Tasx

tasx — file-based tasks, folder = state

Markdown files are the source of truth; the folder a file sits in IS its state. No database, no daemon: ls tasks/in-progress/ is a kanban, git mv is a status change, and the board UI is just a viewer/editor over the same files. The CLI lives at scripts/tasx (installed on PATH as tasx).

Layout

tasks/               INBOX — new/untriaged tasks as *.md
  README.md          the convention doc (written by `tasx init`)
  in-progress/       actively worked — set `owner:` to your agent-chat name
  waiting/           blocked — `waiting-on:` header says on what (see below)
  done/              completed (kept; `tasx archive` rolls old ones into done/YYYY-MM/)
  cancelled/         dropped — say why in the file
  decisions/         open choices: `option:` lines; picking writes `choice:` → decided

Task file = optional header lines (id:, kind:, owner:, group:, seq:, title:, waiting-on:, option:, choice:), blank line, then a markdown body (What / Why / Next steps / Refs), optionally ending with ## Comments. Headerless files work too (id = filename, title = first # heading) — legacy myhdd-style folders parse as-is, blocked/ is read as waiting/. Full format spec: references/format.md.

Commands

tasx init                          # create tasks/ tree + README + CLAUDE.md/AGENTS.md pointer
tasx new "Title" [--owner me] [--group area] [--body "..."] [--id slug]
tasx new "Which renderer?" --decision --option "a | raster" --option "b | webgl"
tasx list [--all]                  # kanban to stdout (needs-you / in-progress / inbox / waiting)
tasx move <id> <state> [--waiting-on "you|decision <id>|external: <thing>"]
tasx comment <id> "text" --as <your-agent-name>    # timestamped; nudges the owner
tasx doctor                        # staleness + hygiene report — run at session start
tasx archive [--days 30]           # roll old done/ files into done/YYYY-MM/
tasx serve [--port N]              # board UI on a stable per-project port (127.0.0.1)

Commands find the nearest tasks/ walking up from cwd ($TASX_TASKS overrides).

Agent contract

Working in a repo that has a tasks/ folder means playing by these rules:

1. Session start: run tasx doctor. Reconcile what you own — finished things go to done, stalled things get a comment saying why, tasks owned by a dead agent get taken over or moved back to the inbox.
2. Starting work: tasx move <id> in-progress and set owner: to your agent-chat name (edit the header or create the task with --owner). One owner per task. Don't work on tasks owned by a live agent — agent-chat status <name>.
3. Blocked on the user: don't ask and idle. Move it: tasx move <id> waiting --waiting-on you, put the question in the file (or tasx comment). If it's a real fork in the road, create a decision (tasx new ... --decision --option ...) and set --waiting-on "decision <id>" — the user answers with one click on the board and you get nudged via agent-chat.
4. Finishing: tasx move <id> done + a short tasx comment <id> "what shipped, where" --as <name>. Cancelling requires a why.
5. New work discovered mid-task: don't silently expand scope — tasx new it into the inbox and continue.
6. Session end: nothing should be in in-progress/ under your name unless you (or a respawn of you) are actually continuing it. Move it back or comment its state.
7. Keep files lightweight; body sections What / Why / Next steps / Refs. Comments are append-only with author + timestamp — never rewrite others' comments.

The board (tasx serve)

For the human, not for agents (agents use the CLI/files). Serves scripts/board.html on a stable per-project port. Reads nothing but the md files; every edit writes back: status dropdown physically moves the file (git mv-aware), decision radios write choice:, comments append to ## Comments — and each of those nudges the owning agent through agent-chat send --to <owner> (when agent-chat is installed), so answering on the board resumes the work. Comments send only on an explicit button or ⌘⏎ (drafts survive re-renders, nothing fires on blur). On needs-you cards the primary action is answer & resume: append the answer, move the task back to in-progress (inbox if unowned), clear waiting-on:, one nudge — the human never needs to touch the state dropdown just to hand a question back. UI shows: pinned Needs you section + count in the tab title, in-progress with owner liveness dots (agent-chat status) and freshness ages, ⚠ stale chips (in-progress >24h, waiting-on-you >3d, inbox >14d), a "since your last visit" digest, and a collapsed done/cancelled archive grouped by day.

Run it in background when asked: nohup tasx serve > /tmp/tasx-board.log 2>&1 & then read the port from the log's first line.

Install (new machine / new agent CLI)

Symlink the skill folder into the agent's skill dirs and put tasx on PATH:

ln -s ~/github/skills-creation/skills/tasx ~/.claude/skills/tasx
ln -s ~/github/skills-creation/skills/tasx ~/.codex/skills/tasx
ln -s ~/github/skills-creation/skills/tasx ~/.cursor/skills/tasx
ln -s ~/github/skills-creation/skills/tasx/scripts/tasx ~/bin/tasx   # or any PATH dir

Python 3.8+ stdlib only. agent-chat integration is optional and auto-detected. Optionally export TASX_USER=<yourname>: board comments get authored with it, and waiting-on: <yourname> counts as needs-you alongside the generic you/user/human.