# A real MoA+ team task The [current walkthrough](media/moa-plus-workflow.mp4) records the updated `main` interface opening a retained Note Bundle task. That task was executed during **MoA+ v0.2.0 release preparation**, using the matching DSH 0.2.1-rc.3 runtime and a real DeepSeek-V4.1-Flash model with High reasoning. The recording revisits its original conversation, Worker messages, and results; it does not start a new model task. The GIF and MP4 use sampled native UI clips edited into a walkthrough. They are not an uninterrupted screen recording, and the edit length is not a task-duration or comparative-performance claim. The published [v0.2.0 package](https://github.com/Burns1028/moa-plus/tree/v0.2.0) predates the wordmark control shown in the current interface. ## What the walkthrough shows The opening shows the **44 × 12 px MoA+ wordmark** in a new draft: gray off, green on, with no checkbox. It can be selected before the first message; a started MoA+ conversation keeps it as a fixed identity. Active letters use electric lime in dark themes and darker green in light themes. The recording then opens the retained task and its capsules, enters a Worker conversation and returns with Back, expands original Agent messages, and shows the Workers timeline, `notes.md`, and final results. The UI shots are newly captured; conversation content, Worker state, and tool results are displayed from the unmodified retained history. The 60-second edit uses 66 new native browser captures across six chapters. Chapter captions occupy a separate strip above the application; no conversation text or UI content is replaced. The refreshed control stays transparent even while hovered. ## What happened in the task The tool names below describe the historical v0.2.0 execution, not the current API. 1. The Coordinator read the Note Bundle task and started **Exporter** and **Docs** with `Task`, in a shared workspace with separate file ownership. 2. The Workers exchanged the CLI contract directly with `SendMessage`, including Request and explicit replies. 3. A human opened Docs from its capsule and asked for copyable commands and exact output headings. 4. Back in the main conversation, the Coordinator checked Docs' current state and relevant human/AI history with `QueryAgent`, then brought both Workers into line. 5. Workers returned complete final replies; the Coordinator checked actual output, saved reusable knowledge, and finished through the independent completion-review flow. ## Reproduce on v0.2.10 Copy [examples/note-bundle](../examples/note-bundle) into a new Git workspace. Prepare the [matching runtime](quickstart.md), create a new conversation, and select **MoA+** before its first message. Use separate source files for each Worker so the shared workspace has clear ownership. The fixture intentionally starts with one failing feature test. The prompt below uses the current tool names; it is not a claim that this guide has been rerun with v0.2.10. Send the Coordinator: > Finish TASK.md using two Workers in this shared workspace. Exporter owns note-bundle.mjs and tests; Docs owns docs/usage.md. Use moa_dispatch to delegate the two parts with workspace: "shared". Have the Workers agree the --sort title interface through moa_message. Save the confirmed CLI contract and test command in shared context. Run npm test and npm run demo before reporting the result. Keep it local: no new dependencies or commits. Use concise English. Open Docs and send: > Please include copyable commands for both filename ordering and --sort title, plus the exact headings each produces. Confirm the interface directly with Exporter, and save the accepted command in shared context. Keep the guide concise. Use **Back** and send the Coordinator: > I added a documentation requirement directly to Docs. Check its current status and relevant human/AI messages, keep both Workers aligned, and finish the existing task. Use native **Steer** when a message should interrupt a current wait; native Queue leaves it for the next turn. Check that the tests pass, `dist/notes.md` contains `# apples`, `# Bananas`, and `# Pears` in that order, and `docs/usage.md` documents both ordering modes. Record the actual outcome and any blocker in the [trial feedback form](https://github.com/Burns1028/moa-plus/issues/new?template=project-mode-trial.yml). ## Historical result The original v0.2.0 run passed **10 tests**, including default filename ordering, title ordering, heading fallback, tie-breaking, option errors, and output-file generation. The final `dist/notes.md` contains `# apples`, `# Bananas`, `# Pears` in that order. `docs/usage.md` includes commands and exact heading lines for both default and title sorting. The accepted contract and test command are indexed under `context/`. The earlier [v0.1.23 recording](demo-workflow.md) remains available separately. ## Original execution provenance The original task took about 10 minutes 36 seconds and ended normally after two native Verifier attempts. The first inspected the artifacts and ran the tests successfully, but prefaced its fenced JSON verdict with prose; the candidate's strict parser rejected that format. The task continued without a bypass, and the second attempt returned the accepted JSON verdict. The Coordinator completed 9 ms after that Verifier turn ended. v0.2.0 fixes this formatting mismatch; the original execution includes both attempts. After installing the completion-format fix, the same retained conversation was restarted and given a short follow-up to check saved context and outputs without new implementation work. Both Workers stayed idle. The Coordinator rechecked the existing outputs and 10 passing tests; an independent Verifier read the rubric and candidate, inspected files, and ran 20 Bash commands. It returned a plain JSON Pass, and the Coordinator completed 30 ms later. This follow-up took about 84 seconds. The prose-plus-fenced-JSON parser case is covered by the native regression suite; this follow-up did not reproduce that output format.