4.8 KiB
name, description
| name | description |
|---|---|
| local-brain | Delegate code reviews, document analysis, and planning tasks to local Ollama LLM models to reduce context usage. Supports lightweight hooks (ai, ai-cmd, ai-explain) for quick operations and heavyweight agent for multi-file reviews. Use when users request code reviews, design document summaries, ticket/issue triage, documentation analysis, planning, or routine pattern matching. Ideal for routine analysis that doesn't require cloud-scale reasoning. Do NOT use for complex multi-step reasoning requiring extensive codebase context or security-critical decisions. |
Local Brain - Context Offloading Skill
Tiered system for offloading work to local Ollama models, preserving main agent context.
Tiers
Tier 1 - Hooks (fastest, direct bash):
ai- Quick Q&Aai-cmd- Command generationai-explain- Explain last command
Tier 2 - local-brain binary (structured reviews):
- Single/multiple file reviews
- Directory reviews with patterns
- Git diff reviews
- Structured Markdown output
Tier 3 - Subagent (heavyweight, multi-file):
- Orchestrates multiple local-brain calls
- Handles complex multi-file analysis
- Coordinates multiple review tasks
Decision Logic
Use this flowchart to select the right tier:
User request
↓
Is it a quick question/explanation?
→ YES: Use Tier 1 (hooks)
→ NO: Continue
↓
Is it 1-3 files for review?
→ YES: Use Tier 2 (local-brain binary directly)
→ NO: Continue
↓
Multiple files OR multiple review tasks?
→ YES: Use Tier 3 (spawn subagent)
Prerequisites
- Ollama running locally with at least one model
- local-brain binary installed
- Hooks defined in
~/.zshrc(ai, ai-cmd, ai-explain)
Check prerequisites: which local-brain && ollama ps
See CLI_REFERENCE.md for installation and HOOKS.md for hook details.
Tier 1: Lightweight Hooks
When to Use
- Quick factual questions
- Command generation
- Explaining last command/output
- NO file reading needed
Usage
Quick Q&A:
ai "brief question"
Command generation:
ai-cmd "task description"
Explain last command:
ai-explain
See HOOKS.md for detailed hook documentation.
Tier 2: Direct local-brain Binary
When to Use
- Review 1-3 specific files
- Single directory review
- Single git diff review
- Want structured Markdown output
Usage
IMPORTANT: Do NOT read file contents first - that defeats the purpose of context offloading.
- Verify files exist:
ls path/to/file(do NOT use Read tool) - Run local-brain directly:
# Single file
local-brain --files path/to/file
# Multiple files
local-brain --files path/file1,path/file2
# Directory
local-brain --dir src --pattern "*.rs"
# Git diff
local-brain --git-diff
# With task type
local-brain --task quick-review --files path/to/file
- Parse and present the Markdown output sections:
- Issues Found
- Simplifications
- Consider Later
- Other Observations
Tier 3: Heavyweight Subagent
When to Use
- Multiple directories to review
- Multiple separate review tasks
- Need to coordinate multiple local-brain calls
- Complex multi-step analysis
Usage
Spawn subagent using Task tool with subagent_type=general-purpose and model=haiku:
Example prompt:
Review multiple files using local-brain without reading them into context.
IMPORTANT: Do NOT read file contents - offload to local-brain.
Prerequisites verified:
- local-brain: [path]
- Ollama: [status]
Tasks:
1. Review [file1] with local-brain --files [file1]
2. Review [file2] with local-brain --files [file2]
3. Review [dir] with local-brain --dir [dir] --pattern "*.ext"
For each review:
- Execute local-brain command
- Parse Markdown output
- Extract key findings
Return consolidated summary:
1. Critical issues across all files
2. Common patterns found
3. Recommended priority actions
Return complete analysis in final message.
Subagent Responsibilities
- Execute multiple local-brain commands
- Parse each Markdown output
- Consolidate findings
- Return structured summary
Output Handling
All tiers produce different outputs:
Tier 1 (hooks): Plain text responses Tier 2 (binary): Structured Markdown with sections Tier 3 (subagent): Consolidated cross-file analysis
After receiving results:
- Highlight critical items from "Issues Found"
- Summarize simplification opportunities
- Distinguish urgent vs. later improvements
- Ask if user wants to address specific findings
References
- CLI_REFERENCE.md - Installation, flags, troubleshooting
- HOOKS.md - Detailed hook documentation and usage