rsi

rsi command [options]

rsi is the shared command-line workflow of RSI-Anything, a research framework that turns world-model predictions and interaction feedback into recursive self-improvement. The paper applies the same propose → predict → select → evaluate → learn → retain cycle to code repair and GUI interaction (and, in the paper, laboratory planning, reaction optimization, and immune science). The CLI covers the two implemented task adapters: code_repair and gui.The console script is `rsi_anything.cli:main` (package rsi-anything 0.2.0). Install from a source checkout with `pip install -e .` (Python 3.10 or newer). The runner uses only the Python standard library; each task runs in its own interpreter named by the recipe's python field (`current` means the process that launched `rsi`).Code repair, full: verify assets → update initial / semantic / expanded / incumbent world models → evaluate. Evaluation-only recipes use `configs/code_repair/evaluate.json`. GUI, full: acquire models → check environment → freeze protocol → browser and world preflights → adapt and evaluate. GUI, smoke: real browser episodes with the rule controller.Each run lives under `run_root`: `config.json`, `state.json`, `summary.json`, `logs/workflow/`, and task-native checkpoints. Resume requires the same recipe and source identity. Incomplete code-repair outputs are moved into `attempts/` before retry.

tasks

Print a JSON object of tasks and recipe modes: `code_repair` (`full`, `evaluate`, `verify`) and `gui` (`full`, `smoke`).

Download or verify versioned experiment assets for code repair. Default --output is `assets`, default --profile is `evaluate`. --skip-models installs data and checkpoints only. --verify checks selected release files without downloading. No `gui` target exists on this subcommand.

Load the JSON recipe and print the resolved config, stage list, and generated files. Does not create a run directory or start workers.

Execute the recipe's stages in order. Without --resume, runroot_ must be empty (aside from the workflow lock). With --resume, completed stages are skipped after their recorded output hashes match; a changed recipe or source identity is rejected. A filesystem lock prevents two runners from owning the same run.

Print `state.json` from dir (stage status, attempts, output hashes).

Combine `state.json` with the task adapter's result files and print task, mode, status, stages, and metrics.

A recipe is a JSON file. Relative paths resolve beside that file. Allowed keys:task

`code_repair` or `gui`. Required.

Directory for this run. Required. Must not be the source checkout or a source/config/task subdirectory (`src`, `tasks`, `configs`, `docs`, `scripts`, `tests`, `.git`).

Code repair: `full`, `evaluate`, `verify`. GUI: `full`, `smoke`. Default `full`.

Worker interpreter: a path, an executable name, or `current`. Virtualenv paths keep their symlink so the selected environment stays active.

Path to the RSI-Anything source tree. Default is the installed package's checkout. Must contain `tasks/<task>/`.

Task-specific object. Code repair: `assets`, `traingpu`, `gpus`, `inputs`, `judgeworkers`, optional `checkpoint`. GUI: `worldmodel`, `policymodel`, `assets`, `policyport`, `resourceplan`, `downloadmodels`, smoke `episodes`. Path-valued keys (`assets`, `checkpoint`, `worldmodel`, `policy_model`) are resolved beside the recipe.

This is a research workflow, not a general-purpose crawler or trainer. Full recipes expect GPUs, large Hugging Face assets, and the task-specific virtualenv (`.venv-code-repair` or `.venv-gui` in the upstream guides). GUI workers need Playwright Chromium (`python -m playwright install chromium`) and the `rsianythinggui` extra.`rsi assets` only accepts `coderepair`. Changing scientific inputs needs a new `runroot`. Two processes cannot share a run directory. The CLI is installed from the GitHub tree; it is not a separate PyPI distribution under the name `rsi`.

RSI-Anything (Cui, Zeng, Xu, and others, 2026, arXiv:2610.13912) builds on JEPA-Anything. The `rsi` console script is the shared source-checkout interface that runs the code-repair and GUI recipes from that paper.

python(1), pip(1), playwright(1), mlflow(1)

Source code · Documentation