rlmlisted

# RLM — Recursive Language Model Workflow ## Arguments (from $ARGUMENTS) | Arg | Required | Default | Purpose | |-----|----------|---------|---------| | `query=<question>` | yes | — | What to answer | | `context=<path>` | no | — | Path to large context file | | `limit=<int>` | no | 10 | Max KB atoms to retrieve | | `chunk_chars=<int>` | no | 200000 | Characters per chunk | If arguments are missing, ask for `query` and optionally `context` before proceeding. ## Procedure ### Step 1 — KB retrieval Call `willow_knowledge_search` with the query and limit. Collect the returned atoms. Assess coverage: do the atoms answer the query fully? - **Yes** → synthesize from atoms and stop. Do not proceed to file chunking. - **No, and no `context` path given** → surface what's missing. Ask the user for a context file path. Stop here until they provide one. - **No, and `context` path given** → proceed to Step 2. ### Step 2 — Chunk the context file Write chunk files to `.claude/rlm_state/chunks/`. Create the directory if it doesn't exist. Run this inline Python to chunk the file: ```bash python3 - <<'PYEOF' import sys, os, pathlib context_path = "<CONTEXT_PATH>" # substitute $ARGUMENTS[context] chunk_chars = <CHUNK_CHARS> # substitute with integer, no quotes (e.g. 200000) import subprocess repo_root = pathlib.Path(subprocess.check_output(["git", "rev-parse", "--show-toplevel"]).decode().strip()) out_dir = repo_root / ".claude/rlm_state/chunks" out_dir.mkdir(parents=True, e