Smart Handoff System

Overview

This skill manages context-aware model switching via structured handoff manifests. It enables seamless work continuation across model switches by preserving project state, bug context, and model-specific instructions.

Model Fallback Chain

Policies:

• •on_context_limit → handoff and resume
• •on_reasoning_failure → escalate after 3 failed attempts

Context Degradation Signals

Watch for these throughout every session:

• •Conversation history becoming very long (many exchanges)
• •Needing to re-read files you already viewed earlier
• •Losing track of decisions made earlier in the conversation
• •Repeating analysis you already performed
• •3+ failed attempts at the same task (reasoning loop)

When Degradation Is Detected

Step 1: Stop and Alert

⚠️ Context degradation detected — recommending handoff.

Step 2: Generate Manifest

Read the template at ~/.antigravity-configs/templates/handoff_manifest.md and fill in ALL fields with actual session data. Save to:

• •Current conversation artifact dir as handoff_active.md
• •~/.antigravity-configs/handoff_active.md (global)

Step 3: Select Fallback Target

For reasoning loops (3+ failures): always escalate UP a tier.

Step 4: Inject Model Persona

For Gemini targets:

Persona: Multi-File Global Scanner

• •Scan ALL active files before making changes
• •Build cross-file dependency maps
• •Prefer breadth-first multi-file edits
• •Cast a wide net when debugging

For Claude targets:

Persona: Logical Precision & DRY Architect

• •Deep-read specific files from bug tracker first
• •Reason step-by-step before proposing changes
• •Check for DRY violations
• •Consider all edge cases

Step 5: Prompt the User

📋 Manifest: handoff_active.md
🤖 Current: [model] ([role])
➡️ Switch to: [target] ([role])
🎭 Persona: [persona name]

To continue:
1. Open a new chat
2. Switch global model to [target]
3. Type /resume

Context Compression

When conversation gets long but work is still productive:

1. •Identify fully-completed blocks (resolved bugs, finished edits, answered questions)
2. •Compress into context_summary_N.md in the conversation artifact dir
3. •Reference the summary instead of re-reading history
4. •Note: 📦 Compressed completed work into context_summary_N.md

Reasoning Failure Escalation

When you detect you're looping (3+ failed attempts at same task):

1. •Acknowledge it: 🔁 Reasoning loop detected — N failed attempts at: [task]
2. •Document in manifest under ## Reasoning Failure: what was tried, why each failed
3. •Escalate to next tier
4. •The incoming model MUST read the failure section and NOT repeat the same approaches

File Locations

• •Template: ~/.antigravity-configs/templates/handoff_manifest.md
• •Active manifest: ~/.antigravity-configs/handoff_active.md
• •Fallback config: ~/.antigravity-configs/model_fallback.json
• •Archived manifests: ~/.antigravity-configs/handoff_[timestamp].md