Installs brew deps, oh-my-zsh, TPM, clones alacritty themes, and symlinks configs into place, including bin/ tools into ~/.local/bin. Idempotent; rerun after adding a tool or config.

Neovim worktree picker #

Press <leader>gw inside a Git repository to open the Telescope worktree and branch picker. It shows the current and checked-out worktrees, dirty state, upstream divergence, branch age, and the latest CI pipeline status. Press <C-r> in the picker to fetch and refresh Git state.

Pipeline status loads asynchronously and supports GitHub Actions, GitLab CI, Bitbucket Pipelines, and public Tangled Spindle pipelines. The picker uses the origin remote by default. Select another remote per repository with a .env file at that repository's root:

WORKTREE_CI_REMOTE=tangled

Provider credentials can be supplied in the same project-level .env file:

# GitHub, either name is accepted
GH_TOKEN=
GITHUB_TOKEN=

# GitLab
GITLAB_TOKEN=

# Bitbucket, use an access token or username with an API token/app password
BITBUCKET_ACCESS_TOKEN=
BITBUCKET_USERNAME=
BITBUCKET_API_TOKEN=
BITBUCKET_APP_PASSWORD=

GitHub requires an authenticated gh CLI, GitLab requires an authenticated glab CLI, and Bitbucket and Tangled require curl. Tangled status currently supports public repositories only. Provider queries time out after 10 seconds and inspect up to 100 recent runs.

Pipeline states are shown as ✓ passed, ✗ failed, ● running, ○ pending, - none, or ? unknown. The initial … loading state is replaced when the provider query finishes.

Existing shell environment variables take precedence over .env. Only the listed variables are loaded, and they are passed to the provider subprocess without changing Neovim's global environment. Missing tools, credentials, or network access produce an unknown status without blocking the picker. Values are read as simple dotenv literals; variable expansion is not supported. Keep the project .env ignored by Git.

Tools (bin/) #

bar: second tmux status row built from modules. A module publishes its segment into the tmux option @bar_<name> (daemons via bin/bar-lib, one-shots via bar set <name> <text>); bar mount builds status-format[1] from ~/.config/tmux/bar/modules, so no #() runs on the row. Modules: bar-sys (statmon's samples plus cpu and battery), bar-pr (open PRs per repo from ~/.config/tmux/bar/repos), bar-reeds (whisper topics waiting on a human).
researchmon: tracks explicitly selected Codex research threads on the shared local app server. researchmon track <thread-id> adds a thread; untrack removes it, and list prints tracked IDs. Inside a Codex tool shell, the ID defaults to CODEX_THREAD_ID. The tmux bar refreshes a cached summary every status interval; the sampler polls every five seconds. Waiting, idle, unloaded, and error states are shown separately. Idle does not imply research is done. researchmon sample queries live status; status only reads the cache. The sampler starts with tmux and reports disconnected if its cache goes stale. It uses the existing daemon socket without starting or resuming Codex tasks. Override RESEARCHMON_SOCKET for a different Unix socket or RESEARCHMON_HOME for a different state directory. Requires Python 3 and a Codex daemon with the local WebSocket control interface.
statmon: sampler daemon behind the tmux status bar. It writes files under ~/.statmon/ and the bar cats them; samplers are never forked from #().
throb: rainbow throbber for the tmux status bar with room for a passing thought (throb -t "..."); prefix+T toggles it, log in ~/.throb/.
train: merge/release train runner and curses watcher (python3 >= 3.11). A train is a plan of key|label|command steps stored under ~/.local/state/train/<id>/. Step exit 0 marks it done, 75 hands it to a human (needs-user), anything else stalls the train; rerunning train run resumes past done steps. Step stderr streams through while a tail is kept as the failure detail. For steps driven externally via train mark (another agent session), use exit 75 as the placeholder command so train run hands them over instead of completing them. Marks mirror to the reeds whisper network on train.<id>.<key> when the daemon is reachable. The watcher's --whispers [prefix] flag splits the screen 50/50 with a live reeds feed (default prefix train, stacked instead of side-by-side on narrow windows). In tmux, prefix+M pops up that split view, which shows every attended incomplete train as its own section (falling back to the latest train when all are done, idling when there are none) and closes itself when the shown trains complete on its watch; q or Esc dismisses it. Attended means: a fresh runner heartbeat (train run touches one every 15s), or marks within 4h, stretched to 24h while a step waits on a human. A section whose live step has a stale heartbeat shows "runner heartbeat lost". Steps waiting on a human get numbered badges: press the digit to select, then d marks done or s skips (Esc cancels the selection); rerun train run afterwards if the runner had exited. A popup swallows all keys including the tmux prefix, so if one ever refuses to die, run tmux display-popup -C from any other terminal to close it.
glab-merge <project> <iid>: waits for a GitLab MR to become mergeable, then merges it. Exits 75 on human-shaped blocks (conflicts, approvals, tag protection) so it slots straight into a train step. GLAB_MERGE_TIMEOUT caps the wait in seconds (default 600).

A train in one sitting #

train new demo "release demo" <<'EOF'
685|!685 -> main|glab-merge web/app 685
tag|tag rc/demo|git push origin refs/tags/rc/demo
EOF
train run demo      # stops on fail/needs-user; rerun to resume
train watch         # curses table; q quits, exits 0 on completion
train ls

Clone this repository

dots #

Install #

Neovim worktree picker #

Tools (bin/) #

A train in one sitting #