rlm (Recursive Language Model)

Core Principle

Large content stays in the REPL environment, not in your context.

The REPL holds the full file in memory. You write Python to analyze it. Only your print() output returns to your context—never raw file content.

When to Use

• Files >100KB or >2000 lines
• Need to query the same large file multiple times in a session
• Structural or semantic analysis of logs, transcripts, codebases

Quick Start

# 1. Initialize (loads file into REPL memory)
python3 ~/skills/rlm/scripts/rlm_repl.py init <file>
# → Returns state path like: .pi/rlm_state/myfile-20260122-093000/state.pkl

# 2. Explore with Python (zero context cost)
python3 ~/skills/rlm/scripts/rlm_repl.py --state <state_path> exec -c "
hits = grep('ERROR')
print(f'Found {count(hits)} errors')
for item in expand(hits, limit=5):
    print(item['snippet'])
"

# 3. Only escalate to LLM when semantic reasoning is needed

---

The Escalation Ladder

Level 1: REPL Analysis (Default)

Use for: Pattern matching, structure extraction, aggregation, JSON parsing.

Context cost: Only your print() output returns.

python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> exec <<'PY'
import json

# The full file is available as `content`
lines = content.split('\n')
print(f'Total lines: {len(lines)}')

# Find patterns
hits = grep_raw('error|exception', max_matches=50)
print(f'Found {len(hits)} error lines')

# Parse specific lines
for i, line in enumerate(lines[:10]):
    if line.strip():
        data = json.loads(line)
        print(f"Line {i}: type={data.get('type')}")

# Aggregate
sizes = [(len(line), i) for i, line in enumerate(lines)]
sizes.sort(reverse=True)
print(f'Biggest lines: {sizes[:5]}')
PY

When Level 1 is sufficient:

• Finding all occurrences of a pattern
• Counting/sizing content
• Extracting fields from structured data (JSON, logs)
• Computing statistics

---

Level 2: REPL + llm_query()

Use for: Semantic reasoning where you need LLM judgment, but want results to stay in the REPL.

Context cost: Only your print() output returns. The LLM call happens in a subprocess.

python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> exec <<'PY'
# Extract errors with REPL (Level 1)
errors = grep_raw('ERROR', max_matches=10)

# Classify with LLM (Level 2) - reasoning stays in subprocess
for err in errors[:3]:
    snippet = err['snippet'][:2000]
    result = llm_query(f"Classify this error as critical/warning/info:\n{snippet}")
    print(f"Line {err['line_num']}: {result}")
    add_buffer(result)  # accumulate for later synthesis
PY

Batch processing:

python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> exec <<'PY'
# Process multiple items in parallel
chunks = list(Path(session_dir / 'chunks').glob('chunk_*.txt'))
prompts = [f"Summarize:\n{c.read_text()[:50000]}" for c in chunks[:10]]

results, failures = llm_query_batch(prompts, concurrency=5)
for i, result in enumerate(results):
    if "[ERROR:" not in result:
        add_buffer(f"Chunk {i}: {result}")
        print(f"Chunk {i}: {result[:100]}...")
PY

When to use Level 2:

• Classifying or categorizing content
• Summarizing sections
• Semantic search ("find discussions about X")
• Any task requiring judgment, not just pattern matching

---

Level 3: Subagent Synthesis

Use for:

1. Final answer generation after accumulating findings
2. Protecting main context when you'll query the same file many times
3. When synthesis itself is complex enough to warrant a fresh context

Context cost: ~5KB max per subagent (enforced by max-output-chars).

{
  "agent": "rlm-subcall",
  "task": "Query: <user's question>\nChunk file: /absolute/path/to/chunk_0001.txt"
}

For final synthesis of accumulated buffers:

# First, export buffers to a file
python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> export-buffers > findings.txt

# Then use subagent to synthesize
{
  "agent": "rlm-subcall",
  "task": "Synthesize these findings into a structured report:\n$(cat findings.txt)"
}

Parallel chunk analysis (when Level 2 isn't sufficient):

{
  "tasks": [
    {"agent": "rlm-subcall", "task": "Query: Find security issues\nChunk file: /path/chunk_0000.txt"},
    {"agent": "rlm-subcall", "task": "Query: Find security issues\nChunk file: /path/chunk_0001.txt"}
  ]
}

Limits:

• Max 8 parallel tasks per batch
• Expected output: ~2KB per chunk (JSON)
• Total subagent returns should stay under 400KB

---

Decision Tree

Is this a structural query? (find X, count Y, extract fields, parse JSON)
    └─ YES → Level 1: REPL
    └─ NO ↓

Do I need LLM judgment? (classify, summarize, interpret meaning)
    └─ YES → Does it need to return to my context immediately?
        └─ NO → Level 2: llm_query() in REPL
        └─ YES → Level 3: Subagent
    └─ NO → Level 1: REPL

Am I synthesizing final results from accumulated findings?
    └─ YES → Level 3: Subagent (protects your context for future queries)

Will I query this file multiple times in this session?
    └─ YES → Prefer Levels 1-2 (keep main context free for multiple queries)

---

REPL Reference

Initialization

python3 ~/skills/rlm/scripts/rlm_repl.py init <context_path>
python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> status
python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> reset

Environment Variables (available in exec)

• content - Full file content as string
• state_path - Path to state.pkl
• session_dir - Path to session directory
• buffers - List of accumulated text

Content Exploration

Function	Returns	Description
peek(start, end)	str	View slice of raw content
grep(pattern)	handle	Regex search, returns handle stub
grep_raw(pattern)	list[dict]	Raw results with line_num, snippet
write_chunks(out_dir)	list[str]	Write chunks to disk
add_buffer(text)	None	Accumulate text for synthesis

Handle System

Function	Returns	Description
count(handle)	int	Count items
expand(handle, limit)	list	Materialize items
filter_handle(handle, pattern)	handle	Filter results
last_handle()	str	Most recent handle name

LLM Queries (Level 2)

Function	Returns	Description
llm_query(prompt)	str	Single LLM call in subprocess
llm_query_batch(prompts)	(list, dict)	Parallel calls (max 5 concurrent)

Finalization

Function	Description
set_final_answer(value)	Mark JSON-serializable result
export-buffers (CLI)	Dump accumulated buffers

---

Chunking (when needed)

For very large files where you need to process sections:

python3 ~/skills/rlm/scripts/rlm_repl.py --state <path> exec <<'PY'
paths = write_chunks(str(session_dir / 'chunks'), size=200000)
print(f"Created {len(paths)} chunks")
PY

# Read the manifest
cat <session_dir>/chunks/manifest.json

Use manifest hints to skip irrelevant chunks before processing.

---

Context Protection

Budget: Assume 200K tokens (~800KB). Reserve:

• 50K for system prompt
• 50K for your reasoning
• 100K for tool returns (~400KB)

Warning signs:

• Single subagent returned >10KB → it misbehaved
• Total returns >400KB → stop and synthesize

If overwhelmed:

1. Stop dispatching more subagents
2. Synthesize from what you have
3. Use smaller batches (4 instead of 8)

---

Anti-Patterns

❌ Reading full chunks into your context with read tool ❌ Jumping straight to subagents for structural queries
❌ Dispatching subagents before exploring with REPL ❌ Ignoring manifest hints (processing all chunks blindly)

✅ Always start with REPL exploration ✅ Use grep to find relevant sections first ✅ Escalate only when semantic reasoning is needed ✅ Use subagents for synthesis, not initial analysis