AfterPause

Inspiration

An interruption leaves two questions: what did I finish, and what was I about to do? AfterPause saves the exact next step alongside a short handoff note. Its central rule: elapsed time is never evidence of physical completion.

What it does

Create an ordered everyday task, explicitly confirm completed steps, and pause with a note. Return through the browser or a separate MCP client to the same persistent checkpoint. A reminder can expire while paused without advancing the checklist. The demonstration uses synthetic packing tasks, not physical-world verification.

How we built it

Python and the official MCP SDK expose four tools over Streamable HTTP: list_tasks, create_task, get_checkpoint and update_task. A real client negotiated protocol 2025-11-25. SQLite stores task state and events. Revision checks reject stale or duplicated updates; clients must reread after a conflict. The web interface and independent reference client share the store.

The reference client asks for explicit confirmation before advancing. The web action is labeled “I completed this step.” Other MCP clients must honor the same confirmation contract; the server cannot verify real-world actions.

Challenges and lessons

Resume must preserve context, not merely reopen a checklist. We separated timer state from completion state and protected updates with revisions. A local proxy initially caused loopback client requests to fail with HTTP 502; disabling inherited proxy configuration for that local client resolved it. This was an environment issue, not an Amazon SDK defect.

Accomplishments

Six core tests cover recovery, stale writes, reminder semantics, isolation, terminal states and rollback. A live transport test checks tool discovery, protocol negotiation, shared state and cross-origin rejection. The recorded demo shows actual browser operation and actual MCP responses. A reproducible demo client and test instructions are in the MIT-licensed repository.

Current limits

This is a local, single-user prototype for ordinary low-risk tasks. It has no Alexa device integration, speech recognition, remote identity, OAuth or background notifications. It is not a safety-critical reminder. No user adoption or time-saving outcomes have been measured.

What's next

Validate interruption-and-return workflows with users, then explore authenticated remote access and an assistant integration while retaining explicit completion and conflict handling. The submission demonstrates the self-hosted MCP implementation and reference clients, not an actual Alexa session.

Built With

• python
• sqlite