| name | areno-build-agentic-workflow |
| description | Create or debug an AReno multi-turn agentic dataset, run_agent implementation, tool schemas, tool execution, reward, loss masks, interactive TUI game with OpenAI-compatible LLM inference, or agentic training example. Use for text or multimodal agentic workflows, not ordinary single-turn rollout. |
Build an AReno Agentic Workflow
Inspect the nearest example under examples/agentic/ and public types in areno/api/agentic.py before implementation.
Develop on a local branch and pull committed changes into remote validation
hosts. Use ModelScope for remote checkpoint and dataset references.
python .agents/skills/areno-build-agentic-workflow/scripts/validate_transcript.py transcript.json
python .agents/skills/areno-run-training/scripts/inspect_dataset.py \
--dataset-path data.jsonl --loader examples/agentic/<name>/dataset_loader.py --algo gspo
Workflow
- Define one deterministic environment/game domain and valid dataset records.
- Keep dataset loaders processor/tokenizer independent.
- Define strict JSON tool schemas and bounded execution. Read references/message-contract.md.
- Build trajectories using AReno public agentic types. Preserve assistant tool-call and tool-result ordering.
- Reward the intended outcome and process separately; test invalid, partial, and optimal paths.
- Verify concurrency, timeout, cancellation, context truncation, multimodal content, and loss masks.
- Run a bounded agentic rollout, inspect one full transcript, then one real training step.
Do not fabricate missing tool calls or alter raw model text to make a trajectory appear valid. Surface parser/capability failures.