Carousel revision routing in chat

The bug

Asking the chat to change one slide of a carousel produced a brand-new carousel on a nonsense topic, or rewrote the whole deck.

Two defects stacked:

1. The revision tool was never mounted. isCarouselIntent matched the word "revise" too, so the harness forced allowedTools = ['generate_carousel'] and single-tool mode. revise_carousel — the tool the request named — is not in CHAT_AGENT_MOUNT, so the model physically could not call it. It fell back to generate_carousel, which substituted the edit message itself as the research topic (ctx.carouselRequest), producing an unrelated deck under a new artifact id. The original was never touched.
2. Ids travelled as prose. The UI composed the ids into the message text (carousel.revisionMessage), which the backend regex-parsed — fragile, and it was exactly what poisoned the research topic.

A third, latent defect: revise_carousel had no way to know which slide was clicked, and carousel-generation.service.ts silently fell back to a whole-deck rewrite when it could not resolve a target slide from the instructions.

The fix

Structural routing, not prose. The client now sends a carouselRevision object, not ids baked into the message:

// packages/ai-tools/.../useAgentChat.ts
carouselRevision: { artifactId: string, slideIds: string[] }

POST /api/v1/agent/chat validates it (ChatRequestSchema.carouselRevision) and resolveCarousel(body, requestedTools) routes it:

• carouselRevision present → allow only revise_carousel, inject a revision harness prompt, pass the ids through toolContext.carouselTarget.
• creation verbs + "carousel" (and no revision) → allow only generate_carousel.
• otherwise → the normal allowed-tools path.

isCarouselIntent no longer matches "revise"/"update", so an edit verb can never be mistaken for a generation.

The tool prefers the structural target. revise_carousel reads ctx.carouselTarget.slideIds when the model passed none, so the clicked page is always the target.

No silent whole-deck rewrites. prepareRevision now rejects an ambiguous request that names a slide ("slide 2", "the second page") but resolves none, instead of shipping the whole deck back to the model.

Platform is preserved. The create workflow stores platform on the artifact output; revise returns the stored platform instead of hardcoding instagram.

The UI also gained a page picker next to "Request changes", so a specific slide can be targeted explicitly (empty = whole deck).

Gotchas

• MAX_TOOLS_PER_TURN is 7 and CHAT_AGENT_MOUNT already sits at 5 + 2 framework tools. revise_carousel is deliberately not added to the mount — it is reached only through the revision harness, which mounts a single tool.
• The harness prompt still lists Original user request / Requested changes for the model to read; ids must never be encoded there again.
• Whole-deck regeneration is still valid when the user explicitly asks for it (no slide is named), so the merge path is chosen by whether a target resolved.

Verify

1. Generate a carousel, pick page 2, ask for a change → only page 2 changes; every other slide is byte-identical; the same artifact id/version advances.
2. Ask "make it punchier" with whole deck selected → full regenerate is allowed.
3. Ask "change the slide about pricing" (no ordinal) → clear error asking which slide, never a silent rewrite.
4. pnpm --filter @local-monorepo/agent test passes.