Multi LLM

Multi-LLM Routing

Route tasks to the optimal LLM provider for cost efficiency and capability matching.

Quick Reference

Provider Status

Check installed providers and authentication:

~/.claude/scripts/diagnostics/llm-status.sh

Scripts

llm-route.sh - Routing Decisions

# Auto-detect best provider
~/.claude/scripts/automation/llm-route.sh "analyze this large log"

# Force specific provider
~/.claude/scripts/automation/llm-route.sh -p gemini "summarize"

# With file input (checks size for routing)
~/.claude/scripts/automation/llm-route.sh -f large.log "what errors?"

# List providers with auth status
~/.claude/scripts/automation/llm-route.sh --list

llm-delegate.sh - Execute with Fallback

# Delegate to Gemini (falls back if unavailable)
~/.claude/scripts/automation/llm-delegate.sh gemini "summarize 500KB log"

# With timeout
~/.claude/scripts/automation/llm-delegate.sh -t 180 gemini "analyze codebase"

# Pipe content
cat large.log | ~/.claude/scripts/automation/llm-delegate.sh gemini "summarize"

# Disable fallback (fail if provider unavailable)
~/.claude/scripts/automation/llm-delegate.sh --no-fallback codex "generate API"

# Large prompts (>100KB) automatically use temp files to avoid ARG_MAX
~/.claude/scripts/automation/llm-delegate.sh gemini "$(cat large-codebase.py)"

# Multi-file content: use semicolons (NOT brace groups with newlines)
cat file1.py file2.py | ~/.claude/scripts/automation/llm-delegate.sh gemini "review"

# Or use command substitution for complex gathering
~/.claude/scripts/automation/llm-delegate.sh gemini "analyze: $(cat src/*.py)"

llm-logging.sh - Routing Analytics

# View recent routing decisions
~/.claude/scripts/lib/llm-logging.sh recent 10

# Today's stats by provider
~/.claude/scripts/lib/llm-logging.sh stats

# Rotate logs if >10MB
~/.claude/scripts/lib/llm-logging.sh rotate

Fallback Chain

When preferred provider fails:

Best-fit → Claude → Gemini → Codex → Copilot

Provider Capabilities

Gemini

gemini "summarize this" --output-format json
gemini -m gemini-2.5-pro "complex task"

Codex

codex exec --json "generate REST API"
codex exec --full-auto -s workspace-write "run tests"
codex exec -o output.txt "generate code"  # Save last message to file

Copilot

copilot -p "explain find -exec" --allow-all-tools
copilot -p "how to grep recursively" --allow-all-tools
copilot -i "interactive session"  # Start interactive with initial prompt

Claude

claude -p "review security" --output-format text

Cost Comparison

Savings Strategy

• Large context → Gemini: 10x cheaper on input
• Boilerplate → Codex: Lower per-token cost
• Reasoning → Claude: Best quality for complex tasks

Usage Examples

Large Log Analysis

# Gemini handles large context efficiently
~/.claude/scripts/automation/llm-delegate.sh gemini "Analyze errors in:
$(cat /path/to/large.log)"

Boilerplate Generation

# Codex optimized for code generation
~/.claude/scripts/automation/llm-delegate.sh codex "Generate TypeScript REST API for User model with CRUD, validation, OpenAPI docs"

Shell Command Help

# Copilot excels at shell explanations
~/.claude/scripts/automation/llm-delegate.sh copilot "explain: find . -type f -exec grep -l 'TODO' {} +"

Hybrid Approach

1. Delegate boilerplate to Codex
2. Review generated code with Claude
3. Apply Claude's security insights
4. Combine for best result

Configuration

Environment Variables

# Delegation timeout (seconds)
export LLM_DELEGATE_TIMEOUT=120

# Provider-specific models
export GEMINI_MODEL="gemini-2.5-pro"
export CODEX_MODEL="gpt-5.1-codex-max"

Provider Installation

# Gemini (npm)
npm install -g @google/gemini-cli

# Codex (bun or npm)
bun install -g @openai/codex
# or: npm install -g @openai/codex

# Copilot (standalone binary)
# Download from: https://github.com/github/copilot-cli/releases

# Claude
npm install -g @anthropic-ai/claude-code

Log Files

Log Schema

{
  "timestamp": "2026-01-01T12:00:00Z",
  "provider": "gemini",
  "prompt": "summarize this 500KB log...",
  "status": "success",
  "reason": "large_context_512000_bytes",
  "latency_ms": 3500
}

Troubleshooting

Provider Not Found

# Check all providers
~/.claude/scripts/diagnostics/llm-status.sh

# Verify specific provider
command -v gemini && gemini --version

Authentication Failed

# Check environment variables
echo $GEMINI_API_KEY
echo $OPENAI_API_KEY
echo $GH_TOKEN

# Or check config files
cat ~/.gemini/.env

Delegation Timeout

# Increase timeout
~/.claude/scripts/automation/llm-delegate.sh -t 300 gemini "large task"

# Or set globally
export LLM_DELEGATE_TIMEOUT=300

View Routing History

# Recent decisions
~/.claude/scripts/lib/llm-logging.sh recent 20

# Today's stats
~/.claude/scripts/lib/llm-logging.sh stats

Shell Syntax Errors

If you see {: command not found or similar errors when piping to llm-delegate:

Problem: Multi-line brace groups { ... } don't work in Claude's Bash tool.

Solutions:

# Use cat with multiple files
cat file1.py file2.py | llm-delegate.sh gemini "review"

# Use command substitution
llm-delegate.sh gemini "$(cat file1.py; echo '---'; cat file2.py)"

# Use semicolons for multiple commands
(echo "Header"; cat file.py; echo "Footer") | llm-delegate.sh gemini "analyze"

Integration

• batch-operations: Delegate multiple tasks in parallel
• context-optimizer: Suggest delegation when context bloated
• using-tmux: Foundation for CLI delegation