In this initiative
Offwork began with a small wish: to stop working without having to keep the entire task in my head. Its name means leaving work on time. For someone working with AI agents, a useful place to start is the handoff between today’s session and the next one.
The Moment I Want to Improve
A session ends, a context window fills up, or I switch to another agent. The files are still there, but the reasoning is scattered through an old conversation. Which change was intentional? Did the test actually run? What was the unresolved question I meant to return to? Reconstructing that context becomes another task.
I started exploring Offwork for people doing local development with AI agents. The question I want the product to answer is simple: can someone with no access to the old chat look at this handoff and choose a sensible first step?
A Small Handoff, with Enough Context
At the end of a session, Offwork packages the task context into a Capsule: the goal, the agent’s account of progress, the project state, the results of configured checks, the open questions and the proposed next step. The saved record stays fixed, so it can be compared with what is true later.
On return, it presents a readable handoff and checks whether the relevant Git workspace has changed. The next agent can use that information to decide whether to continue, verify something first, or stop. Reopening the handoff leaves the actual next action for the user and agent to choose.
What I Chose to Keep Separate
The most useful design choice was to keep the agent’s description beside the checks that actually ran. In the repository demo, an agent says that every test passed while a configured check fails. Seeing both statements together gives the next person a concrete place to begin.
A saved record can also be intact while the working files have changed. And a passing test leaves a separate question: has the user reviewed this work? Keeping these judgments distinct makes the handoff more useful than a single ‘done’ label.
I kept the first version local and focused on a small CLI, with a Codex plugin as a simpler entry point. The product work is in making the stopping point legible: enough detail to resume, without asking someone to reread an entire session.
Where the Project Is Today
The public repository contains the Python CLI, the Codex plugin and a four-frame terminal walkthrough. It also includes recorded exercises in which fresh agent sessions chose to continue, verify or stop from the handoff they were given. These give me a technical prototype to examine and improve.
The recorded verification is on macOS. The workspace comparison covers the selected Git project; changes in an external service or database still need their own checks. Whether this helps people return to real work with less effort is the next product question.
What I Want to Learn Next
I want to use Offwork across ordinary interruptions: an evening break, a different agent, or a task whose files changed overnight. I want to see how long it takes to reach a useful first action, what still needs explaining, and whether the handoff helps someone notice an unresolved check before continuing.
That would give me a practical way to improve the product. The ambition behind 准点下班 is to make a stopping point feel usable—to leave enough clarity behind that tomorrow does not begin with reconstructing yesterday.