| name | rlm |
| description | Recursive Language Model (RLM) pattern for processing large contexts. Use when dealing with massive files, logs, repos, or data that exceeds context windows. Trigger phrases: "analyze large file", "process huge log", "scan entire repo", "recursive context", "RLM", "context compilation", "unbounded context", "too big for context". |
RLM — Recursive Language Model Protocol
Based on MIT's RLM paper (arXiv:2512.24601) and DSPy's structured REPL pattern. Instead of stuffing data into the token window, explore it programmatically through a structured protocol. Only printed results enter context.
Tokens are CPU, not storage. Never dump raw data into context. Write code to extract what matters, print only the summary.
When to Use
- File/data too large for context window (logs, databases, binaries)
- Codebase-wide analysis (grep across 100+ files, dependency graphs)
- Multi-step data extraction where each step depends on prior results
- Any task where raw data would burn tokens without adding value
Decision Logic
| Size | Protocol |
|---|
| < 5KB | Read directly — no RLM needed |
| 5KB–500KB | Steps 1-3 only (METADATA, PEEK, SEARCH) |
| 500KB+ | Full protocol steps 1-6 with sub-agent decomposition |
The 6-Step Protocol
Follow these steps IN ORDER. Each step uses python3 -c (or python -c on Windows) via Bash/shell. Raw data never enters context — only stdout does.
Windows note: Use python instead of python3. PowerShell commands like Get-Content, Select-String are also intercepted by the RLM hook — prefer python scripts over PowerShell for data processing.
Step 1: METADATA
Assess the file before touching it.
For multi-file discovery: Use Glob (Claude Code) or glob tool (OpenCode) to find files by pattern. Never use find via Bash — Glob is faster and keeps output compact.
WebFetch is blocked. Never use WebFetch/fetch to pull remote data into context. Instead, download via python3 -c using urllib/requests, save to a local file, then process that file through the protocol.
python3 -c "
import os
path = '/path/to/file'
size = os.path.getsize(path)
print(f'File: {path}')
print(f'Size: {size:,} bytes ({size/1024/1024:.1f}MB)')
print(f'Type: {os.path.splitext(path)[1] or \"unknown\"}')
with open(path, 'rb') as f:
head = f.read(200)
try: preview = head.decode('utf-8', errors='replace')
except: preview = repr(head)
print(f'Preview: {preview[:200]}')
try:
with open(path) as f:
lines = sum(1 for _ in f)
print(f'Lines: {lines:,}')
except: pass
"
Log: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_metadata','FILE_PATH',SIZE_BYTES,1)"
Step 2: PEEK
Sample strategically — head, tail, random slices, structure detection.
python3 -c "
with open('/path/to/file') as f:
lines = f.readlines()
print('=== HEAD (first 20 lines) ===')
for l in lines[:20]: print(l.rstrip())
print(f'\n=== TAIL (last 10 lines) ===')
for l in lines[-10:]: print(l.rstrip())
print(f'\n=== SAMPLE (every {max(1,len(lines)//10)}th line, 10 samples) ===')
step = max(1, len(lines)//10)
for i in range(0, len(lines), step):
print(f'L{i}: {lines[i].rstrip()[:120]}')
"
Log: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_peek','FILE_PATH',SIZE_BYTES,1)"
Step 3: SEARCH
Targeted extraction based on what PEEK revealed.
python3 -c "
import re
with open('/path/to/file') as f:
content = f.read()
# Adapt search to what you're looking for:
matches = re.findall(r'PATTERN', content)
print(f'Found {len(matches)} matches')
for m in matches[:30]:
print(m[:200])
"
Log: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_search','FILE_PATH',SIZE_BYTES,2)"
Step 4: ANALYZE (500KB+ only)
Decompose into sub-queries. Max 15 sub-queries.
Both Claude Code and OpenCode support sub-agents for parallel analysis.
For each chunk identified in SEARCH, spawn a sub-agent:
- Claude Code: Use the Agent tool to spawn sub-agents per chunk
- OpenCode: Use @explore sub-agents or session agents per chunk
- Pass ONLY the chunk + the specific question (never the full file)
- Each sub-agent returns a focused summary (max 1,000 chars)
To extract a chunk for a sub-agent:
python3 -c "
with open('/path/to/file') as f:
lines = f.readlines()
chunk = lines[START:END]
print(f'=== Chunk N ({len(chunk)} lines) ===')
for l in chunk: print(l.rstrip())
"
Log each sub-agent: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_analyze','FILE_PATH',CHUNK_SIZE,2)"
Sub-query types:
- Chunk analysis: split large file by sections
- Cross-reference: find callers/references across files
- Semantic filter: narrow down too many SEARCH results
- Recursive drill: sub-query reveals deeper structure to explore
Step 5: SYNTHESIZE
Combine findings from all sub-queries. Cross-reference. Resolve conflicts. This is reasoning — no code needed unless aggregating data.
Log: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_synthesize','FILE_PATH',SIZE_BYTES,2)"
Step 6: SUBMIT
Always end with an explicit SUBMIT block:
=== RLM SUBMIT ===
Query: [original question]
Confidence: [high/medium/low]
Protocol: [steps executed, e.g. METADATA->PEEK->SEARCH->ANALYZE->SYNTHESIZE]
Sub-queries: [N spawned, N completed]
Data processed: [size of original file]
Context used: [estimated tokens that entered context]
[Final structured answer here]
=== END ===
Log: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_submit','FILE_PATH',SIZE_BYTES,2)"
Confidence levels:
- high: All relevant data found, cross-referenced, no ambiguity
- medium: Answer found but some sections couldn't be fully analyzed
- low: Iteration budget exhausted or data was ambiguous
Iteration Budget
| Parameter | Limit |
|---|
| Max REPL iterations | 20 |
| Max output per step | 15,000 chars |
| Max sub-queries | 15 |
If you hit max iterations without resolving, SUBMIT with confidence: low.
Massive Data (50MB+)
Use rlm-cli which adds recursive sub-LLM decomposition:
rlm-cli query "Find all security issues" --file /path/to/large.log --backend openai --model qwen3:8b --base-url http://localhost:11434/v1
rlm-cli query "Find bugs" --repo /path/to/repo --backend openai --model Qwen/Qwen3-8B --base-url http://localhost:8000/v1
rlm-cli query "Analyze architecture" --repo /path/to/repo --backend anthropic --model claude-sonnet-4-6
Log: python3 -c "import sys; sys.path.insert(0,'${CLAUDE_PLUGIN_ROOT}/src'); from stats import log_event; log_event('rlm_cli','FILE_PATH',SIZE_BYTES,3)"
References
