Turn a paragraph into a reviewable backlog

A feature request arrives as a paragraph from a sales call. Turning it into an epic with the right children — correctly typed, roughly estimated, sensibly prioritised — is an afternoon of clerical work that nobody enjoys and everybody postpones. This is the decision PlanMagnet automates first, because it is the one where the model's output is easy for a human to check in seconds.

The request goes to a work-item breakdown call in the PlanMagnet service. The model is asked for a strict JSON array and nothing else, at a low temperature, with an explicit ceiling on how many items it may propose. What comes back is parsed and validated rather than trusted: item types are checked against PlanMagnet's real work-item types, priorities against its real priority levels, estimates must be finite and non-negative, and any suggestion missing a title fails the whole parse. A response the service cannot parse into valid suggestions produces a clear error — never a partial or invented list.

The most important property is what the call cannot do. It returns suggestions and writes nothing. There is deliberately no auto-create path: a reviewer reads the list, edits or discards what is wrong, and the ordinary work-item creation flow persists the survivors — the same flow that already handles project lookup, key assignment, permission checks, and the audit trail. So the backlog never fills with items nobody agreed to, and the AI path does not get its own privileged write route into your plan.

Every call is written to an append-only ledger row before the result is returned, capturing the provider, the model, token counts, latency, and the parsed suggestions themselves. Failures are recorded too, including the case where the provider answered but the answer could not be parsed. What the ledger deliberately does not store is your raw prompt — only a short non-identifying summary of what was asked.

The backend for this is live in the PlanMagnet service today, behind a permission check enforced at the API gateway. The reviewer UI that turns it into a two-click flow inside the app is committed roadmap, so today this decision is reachable through the API, the CLI, and agent tooling rather than a button. This story is narrative-only today; the interactive app walkthrough is not yet available.