: omit --inject; track with task-create if needed, deliver spec via orca terminal send --terminal <handle> --text "<spec>" --enter --json.

WAIT(types, timeout)

orca orchestration check --wait --types worker_done,escalation,decision_gate,merge_ready --timeout-ms <n> --json. Returns ONE message at a time — loop N times for N concurrent finishers. Timeout / {count:0} = checkpoint, not failure. Heartbeats and terminal activity = alive, not done — never kill a live worker. Rolling 15–60 min windows; on timeout inspect task-list, terminal read, or terminal wait --for tui-idle as liveness before retrying.

INSPECT() — non-destructive

orca orchestration task-list --json · task-list --ready --json · orca orchestration inbox --limit <n> --json · orca orchestration dispatch-show --task <id> --json. (These do NOT mark messages read. check --all also exists on newer CLIs — use if available, don't depend on it. Reserve check --unread/default, which MARKS read, for when you intend to consume.)

WORKER_DONE(...) — exactly once, even on failure — SYNTAX-TOLERANT

Must carry taskId + dispatchId + a short summary + files modified. Try one form; if rejected, the other:

  • payload: orca orchestration send --to <coordinator> --type worker_done --subject "<short>" --body "<what·found·remains>" --payload '{"taskId":"<id>","dispatchId":"<id>","filesModified":["<path>"],"reportPath":"<opt>"}' --json
  • flags: same send … --type worker_done --subject --body with --task-id <id> --dispatch-id <id> --files-modified "<path>[,…]" --report-path "<opt>". Heartbeat likewise (payload {"taskId","dispatchId","phase"} OR --task-id --dispatch-id --phase). Target the CONCRETE coordinator handle from the live preamble — never a group (@all, @idle, @claude, @codex, @worktree:<id>, … are broadcast-only).

Review-only worker_done: when the worker is reviewer-only, completion reports findings — it does not authorize the coordinator to edit. Synthesize, gate if needed, dispatch fixes to the builder role (or orca-cli handoff when the plan names a next owner).

ASK / REPLY (worker blocking question / coordinator answer)

  • Worker: orca orchestration ask --to <coordinator> --question "<q>" --options "<a,b>" --timeout-ms 600000 --json (consume inline, e.g. | jq -r .answer).
  • Coordinator: orca orchestration reply --id <msgId> --body <answer> --json. Use gate-create/gate-resolve ONLY for your own DAG decisions, never to answer a worker's ask.

OPEN_PR / MERGE_PR(conflict-aware) / CLEANUP

  • OPEN_PR: gh pr create --base BASE --head <BRANCH_PREFIX><slug>-<run_short> --title "<title>" --body "<body>".
  • MERGE_PR: check conflicts first (gh pr view <n> --json mergeable,mergeStateStatus or a trial rebase). If conflicts: git fetch origin BASE && git rebase origin/BASE, resolve, re-test green, re-review if logic changed, force-push. Then gh pr merge <n> --merge --delete-branch (merge commit, commits preserved, NEVER --squash).
  • CLEANUP (WT_CLEAN gate): verify MERGED + branch-deleted FIRST. Apply core engine guard clauses — NEVER remove the active worktree; NEVER remove an unmerged or dirty worktree. Version-tolerant: orca worktree remove <id> or orca worktree archive <id> (try X, fall back to Y). Set task-row WT_CLEAN=true in the ledger; pull BASE.

SYNC_TASK_STATE(task, status)

orca orchestration task-update --id <taskId> --status <ready|dispatched|completed|failed|blocked> [--result '{"reason":...}'] --json on every lifecycle change, so task-list --ready/--status stay aligned with the file ledger (which remains source of truth).

TERMINAL TOOLKIT (coordinator)

orca terminal list [--worktree <selector>] --json
orca terminal create [--worktree <selector>] [--title <text>] [--command <cmd>] --json
orca terminal split --terminal <handle> [--direction horizontal|vertical] [--command <cmd>] --json
orca terminal wait --terminal <handle> --for tui-idle --timeout-ms <n> --json
orca terminal read --terminal <handle> --json
orca terminal send --terminal <handle> --text <text> --enter --json

GATES AND orchestration run

orca orchestration gate-create --task <id> --question <text> [--options <json>] --json
orca orchestration gate-resolve --id <gate_id> --resolution <text> --json
orca orchestration run --spec <text> [--max-concurrent <n>] [--worktree <sel>] --json
orca orchestration run-stop --json

Worker blocking questions → ask + coordinator reply (creates decision_gate message). gate-create is for coordinator DAG decisions only, not worker asks.

DIAGNOSTICS

  • Worker seems done but sent no worker_done: dispatch-show --task <id> --preamble --json; terminal read / terminal wait --for tui-idle as liveness. Re-send via terminal send if preamble was lost — never kill a live worker.
  • Inherited stale preamble from terminal history or a prior full handoff → treat as absent unless the current fleet prompt re-attaches a live coordinator loop.
  • NEVER orca orchestration reset during a fleet run (recovery-only when abandoning state).
  • Orca circuit-breaks a dispatch after 3 consecutive failures → reassign per engine, not stop.

RUNTIME GOALS (Orca — ledger-only)

Orca has no /goal API. Primitives 9–12 map to the file ledger only:

  • SET_GOAL: Write ## Runtime goal + CONDITION: in the program/mission ledger (documentation for humans and handoff). The coordinator check --wait loop IS the enforcement harness.
  • UPDATE_GOAL: Append progress to ledger LAST_UPDATE; optional send heartbeat to coordinator.
  • GOAL_COMPLETE: PHASE: DONE in ledger + FINAL report after TERMINATE checks.
  • GOAL_BLOCKED: escalation message + fleet-outcome.status: blocked.

Headless driver compatibility

scripts/run-mission-headless.sh and scripts/run-campaign.sh accept grok, claude, and codex CLIs only — not the Orca app. Production fleet runs on Orca use this adapter interactively in the Orca IDE: multi-terminal, cross-vendor roles, structural build-blindness. Headless scripts are a CI/contributor path for other adapters; Orca distribution does not depend on them.

ORCA NOTES

  • Primary coordinator mode: manual loop (task-create → spawn → dispatch --inject → check --wait) to keep file-ledger boolean-gate control — matches production directives. orchestration run is fallback only when the coordinator repeatedly stalls on context limits.
  • Group addresses (@all, @idle, @claude, @codex, …) are for broadcasts only — never for dispatch lifecycle messages.
  • Dependency chains ≤3–4 deep; one in-flight task per hot file; retire each worktree on merge.

RESUMABILITY + REVIEWER ISOLATION (Wave 3 contract)

Follow the single-sourced contract in autonomous-fleet-core → references/adapter-contract.md (issue #89 — do NOT re-inline it here; the drift lint fails copies). It covers run_short branch/worktree namespacing, the CONTINUE_WORKER resume discipline (RESUME_COUNT / MAX_RESUME_ATTEMPTS), and reviewer write-isolation. This adapter's only runtime-specific binding:

  • CONTINUE_WORKER binding: none documented -> ALIAS to SPAWN_WORKER (idempotent relaunch)