Task Harbor

Task Harbor is a command-line tool that runs coding tasks with Claude Code or Codex. It reads a Markdown backlog, plans each task, makes changes in separate Git worktrees, and checks the results before merging them into a local branch.

It does not push changes or merge them into your original branch. You review the result on manager/integration/<run-id>.

Check the work before accepting it

Configure tests once, then inspect the exact commands, checkpoint, and full diff for each task. Task agreements capture scope and unanswered questions. A failed check stops approval; a missing environment pauses work before more model calls.

• Set up checks and task agreements
• Compare Harbor with a direct agent

task-harbor preflight
task-harbor review --run-id RUN_ID --html review.html

In the dashboard, press E for task evidence or P for suggested checks.

Install

You need Python 3.11+, Git, and either Claude Code or Codex installed and signed in. The repository you want to work on must have at least one commit and a configured Git author.

Install the Python package directly from this repository (no clone required):

python -m pip install "git+https://github.com/rhit-nayeems/task-harbor.git"
task-harbor --version
task-harbor guide

The dashboard is included. This project is not published to PyPI yet; do not use pip install task-harbor expecting this repository's package.

pipx install "git+https://github.com/rhit-nayeems/task-harbor.git"
pipx ensurepath

python -m pip install --upgrade "git+https://github.com/rhit-nayeems/task-harbor.git"

First run · Commands · Configuration

Terminal dashboard

These are captures of the actual terminal interface. The task dashboard uses sample run data. Open it inside your project:

task-harbor

Running task-harbor (or python -m task_harbor) without a command opens the dashboard for the current directory. --help lists the other commands.

The dashboard shows saved runs, task status, dependencies, recent events, agent calls, and reported costs. Select a run with the picker and use the arrow keys to inspect tasks. Tab changes panes, F1 opens help, r refreshes, and q exits. It also refreshes automatically every two seconds.

A run is one execution of your backlog. The run picker appears only after a run exists in the selected project.

Use Set up to create the configuration and sample backlog, edit those files, then click Validate. New run and Resume show the provider, model, and call limit before you confirm agent work. Stop requests a pause after current calls finish. Closing the dashboard does not stop a launched run.

Viewing progress does not consume tokens. Starting or resuming work does. Launch output is saved in .manager/launch-logs/ (or your configured manager directory). Launch errors appear above the task list. Unreported costs are unknown. Use tui --run-id <id> to open a specific saved run. All commands below still work without the dashboard.

Start a run

Replace path/to/your-app with the repository you want Task Harbor to work on. Use --repo before the command.

task-harbor --repo "path/to/your-app" init

This creates .manager/config.json and a sample BACKLOG.md in that repository. An existing backlog is kept. Replace the sample task with your own work:

## BUG-001 — Fix empty search results
- **Priority:** critical_bug
- **Status:** UNPLANNED
- **Dependencies:** —
- **Acceptance criteria:**
  - An empty search shows a message instead of crashing
  - Existing search tests pass
- **Next action:** Plan the task

Use a different ID for each task. List dependency IDs separated by commas. See BACKLOG.example.md for a longer example.

Choose Claude or GPT

Edit .manager/config.json before starting. Claude is the default:

{
  "provider": "claude",
  "model": "opus",
  "effort": "medium"
}

For GPT, sign in with codex login and change these fields:

{
  "provider": "codex",
  "model": "gpt-6.1-sol",
  "effort": "medium",
  "max_run_cost_usd": null,
  "max_call_cost_usd": null
}

Use a model available to your account. Provider choice applies to the whole run; you cannot switch providers when resuming it.

Set checks and limits

Add your project's test or build commands to the configuration. For example:

{
  "verification_commands": ["python -m pytest"],
  "max_agent_calls": 100,
  "max_task_agent_calls": 20
}

These call limits are examples; the defaults are unlimited. Both providers support call limits and timeouts. Only Claude supports estimated dollar caps. Codex requires both dollar-budget settings to be null.

Now validate the backlog, check the selected agent, and start:

task-harbor --repo "path/to/your-app" validate
task-harbor --repo "path/to/your-app" doctor
task-harbor --repo "path/to/your-app" manage

doctor makes a model call. manage uses model calls throughout the run and prints a run ID. Save that ID to check progress or resume later.

CLI commands

Before spending model usage, run task-harbor check. It checks Git, the backlog, configuration, agent executable availability, required environment, and whether verification and call limits are configured. It reads local files without creating configuration, running project commands, or testing provider authentication. Errors return exit code 2; warnings explain missing safeguards. Use check --json for structured results.

Find saved run IDs with task-harbor runs, or filter with task-harbor runs --status paused. runs --json includes task counts, blocked counts, stop reasons, and integration branches. Unreadable runs are reported without hiding healthy ones. These are persisted snapshots: “unfinished” does not mean a process is currently running.

All commands accept --repo before the command. Replace RUN_ID below with the ID printed by manage.

Status, reports, stop, retry, and cleanup do not call a model. Reports are snapshots; generate a new one to see updates. Retry and resume keep existing usage counts.

What happens during a run

Task Harbor plans up to four tasks at once, but implements them one at a time. Each implementation gets a local checkpoint commit. Configured checks run first, then a separate agent checks the acceptance criteria. Passing work is merged into the integration branch, with code review every three passes by default.

A failed task gets two repair attempts in the same session. Further failures trigger diagnosis and a fresh session, up to a configured limit.

Run state, plans, checks, and logs are saved under .manager/runs/<run-id>/. Worktrees are stored beside your repository. Edit task requirements before starting a new run; resume uses saved requirements, not later backlog edits.

Agents execute local commands. The permission settings and worktrees are not a substitute for an isolated machine when working with untrusted code. After a hard crash, stop any remaining agent processes before resuming.

Development

python -m pip install -e ".[dev]"
python -m pytest

Tests use simulated agent responses and real temporary Git repositories. CI runs on Windows and Linux with Python 3.11 and 3.14. These tests do not establish live model quality or account compatibility; see EVALUATION.md for opt-in live evaluations.

• Configuration and provider details
• Architecture
• Workflow
• Evaluation
• Changelog
• Technical status report

Website

The landing page lives in website/. It uses static HTML, CSS, and JavaScript, with an interactive example workflow and installation guide.

License

Task Harbor is available under the MIT License.