Continues AI Session Transfer

continues-ai-session-transfer

Skill by ara.so — Devtools Skills collection

What it does

continues solves the context-switching problem when AI coding tools hit rate limits or need to be switched mid-session. It extracts conversation history, file changes, tool activity, and working state from one AI coding assistant and transfers it to another. Supports 16 tools with 240+ cross-tool handoff paths.

Supported tools: Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, Cursor, Amp, Cline, Roo Code, Kilo Code, Kiro, Crush, OpenCode, Factory Droid, Antigravity, Kimi CLI, Qwen Code

Installation

No installation required with npx:

npx continues

For permanent installation:

npm install -g continues

This gives you both continues and cont as global commands.

Requirements: Node.js 22.5+ (uses built-in node:sqlite)

Core commands

Interactive session picker

Launch the TUI to select any session across all tools:

continues

Automatically prioritizes sessions from the current project directory.

Quick resume in same tool

Resume the most recent (or Nth most recent) session from a specific tool:

continues claude          # Latest Claude Code session
continues codex 3         # 3rd most recent Codex session
continues cursor          # Latest Cursor session
continues gemini          # Latest Gemini CLI session
continues cline           # Latest Cline session
continues amp             # Latest Amp session

Works for all 16 supported tools. Uses native resume with full history.

Cross-tool handoff

Transfer a session from one tool to another:

# Basic handoff by session ID
continues resume abc123 --in gemini

# With tool-specific flags
continues resume abc123 --in codex --yolo --search

# Pass extra directories to the target tool
continues resume abc123 --in cursor --add-dir /tmp/shared

# Inspect the handoff prompt without launching the tool
continues resume abc123 --in gemini --debug-prompt

List sessions

# Table view
continues list

# Filter by source tool
continues list --source claude

# JSON output for scripting
continues list --json

# JSONL output (one session per line)
continues list --jsonl

# Limit results
continues list -n 10

Inspect session details

View detailed diagnostic information about a session:

# Full diagnostic view
continues inspect abc123

# Export to markdown
continues inspect abc123 --preset full --write-md handoff.md

# Compact one-liner view
continues inspect abc123 --truncate 50

Bulk export sessions

Export all sessions to files for backup or analysis:

# Export all sessions to markdown
continues dump all ./sessions

# Export specific tool's sessions
continues dump claude ./sessions/claude
continues dump cursor ./sessions/cursor

# Export as JSON
continues dump all ./sessions --json

# Control detail level
continues dump all ./sessions --preset full

# Limit number of exports
continues dump all ./sessions --limit 50

Output files: {source}_{id}.md or {source}_{id}.json

Scan and rebuild index

# Show discovery stats
continues scan

# Force rebuild session index
continues rebuild
continues scan --rebuild

Session index cached at ~/.continues/sessions.jsonl (5-minute TTL).

Configuration

Verbosity presets

Control how much context gets transferred:

continues resume abc123 --preset verbose --in gemini
continues inspect abc123 --preset full
continues dump all ./backup --preset minimal

YAML configuration

Create .continues.yml in your project root:

preset: verbose
recentMessages: 15
shell:
  maxSamples: 10
  stdoutLines: 20
files:
  maxSamples: 8
thinking:
  maxChars: 2000
keyDecisions:
  maxCount: 5

Configuration resolution order:

1. --config <path> flag
2. .continues.yml in current directory
3. ~/.continues/config.yml (global)
4. standard preset (default)

Environment variable overrides

Customize tool data directories:

export CLAUDE_CONFIG_DIR="$HOME/.config/claude"
export CODEX_HOME="$HOME/.local/codex"
export GEMINI_CLI_HOME="$HOME/.gemini-custom"
export XDG_DATA_HOME="$HOME/.local/share"

Code examples

TypeScript: Programmatic session access

import { SessionRegistry } from 'continues';
import { ClaudeParser } from 'continues/parsers';

// Initialize registry
const registry = new SessionRegistry();
await registry.scan();

// Get all sessions from Claude Code
const claudeSessions = registry.getSessions('claude');

// Find sessions for current project
const projectSessions = registry.getSessionsForProject(process.cwd());

