GitHub - PiemonteF/t: Task-centric git worktrees with durable, multi-agent conversations

Conductor, but it lives in your terminal.

A small worktree manager that never loses a conversation. You name a task; t gives that task a git worktree, a branch, and a permanent link to every chat any agent has ever had inside it — Claude, Codex, Grok, Gemini, in any mix.

Delete the worktree six months later. The conversation still resumes.

And it stays out of your way. tmux, herdr, Zellij, iTerm splits, plain terminal tabs — whatever you already use, keep using it. t never splits, spawns, or rearranges a single pane. It has one job and does not want yours.

$ t new "fix vault sync"      # worktree + branch, cds this pane into it
  <prefix+v>                  # you split, however you like
$ t claude                    # launches here, linked to the task
  <prefix+v>
$ t codex                     # same worktree, same task, different agent

Why `t`

It is short for task. It is also the entire name, which is the design in miniature: you will type it fifty times a day, so it costs one keystroke and asks nothing of you.

The working title was task-centric-agentic-worktree-orchestrator. It did not survive first contact with the keyboard.

The problem

Every agent decides where to put your transcript based on the working directory — and each one does it differently. Claude Code bakes the path into a directory name, so deleting a worktree orphans the conversation: it is still on disk, but claude --resume can never see it again.

On the machine this was built against, 55 of 109 Claude conversations were already unreachable that way. Every byte intact. The link gone.

t stops deriving the link and writes it down instead:

~/work/myrepo/fix-vault-sync/.agent/
  task.json                    repo, branch, base, worktree, created
  sessions/
    claude-f546376a….json      provider, session id, transcript path, pane
    codex-019cd415….json
    grok-019fd204….json

One file per session, and a session only ever touches its own — ten agents starting at once produce ten files and zero races.

For Claude and Grok the id is assigned before launch (--session-id), so the link exists before the conversation does. Codex and Gemini have no such flag, so t stamps the launch time and matches the session whose recorded cwd is this worktree.

If you know Conductor

I really liked the Conductor experience. I just love my own setup more.

That is the whole reason t exists. Conductor showed that the link between a worktree and a conversation should be written down rather than re-derived from a path — that is the insight, and t takes it wholesale. What it does not take is your terminal: no app window, no managed layout, no separate place you have to go to read your own history.

	Conductor	`t`
interface	a Mac app	your terminal, unchanged
storage	its own database	the vendor transcript, verbatim
resuming	in the app	hands the id back to the real CLI
pane layout	manages it for you	never touches it

Same idea, different room.

Commands


`t new "short description"`	worktree + branch, cd this pane into it
`t claude \| codex \| grok \| gemini`	launch an agent here, linked to this task
`t`	status of the task this pane is in
`t sessions`	pick a past conversation and resume it
`t resume <# \| id>`	resume without the picker
`t ls`	every task, with git and session state
`t open <task>`	cd this pane into another task
`t find <query>`	search every conversation, every provider
`t done [task]`	archive this task's conversations to the vault
`t doctor`	check the environment
`t import`	(deferred) pull Conductor's history into the vault

Picking a conversation

t claude always starts a new chat. To go back to an old one:

t sessions           # arrow-key picker, enter resumes
t sessions --all     # every session on the machine, not just this task
t sessions --list    # plain output, for scripting

 8 session(s)
▸ claude  fix-vault-sync   2026-08-05  the sync drops every record with a null tenant
  codex   redesign-page    2026-08-04  the build broke after the upgrade
  claude  old-migration    2026-07-21  how do I restrict this to a subset of repos

That third one is three weeks old and its worktree is long gone. It still resumes.

↑↓/jk to move, / to filter, enter to resume, q to quit. The picker is stdlib curses — no fzf needed.

Sessions whose worktree no longer exists are listed too. Choosing one copies its transcript into the current worktree (with subagents/ and tool-results/) and then hands the id back to the CLI that wrote it, so the conversation genuinely continues instead of being replayed read-only. It is then linked to the task you resumed it in.

Same thing without the picker:

t resume 3           # by row number from the last listing
t resume dead1111    # by session id prefix
t claude -c          # continue this task's most recent claude session

Install

git clone https://github.com/PiemonteF/t.git ~/.local/share/t/src
cd ~/.local/share/t/src

./install.sh            # config + work root, prints the .zshrc line
./install.sh --shell    # also append the wrapper to ~/.zshrc
./install.sh --hooks    # also register the capture hook (see below)

Clone it wherever you like — the installer records that path as T_HOME, so t keeps working from there. Nothing is copied outside ~/.config/t and your .zshrc.

The shell wrapper exists because a child process cannot change its parent's working directory. t new and t open write the cd they want to a temp file and the wrapper sources it.

The --hooks step is optional and merges into your existing hook config — hooks it did not write are left alone, and the file it edits is backed up first. It covers the case where you split a pane and just type claude out of habit: the agent hands its session id to the hook, so the link is recorded anyway. With it installed, there is no way to run an agent inside a task worktree without it being written down.

Config

~/.config/t/config.json

{
  "work_root": "~/work",
  "vault": "~/.local/share/t/vault",
  "repos": {
    "myrepo": { "base": "origin/dev", "copy": [".env"], "setup": "npm i" }
  }
}

copy are the files git will never carry across; setup runs once in the new worktree.

Bring your own multiplexer

There is nothing to configure here. t is a plain CLI that acts on the current directory, so tmux, Zellij, herdr, iTerm splits, WezTerm, or four bare terminal tabs all work identically. Split however you like; run t claude in whichever pane you landed in.

The only thing worth doing is making new panes inherit the current directory — tmux split-window -c "#{pane_current_path}", or the equivalent in your tool — so a fresh pane is already inside the worktree.

herdr

If herdr — a terminal pane manager — happens to be on PATH, t also:

reads HERDR_PANE_ID / HERDR_WORKSPACE_ID and records which pane a session is running in
calls herdr pane report-agent-session --agent-session-id --agent-session-path so herdr's own sidebar and herdr agent list know the session too
renames the pane to the task name

With new_cwd = "follow" in the herdr config, every pane you split is already inside the worktree. All of it is best-effort; t behaves the same without it.

Requirements

git, python3, ripgrep (for t find), and at least one agent CLI. No multiplexer, no daemon, no database.

See DESIGN.md for the storage layout of each provider and why the design is shaped this way.

License

MIT

Why t