Chat products love "suggested next steps." Teams often stuff them into a fenced JSON blob inside the assistant message. Then the stream truncates, the fence breaks, and the UI shows nothing, or a half-parsed ghost.
Why the JSON-in-markdown pattern breaks
The assistant's reply is a stream meant for a human to read, and it gets edited in flight: markdown renderers reflow it, streaming can cut off mid-token, and some UIs summarize or truncate long replies before showing them. A JSON fence surviving all of that intact is a coincidence, not a guarantee. When it breaks, the failure is silent: no error, just a button that never appears or a stray {"action": leaking into the visible text.
There's a second problem underneath the parsing one: a JSON blob the model wrote from scratch has no relationship to what actually happened in the turn. Nothing stops it from suggesting an action against a record the turn never looked up.
A typed side channel
Emit suggestions as a real tool call, not prose:
The server validates the schema before the UI ever sees it: required fields, known action enum, targetId shaped right. A malformed call fails the tool invocation, not the render.
Validate against what the turn actually touched
The schema check catches shape; a second check catches truth. Keep a set of IDs the turn actually fetched, created, or modified, and reject any suggested action whose targetId isn't in that set. This kills the class of bug where the model suggests acting on something it never actually saw: hallucinated ticket numbers, IDs from a previous turn, records that don't exist.
Fewer, clearer suggestions beat three clones of "look into this." Structured side channels survive streaming. Prose parsers don't.