// Parse a specific session
const parser = new ClaudeParser();
const sessionData = await parser.parse('~/.claude/projects/abc123/session.jsonl');

console.log(`Messages: ${sessionData.messages.length}`);
console.log(`Files touched: ${sessionData.files.length}`);
console.log(`Shell commands: ${sessionData.shell.length}`);

TypeScript: Custom handoff generation

import { HandoffGenerator } from 'continues/handoff';
import { SessionParser } from 'continues/parsers';

// Parse source session
const parser = SessionParser.forTool('claude');
const session = await parser.parse(sessionPath);

// Generate handoff with custom config
const generator = new HandoffGenerator({
  preset: 'verbose',
  recentMessages: 20,
  shell: { maxSamples: 15 }
});

const handoffDoc = generator.generate(session, {
  sourceId: 'abc123',
  sourceTool: 'claude',
  targetTool: 'gemini'
});

// Write handoff file
await fs.writeFile('handoff.md', handoffDoc);

TypeScript: Adding a custom parser

// src/parsers/mytool.ts
import type { SessionData, BaseParser } from '../types';

export class MyToolParser implements BaseParser {
  readonly name = 'mytool';
  readonly format = 'json';
  
  getSessionPaths(): string[] {
    return [`${process.env.HOME}/.mytool/sessions/*.json`];
  }
  
  async parse(filePath: string): Promise<SessionData> {
    const raw = JSON.parse(await fs.readFile(filePath, 'utf-8'));
    
    return {
      id: raw.id,
      source: 'mytool',
      projectPath: raw.workspace,
      timestamp: new Date(raw.created_at),
      messages: raw.chat.map(msg => ({
        role: msg.role,
        content: msg.text,
        timestamp: new Date(msg.ts)
      })),
      files: raw.files.map(f => ({
        path: f.path,
        operation: f.op,
        content: f.content
      })),
      shell: raw.commands.map(cmd => ({
        command: cmd.exec,
        output: cmd.stdout,
        exitCode: cmd.exit,
        timestamp: new Date(cmd.ts)
      })),
      metadata: {
        model: raw.model,
        tokensIn: raw.usage?.input,
        tokensOut: raw.usage?.output
      }
    };
  }
}

// src/parsers/registry.ts
import { MyToolParser } from './mytool';

export const PARSERS = {
  // ... existing parsers
  mytool: MyToolParser,
} as const;

// src/types/tool-names.ts
export const TOOL_NAMES = [
  'claude', 'codex', 'cursor', 'gemini', 'copilot',
  // ... other tools
  'mytool'
] as const;

Session storage locations

Where continues looks for session data:

// Claude Code
~/.claude/projects/*/session.jsonl

// Codex
~/.codex/sessions/*.jsonl

// Cursor
~/.cursor/projects/*/agent-transcripts/*.jsonl

// Gemini CLI
~/.gemini/tmp/*/chats/*.json

// Copilot
~/.copilot/session-state/*.{yaml,jsonl}

// OpenCode (SQLite)
~/.local/share/opencode/storage/sessions.db

// Crush (SQLite)
~/.crush/crush.db

// Amp
~/.local/share/amp/threads/*.json

// Kiro
~/Library/Application Support/Kiro/workspace-sessions/*.json

// Cline (VS Code extension)
<vscode-data>/globalStorage/saoudrizwan.claude-dev/tasks/*/

// Roo Code (VS Code extension)
<vscode-data>/globalStorage/rooveterinaryinc.roo-cline/tasks/*/

// Kilo Code (VS Code extension)
<vscode-data>/globalStorage/kilocode.kilo-code/tasks/*/

All reads are read-only — continues never modifies source session files.

