Content Pipeline & Chat Fixes (media caps, carousel revision, ideas, empty drafts)

2026-10-17 — five production complaints across publishing and the agent chat, their root causes, and what changed.

Why

A round of user reports pointed at the content pipeline:

Report	Symptom
Carousel posts	A multi-image post failed on Twitter/X, Bluesky, Facebook, Google Business
Chat carousel edit	"revise_carousel is not available in my current set of tools"
Chat "write post"	Output did not look like /app/posts/new (connected platforms should be on the left)
Content board "Find ideas"	Always "the agent found no ideas for this topic"
Card "Write now"	Returned success but produced no article at all

They are five separate defects.

What changed

1. Multi-image posts are clipped, never rejected

maxImages was defined once in packages/shared/shared/platformConstants.ts but only the client and the CLI/MCP validators ever read it — plugins were handed the whole postDetails.assets array. Bluesky (>4), Facebook (>10) and Google Business (>1) sent them all and the Graph API rejected the publish; X/Instagram/Threads hard-coded their own slice; LinkedIn/Pinterest/Snapchat silently dropped the extras.

AutoPostService.clipAssetsForPlatform() now trims a post's images to the target platform's maxImages before publishing (videos are never clipped; maxImages: 0 means "leave alone"). When images are dropped the user gets a warning notification naming the platform, so the cap is announced rather than silent. The clip happens on a copy, so other platforms in the same batch still get the full set.

2. Carousel edits work from plain chat

The revision harness only activated when the client sent carouselRevision, which the UI sets only via the page picker. Typing "change the slide count" fell through to the normal chat mount — which deliberately has no revise_carousel (the mount sits at the 8-tool budget) — so the model refused.

chat.post.ts now detects an edit verb + carousel noun (isCarouselRevisionIntent), resolves the business's most recent carousel artifact (contentArtifactService.listArtifacts → first outputKind: 'carousel') and routes the turn to revise_carousel in single-tool mode with that artifact. A structural request (isSlideCountIntent) gets a harness prompt that allows the whole deck to change, so "change the slide count" is not blocked by the "never rewrite the whole deck" rule.

3. "Find ideas" always returns cards

generateContentIdeas had a brief-derived fallback, but it only ran when the model responded with zero valid ideas. When the brief yielded nothing too, the batch was empty and the UI said "no ideas".

deriveGroundedIdeas() (new) keeps the brief-derived angles when there are any (so existing behaviour and counts are respected) and, when the brief yields nothing, tops up to at least MIN_IDEAS = 5 with evergreen angles grounded in the business name/industry/location (angleTemplates). Research failure now degrades to a topic-only brief instead of failing the scan.

Deliberate non-change: a downed provider still fails loud (CONTENT_INCOMPLETE) — generateIdeasOrFallback does not catch completion errors. tests/slice-b.test.mjs pins that: a provider outage must be diagnosable, not masked by generic template ideas.

4. "Write now" can no longer "succeed" empty

writePost/writeArticle returned success: true for any reply, including ''/{}; the artifact contract defaults caption/article to ''; and the UI toasted success unconditionally. So a blank completion produced a valid, empty artifact and a green toast.

The guard now sits at the composed-draft boundary (capabilities/content.ts draftHasContent): if the final draft has no caption, no article and no variants, content.write fails with EMPTY_OUTPUT and useContentEditor.write() shows a failure toast instead of success.

Placing the guard in writePost was wrong: an empty caption is legitimate when the article or platform variants carry the content (content-api.test.mjs "falls back to the platform body when the card has no short caption"). Guard the whole draft, not one field.

5. Chat "write post" embeds the real composer on the left

Covered separately in Chat post composer and the composer docs — PostModalContent gained a platformsPosition: 'left' mode (a vertical PostPlatformSelector column), ChatPostComposer gained an inline variant, and the chat page renders it inline in the thread instead of a modal.

How to verify

pnpm --filter @local-monorepo/agent test          # content chain + capabilities
cd packages/site && pnpm exec tsc --noEmit -p tsconfig.json
pnpm site:build                                    # compiles the Vue templates

Manual:

1. Compose a 6-image post, target Twitter + Facebook → it publishes to both; a "some images were skipped" warning names Twitter.
2. Generate a carousel, then type "change the slide count to 6" in chat → the deck is revised (no "tool not available").
3. On an idea card, click Find ideas for a thin topic → at least 5 cards.
4. Click Write now with the AI provider misconfigured → a failure toast, not a silent empty card.
5. Click Edit & schedule on a chat draft → the composer appears with the connected platforms on the left.

Related

• Provider Token Error Triage
• Carousel revision routing
• packages/scheduler/server/services/AutoPost.service.ts, packages/agent/server/content-intelligence/generate.ts, packages/agent/server/capabilities/content.ts

Chat / content tool & skill audit (2026-10-18)

Driving /app/chat and the content board surfaced two wiring gaps and one model-default bug:

• schedule_direct_posts was unmounted from the catalog. It is in CHAT_AGENT_MOUNT and defined in delivery.tools.ts, but was missing from AGENT_TOOL_CATALOG. Because AGENT_TOOL_NAMES derives from the catalog, the oversight/options UI never listed it and the chat route filtered it out of any client-supplied allowedTools. Added to the catalog (delivery group).
• The content-ops skill referenced the retired subagent tool. Delegation is Flue's built-in task; subagent no longer exists. The skill's spec and frontmatter also disagreed. Both now list board_list, board_move, board_add_cards, list_media, delete_media.
• The system default model was out of sync with the catalog.SYSTEM_DEFAULT_MODEL was gemini-2.5-flash, but models.json lists only Gemini 3.x (gemini-3.8/3.6/3.5-flash, 3.1-flash-lite, 3.1-pro-preview). Git history shows the catalog was bumped to 3.x while the constant stayed at 2.5 — exactly the drift the comment above it warns about. New users (no LLM config) therefore resolved a model the catalog does not list, which is what produced the "No response received" / "found no ideas" / empty-draft reports. The default is now gemini-3.5-flash.

Remaining environment finding: with the Google default, the model calls a tool and the tool runs, but the tool-result follow-up call returns 400 status code (no body) from Google's OpenAI-compatible endpoint. The same turn on openai/gpt-4o-mini completes with a real answer. The Google compat tool round-trip is the outstanding item for a clean chat reply.

Evidence: packages/doc/public/img/screenshots/app-chat/ (open, tools, message, reply, and a retained 04-chat-run.png showing a completed board_list) and app-business-id-content/01..09 (the manual board lifecycle).