AI Agentsai-coding-assistantclaude-codeclicontext-transfercursor
Installation
# Add to your Claude Code skills
git clone https://github.com/yigitkonur/cli-continues
README.md
continues
You hit the rate limit mid-debug. 30 messages of context — file changes, architecture decisions, half-finished refactors — and now you either wait hours or start fresh in another tool. continues grabs your session from whichever AI coding tool you were using and hands it off to another one. Conversation history, file changes, working state — all of it comes along.
Handoff — generates a structured context doc and injects it into the target tool
The handoff document is designed so the receiving agent immediately understands what you were doing, what files were touched, what commands ran, and what's left to do.
Usage
Interactive (default)
Just run continues. It finds all your sessions, lets you pick one, and asks where to continue:
┌ continues — pick up where you left off
│
│ Found 1842 sessions across 14 CLI tools
│ claude: 723 codex: 72 cursor: 68 copilot: 39 ...
│
◆ Select a session
│ [claude] 2026-02-19 05:28 my-project Debugging SSH tunnel config 84a36c5d
│ [copilot] 2026-02-19 04:41 my-project Migrate presets from Electron c2f5974c
│ [codex] 2026-02-18 23:12 my-project Fix OpenCode SQLite parser a1e90b3f
│ ...
└
◆ Continue in:
│ ○ Gemini ○ Codex ○ Amp ○ Kiro ...
└
When you run from a project directory, sessions from that directory are prioritized.
Quick resume
Skip the picker entirely — resume the Nth most recent session from a tool:
Works for all 14 tools. This uses native resume — same tool, full history, no context injection.
Cross-tool handoff
This is the main thing. Start in one tool, finish in another:
# Hit the Claude rate limit? Hand it off to Gemini:
continues resume abc123 --in gemini
# Or pass flags through to the destination tool:
continues resume abc123 --in codex --yolo --search --add-dir /tmp
continues maps common flags (model, sandbox, auto-approve, extra dirs) to the target tool's equivalent. Anything it doesn't recognize gets passed through as-is.
Scripting & CI
continues list # table output
continues list --source claude --json # JSON, filtered
continues list --jsonl -n 10 # JSONL, last 10
continues scan # discovery stats
continues scan --rebuild # force re-index
Inspect (for debugging)
See exactly what gets parsed and what ends up in the handoff:
Export all sessions to files for backup, analysis, or archival:
# Export all sessions to markdown (default)
continues dump all ./sessions
# Export specific tool's sessions
continues dump claude ./sessions/claude
continues dump gemini ./sessions/gemini
# Export as JSON instead of markdown
continues dump all ./sessions --json
# Control verbosity with presets
continues dump all ./sessions --preset full
# Limit number of sessions
continues dump all ./sessions --limit 50
File naming: {source}_{id}.md or {source}_{id}.json
Verbosity control
Not every handoff needs to be a novel. Four presets control how much detail goes in:
Resolution order: --config <path> → .continues.yml in cwd → ~/.continues/config.yml → standard preset. See .continues.example.yml for the full reference.
What gets extracted
Every tool stores sessions differently — different formats, different schemas, different paths. Here's what continues reads:
All reads are read-only — continues never modifies your session files. Index cached at ~/.continues/sessions.jsonl (5-min TTL, auto-refresh).
Tool activity in handoffs
The handoff document includes a Tool Activity section so the target agent knows what was done, not just what was said:
## Tool Activity
- **Bash** (×47): `$ npm test → exit 0` · `$ git status → exit 0` · `$ npm run build → exit 1`
- **Edit** (×12): `edit src/auth.ts` · `edit src/api/routes.ts` · `edit tests/auth.test.ts`
- **Grep** (×8): `grep "handleLogin" src/` · `grep "JWT_SECRET"` · `grep "middleware"`
## Session Notes
- **Model**: claude-sonnet-4
- **Tokens**: 45,230 in / 12,847 out
- 💭 Need to handle the edge case where token refresh races with logout
This works for all 14 tools — bash commands, file reads/writes/edits, grep/glob, MCP tool calls, thinking blocks, subagent dispatches, token usage, model info. The shared SummaryCollector keeps the format consistent regardless of source.
Every handoff also includes the full file path of the original session, so the receiving tool can trace back to the raw data if needed.
Commands reference
| Command | What it does |
|:--------|:-------------|
| continues | Interactive TUI picker |
| continues list | List sessions (--source, --json, --jsonl, -n) |
| continues resume <id> | Resume by ID (--in <tool>, --preset) |
| continues inspect <id> | Diagnostic view (--truncate, --write-md, --preset) |
| continues dump <source\|all> <dir> | Bulk export sessions (--json, --preset, --limit) |
| continues scan | Discovery stats (--rebuild) |
| continues rebuild | Force-rebuild session index |
| continues <tool> [n] | Quick-resume Nth session from any of the 14 tools |
Global flags: --config <path>, --preset <name>, --verbose, --debug
Community contributions
This started as a 7-tool project and grew fast thanks to contributors:
Factory Droid support — #1, first community parser
Cursor AI support — #4 by @Evrim267, with smart slug-to-path resolution
Single-tool error handling — #3 by @barisgirismen, clear error when only one CLI is installed
Env var overrides — #14 by @yutakobayashidev, respects CLAUDE_CONFIG_DIR, CODEX_HOME, GEMINI_CLI_HOME, XDG_DATA_HOME
The latest batch — Amp, Kiro, Crush, Cline, Roo Code, Kilo Code, and Antigravity — was added by reverse-engineering mnemo's Go adapters and adapting the schemas for TypeScript. Along the way we also improved token/cache/model extraction for the existing Claude, Codex, Cursor, and Gemini parsers.
Bugs fixed in this round:
Symlink traversal — fs.Dirent.isDirectory() returns false for symlinks; fixed with isSymbolicLink() && statSync() fallback
Zero-token display — no longer shows "0 in / 0 out" when a session has no token data
Key Decisions count — now respects the verbosity config instead of being hardcoded to 5
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
