exp: general-agent

[mikasenghaas]

Summary

• Add general-agent to pyproject.toml (envs list, workspace members, uv sources) and pull pytest-asyncio into the dev group so the env's tests are runnable.
• Add public configs/general_agent/ with three RLM configs using general-agent-solver-rlm, all logging to the general-agent-debug wandb project:
  • rl_qwen3_0p6b.toml — single-GPU smoke test
  • rl_qwen3_4b.toml — 4 train + 4 infer GPUs, max_steps=200
  • rl_qwen3_30b_a3b.toml — multi-node (1 train + 1 infer, dp=2 / tp=4), max_steps=400
• Bump deps/research-environments to c752781 (was origin/main at time of write; main has since advanced — re-bump before merge if a refresh is wanted). Env version bumps:
  • ddbc 0.1.1 → 0.1.2
  • ddbc_rlm 0.1.5 → 0.1.6
  • deepdive 0.2.7 → 0.2.9
  • deepdive_rlm 0.2.11 → 0.2.13
  • general_agent 0.1.0 → 0.1.4
  • opencode_deepdive 0.1.15 → 0.1.16
  • rlm_deepdive 0.2.3 → 0.2.4
  • rlm_swe 0.3.4 → 0.4.2
• Bump configs/private submodule pointer (now a merge commit 70c3503 that joins the PR's behavior-learning RESULTS writeups with main's rlm5 X-Session-ID header cleanup).
• Skip vf-eval-style TOMLs (detected via top-level eval list) in tests/unit/test_configs.py::test_load_configs so non-entrypoint configs don't fail validation.
• Document exp/ as the branch prefix for experiment branches in AGENTS.md.

Verification

• uv sync --all-extras rebuilds general-agent==0.1.4; entry point general-agent-solver-rlm resolves and vf.load_environment("general-agent-solver-rlm") returns a ComposableEnv.
• uv run pytest tests/unit/test_configs.py — 106 passed (covers all three new configs/general_agent/*.toml).

---

Note

Low Risk
Mostly new TOMLs, dependency wiring, and a targeted config-test skip; no changes to core training or auth paths in this diff.

Overview
Wires the general-agent research environment into the repo and adds RL experiment configs for general-agent-solver-rlm on Qwen3 at 0.6B (smoke), 4B, and 30B-A3B scales, all targeting the general-agent-debug W&B project.

Packaging: general-agent is added to the envs extra, uv workspace members/sources, and uv.lock (new editable general-agent==0.1.4; lock also bumps deepdive / opencode-deepdive versions shown in the diff). Dev deps gain pytest-asyncio for async env tests.

Configs: New configs/general_agent/rl_qwen3_{0p6b,4b,30b_a3b}.toml tune steps, seq length, GPU/deployment layout, orchestrator batch/rollouts, and inference parallelism for each model size.

Tests / docs: tests/unit/test_configs.py skips TOMLs with a top-level eval list (vf-eval, not prime-rl entrypoints). AGENTS.md documents exp/ branch prefix for experiment work.

^{Reviewed by Cursor Bugbot for commit 278ed64. Bugbot is set up for automated code reviews on this repo. Configure here.}