context-unlimited

name	context-unlimited
description	Evidence-backed context-unlimited probing using a persistent local Python REPL.
compatibility	cursor,opencode,claude
metadata	{"audience":"developers","mode":"local-probing","outputs":"json,evidence"}

What I do

I answer repo-wide questions by using deterministic local tools instead of stuffing the entire repo into the model context.

• Search with rg / fd
• Extract with Python (AST/regex) via a persistent REPL harness
• Emit machine-checkable JSON with evidence when requested: {path, excerpt, sha256}

How to use me

1) Preferred workflow (mechanically verifiable)

When the user asks for “find X / list Y / return JSON”:

1. Define a strict target output schema (JSON-only if requested).
2. Use local extraction:
   • rg -n "<pattern>" <paths...>
   • python -u integrations/claude_code/rlm_runner.py exec --code '...'
3. Store intermediate results in the REPL state (STATE[...]).
4. Retrieve final structured output:
   • python -u integrations/claude_code/rlm_runner.py get --var <var>
5. Return JSON only (no prose) unless explicitly asked.

2) Evidence format

If you claim anything about a file, include evidence records:

{
  "path": "path/to/file.py",
  "excerpt": "literal excerpt",
  "sha256": "sha256(excerpt)"
}

3) Safety constraints

• No destructive commands (no rm -rf, no mass rewrites).
• No network access.
• Prefer AST parsing over vibes.

Handy commands

Reset state:

python -u integrations/claude_code/rlm_runner.py reset

Get a variable:

python -u integrations/claude_code/rlm_runner.py get --var envvar_gold