Skip to content

AGENTS.md

Latest commit

 

History

History
69 lines (48 loc) · 3.22 KB

AGENTS.md

File metadata and controls

69 lines (48 loc) · 3.22 KB
Raw

Meldog RL — Agent Guide (source of truth)

Model-agnostic brain for every agent (Claude/Opus, Fable, Gemini). CLAUDE.md / GEMINI.md are thin pointers here — put shared knowledge in this file.

Thesis (KNR robotics): reconstruct a robot-centric terrain map from partial RGBD on a 12-DOF quadruped (Meldog). Sim-first (Isaac Lab); a real robot exists — keep sim-to-real in mind.

1. Running Isaac Lab commands

Do not call bare python / conda run. Every training/eval/collection script must be wrapped in this exact sequence (exposes isaaclab + binds Vulkan/env vars):

bash -c 'source ~/.bashrc; eval "$(conda shell.bash hook)"; conda activate isaac-sim; cd /home/frydjak/IsaacLab; source _isaac_sim/setup_conda_env.sh; cd /home/frydjak/KNR/Meldog/LRT_meldog_RL; <YOUR_PYTHON_COMMAND>'

Example (short sanity training):

bash -c '...prefix...; python scripts/locomotion/train_locomotion.py --task Meldog-RL-Locomotion-Rough-Sim-v0 --num_envs 2 --max_iterations 10'

2. Where things are

See directory.md — curated map of local paths, registered task ids, and the external Isaac Lab reference configs (anymal / unitree, flat + rough). Read it before searching; agents cannot discover the external /home/frydjak/IsaacLab paths on their own.

3. Model tiering — who does what

Tag every task in a plan [FABLE] or [EXEC]. Heuristic:

If the hard part is "what should we do?" → [FABLE]. If it is "do this, precisely specified" → [EXEC] (Opus / Gemini).

Class Model(s) Use for
[FABLE] Fable 5 (claude-fable-5) Open-ended design, metric/architecture design, research & literature, thesis review, ambiguous judgment where being wrong is expensive.
[EXEC] (driver) Opus 4.8 (claude-opus-4-8) Implementing a specified change, git surgery, code review, wiring scripts.
[EXEC] (pipeline) Gemini Long-running training / eval / dataset-collection orchestration.
cheap Haiku 4.5 (claude-haiku-4-5-20251001) Mechanical file moves, boilerplate, log scraping.

[FABLE] tasks produce plans/designs; they do not implement or train. [EXEC] tasks execute against a concrete spec (exact file paths + current→proposed values). There is no separate fable.md — Fable is Claude and reads CLAUDE.md.

Prompts that put this into practice live in agentic/prompts.md.

4. Agentic workflow & shared memory

Cross-agent collaboration and rationale must be version-controlled in the repo (not in any single tool's private memory — Gemini can't see Claude's memory and vice versa).

  • Active tasks / experiments: agentic/<YYYY-MM-DD>_Task_*.md or _Experiment_*.md.
  • Archiving: move finished files to agentic/archive/. Do not read agentic/archive/ wholesale — it blows out context. Read active agentic/*.md to assess project state.
  • Reuse, don't re-derive: existing analysis (e.g. agentic/*Locomotion*, prior experiment logs) is authoritative — reference it instead of redoing it.

5. Manual maintenance

agentic/manual.md is the framework's user manual. Update it whenever you add or change scripts, CLI tools, task ids, or repo structure. Keep docs in sync with code.