README.md

LAB + OpenCode adapter for nanoRSI

Runs the Harness² paper's LAB benchmark (harveyai/harvey-labs, MIT) through the paper's executor substrate (OpenCode) from the nanoRSI v0.5 loop, using nothing but the core's two command protocols:

• runner.py — NANORSI_REQUEST → stages the nano harness in the layout LAB's own opencode/run_opencode.py --harness-dir expects (the entry point the harness2 code itself calls), runs one OpenCode rollout, renders the session stream into a trajectory, text-extracts deliverables for the contrast step and keeps the binary deliverables under <run_dir>/output → NANORSI_RESULT.
• judge.py — NANORSI_JUDGE_REQUEST → LAB's evaluation.run_eval per-criterion rubric judge on <run_dir>/output → NANORSI_JUDGE_RESULT.
• prepare.py — writes a workspace: nanorsi.toml (command runner + judge, model, budget), lab_opencode.json (checkout paths, executor model, judge model, isolated OpenCode config dir, GNU-timeout shim), tasks/manifest.json from LAB task.json files, and the base harness = LAB's own system prompt with empty guardrails / memory / tool guidance — the same starting point the harness2 code uses.

Prerequisites are the ones recorded for the faithful harness2 run: a harness2 checkout with harness2-setup lab done (OpenCode source + bun + harvey-labs venv), the OpenCode provider file in an isolated XDG_CONFIG_HOME, the custom/ judge provider patch, coreutils (gtimeout as timeout on PATH), and the API key file outside the workspace. See examples/results/harness2-lab-opencode/setup/.

python examples/lab_opencode/prepare.py ./lab-ws --tasks <id>,<id>,...
nanorsi doctor --workspace ./lab-ws --check-model
nanorsi evolve --workspace ./lab-ws
nanorsi freeze --workspace ./lab-ws && nanorsi judge --workspace ./lab-ws
nanorsi report --workspace ./lab-ws