Session continuity and context recovery for Claude Code. Auto-tracks your work via hooks, survives autocompact, and generates rich continuation prompts.

Note: this is an experimental plugin born from my config. It does work pretty well, but beware of bugs.

• Automatic Tracking: Hooks capture file edits, todos, plans, and user decisions as you work
• Session Handoffs: Create explicit checkpoints with /handoff for seamless continuation
• Autocompact Survival: Session markers survive context summarization for automatic recovery
• Rich Context Recovery: /start gathers project state and can restore from handoffs
• Optional Integrations: Enhanced context from beads, claude-harness, claude-mem, and agent-mail

# Add the marketplace (one-time setup)
/plugin marketplace add sethyanow/session-context

# Install the plugin
/plugin install session-context@sethyanow

Gather project context and recover from previous sessions.

/start              # Normal session start
/start handoff=abc  # Restore specific handoff

Create an explicit checkpoint before ending a session or when context is low.

/handoff

This generates a continuation prompt you can paste into your next session.

At session start, the plugin's SessionStart hook automatically gathers and injects session context directly into the conversation. This means:

• No MCP call needed - Session data is pre-loaded before /start runs
• Faster startup - Context is immediately available
• Token efficient - Eliminates round-trip to MCP server

The hook gathers:

• Git status (branch, uncommitted files, recent commits)
• Beads triage (if .beads/ exists)
• Harness state (if .claude-harness/ exists)
• Agent Mail status (if configured)
• Recovery checkpoint info (if available)

When you run /start, it uses this pre-loaded data directly instead of calling the MCP.

The plugin installs hooks that capture your work:

Data is stored in ~/.claude/session-context/handoffs/.

1. During work, hooks maintain a rolling checkpoint with your current state
2. When ready to end a session, run /handoff
3. Plugin generates a continuation prompt with:
   • What you were doing
   • Key decisions made
   • Next steps to take
4. Copy the prompt for your next session
5. New session detects the session marker and auto-recovers

The continuation prompt includes a session marker:

<!-- session:abc123 -->

This marker survives Claude's context summarization. When detected after autocompact, the plugin automatically loads full context from storage.

The plugin provides three MCP tools:

Gather session startup data with optional handoff recovery.

Note: Session data is typically pre-loaded by the SessionStart hook, making this call unnecessary for normal startup. Use this MCP tool as a fallback or when you need to restore a specific handoff.

get_session_status({
  level: "full",      // minimal | standard | full
  handoff: "abc123"   // Optional: restore specific handoff
})

Levels:

• minimal: project, harness, beads counts
• standard: + beads_triage, agentmail
• full: + skill_usage, claude_mem references

Create explicit checkpoint with continuation prompt.

create_handoff({
  task: "Implementing user auth",
  summary: "Added login endpoint, working on JWT",
  nextSteps: ["Add token refresh", "Write tests"],
  decisions: ["Using JWT over sessions"],
  blockers: []
})

Update rolling checkpoint (called by hooks).

update_checkpoint({
  files: ["src/auth.ts"],
  todos: [...],
  plan: { path: "...", content: "..." }
})

The plugin auto-detects and enhances with optional integrations:

Create ~/.claude/session-context/config.json:

{
  "checkpointTTL": "24h",
  "maxHandoffs": 10,
  "integrations": {
    "beads": true,
    "harness": true,
    "claudeMem": true,
    "agentMail": true
  },
  "tracking": {
    "files": true,
    "todos": true,
    "plans": true,
    "decisions": true
  }
}

Data is stored in ~/.claude/session-context/:

~/.claude/session-context/
├── config.json           # Plugin configuration
└── handoffs/
    ├── {hash}-current.json  # Rolling checkpoint (per-project)
    └── {id}.json            # Explicit handoffs

# Run the MCP server directly
cd mcp
bun run start

# Type check
bun run typecheck

# Build for distribution
bun run build

This plugin builds on and integrates with several excellent projects:

• Beads by @steveyegge - Dependency-aware issue tracking and task management
• Claude Harness by @panayiotism - Feature tracking and agentic loop management
• Claude-Mem by @thedotmack - Persistent memory and observation storage
• MCP Agent Mail by @Dicklesworthstone - Multi-agent coordination and messaging

Thanks to all these maintainers for their work on the Claude Code ecosystem!

MIT