taskproof — a conveyor belt for AI coding agents

A conveyor belt for AI coding agents.

给 AI 编码 agent 用的传送带。

It does not decide what to do. It guarantees that whatever gets dispatched is independently verified and recorded — so you never have to trust an agent's own claim that it finished.

它不决定做什么。它保证派出去的活被独立验收并记账 —— 从此不必相信 agent 自称"已完成"。

Your agent says “done”.The acceptance command decides.

agent 说"完成了"。算数的是验收命令。

Download 0.3.0下载 0.3.0 GitHub Design docs设计文档

A card is dispatched, the agent claims it is done, and the machine disagrees. — A card is dispatched. The agent says it finished. The record says it changed nothing — the acceptance command is what decides.

How it works

怎么运转

1 · card

1 · 卡

A brief plus the command that will decide it. That command is the only thing believed.

一段任务 + 裁决它的命令。只有这条命令算证据。

2 · lane

2 · 道

It runs on a lane: your choice of codex, claude, gemini, opencode — or custom:<shell command>.

落在一条道上跑：codex / claude / gemini / opencode，或 custom:<任意命令>。

3 · ledger

3 · 账本

Either way it is recorded: what changed, the verdict, the full event stream. Local SQLite.

无论成败都记账：真实改动、裁决、完整事件流，本地 SQLite。

What the ledger holds

账本里有什么

The board: a failed card beside a completed one. — The board, served from the same renderer the desktop console loads. A card that claimed success and changed nothing sits next to one that actually did the work.

Recorded记录项	Why it matters为什么重要
change surface改动面	“it changed 0 files” is a fact, not an opinion."它改了 0 个文件"是事实，不是观点。
verdict裁决	An acceptance command, never the agent's summary. “Did not run” is never recorded as “passed”.验收命令说了算，不是 agent 的自述；"没跑"永远不会记成"通过"。
event stream事件流	Every run is auditable afterwards, including which budget stopped it.每次运行事后可审计，包括是哪个预算把它停下的。

Install

安装

# command line
pipx install taskproof

# latest development version
#   pipx install git+https://github.com/HaoHaoP/taskproof

# desktop console (macOS / Windows / Linux)
#   installers live on the Releases page

Python 3.11+, standard library only, no API keys, no LLM inside the tool itself — the judge must not be another model.

Python 3.11+，只用标准库，无需 API key，工具内部不含 LLM —— 判官不能是另一个模型。