Have you ever started a CLI project, forgot to note down the resume command, then burned a stack of tokens making the agent re-read the repo just to get its bearings? Same.

stash is an interactive launcher that resumes claude, codex, and opencode sessions across every project you've ever opened. No more copy-pasting your "here's where we left off" recap into a fresh chat. No more hunting through ~/.claude/projects/<encoded-path>/<uuid>.jsonl for the right session id. Just stash, pick a project, pick the conversation, and the right tool launches in a new terminal window in the right directory with the right --resume <id> flag already filled in.

$ stash
  ┏━┓╺┳╸┏━┓┏━┓╻ ╻
  ┗━┓ ┃ ┣━┫┗━┓┣━┫
  ┗━┛ ╹ ╹ ╹┗━┛╹ ╹
  resume across claude · codex · opencode

┌  stash  resume across tools
│
◆  Pick a project
│  ★ api                       2m ago  [claude·3]
│    web-app                  31m ago  [claude·2, codex·2]
│    deploy-scripts            9h ago  [claude·3, opencode·2]
│    + Register a new project…
│    ✗ Delete a project (purge all sessions)…
│    🧹 Sweep clutter (8 hidden / stale)…
│
│  ↑↓ navigate · enter select · space toggle · ctrl+c exit
└

What's new in 0.3.1

• Fork-on-resume. A single ⤴ Fork on resume: ON/off toggle in the session menu (space-toggleable like the other rows). Branches every resume into a fresh session id so the parent stays untouched. Wires through to claude --fork-session, codex fork <id>, and opencode --fork --session <id> transparently. Also exposed as stash <name> --fork on the command line.
• Rename-on-miss. When stash <name> finds the registered directory no longer exists, it searches every indexed session for a live directory with the same basename and asks you to confirm the move (or pick between several matches). Never auto-decides.

What's new in 0.3

• stash search — fuzzy-match across every session you've ever opened in any of the three tools, with a live preview pane showing the last user prompt and assistant reply for the focused row. No more guessing which "refactor auth" session is the live one.
• First-prompt indexing — search hits content, not just titles. If your conversation 11 days ago was about "fanvue nsfw" and the AI never generated a title, stash finds it anyway.
• stash s / stash find aliases for the impatient.

Demo

$ stash search docker
┌  stash  search across tools
◆  stash search   › docker_
│
│  ─ results (3/3) ──────────────────
│  ▸ claude   optima-docker  Read the markdown file       2h ago
│    opencode optima-docker  Deploying ARM Docker on AWS  29d ago
│    claude   Lycan          docker compose troubles      2d ago
│
│  ─ preview ────────────────────────
│  469 msgs · ~344.6k tok · ~/Documents/Projects/Almas/optima-docker
│
│  you: Ok we can test it on a x86 24.04 lts tomorrow…
│
│  claude: Done. launch.md now has a Test plan section with a full runbook…
│
│  type to filter · ↑↓ navigate · enter resume · ctrl+c exit

What's new in 0.2

• stash doctor — three checks across claude / codex / opencode that validate each tool's on-disk format still parses. Run it after upgrading any of the three to catch schema drift before it surprises you.
• Message counts and (where the tool exposes them) real token totals on every picker row, so two sessions with similar titles aren't a coin flip anymore.
• "Tested against" version table in the docs so you know which CLI releases the parser was built for, and doctor will tell you when yours diverge.

What it actually solves

• Lost resume commands. Every tool has a --resume <id> (or equivalent), but the id is a UUID buried in a log somewhere. stash finds it for you.
• Wasted tokens. Resuming the right conversation instead of starting fresh means the agent already knows your codebase, decisions, gotchas. No more "read the repo, then we'll begin."
• Three tools, one mental model. claude / codex / opencode each store sessions differently (JSONL files, rollout files + index, a SQLite db). stash normalises them into one picker.
• Multi-project sprawl. If you bounce between 30 projects like me, you forget what's in each one. stash groups every session by project directory and tags it with the last AI-generated title.
• Clutter from accidental invocations. cd ~/.cache/huggingface/hub, ran claude once by mistake → it's now a "project" forever. stash hides obvious clutter and offers a one-tap Sweep to nuke it.

Install

brew install SectorOPS/Stash/stash

brew install stash (unqualified) matches the unrelated Stash.app proxy tool's cask, not this CLI. Always use the fully-qualified SectorOPS/Stash/stash form. If you'd previously installed the Cask, you may also need brew link --overwrite stash after installing ours.

To upgrade:

brew update && brew upgrade SectorOPS/Stash/stash

Install from source

git clone https://github.com/SectorOPS/Stash.git
cd Stash
./install.sh

Requires bun (brew install oven-sh/bun/bun). The installer picks the best writable directory on your $PATH (/opt/homebrew/bin on a Homebrew Mac, ~/.local/bin as a fallback) and drops a stash symlink there. PREFIX=/usr/local ./install.sh to override.

Works on macOS and Linux. The new-terminal opener detects iTerm2, Terminal.app, Ghostty, WezTerm, Warp, kitty, Alacritty, gnome-terminal, konsole, xfce4-terminal, foot, and xterm. If you're already inside tmux it just opens a new tmux window.

