agent-watchtower-core 0.1.1

Tiny local CLI for coding-agent continuity.

Agent Watchtower Core

关掉终端再打开，Agent 记得做到哪了。

You use Codex, Claude Code, Hermes, CodeWhale, or another coding agent for a few hours. Then the terminal closes, the session gets interrupted, or the next window starts cold.

Agent Watchtower keeps the small local trail that matters, so the next session can continue instead of starting over:

• what the goal is
• what the next safe action is
• where the last artifact is

It is for people who have felt this pain:

"My AI agent was working, then it stopped. Now I don't know what it finished, what it planned next, or where the useful output is."

It stores plain local files:

• goals.json
• work-queue.json
• run-receipts.json
• work-artifacts/*.md

In one sentence: Agent Watchtower is a tiny local CLI that lets coding agents leave behind goals, next actions, run receipts, and artifacts so interrupted work can resume without rebuilding context from chat.

Quick Demo

ROOT="$(mktemp -d)"
agent-watchtower --root "$ROOT" init
agent-watchtower --root "$ROOT" worker-status
agent-watchtower --root "$ROOT" worker-run
agent-watchtower --root "$ROOT" artifact-path

You should see one runnable task, one completed worker run, and one markdown artifact path.

Close the terminal. Come back later. The local Watchtower files still say what happened and where the next session should resume.

For a clearer before/after story, see docs/interrupted-recovery-demo.md. 中文对比演示见 docs/interrupted-recovery-demo.zh-CN.md。

Two Readers

There are two different readers:

• Humans read this README.md to understand what the tool does and how to try it.
• Coding agents read AGENTS.md to learn exactly when and how to call the five public commands.

The README explains the value. AGENTS.md explains the operating rules.

For a plainer Chinese introduction, see docs/landing-page.zh.md.

中文说明见 README.zh-CN.md。

Install

Requirements:

• Python 3.11+
• A writable local runtime directory
• No background service, external account, or network dependency

From PyPI

python3 -m pip install agent-watchtower-core

Homebrew

brew tap un-n-smith/tap
brew install agent-watchtower

From GitHub

python3 -m pip install "agent-watchtower-core @ git+https://github.com/un-n-smith/agent-watchtower.git"

Pipx And Uv

For isolated CLI installs, pipx is also clean:

pipx install agent-watchtower-core

Or with uv:

uv tool install agent-watchtower-core

From A Local Checkout

python3 -m pip install .

After install, the CLI is:

agent-watchtower --help

For development without installing:

PYTHONPATH=src python3 -m agent_watchtower.cli --help

60 秒看到价值

For the default local runtime:

agent-watchtower init
agent-watchtower worker-status
agent-watchtower worker-run
agent-watchtower artifact-path

For a disposable demo runtime:

ROOT="$(mktemp -d)"
agent-watchtower --root "$ROOT" init
agent-watchtower --root "$ROOT" worker-status
agent-watchtower --root "$ROOT" worker-run
agent-watchtower --root "$ROOT" artifact-path

What to look for:

• worker-status shows whether one task is runnable.
• worker-run writes one markdown artifact and records one run receipt.
• artifact-path prints the latest artifact path.
• run-receipts.json records the completed bounded step and next safe action.

After this, close the terminal and come back later. The point is that the local files still say what happened and where to resume.

If you want a stricter release check, run:

./scripts/release_preflight.sh

For a scripted interruption demo from a source checkout, run:

./scripts/demo_interrupted_recovery.sh

Commands

Only these commands are part of v0:

init
task-add
worker-status
worker-run
artifact-path

Runtime

By default, data is stored under:

~/.agent-watchtower

Use --root <dir> for tests, demos, or project-local state.

For install, acceptance, and cleanup notes, see PACKAGING.md.

Feedback

This is an alpha project. Real interrupted-agent stories are more useful than abstract feature requests.

Please open an issue if:

• your agent lost track of work and Watchtower helped
• your agent could not understand the Watchtower state
• the install or quick demo failed
• you found a clearer way to explain this problem

User feedback is the main way this project gets better.

Uninstall

python3 -m pip uninstall agent-watchtower-core

To remove local runtime data:

rm -rf ~/.agent-watchtower

Boundaries

v0 is only a local continuity layer. It does not run a background service, create accounts, call external APIs, or claim to finish complex projects by itself.