Multi-LLM Routing Skill

Multi-LLM Routing

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

Quick Reference

| Trigger | Provider | Why | |---------|----------|-----| | Input >100KB | Gemini | 1M token context | | "entire codebase" | Gemini | Large context | | generate/scaffold/CRUD | Codex | Cost-optimized | | security/architecture/debug | Claude | Best reasoning | | shell/command/CLI | Copilot | Shell expertise | | Default | Claude | Primary tool |

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"

# Code review with Codex (auto-detected from prompt keywords)
~/.claude/scripts/automation/llm-delegate.sh codex "review uncommitted changes"

# Force code review mode (uses 'codex review --uncommitted')
~/.claude/scripts/automation/llm-delegate.sh -r codex "check the hooks"

# 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

| Attribute | Value | |-----------|-------| | Binary | gemini | | Context | 1M tokens (2M coming) | | Auth | GEMINI_API_KEY or GOOGLE_API_KEY | | Best for | Large files, whole codebase, long documents |

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

Codex

| Attribute | Value | |-----------|-------| | Binary | codex | | Model | GPT-5 (default), GPT-5.2-Codex | | Context | 128K tokens | | Auth | OPENAI_API_KEY | | Best for | Code generation, CRUD, boilerplate, code review, refactoring |

# Code generation
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

# Code review (via llm-delegate.sh)
~/.claude/scripts/automation/llm-delegate.sh -r codex "review"
# Or auto-detected:
~/.claude/scripts/automation/llm-delegate.sh codex "review uncommitted changes"

Copilot

| Attribute | Value | |-----------|-------| | Binary | copilot (standalone) | | Auth | GH_TOKEN or GITHUB_TOKEN | | Best for | Shell commands, quick explanations |

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

| Attribute | Value | |-----------|-------| | Binary | claude | | Context | 200K tokens | | Best for | Architecture, security, debugging, review |

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

Cost Comparison (January 2026)

| Provider | Input (per 1M) | Output (per 1M) | Context | |----------|----------------|-----------------|---------| | Claude Opus 4.5 | $5 | $25 | 200K | | Claude Sonnet 4/4.5 | $3 | $15 | 200K | | Claude Haiku 4.5 | $1 | $5 | 200K | | Gemini 2.5 Pro (≤200K) | $1.25 | $10 | 1M | | Gemini 2.5 Pro (>200K) | $2.50 | $15 | 1M | | GPT-5 | $1.25 | $10 | 128K | | GPT-5 mini | $0.25 | $2 | 128K | | Copilot | Subscription | - | Limited |

Cost-Saving Options

Batch API: 50% discount (Claude, Gemini) for non-urgent tasks
Context caching: Up to 90% savings on repeated content

Savings Strategy

Large context → Gemini: 1M token context, cost-efficient for big files
Boilerplate → GPT-5 mini: Cheapest for code generation
Reasoning → Claude Opus 4.5: 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

Delegate boilerplate to Codex
Review generated code with Claude
Apply Claude's security insights
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-5.2-codex"  # or gpt-5 for default

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

| File | Content | |------|---------| | ~/.claude/data/logs/llm-routing.jsonl | Detailed routing decisions | | ~/.claude/data/hook-events.jsonl | Summary events |

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

Should NOT Attempt

Routing sensitive data to external providers without user awareness
Using external providers for tasks Claude handles well (unnecessary cost)
Delegating without checking provider availability first
Ignoring fallback chain when primary provider fails
Using Gemini for small tasks (overhead not worth it)

Escalation

Provider consistently failing → Check authentication and quotas
All providers unavailable → Fall back to Claude-only workflow
Routing decisions wrong → Adjust routing rules in llm-route.sh

When Blocked

If multi-LLM routing fails:

Run ~/.claude/scripts/diagnostics/llm-status.sh to check provider status
Verify API keys are set correctly
Try the fallback chain manually
If all providers fail, proceed with Claude only
Report which providers failed and why

Agent Skills: Multi-LLM Routing

Skill Files