Model Escalation

Canonical implementation: Tools/model_escalator_v2.py

Complexity-based model switching with prefix tracking for OpenClaw agents.

Quick Start

from Tools.model_escalator_v2 import model_escalation_hook_v2 as model_escalation_hook

# Determine model from conversation context
result = model_escalation_hook("main-session", {
    "messages": [],
    "current_message": "Explain this API integration",
    "token_count": 0,
})
session_status(model=result.model)

Complexity Thresholds

| Range | Model | Prefix | Use Cases | |-------|-------|--------|-----------| | 0-0.25 | Haiku | [H] | Simple questions, confirmations, chat | | 0.25-0.75 | Sonnet | [S] | Code, debugging, API integration | | 0.75-1.0 | Opus | [O] | Architecture, complex algorithms, strategic planning |

High Complexity Indicators (Opus)

• System architecture design
• Distributed systems
• Multi-step technical implementations
• Deep code reviews
• Strategic planning with tradeoffs
• "Implement X mechanism/system" requests

Prefix Tracking

Ensures conversation prefix matches active model:

from Tools.model_prefix_tracker import ModelPrefixTracker

tracker = ModelPrefixTracker()

# Validate and correct prefix
correction = tracker.correct_prefix(
    session_id='main',
    current_prefix='[S]',
    current_model='anthropic/claude-opus-4-5'
)
# Returns '[O]' if mismatch detected

Logs all discrepancies to SQLite for analysis.

Integration with AGENTS.md

Add to your agent's AGENTS.md:

## Model Self-Check

At response START, assess complexity:
- High (0.75+) → `session_status(model="anthropic/claude-opus-4-5")`
- Medium (0.25-0.75) → `session_status(model="anthropic/claude-sonnet-4-5")`
- Low (<0.25) → Stay on Haiku

Prefix responses with [H], [S], or [O] to match model.

Configuration

Edit scripts/config.json to customize:

{
  "thresholds": {
    "low": 0.25,
    "high": 0.75
  },
  "models": {
    "haiku": "anthropic/claude-3-5-haiku-20241022",
    "sonnet": "anthropic/claude-sonnet-4-5",
    "opus": "anthropic/claude-opus-4-5"
  },
  "cooldown_minutes": 5
}

Database Schema

Prefix tracking logs stored in SQLite:

-- prefix_log table
timestamp REAL,
session_id TEXT,
detected_prefix TEXT,
actual_model TEXT,
correction_applied INTEGER

Query discrepancies:

sqlite3 model_prefix_state.db "SELECT * FROM prefix_log WHERE correction_applied=1;"