RLM Process

Process large contexts using Recursive Language Model patterns from arXiv:2512.24601.

When to Use

Use /rlm-process when you need to:

• Analyze entire codebases without context overflow
• Process all pairs/instances in a dataset
• Aggregate information across many files
• Search through millions of tokens effectively

How It Works

┌─────────────────────────────────────────────────────────────┐
│  RLM Processing Pipeline                                     │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  1. ASSESS: Measure context size                            │
│     └─> Determine if RLM needed (>10K tokens)               │
│                                                             │
│  2. FILTER: Search without loading                          │
│     └─> Use grep/glob to narrow relevant content            │
│                                                             │
│  3. CHUNK: Break into processable pieces                    │
│     └─> By file | By section | By line batch                │
│                                                             │
│  4. PROCESS: Recursive sub-calls                            │
│     └─> Each chunk → sub-agent → structured result          │
│                                                             │
│  5. AGGREGATE: Combine results                              │
│     └─> Store in REPL state → synthesize final answer       │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Usage

Basic Usage

/rlm-process <query> <context-reference>

Examples

Codebase Analysis:

/rlm-process "Count all functions that handle authentication" src/

Cross-File Search:

/rlm-process "Find all pairs of files that import the same module" packages/

Large File Processing:

/rlm-process "Summarize each section of this document" large-doc.md

Data Aggregation:

/rlm-process "What labels are most common?" data/questions.jsonl

Invocation

When this skill is invoked, spawn the RLM processor agent:

Task tool:
  subagent_type: "rlm-processor"
  model: "sonnet"
  prompt: |
    ## RLM Task
    Query: <user's query>
    Context Reference: <path or reference>

    Use RLM patterns to process this context efficiently.
    Store results in REPL state.

Chunking Strategies

| Strategy | Use When | Example | |----------|----------|---------| | by-file | Processing codebase | Each .py file separately | | by-section | Document with headers | Each ## section | | by-line-batch | Line-oriented data | 100 lines per batch | | pairwise | Need to compare pairs | All (A,B) combinations | | adaptive | Unknown structure | Let RLM decide |

Results

Results are stored in REPL state:

# Check results
python3 ~/.claude/scripts/repl_state.py get rlm_result

# See all chunks
python3 ~/.claude/scripts/repl_state.py list | grep chunk_

When NOT to Use

• Small contexts (<10K tokens): Direct read is faster
• Simple lookups: Just use grep/read
• Real-time requirements: RLM has overhead

Integration

The /rlm-process skill integrates with:

• REPL State (/repl): Stores intermediate and final results
• Scout Agent: Can be used for chunk exploration
• Worker Agents: Spawned for parallel chunk processing

Reference

Based on "Recursive Language Models" (arXiv:2512.24601)

• MIT CSAIL: Alex L. Zhang, Tim Kraska, Omar Khattab
• Key insight: Treat context as environment variable, not neural input

Session Implementation

When the user invokes this skill:

1. Parse the query and context reference from args
2. Check context size using file stats or line counts
3. Spawn rlm-processor agent with the task
4. Report results from REPL state when complete

Example spawn:

# Orchestrator spawns RLM processor
Task(
    subagent_type="rlm-processor",
    model="sonnet",
    prompt=f"""
## RLM Task
Query: {user_query}
Context Reference: {context_ref}
Context Size: {estimated_size}

## Constraints
- Max tokens to read directly: 50000
- Output format: Markdown with citations

Process this context using RLM patterns.
Store final result in REPL state key: rlm_result
""",
    run_in_background=False
)