strait
Sasitilak/strait
Move your AI agent sessions between Claude Code, Codex, and OpenCode
Instalação
npx convez-ai Sasitilak/strait1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
strait
Move your AI agent sessions between Claude Code, Codex, and OpenCode.
Hit a Claude rate limit mid-task? Want to swap to Codex without losing context? strait translates sessions between Claude Code, Codex, and OpenCode — with full history preserved — and gives you one CLI to search, resume, and track every conversation across all three.
Status: alpha. All three runtimes (Claude Code, Codex, OpenCode) are bidirectional and resume-verified end-to-end.
strait — session portability for AI agents
? What do you want to do? Sync Claude → Codex
? Pick a claude session: 5f1e23a6 2026-04-28 (103 msgs) # the build plan
? Where should the translated session be written? Real
✔ Translated 103 messages ↦ 44 tool calls
⚠ dropped 10 thinking blocks (no Codex equivalent)
Claude ~~~~~~~~~~~⛵~~~~~~~~~~~ Codex
✓ Done. Resume command: codex resume f2fca69c-…
? What now? Run "codex resume f2fca69c-…" now (in /Users/me/proj)
Or skip the menu and pass flags directly: strait sync claude codex --latest --dry-run.
Install
npm install -g strait-cli
strait # interactive menu
strait sync claude codex --latest # or use flags
(The CLI command is strait even though the package is strait-cli — the bare name was already taken on npm.)
From source
git clone https://github.com/Sasitilak/strait.git && cd strait
npm install
npm run dev
From a GitHub tarball (latest main)
npm install -g https://github.com/Sasitilak/strait/archive/refs/heads/main.tar.gz
Requires Node 18+. Pure JavaScript / WASM only — no native compilation needed.
Usage
# pick a session interactively (recommended)
strait
# or specify everything via flags
strait sync claude codex --latest [--dry-run] [--verbose]
strait sync codex claude --latest
strait sync opencode claude --latest
strait sync opencode codex --latest
strait sync <from> <to> --session <id> # specific session
strait list claude # 10 most recent, color-coded UUIDs
strait list codex
strait list opencode
--dry-run writes to ./tmp/ instead of the real target dir so you can inspect output before committing it. --verbose logs each message as it's translated.
After a sync, strait can launch the resume command for you (in the original session's cwd) — pick "Run resume now" in interactive mode.
Discovery & history
strait status # what's installed, session counts, disk usage
strait list-all # sessions from every runtime, merged, newest first
strait search "auth bug" # grep the first user message of every session
strait open <id> # find a session by id prefix and launch resume
strait stats # per-runtime counts + per-direction conversions
strait history # past syncs (logged to ~/.strait/history.jsonl)
search, list-all, and history end with an interactive picker — pick an entry and strait spawns the right runtime (claude --resume / codex resume / opencode --session) for you.
Supported runtimes
| Runtime | Read | Write | Resume verified |
|---|---|---|---|
| Claude Code | ✓ | ✓ | ✓ |
| Codex | ✓ | ✓ | ✓ |
| OpenCode | ✓ | ✓ | ✓ |
All six directions work:
claude ↔ codex
claude ↔ opencode
codex ↔ opencode
What works
- Text user messages, assistant text replies
- Tool calls + tool results (any tool name — passed through verbatim)
- Multi-block assistant messages (text + tool_use + text + tool_use, in order)
tool_resultcontent as either string or array of text parts- Streaming line-by-line read (Claude/Codex) and indexed SQLite read/write (OpenCode)
- Atomic write to OpenCode's SQLite DB with WAL checkpoint, integrity check, and sidecar cleanup — probes the SQLite write lock to refuse a write only when OpenCode is actually holding it (no false positives from Spotlight or Finder)
- Auto-cd into the imported session's original cwd when launching resume
- Auto-launch with the right resume flag per runtime (
codex resume <id>,claude --resume <id>,opencode --session <id>) - Filters out Codex's
<environment_context>synthetic user turns
Known limitations
- Thinking blocks are dropped. Both Claude and Codex store reasoning in formats the other can't reconstruct (Codex uses encrypted blobs, Claude requires a valid signature). Tool calls and final outputs are preserved; the model just doesn't see its own prior chain-of-thought.
- Tool names pass through verbatim. Resume will display the original tool calls (e.g.
Bash), but the receiving runtime won't re-execute them unless that name is also one of its registered tools. The transcript is intact; new turns use the receiving runtime's own tools. - Codex resume logs a non-fatal "thread not found" warning on imported sessions. This is a Codex-internal cache issue for rollouts it didn't create itself — the conversation context still loads correctly and new turns persist.
- Cloud-only agents are not supported — Antigravity, Cursor cloud agents, ChatGPT, Devin, Replit Agent, etc. don't store sessions on disk, so there's nothing for
straitto read. - Format is unstable. Schemas change between releases. Current release targets Claude Code 2.1.x, Codex 0.114–0.125, and OpenCode (latest).
- No tests yet. Verified by running against real sessions.
Contributing
Issues and PRs welcome — this is alpha. The most valuable bug reports are real Claude sessions that fail to resume in Codex.
License
MIT.