islinxu/yolo-master-skill/yolo-master-skill

YOLO-Master Agent Skill

Setup

This is a standalone Skill bundle. It requires a separate YOLO-Master checkout and uses its local Ultralytics package and yolo CLI. Set YOLO_MASTER_ROOT before invoking a script unless the bundle is nested in that checkout.

export YOLO_MASTER_ROOT=/path/to/YOLO-Master
python -m pip install -e "$YOLO_MASTER_ROOT"
python scripts/run_yolo_master_skill.py --json \
  '{"skill":"yolo.system","action":"doctor","params":{"ensure_cli":false},"policy":{"dry_run":true}}' --pretty

Supported Skills

• Core: yolo.train, yolo.lora.train, yolo.val, yolo.predict, yolo.track, yolo.export, yolo.benchmark, yolo.tune.
• Inspection and system: yolo.system, yolo.model.inspect, yolo.lora.adapters, yolo.lora.diagnose.
• MoE and evaluation: yolo.moe.diagnose, yolo.moe.prune, yolo.eval.peft_compare, yolo.eval.sparse_sahi_compare.
• Multimodal: yolo.multimodal.infer, yolo.multimodal.evaluate.
• Orchestration: yolo.pipeline.experiment, yolo.release.audit, yolo.job.status, yolo.job.cancel.
• Launchers: yolo.solutions.run, yolo.ui.launch.

Request Contract

{
  "skill": "yolo.train",
  "runtime": {"prefer_cli": true, "prefer_mps": true},
  "inputs": {"model": "yolo11n.pt", "data": "coco8.yaml"},
  "params": {"epochs": 1, "imgsz": 640},
  "policy": {"dry_run": true}
}

Use policy.dry_run=true before any costly operation. params carries task-specific YOLO arguments. On Apple Silicon, the dispatcher selects MPS for compute-heavy jobs when possible and retries once on CPU only when it selected the device automatically and encounters a device-runtime failure.

The dispatcher emits a structured response and a redacted, versioned skill_manifest.json. Experiment artifacts and manifests are written in the YOLO-Master checkout under runs/agent/ by default.

artifacts.project must resolve beneath runs/agent/; use a simple project label such as artifacts.project="trial-a" or an explicit contained path such as runs/agent/trial-a. artifacts.name is one safe path component. These are the only experiment-output controls: do not send params.project, params.name, or params.save_dir. Task-specific output values such as output_dir, output_path, output, and report_name must be relative children of the request's artifact directory; absolute paths and .. escapes are rejected. yolo.export retains params.name only for exporter target selection (for example rk3588), never as an output directory.

Examples

Train plan:

python scripts/run_yolo_master_skill.py --json \
  '{"skill":"yolo.train","inputs":{"model":"yolo11n.pt","data":"coco8.yaml"},"params":{"epochs":1,"imgsz":32},"policy":{"dry_run":true}}' --pretty

Multimodal plan:

python scripts/run_yolo_master_skill.py --json \
  '{"skill":"yolo.multimodal.infer","inputs":{"model":"yolo11n.pt","source":"ultralytics/assets/bus.jpg","prompt":"Identify the important objects."},"params":{"thinking_with_image":true,"structured_output":true,"prompt_template":"vlm_open_world_detection"},"policy":{"dry_run":true}}' --pretty

Release audit:

python scripts/run_yolo_master_skill.py --json \
  '{"skill":"yolo.release.audit","inputs":{"manifest":"runs/agent/experiment/skill_manifest.json"},"policy":{"dry_run":true}}' --pretty

Async job:

python scripts/run_yolo_master_skill.py --json \
  '{"skill":"yolo.train","inputs":{"model":"yolo11n.pt","data":"coco8.yaml"},"params":{"epochs":1,"imgsz":32},"policy":{"async":true}}' --pretty
python scripts/run_yolo_master_skill.py --json \
  '{"skill":"yolo.job.status","inputs":{"job_id":"<job_id>"}}' --pretty

Async jobs persist a redacted request snapshot, runner status, logs, and result.json in the Skill-local ignored logs/async-jobs/ directory. Poll until a terminal state: succeeded, failed, cancelled, or interrupted. yolo.job.cancel requests termination and returns cancelling until the runner confirms the terminal state. API keys must remain in environment variables; inline credentials and callback URLs are redacted and are never dispatched.

Multimodal Rules

yolo.multimodal.infer runs detector evidence first and optionally calls an OpenAI-compatible VLM/LLM service. Set OPENAI_API_KEY, or DASHSCOPE_API_KEY with params.provider="dashscope", only for real multimodal runs.

• Prompt templates live in assets/prompts/.
• structured_output=true enables parsed verdicts.
• use_marked_image=true enables numbered detection evidence.
• visual_search_mode=auto permits crop-and-zoom follow-up requests.
• fusion_mode=preview creates metric-safe fusion proposals; the default fusion_policy=add_only allows only guarded additions.
• fusion_policy=open_world_assist preserves novel unmapped predictions for reasoning, using the bundled LVIS/V3Det taxonomy.

Do not expose API keys in requests or artifacts. The manifest redacts secret-looking fields and inline credentials.

Validation

python scripts/validate_yolo_master_skill.py --suite quick --pretty --summary-only

quick combines fast smoke, dry-run, and contract coverage. Use all only when intentionally running the full non-manual suite. extended contains real CLI probes and is manual-only.

Release Audit

The release audit only consumes evidence explicitly referenced by a manifest, hashes every eligible file, and returns publishable, experimental, or refused. It never copies checkpoints, builds a model, or scans arbitrary run directories.

python scripts/audit_release_manifest.py runs/agent/experiment/skill_manifest.json \
  --output runs/agent/experiment/release_bundle.json --fail-on experimental