| name | autonomous-fleet-core |
| description | The portable, tool-agnostic ENGINE for running fully-autonomous multi-agent engineering jobs. Shipped mission skills (doc-sync, test-coverage, adversarial-review-and-fix) invoke THIS engine plus exactly one ADAPTER (orca, claude-code, grok, or another runtime). Eighteen additional missions are documented under docs/exploratory/missions/ and re-promote on real-run evidence. This core holds everything that does NOT depend on orchestration tool: self-orientation, fully-autonomous coordinator behaviour with file-ledger boolean gates, context-handoff to survive compaction, the worker-placement DECISION LOGIC (dependent vs independent), the PR-per-task pipeline with commits-preserved + conflict-aware merge + worktree cleanup, the empirical risk tiers, safety rails, secret hygiene, and commit/authorship policy. It speaks in PRIMITIVES; the ACTIVE ADAPTER maps each primitive to its tool's real commands. Load with a mission skill and one runtime adapter — do not run alone. |
| license | MIT |
| compatibility | Requires git and gh CLI in the target repository |
| metadata | {"author":"ravidsrk","version":"1.3.4","fleet-component":"core"} |
Autonomous Fleet — Core Engine (tool-agnostic)
You are the COORDINATOR for an autonomous multi-agent run. You are a THIN LOOP-HOLDER: create
tasks, spawn workers, dispatch, wait, answer worker questions from defaults, sequence the
pipeline, decide what runs next and how parallel. You do NOT review, code, or merge yourself — all
dispatched. Context stays light; the source of truth is the ledger FILE, not your memory.
Required composition
Do not run this skill alone. A mission skill defines the work; an adapter skill defines
runtime mechanics.
Instructions
Read and follow references/engine.md in full before coordinating any run — it is the slim ALWAYS-READ core capped at 300 lines. Its trigger map tells you when to load references/engine-autonomy.md, references/engine-workers.md, references/engine-review.md, and references/engine-recovery.md (issue #84).
It contains the engine identity, coordinator loop contract, ledger boolean gates, the 4-layer index, and trigger stubs for full doctrine.
For how missions, worker skills, optional skills, and campaigns compose, read
references/composition.md, references/community-skills.md,
references/fleet-outcome.md, and references/runtime-goals.md
for native /goal / update_goal binding. Per-repo defaults: setup-autonomous-fleet → docs/agents/fleet-config.md.
Mission chains and conditional DAGs use fleet-program — not a second mission loaded alongside
the first.
For review missions emitting structured findings (adversarial-review-and-fix and any future
reviewer phase), see references/review-findings.md: the
JSON schema, the source-verification CLI, the fix-strategy/confidence gating rules. For RUNTIME
enforcement of the engine's EVID/WT_CLEAN/e2e_verified disciplines via a Claude Code Stop hook,
see references/strict-mode.md: opt-in install of the stop-verify
gate that refuses session termination without verifiable evidence on disk.
Every run that emits first-class artifacts (findings, blind-fix files, verifier summaries) leaves a
manifest-audited trail under .fleet/runs/<run_id>/. See
references/run-archive.md for the manifest scheme and the
archive_enabled gate (a status: done outcome is rejected if its archive doesn't validate).
Substrate distribution (skills-install mode). The Python enforcement substrate (the
validators above plus lib/) ships WITH this skill under
assets/substrate/ — on a skills-installed repo it lives at
.agents/skills/autonomous-fleet-core/assets/substrate/ and each CLI runs standalone with
python3 (deps: assets/substrate/requirements.txt; version pinned in
assets/substrate/substrate-manifest.json). In a framework clone, prefer the canonical
scripts/ copies. Shell wrappers (preflight.sh, validate-fleet-outcome.sh,
run-sandboxed.sh) remain clone-only for now. setup-autonomous-fleet records the resolved
SUBSTRATE_PATH in docs/agents/fleet-config.md.
Ports of Agent Orchestrator mechanisms (nudge dedup, stacked PR, hook-signal, review supersede)
without adopting AO's daemon: references/ao-adoptions.md.
Primitives (summary)
The active adapter must implement: PLACE, SPAWN_WORKER, DISPATCH, WAIT, INSPECT,
WORKER_DONE / ASK / REPLY, OPEN_PR / MERGE_PR / CLEANUP, SYNC_TASK_STATE.
When the host supports goal mode, also implement: SET_GOAL, UPDATE_GOAL, GOAL_COMPLETE,
GOAL_BLOCKED, and LOOP_POLL (host-native scheduler; Orca exempt — see runtime-goals.md).
Optional primitive 14, CONTINUE_WORKER (re-attach an existing resumable session for an in-flight
task), is implemented by adapters whose runtime exposes a restore command and ALIASED to
SPAWN_WORKER (idempotent relaunch) otherwise — see engine-workers.md.
