-
-
Actual local interface: the updated East-door briefing is confirmed; unrelated source checks were preserved.
Inspiration
A community organizer preparing a volunteer briefing can be interrupted halfway through. Copying the chat to another assistant does not answer the important questions: which facts were checked, what is unfinished, and whether the venue instructions have changed. We wanted continuity that follows the actual work rather than a conversation summary.
What it does
Agent3 Purser Handoff turns three workshop source documents into a volunteer briefing. The local assistant checks programme details, venue instructions and assignments, saves the extracted facts with source versions, and drafts a cited Markdown file. You can pause and return with another assistant without repeating finished checks. The replaced assistant can no longer write.
If the entrance changes from West door to East door, the app keeps the programme and volunteer checks, rechecks the venue, and invalidates the old briefing. You review the new version and explicitly confirm it before downloading. Nothing is booked, paid or sent automatically.
How we built it
An English browser UI orchestrates the deterministic workflow. A Node.js/Express backend owns the state. Task revisions and server-issued runner permissions prevent conflicting or expired writes. Source version/hash dependencies support incremental recovery. An exclusive writer lock and atomic durable snapshots preserve progress through restarts. Confirmation binds an exact SHA-256 document digest; the download gate rejects outdated or unconfirmed drafts.
The official Model Context Protocol SDK is imported and called at runtime. A real self-hosted Streamable HTTP endpoint negotiates MCP 2025-11-25. An independent official SDK Client executes the complete workflow over HTTP. The same state service powers the UI and tools.
Challenges
A retry is not always proof that the old result is still current. We found that an old confirmation replay could misleadingly return a previously completed task snapshot after a source update. We changed replay responses to return the current task and an explicitly historical originalResult, then added a regression. We also fixed an actual first-run UI rendering fault and isolated cookies by port so independent local test servers do not overwrite each other's pairing. These were our implementation issues, not claimed Amazon platform defects.
Accomplishments
The demonstration performs real source checks, pauses, reloads, takes over, rejects an old runner, generates a cited draft, changes the entrance, updates only the affected source, confirms the regenerated draft, and downloads the actual file. Independent tests cover concurrent takeover, stale/expired permissions, withdrawal, missing/conflicting facts, idempotency and a forced process crash/restart. The final downloaded bytes match the owner's confirmed digest.
What we learned
Continuity needs both saved progress and a rule for invalidating it. A saved checkpoint is useful only if it carries actual output and its source dependencies. The interface should explain what remains and why, while the server decides whether a write or confirmation is still valid.
What's next
A future version could add authenticated cross-device sessions, additional document adapters and a genuinely integrated Alexa experience. Those capabilities are not part of this local prototype.
New work and reuse disclosure
Created on 23 October 2026. Earlier Agent3 Purser work addressed payment authorization and delivery checks. This project retains the name and interruption-recovery concept, but implements a new workshop briefing workflow, incremental source checks, automatic preparation and pause, revision/runner state machine, exact-document approval and MCP transport. No first-party source code or private data was copied from the prior payment implementation. Dependencies retain their licenses; our application is MIT licensed. OpenAI Codex assisted development; its precise model revision was not independently verified. No runtime LLM is required.
Current prototype scope
This is a browser-based Alexa+ experience simulation, not an official Alexa integration. Phone/desktop are runner identities on one local server, not a hosted multi-device service. All workshop input data is synthetic; document creation and safety operations are real. Deterministic label:value parsing supports this bounded scenario. Trusted local clients share owner authority; tool instructions cannot prove a human actually approved. No cloud, payment, message sending, customer adoption or production authentication claims are made.
