Agent Task Loop CLI

Agent Task Loop is a local CLI for running AI coding agent task delivery workflows.

It connects your task trackers (a Feishu Base table today, GitHub Issues optionally) to local coding agents, then drives a task through execution, review, rework, branch publishing, and Pull Request handoff. Tasks stay owned by their source — the CLI and TUI are an integration layer, not a system of record — so writes route back to the backend each task came from.

Local CLI

Install dependencies from the monorepo root:

pnpm install

Then run the local CLI:

npx --no-install @rivus/agent-task-loop --help

From npm, run:

npx @rivus/agent-task-loop --help

Use --no-install when you want to force the repo-local binary:

npx --no-install @rivus/agent-task-loop --help

Commands

• npx agent-task-loop sync
• npx agent-task-loop schema
• npx agent-task-loop schema --apply
• npx agent-task-loop start --task TASK-101
• npx agent-task-loop watch --task TASK-101
• npx agent-task-loop resume --task TASK-101
• npx agent-task-loop complete --task TASK-101

Summary commands support --json for scripts and agents:

npx agent-task-loop sync --json
npx agent-task-loop schema --json
npx agent-task-loop schema --apply --json
npx agent-task-loop cleanup --task TASK-101 --json
npx agent-task-loop complete --task TASK-101 --json

Complete Flow

complete closes a task that is already in 待发布 or 待验收.

The flow is:

1. Reuse the task workspace branch and commit state.
2. Commit any pending workspace changes with an AI-generated commit message.
3. Push the remote branch and verify the remote head.
4. Create or reuse a GitHub Pull Request with gh.
5. Update the Pull Request body with a generated delivery summary.
6. Write branch, commit, Pull Request, and completion metadata back to the task table.
7. Move the task to 已完成.

Local Requirements

• Node.js 20+
• pnpm
• GitHub CLI (gh) authenticated for Pull Request creation (and as the GitHub-Issues token source)
• lark-cli — only when using a Feishu task source
• Locally executable coding agents such as claude, codex, coco, or glm

Config

Config is JSON only and resolved from exactly three places, in order:

1. --config <path> (explicit)
2. AGENT_TASK_LOOP_CONFIG environment variable
3. ~/.agent-task-loop/config.json (the global config; the default)

There is no per-directory task.config.* discovery. The fastest way to create the global config is:

npx agent-task-loop init

init asks which task source(s) to use — GitHub Issues, Feishu Base, or both — and writes ~/.agent-task-loop/config.json. Fill in projects and repositories afterward. See config.example.json for the full shape.

To add sources to an existing config (init refuses to overwrite), use the source command instead of hand-editing:

agent-task-loop source list                                   # show sources + default
agent-task-loop source add --type github --owner you --repo your-repo
agent-task-loop source add --type github --owner you --repo another-repo  # appends to repositories[]
agent-task-loop source add --type feishu --token <base> --table <tableId>
agent-task-loop source remove github:you/your-repo

source add merges into the config without touching unrelated blocks (adding a second GitHub repo appends to repositories[]); source remove keeps at least one source. Run with no flags in a terminal for interactive prompts.

Task sources

Configure at least one of feishu / githubIssues:

• GitHub-only — set githubIssues (owner, repo, optional defaultAgent), omit feishu. The token is resolved from githubIssues.token, then GITHUB_TOKEN, then gh auth token — so a gh-authenticated machine needs no token in config. Tasks created from the TUI become GitHub issues (the issue number/URL link back to the task).

  Multiple repositories — instead of a single owner/repo, list several under repositories; each becomes its own github:<owner>/<repo> source and shows up as a separate option in the TUI's create-form source selector. A top-level token / defaultAgent applies to all; a repo may override defaultAgent:

  {
    "githubIssues": {
      "defaultAgent": "codex",
      "repositories": [
        { "owner": "your-org", "repo": "service-a" },
        { "owner": "your-org", "repo": "service-b", "defaultAgent": "claude" }
      ]
    }
  }

  Which issues become tasks — to avoid adopting every issue in a repo, an issue is treated as a task only when it opts in: it carries the hidden <!-- task-id: ... --> marker (issues created through this tool) or an agent:<name> label (the way you hand off an existing issue). Issues with neither are ignored.

• Feishu-only — set feishu (baseToken, tableId), omit githubIssues.

• Both — tasks are read from both; writes route back to each task's owning backend, defaulting new creates to Feishu.

Manage tasks in the TUI

npx agent-task-loop tui

Press n to open the new-task form. With more than one configured source a selector lets you choose where to publish. When a claude agent is configured, press Ctrl+R on the form to have the AI refine the description before publishing.

When the board spans more than one source (e.g. several GitHub repos), each row shows a compact source tag (the repo short name) and the detail pane shows the full github:<owner>/<repo>. Press s to open the source filter — a multi-select popup (Space toggles, a all, Enter applies) to focus on one or more repos; the active selection appears as a src:… chip in the header. The / text filter also matches source/repository.

Initialize Task Table Schema (Feishu only)

Feishu task tables need a schema; GitHub Issues do not (schema prints a notice and exits when no Feishu source is configured).

Check fields:

npx agent-task-loop schema

Create missing fields:

npx agent-task-loop schema --apply