taskmd 1.3.0

Markdown-native task management for developer projects

task.md

Most task trackers assume a human is driving. But when an LLM agent picks up a task, it needs structured metadata -- priority, status, completion criteria -- not a Jira ticket behind an OAuth wall. And when a human reviews that same task, they don't want to parse JSON. They want to read a file.

taskmd makes tasks that work for both. Each task is a markdown file. All metadata lives in the filename. The body is free-form markdown. The filesystem is the data store, git is the audit trail -- no database, no config, no coordination server.

tasks/
  34001-p1-done--initial-setup.md
  34002-p2-ready--add-feature.md
  34003-p3-blocked--waiting-on-api.md

Three properties make this work:

• Filesystem-native. Tasks are files. Create them with your editor, move them with mv, search them with grep. Every tool you already have works.
• Git-native. Status changes are renames, which are commits. Branches get their own task state. Merges resolve naturally. git log is your audit trail.
• Agent-native. When taskmd detects it's running inside Claude Code, Cursor, Codex, or a dozen other agents, it automatically switches to structured JSON output -- same files, same commands, zero configuration.

Install

pip install taskmd

Or run without installing:

uvx taskmd validate
uvx taskmd list

Quick start

taskmd init                                                                   # create tasks/ with a template
echo "Fix the login redirect loop when JWT is expired." | taskmd new --slug fix-login
cat body.md | taskmd new --slug add-oauth --priority p1                       # body from file
taskmd status 34042 in-progress                                               # change a task's status atomically
taskmd validate                                                               # check filenames for consistency
taskmd fix                                                                    # auto-repair legacy IDs and duplicate IDs
taskmd list                                                                   # list tasks with metadata
taskmd list --status ready --priority p0                                      # filter to what matters

taskmd new is the recommended way to create tasks — it allocates the ID, formats the filename, and writes the file in one atomic step. The body is required on stdin; a task with no description is a placeholder that pollutes triage. taskmd next exists for integrations that need just an ID string, but it's a sharp edge (two concurrent callers can receive the same ID).

All commands auto-detect a single immediate child directory containing _TEMPLATE.md (for example ./tasks or ./tickets) as the default directory. Pass a path to override.

Agent mode

When run inside an LLM coding agent, taskmd auto-detects the environment and switches to JSON output. The agent gets structured data; the human reading the same file gets markdown. No separate "agent API" to maintain.

taskmd --agent validate          # structured JSON envelope
taskmd --agent --help            # self-documenting schema for agents
taskmd --agent --compact --help  # minimal schema (fewer tokens)

Detection is automatic for Claude Code, Cursor, Codex, Windsurf, Aider, Cline, Amazon Q, Gemini Code Assist, Cody, and any tool that sets AGENT=1.

Task file format

Filename: DDNNN-pX-status--slug.md

# Task Title

## Summary
What needs to be done.

## Done When
- [ ] First criterion
- [ ] Second criterion

The filename is the only source of truth. The body is free-form markdown.

Statuses: ready, in-progress, blocked, done, wont-do, brainstorming

Priorities: p0 (critical) through p4 (polish)

To change status: taskmd status <id> <new-status> renames the file. It refuses to clobber an existing target filename and rejects invalid statuses up front.

Task IDs

IDs are 5-digit numbers (e.g., 34042). The first two digits are a hash of (machine_identity, tasks_directory_path) mod 100. Different machines and different worktrees on the same machine land in different buckets, avoiding ID collisions without coordination. Birthday-50% collision point is ~12 concurrent worktrees.

Set TASKMD_MACHINE_ID to override the machine-identity input (any string; replaces hostname in the hash). Useful in ephemeral containers where the hostname is constant across sessions.

Library

taskmd is also a Python library -- the CLI is a thin wrapper around it.

pip install taskmd

from taskmd import validate, fix, next_id, list_tasks

result = validate("tasks")
if not result.ok:
    for err in result.errors:
        print(err)

tasks = list_tasks("tasks")
for t in tasks:
    print(f"{t.id:5s} {t.priority} {t.status:12s} {t.slug}")

n = next_id("tasks")
print(f"Next task: {n}")

Use from a PEP 723 script

#!/usr/bin/env -S uv run
# /// script
# requires-python = ">=3.11"
# dependencies = ["taskmd"]
# ///

from taskmd import validate

result = validate("tasks")
if not result.ok:
    print(f"Task validation failed: {len(result.errors)} errors")

License

MIT