Feature Proposal: Agent Continuity - Persistent Agents Across Session Boundaries

Feature Proposal: Agent Continuity - Persistent Agents Across Session Boundaries

Summary

Proposal for a new feature enabling Claude Code agents to persist and resume across session boundaries. When a session ends unexpectedly (context limit, crash, user exit), agents can be resumed in a new session with full context restoration.

Problem Statement

Currently, when using parallel agents in Claude Code:

1. Session Loss = Agent Loss - Agents exist only in session memory
2. No Recovery Mechanism - When session ends, all context is lost
3. Manual Reconstruction Required - Users must manually restart work
4. Wasted Development Time - Significant work loss in multi-agent scenarios

Real-World Scenario

From a recent use case with 3 parallel agents:

• 3 agents spawned to work on different features in parallel
• Session ran out of context during development
• Compaction failed due to conversation size
• All agents became unreachable
• ~1,800 lines of work (540+670+470 lines) was lost
• No way to recover or resume

Proposed Solution

Agent Continuity - A comprehensive plugin + core infrastructure changes enabling:

1. Automatic State Persistence - Agents save minimal context snapshots (~10 KB)
2. Session Recovery - Full context injection when resuming
3. Orphaned Agent Detection - Automatic notification of agents from previous sessions
4. Intelligent Auto-Resume - Claude automatically detects and resumes previous work
5. Agent Lifecycle Management - Automatic archival and cleanup

Key Innovation: Minimal Context Snapshots

Instead of saving massive transcripts (1-10 MB), save focused snapshots (< 10 KB):

• Current objective
• What's been completed
• Known blockers
• Next steps
• Key file references

This enables fast context injection without massive file I/O.

Implementation Plan

Phase 1: Plugin - Can start immediately

• /resume-agents command
• Auto-detect resume agent
• SessionStart/SubagentStop hooks
• State file management

Phase 2: Core Infrastructure - Blocking for full functionality

• Agent identity system (CLAUDE_AGENT_ID)
• Registry management API with file locking
• State writing API for agents
• Context injection mechanism
• New hook events (PreAgentResume, PostAgentResume, etc.)
• CLI commands for agent management

Phase 3: Integration & Testing

• Full flow testing
• Edge case handling
• Performance verification

Why Now?

1. Multi-Agent Bottleneck - Parallel agents hit context limits frequently
2. Developer Friction - Session loss is unpredictable and costly
3. Feasible Architecture - Existing hook system provides foundation
4. Clear Spec - Complete design already documented

Design Specification

Comprehensive design specification available in the public repository:

📍 https://github.com/richardwhiteii/agent-continuity**

$ claude code
🔄 ORPHANED AGENTS DETECTED
Found 3 agents from previous session...
> /resume-agents agent_001 agent_002 agent_003
✓ All agents context loaded and injected