Usage

Interactive

stash                  # the main event — pick project, pick session

A typical run:

1. Project picker. Every directory where you've ever opened claude/codex/ opencode, sorted by recency. Registered projects get a ★. Clutter (caches, ~/Downloads, /tmp, missing dirs) is hidden by default.
2. Session menu. For the picked project: a row per existing session (showing the AI-generated title and the exact resume command that'll run), plus options to start a fresh session in any of the three tools.
3. Toggle skip-permissions (SPACE on that row) and new terminal window (SPACE on that row). Save defaults per registered project.
4. Launch. Spawns a new terminal window in the project's directory and runs the resume command for you.

Direct

stash <name>           # resume registered project by name (most recent session, default tool)
stash <name> -y        # …with skip-permissions
stash <name> --new     # start a fresh session in that project's default tool
stash <name> --tool codex
stash <name> --session <uuid>
stash <name> --dry-run # print the command, don't run it

Registry management

stash add [dir]              # register cwd (or dir) as a project
stash add --name api --tool claude --dir ~/code/api
stash ls                     # list visible projects
stash ls --all               # include hidden clutter
stash rm <name>              # unregister (keeps sessions on disk)
stash where                  # print registry path
stash edit                   # open registry in $EDITOR

Health check

stash doctor                 # verify each tool's session format still parses

Runs a per-tool check that opens one session, looks for the fields stash depends on, and reports ok / warn / error with a one-line explanation. Exits non-zero on any error so you can wire it into CI or a shell precmd. Run it after upgrading claude / codex / opencode to catch schema drift early.

Flags reference

-y, --yolo          add --dangerously-skip-permissions  (claude/codex)
    --here          run in current terminal instead of new window
-w, --new-window    force a new window even if default is --here
    --new           start fresh session instead of resuming
    --tool TOOL     override default tool (claude|codex|opencode)
    --session ID    resume a specific session uuid
-n, --dry-run       print the command, don't launch
-a, --all           (with `ls`) include hidden clutter

How it finds sessions

Tool	Source
claude	~/.claude/projects/<encoded-path>/<uuid>.jsonl
codex	~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl + session_index.jsonl
opencode	~/.local/share/opencode/opencode.db (SQLite)

Titles come from each tool's own metadata: claude's ai-title (falling back to the last user prompt), codex's thread_name in the session index, opencode's session title column.

Token totals are pulled from each tool's own usage tracking and shown where they're accurate: claude records usage.input_tokens / usage.output_tokens per assistant turn and opencode records tokens.input / tokens.output per message — both are summed into the ~Nk tok you see in the picker. codex doesn't record per-turn usage in its rollouts, so codex rows show only N msgs. The asymmetry is deliberate — better to show a real number where one exists than to fake it everywhere.

Tested against

Stash is tested against the following CLI versions. stash doctor will tell you if your install diverges in a way that breaks parsing.

Tool	Tested up to	Storage format
claude	2.1.x	JSONL with ai-title, last-prompt, user, assistant events
codex	0.129.x	rollout JSONL with session_meta + response_item envelopes
opencode	1.14.x	SQLite schema with session / message / part tables

The encoded-path scheme claude uses (/ and . both map to -) is ambiguous on decode, so stash walks the real filesystem to disambiguate (e.g. it knows -home-user--cache-huggingface-hub is ~/.cache/huggingface/hub, not ~/cache/huggingface/hub).

Registry

~/.config/stash/registry.json (or $XDG_CONFIG_HOME/stash/registry.json):

{
  "version": 1,
  "projects": [
    {
      "name": "api",
      "dir": "/home/you/code/api",
      "defaultTool": "claude",
      "skipPermissions": false,
      "newWindow": true,
      "lastSessionId": "00000000-0000-0000-0000-000000000000",
      "lastTool": "claude",
      "addedAt": "2026-01-01T00:00:00.000Z",
      "updatedAt": "2026-01-01T00:00:00.000Z"
    }
  ]
}

stash edit opens it in $EDITOR if you need to clean it up by hand.

Skip-permissions mapping

-y / --yolo translates per-tool:

• claude → --dangerously-skip-permissions
• codex → --dangerously-bypass-approvals-and-sandbox
• opencode → no CLI flag exposed; the option is silently ignored

Housekeeping

stash gives you three flavours of cleanup, from gentle to scorched-earth:

• ✂ Delete sessions… (inside a project's session menu): multi-select checklist · prune duplicates · keep the most-recent N.
• ✗ Delete a project (purge all sessions)… (in the project picker, or inside a session menu): nuke every session attached to a project across all three tools, plus its registry entry. Asks you to type the project name to confirm.
• 🧹 Sweep clutter… (in the project picker, when there's anything to sweep): auto-finds sessions in obvious junk paths (~/.cache, ~/Downloads, ~/Library, /tmp, /var, etc.) and dirs that no longer exist, previews exactly what'll be removed, then deletes the lot in one confirm.

The on-disk project directory is never touched — only the session files / SQLite rows.

Environment

• CLAUDE_CODE_SESSION_ID — when set (it is whenever stash is launched from inside an active Claude Code conversation), that session is excluded from the picker, since deleting it just gets recreated by the next assistant message.
• NO_BANNER=1 — suppresses the ASCII logo at launch.
• XDG_CONFIG_HOME — overrides the default ~/.config location for the registry.

License

MIT.