Assess, integrate, and validate external agents for Relax resident Agentic rollout.
npx skills add https://github.com/redai-infra/Relax --skill agentic-rollout
Please help me install this Agent Skill. Skill: agentic-rollout Source: https://github.com/redai-infra/Relax/tree/main/skills/agentic-rollout Command: npx skills add https://github.com/redai-infra/Relax --skill agentic-rollout
Connecting an external agent application to the Relax RL framework risks corrupting model-visible context, breaking export and credit semantics, or deadlocking on capacity limits. This Skill provides a staged workflow to assess an agent, bridge it into Relax with --use-agentic-rollout, verify launch readiness, and demonstrate a working resident Agentic rollout with evidence.
Assess my agent at examples/my_agent/ for integration with Relax agentic rollout and tell me whether it is ready to bridge.
High-intent search queries and answers about installing and using this skill.
Enable resident Agentic rollout with --use-agentic-rollout, --agent-command, and --agent-cwd, then wire the agent to the injected RELAX_BASE_URL and RELAX_SESSION_ID environment variables. The agent sends complete normalized messages with stable tools and chat_template_kwargs on every Chat Completions request.
Run the Stage C pre-launch check: resolve effective Agentic parameters after defaults, audit context linearity from wire payloads, verify external executor slot capacity, and confirm the Relax-facing client timeout covers prelaunch and partial-rollout holds. The verdict is PASS, UNSAFE, or UNVERIFIED.
No. Choosing --use-agentic-rollout gives the Agentic layer ownership of train and Eval generation, and any configured --custom-generate-function-path is ignored. Use the legacy custom-generate path instead when Relax-side code must own the model/tool loop.
A common cause is external executor slot exhaustion: incomplete multi-Session Groups never reach the all-Session first-request barrier. The signal is warming_prepare_group_count above zero with unchanged resident counts and recurring idle_heartbeat, requiring agent or platform cancellation to break.
Explicit export is required for any nonlinear history: intentional forks, multiple agents, deleted reasoning content, or changing tools between turns. Implicit export is reserved for audited strictly linear histories, and multiple exported contexts require --agentic-custom-advantage-path plus dynamic batching.
The Relax-facing client timeout applies to each individual Chat Completions request and must span prelaunch, partial-rollout, and fully-async holds. --agent-timeout is a separate active-time budget that terminates a stuck agent process and does not extend any request deadline.
aiagenta2zAccess Baidu Maps SSE tools through the OneKey Router.
DeepExperienceDiagnose stalled Ray distributed training tasks by tracing blocked execution paths.
box755Coordinate distributed PyTorch-native reinforcement learning training with TorchTitan and vLLM.
matlabConfigure Simulink C Function blocks programmatically using the SymbolSpec API.
