End-of-scenario performance review

A performance review that runs at the end of a full scenario, pulling together everything the user did across all stages, objectives, and team conversations into one objective read of the whole arc. It names patterns that no single stage could show: whether the user got faster, where they kept stumbling on the same kind of issue, how they grew in independence, what their writing style says about them.

Per-stage feedback is local: it talks about this submission, on this stage. A whole scenario is a story across many stages, and the platform needs to step back and tell that story honestly at the end. The end-of-scenario review names what the user did well across the arc, where they kept getting stuck, what they should push on in the next scenario, and an XP breakdown across the dimensions of the work. It is also the moment the platform reads each team persona's perspective on the user separately, so the final picture is not flattened into one tone.

A FastAPI WebSocket that gathers three contexts in parallel through asyncio.gather: prior objective rollups, per-persona conversation logs, and the user's own understanding artefacts. The contexts fan into multiple LLM calls. Per-persona perspective runs as separate parallel calls so the manager, the senior peer, and the cross-functional partner each keep their own voice instead of collapsing into one. Scenario scoring runs through Gemini with an OpenAI fallback for resilience, and a normalization step rescales the XP breakdown if the per-dimension sum drifts off the expected total. The output is validated against a strict JSON schema before it ships, so the rest of the platform never has to handle a malformed review.

Sole author of the WebSocket review pipeline, the parallel context fan-out, the per-persona perspective generators, the JSON validators, and the XP normalization step.