Add together-rl skill (GRPO with sandboxed code-execution rewards)

[necoline]

Summary

Adds a new together-rl skill: a GRPO reinforcement-learning post-training loop driven by the Together RL training API (client.beta.rl — sessions + training.sample/forward_backward/optim_step), with rewards computed by executing model output in an isolated together-sandbox (write the candidate solution, run its test suite, exit 0 → reward 1.0).

This is the integration the rl-cookbook/grpo_train.py reference loop omits — that demo scores GSM8K with a local \boxed{} string match. Real coding/agentic RL has to run the model's output to score it, which must happen in a sandbox.

Files added

File	Purpose
skills/together-rl/SKILL.md	Routing, GRPO workflow, sync-RL/async-sandbox bridging rules, beta status
skills/together-rl/references/grpo-loop.md	Session lifecycle, the three operations + polling, GRPO sample schema
skills/together-rl/references/sandbox-rewards.md	Reward integration; hands off to the together-sandbox skill
skills/together-rl/scripts/grpo_sandbox_reward.py	End-to-end loop with code-execution rewards
skills/together-rl/agents/openai.yaml	UI metadata
quality/trigger-evals/together-rl.json	3 positive / 3 negative trigger cases

AGENTS.md / README.md / cursor plugin regenerated via publish.sh.

Design notes

• Doesn't re-document the API. The RL half points to the RL API docs + the /rl/training-sessions OpenAPI surface; the reward half hands off to the together-sandbox skill rather than restating the sandbox API.
• RL SDK is gated. client.beta.rl is in beta and not yet in the public together package (public together exposes beta.clusters/beta.jig only) and needs a service-specific base_url. The script targets the public surface and is marked beta; full end-to-end RL is deferred until that release.

Test plan

• quick_validate.py skills/together-rl passes
• quality_check.py passes
• publish.sh --check clean (AGENTS.md/README.md/cursor up to date)
• Sandbox reward path verified live — compute_rewards over a known-good and known-bad candidate returned [1.0, 0.0] via parallel fan-out
• Full RL end-to-end — blocked on client.beta.rl shipping in public together

Dependency

Hands off to the together-sandbox skill (#16) for the reward-execution step; merge that first.

🤖 Generated with Claude Code