taskmanager-exe 0.2.1

Version-controlled task management for AI agents

TaskManager.exe

Version-controlled task management for AI agents. Agents use familiar file editing tools; versioning and sync happen transparently via jj (jujutsu).

Problem

AI agents using file-based task systems lose work when:

• Multiple agents edit the same file
• Context resets mid-task and overwrites with stale state
• No history to recover from
• Agents go in circles - after context reset, they repeat mistakes because they don't know what was already tried

Installation

Requires Python 3.11+ and jj.

pipx install taskmanager-exe
# or
uvx taskmanager-exe

Quick Start

# Initialize in your repo
taskman init

# Install MCP server config
taskman install-mcp claude    # or: cursor, codex

# Install Claude Code skills (optional)
taskman install-skills

To create a worktree with its own .agent-files/:

taskman wt my-feature    # creates worktrees/my-feature/ with git worktree + jj clone

How It Works

Agent
  │
  ├── Edit tool ────────► .agent-files/ (jj repo)
  │   (file ops)                │
  │                        push/pull
  └── MCP Server ───────────────┼──────────────────►
      (batch/sync)              ▼
                        .agent-files.git/ (bare origin)

• Agents edit files with their normal Edit tool
• jj auto-snapshots every change (no explicit commit needed)
• MCP tools handle sync and history queries
• Bare git origin serializes concurrent access across worktrees

CLI Commands

taskman init                    # create .agent-files.git/ + .agent-files/
taskman wt <name>               # create worktrees/<name>/ with git worktree + jj clone
taskman install-mcp <agent>     # install MCP config (claude, cursor, codex)
taskman install-skills          # install skill files to ~/.claude/commands/
taskman uninstall-mcp <agent>   # remove MCP config
taskman uninstall-skills        # remove skill files

taskman describe <reason>       # create named checkpoint
taskman sync <reason>           # full sync: describe + fetch + rebase + push
taskman history-diffs <file> <start> [end]    # diffs across revision range
taskman history-batch <file> <start> [end]    # file content at each revision
taskman history-search <pattern> [file] [limit]  # search history

taskman stdio                   # run MCP server (stdio transport)

MCP Tools

When installed, these tools are available to agents:

Tool	Description
describe(reason)	Create named checkpoint
sync(reason)	Full sync workflow
history_diffs(file, start, end)	Aggregate diffs across range
history_batch(file, start, end)	File content at all revisions
history_search(pattern, file, limit)	Search history for pattern

Direct jj Commands

Agents can also use jj directly for simple operations:

jj status                    # current state
jj log                       # view history
jj diff                      # see changes
jj restore --from <rev> <file>  # restore file from revision

Task File Structure

.agent-files/
  STATUS.md           # Task index, session state
  LONGTERM_MEM.md     # Architecture (months+)
  MEDIUMTERM_MEM.md   # Patterns, gotchas (weeks)
  tasks/
    TASK_<slug>.md    # Individual tasks
    _archive/         # Completed tasks

Sync Model

Sync at task boundaries:

• Session start: pull
• /continue: pull
• /handoff: push
• Task complete: push

On conflict, agent resolves with Edit tool, then calls sync again.

License

MIT