by 1337
What it does
Spec Rescue turns vague product ideas into reviewed MVP specifications. Three Mozaik agents independently analyze product scope, QA risks, and technical architecture in parallel. QA cross-examines their actual drafts, and Product revises the plan with acceptance criteria, implementation tasks, explicit assumptions, and open questions. The browser exposes original drafts, critique, final specification, and real timestamped runtime events. Markdown and JSON exports preserve the collaboration evidence. Runs on a local or hosted OpenAI-compatible model; genuinely tested with local Qwen, with reproducible setup and automated tests in the public repository.
How the agents run concurrently
Each request creates an isolated defineRuntime<RescueState>() session with three distinct createAgent participants. A brief.submitted event starts three non-blocking Mozaik runLoop calls; their model requests overlap before any draft completes. Typed shared state stores the drafts and review. Only after all three draft.ready events does QA receive the original brief and all drafts to critique contradictions and unsupported assumptions. review.ready then triggers Product synthesis, for five model calls in one bounded cycle. Agents coordinate through Mozaik semantic events and shared state; inference only occurs through Mozaik loops. A runtime observer records native inference.started/completed and model.answer events with loop IDs. Drafts, review, final artifact and timeline are inspectable and exported. Real local-model traces and test evidence are in docs/evidence; GPU execution may queue internally even while agent loops overlap.