Common patterns

Rate limit recovery workflow

# Hit Claude rate limit mid-session
# Switch to Gemini with full context
continues resume $(continues list --source claude -n 1 --json | jq -r '.[0].id') --in gemini --preset verbose

# Or use interactive picker
continues
# Select your Claude session, choose Gemini as target

Daily session backup

#!/bin/bash
# Backup all sessions daily
DATE=$(date +%Y-%m-%d)
continues dump all "./backups/$DATE" --preset full --json

Project-specific handoff script

# .continues.yml in your project
preset: verbose
recentMessages: 25
shell:
  maxSamples: 15

# Quick handoff script
#!/bin/bash
# switch-to-codex.sh
LAST_SESSION=$(continues list -n 1 --json | jq -r '.[0].id')
continues resume "$LAST_SESSION" --in codex --yolo

Cross-tool testing

# Start in Claude
claude start "Implement auth middleware"

# After some work, transfer to Cursor to test
continues resume $(continues list --source claude -n 1 --json | jq -r '.[0].id') --in cursor

# Verify in Gemini
continues resume $(continues list --source cursor -n 1 --json | jq -r '.[0].id') --in gemini --debug-prompt

What gets transferred

Every handoff document includes:

• Recent messages: User/assistant conversation (3-50 messages based on preset)
• File changes: Edits, reads, writes with full paths
• Tool activity: Shell commands with output/exit codes, grep/glob results, MCP calls
• Key decisions: Architectural notes, debugging insights from thinking blocks
• Session metadata: Model, tokens, cache usage, timestamps
• Original session path: Full file path for traceability

Example handoff structure:

## Session Context
**Source**: claude (session: abc123)
**Project**: ~/dev/my-app
**Started**: 2026-02-19 05:28
**Model**: claude-sonnet-4
**Tokens**: 45,230 in / 12,847 out

## Recent Conversation
[Last 10 messages with timestamps and roles]

## Tool Activity
- **Bash** (×47): `npm test → exit 0`, `git status → exit 0`
- **Edit** (×12): `edit src/auth.ts`, `edit tests/auth.test.ts`
- **Grep** (×8): `grep "handleLogin" src/`

## Key Decisions
- Need to handle token refresh race with logout
- Using JWT short-lived tokens (15min) + refresh tokens
- Middleware order: auth → rate-limit → routes

## Files Modified
- src/auth.ts (edited)
- src/api/routes.ts (edited)
- tests/auth.test.ts (created)

Troubleshooting

No sessions found

# Check discovery
continues scan

# Verify tool installation
which claude codex cursor gemini

# Force index rebuild
continues rebuild

Session parsing fails

# Inspect raw session
continues inspect abc123 --verbose

# Check for corrupt files
file ~/.claude/projects/abc123/session.jsonl

# Try with minimal preset
continues resume abc123 --in gemini --preset minimal

Target tool doesn't receive context

# Debug the handoff prompt
continues resume abc123 --in cursor --debug-prompt

# Check handoff file location
ls -la ~/.continues/handoffs/

# Verify target tool accepts stdin
echo "test" | cursor

VS Code extension sessions not found

# Set XDG_DATA_HOME if VS Code uses custom location
export XDG_DATA_HOME="$HOME/.config"

# On macOS, Cline/Roo/Kilo default to:
~/Library/Application Support/Code/User/globalStorage/

# Force rescan
continues rebuild

Symlink directories skipped

The scanner now handles symlinks correctly, but if you have custom symlinked session directories:

# Verify symlink target is readable
ls -la ~/.claude/projects
readlink ~/.claude/projects

# Check permissions
stat ~/.claude/projects

Zero token counts showing

If sessions show "0 in / 0 out", the source tool doesn't track token usage in that session format. This is normal for some tools. Use --verbose to see other metadata:

continues inspect abc123 --verbose

Global flags

--config <path>      # Custom config file path
--preset <name>      # Verbosity preset (minimal|standard|verbose|full)
--verbose            # Enable debug logging
--debug              # Extra diagnostics
--in <tool>          # Target tool for handoff
--debug-prompt       # Print handoff prompt without launching tool

Session index

continues maintains a local index at ~/.continues/sessions.jsonl:

• Auto-refreshes every 5 minutes
• Rebuilt with continues rebuild or --rebuild flag
• Stores session ID, source, path, timestamp, project path
• Read-only — never modifies source session files

Development notes

When adding new tool support:

1. Create parser in src/parsers/<tool>.ts implementing BaseParser
2. Add tool name to src/types/tool-names.ts
3. Register in src/parsers/registry.ts
4. Compile-time completeness check ensures no missing parsers

The registry enforces that every tool in TOOL_NAMES has a corresponding parser entry.