yatd, yet another td

There are many tds. This one is mine. It's in Rust, very fast, very small, fairly simple, and includes a skill. I intend it to be the bare minimum for something like a repo-specific issue tracker and possibly complementary to tools like OpenSpec.

Install with mise use -g cargo:https://git.secluded.site/yatd@latest or cargo install --git https://git.secluded.site/yatd or by cloning and running make install. Tell your agent how/when to use td by first installing the skill with td skill, then somehow referring to td when telling the agent to do something involving td. It shouldn't invoke the skill unless you mention td, allowing your agent to use other todo/issue tools in other repos even with this global skill.

Inspired by alosec/td.

Bootstrapping

When syncing a project to another machine, do not run td init on the other machine. Initialize just once on the first machine, then bootstrap others by running td sync on the first machine, then td sync wormhole-code on another.

# Machine A (already initialized)
td sync

# Machine B (same repo checkout, no td project yet)
td sync 5-lurid-gecko

Running td init on both machines creates different project_id values and prevents sync from merging them.

Usage

There are many ways to use something like this and I won't say any one is better than another. However, I tend to use it in a particular way and that may lead its design to faciliate that way particularly well.

td on its own (click to expand)

I first think of a feature, then tell the agent about it and ask it to interview me about any gaps. We go back and forth, me nitpicking things about what it said, it nitpicking things about what I said, until it feels right. Then I say something like "let's think about how we can break this up into the smallest units of work and create tasks with appropriate dependency relationships to communicate the order in which we must tackle them." It creates a bunch of tasks, links them, and probably describes what it did. I run td next to see what's bubbled up to the top, read through it, and decide whether it needs more work or is good to go as-is. If it needs more work, I start a new session with "let's think about {id} and plan it out some more, then log those decisions to the task". If it's ready, the new session begins with "let's get started on {id} by having a look around at the relevant code then breaking it down into smaller child tasks". Then depending on the context window and task scale, I either have it get started in this same session or a new one.

I said it's complementary to things like OpenSpec at the beginning, but realise my workflow almost entirely obviates OpenSpec 🥴 I do intend to use them together soon and will describe whatever workflow I end up with then.

$ td --help
Todo tracker for AI agents

Usage: td [OPTIONS] <COMMAND>

Commands:
  init      Initialize a central project and bind the current directory to it
  use       Bind the current directory to an existing project
  projects  List all known projects in central storage
  create    Create a new task [aliases: add]
  list      List tasks [aliases: ls]
  show      Show task details
  log       Append a work log entry to a task
  update    Update a task
  done      Mark task(s) as closed [aliases: close]
  rm        Delete task(s)
  reopen    Reopen task(s)
  dep       Manage dependencies / blockers
  label     Manage labels
  search    Search tasks by title or description
  ready     Show tasks with no open blockers
  next      Recommend next task(s) to work on
  stats     Show task statistics (always JSON)
  compact   Vacuum the database
  export    Export tasks to JSONL (one JSON object per line)
  import    Import tasks from a JSONL file
  sync      Sync project state with a peer via magic wormhole
  skill     Install the agent skill file (SKILL.md)
  help      Print this message or the help of the given subcommand(s)

Options:
  -j, --json               Output JSON
      --project <PROJECT>  Select a project explicitly (overrides cwd binding)
  -h, --help               Print help
  -V, --version            Print version

Contributions

I'm trying the Jujutsu VCS out for this project. I'm enjoying it and LLMs seem to do pretty well with it too. The collaboration story is a bit less convenient, especially since I'm also trying pr.pico.sh. It works very well with git projects, but jujutsu is missing some things git has which pr.pico.sh relies on. When cloning this repo, do so with jj git clone --colocate git@git.secluded.site:yatd.git and the relevant git commands should work fine.

Patch requests are in amolith/llm-projects on pr.pico.sh. You don't need a new account to contribute, you don't need to fork this repo, you don't need to fiddle with git send-email, you don't need to faff with your email client to get git request-pull working...

You just need:

Git
SSH
An SSH key

If you're using LLM agents, you might instead want to give them my pr.pico.sh skill.

# Clone this repo
jj git clone --colocate git@git.secluded.site:yatd.git

# Create a new change and describe what it does
jj new -m "Add fancy new thing" # Imperative, kernel-style commits, not Conventional Commits

# When ready, create a new patch request
git format-patch origin/main --stdout | ssh pr.pico.sh pr create amolith/llm-projects

# After potential feedback, revise and submit a new patchset
jj amend
git format-patch origin/main --stdout | ssh pr.pico.sh pr add {prID}

# List patch requests
ssh pr.pico.sh pr ls amolith/llm-projects --mine

See "How do Patch Requests work?" on pr.pico.sh's home page for a more complete example workflow.

README.md

yatd, yet another td

Bootstrapping

Usage

Contributions