AGENT SKILL

reasoning-agent-workflows

Enables future agents to operate RLinf reasoning, agentic RL, coding online RL, rollout backend, SFT, reward, and service workflows.

License 许可
Apache-2.0
最近更新
2026-10-08
Stars
★ 11
Forks
3

01 前言 Front Matter

name
reasoning-agent-workflows
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Enables future agents to operate RLinf reasoning, agentic RL, coding online RL, rollout backend, SFT, reward, and service workflows.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

RLinf reasoning and agent workflow router

Use this sub-skill when a task involves RLinf text reasoning, VLM reasoning, agentic RL, coding online RL, SearchR1/rStar2/WideSeek-R1 style tool workflows, AgentLightning, rollout backends, SFT/VLM SFT, or reward/service integration.

Route first

  • Install, Ray startup, node groups, generic placement syntax, and cluster health belong to the setup-and-cluster sub-skill.
  • Embodied simulator, robot, action-space, and VLA environment details belong to embodied-workflows.
  • Metrics, checkpoints, standalone evaluation operations, profiling, CI selection, and runtime log triage belong to operations-evaluation-debugging.
  • Registering a new model, environment, reward, algorithm, runner, worker, or tool parser belongs to extension-development.

Operating loop

  1. Classify the workflow: math/VQA reasoning RL, PPO with critic, multi-turn agentic RL, coding online RL, AgentLightning, SFT/VLM SFT, or reward-model training.
  2. Inspect the user-supplied YAML with scripts/inspect_agentic_config.py before launching any expensive run.
  3. Verify prerequisites: model/tokenizer paths, dataset format, required service endpoints, judge/search/code-execution credentials, and rollout backend memory budget.
  4. Build or edit only the user’s target config/launcher. Use the bundled references below rather than asking future agents to open source examples or docs.
  5. Escalate to setup/operations/extension sub-skills for out-of-scope cluster, evaluation, or source-code-extension work.

Bundled references

Safe helper

  • scripts/inspect_agentic_config.py reads one or more YAML files and reports task type, actor/rollout backends, length budget, algorithm settings, external services, and risky missing fields. It is static and non-mutating.

Evidence basis: distilled from RLinf v0.4.0 public configuration examples, English guides, runner/worker source, reasoning/VLM dataset loaders, reward registry, agent tool loops, rollout workers, and e2e configuration families.

04 同仓库更多 Skills