dev-rlm（开发 RLM）

Recursive Language Model (RLM) Skill

Core Philosophy

"Context is an external resource, not a local variable."

Three principles:

1. Never load what you can query — Filesystem is a database. Use rlm.py to query it.
2. The model decides the strategy — No fixed modes. Assess the task, pick the approach.
3. Recurse when complexity demands it — If a sub-task is too complex for one agent, that agent spawns its own sub-agents.

Context Engine (rlm.py)

The streaming query engine for filesystem interaction. Never loads all files into RAM.

# Codebase overview (no file reads)
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py stats
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py stats --type py

# Regex search across files (streaming)
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py grep "pattern" --type py

# Substring search with context window
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py peek "error_handler" --context 300

# Read single file or line range
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py read src/auth/login.py --lines 50-100

# Partition files for agent distribution
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py chunk --type py --size 15 --output /tmp/rlm_chunks.json

All commands support --output /path/to/file.json to write results to file.

Fallback: If rlm.py unavailable, use native tools: Grep, Glob, Read, rg, find.

Pipeline: Index → Filter → Map → Reduce

1. Index

Discover structure without reading file content.

python3 ~/.claude/skills/rlm/scripts/dev-rlm.py stats
python3 ~/.claude/skills/rlm/scripts/dev-rlm.py stats --type py

2. Filter

Narrow candidates programmatically.

python3 ~/.claude/skills/rlm/scripts/dev-rlm.py grep "TODO|FIXME|HACK" --type py
rg -l "error" --type py

3. Map (Parallel Agents)

Distribute filtered work across agents. See Strategy Selection below.

4. Reduce

Aggregate results from /tmp/.

jq -s '.' /tmp/rlm_*.json > /tmp/rlm_report.json
jq -s '[.[].findings] | add | group_by(.severity)' /tmp/rlm_*.json

Strategy Selection

Assess the task. Pick the strategy that fits. Combine strategies within a single analysis.

Strategies

Selection Logic

1. Run Index (stats). How many candidate files?
2. < 5 files: Peek or Grep+Read. Handle in main context. No agents needed.
3. 5-50 files: Fan-out Explore (questions) or Partition+Map (analysis).
4. 50-200 files: Partition+Map with coarse grouping. Consider Recursive Decompose if partitions remain complex.
5. 200+ files: Recursive Decompose. Split into domains at depth 0, let workers decide depth-1 strategy.

Do NOT pick a strategy before running Index. Let the data decide.

Agent Patterns

Fan-out Explore

Deploy Explore agents with complementary perspectives.

Task(
  description="Trace error propagation paths",
  prompt="Search for error handling patterns in this codebase.
  Focus on: try/catch, error types, propagation chains.
  Write summary to /tmp/rlm_errors.md",
  subagent_type="Explore",
  run_in_background=true
)

Assign each agent a distinct angle: architecture, patterns, specific modules, tests, dependencies.

Partition + Map

Split files into groups. Each general-purpose agent processes a partition.

Task(
  description="Analyze auth module (partition 1/4)",
  prompt="Analyze these files for security issues:
  [file list from rlm.py chunk output]

  Write findings to /tmp/rlm_p1.json as JSON:
  {\"partition\": 1, \"findings\": [{\"file\": \"\", \"line\": 0, \"issue\": \"\", \"severity\": \"\"}]}",
  subagent_type="general-purpose",
  run_in_background=true
)

Partition sources: rlm.py chunk output, directory boundaries, file type grouping.

Collect Results

TaskOutput(task_id=<agent_id>, block=true, timeout=120000)

Recursive Decomposition

When a sub-task is too complex for a single agent, that agent spawns its own sub-agents. Only general-purpose agents can recurse (Explore agents cannot spawn agents).

When to Recurse

An agent should recurse when:

• Its assigned partition has 50+ files and the analysis requires understanding, not just scanning
• It discovers distinct sub-problems (e.g., "this module has 3 independent subsystems")
• The prompt explicitly allows recursion

Depth Control

Hard limits:

• Max recursion depth: 2 (main → worker → leaf)
• Max total agents: 15 across all levels
• Leaf agents MUST NOT spawn sub-agents

Recursive Agent Prompt Template

Include these instructions when spawning agents that may recurse:

"You are analyzing [SCOPE]. You may spawn up to [N] sub-agents if needed.

RECURSION RULES:
- Current depth: [D]. Max depth: 2.
- If depth=2, you are a leaf. Do NOT spawn agents.
- Only recurse if your scope has 50+ files or distinct sub-problems.
- Each sub-agent writes to /tmp/rlm_d[D+1]_[ID].json
- After sub-agents complete, merge their results into your output file."

Output Routing

Guardrails

Limits

Iteration Control

• Each round of agent spawning should have a clear purpose
• If 2 rounds produce no new information, stop
• Never "try again" — refine the query or change strategy

Token Protection

• Agents write to /tmp/rlm_*, not to main context
• Main context reads only summaries and aggregated JSON
• Never cat agent output files raw into main context

Constraints

Never

• cat * or load entire codebases into context
• Spawn agents without running Index (stats) first
• Skip Filter stage for 50+ file codebases
• Exceed depth or agent limits
• Load rlm.py output raw into main context for large results

Always

• Use rlm.py stats before choosing strategy
• Filter with rlm.py grep or rg before spawning agents
• Write agent outputs to /tmp/rlm_*
• Include recursion depth and limits in recursive agent prompts
• Clean up /tmp/rlm_* after delivering results

Fallback (Without rlm.py)

If rlm.py is unavailable, use native Claude Code tools:

The pipeline and strategy selection remain the same. Only the tooling changes.

Integration

• rlm.py for Index/Filter stages
• Explore agents for fan-out investigation
• general-purpose agents for partition+map and recursive decomposition
• rg/grep as rlm.py fallback for Filter
• cli-jq for Reduce stage (merge and filter results)

Quick Reference

See quick-reference.md for decision tree and command patterns.

Credits

Based on the RLM paradigm (arXiv:2512.24601). Original skill by BowTiedSwan.