v1.4.0

🎉 GoSlide v1.4.0

What's New

🧠 LLM transformer inside api components (experimental)

The api component accepts a new render item of type llm. Fetched
JSON is substituted into a user-authored prompt via {{data}} and the
model's reply renders inline alongside the chart, table, or metric.

~~~api
endpoint: /api/sales
fixture: sales.json           # optional; used by goslide build
render:
  - type: chart:bar
    label-key: quarter
    data-key: revenue
  - type: llm
    prompt: |
      Write 2 analyst bullets on these numbers:
      {{data}}
~~~

Control model (three layers, not configurable — that's the point):

• Cache-first — identical (model, prompt, data) triples call the
  LLM at most once. Results land in .goslide-cache/<sha256>.json
  (human-readable, commit-safe). Canonical JSON keying means logically
  equal data (different key order, whitespace) hits the same entry.
• Click-to-call — goslide serve shows a Generate ✨ button on
  cache miss. Page load never triggers an LLM call automatically.
  Localstorage caches per-browser.
• Build-lock — goslide build inlines cached results as a
  data-llm-bakes attribute on the api component. The exported HTML
  never contacts an LLM at view time.

# Warm cache via the dev loop:
goslide serve talk.md        # click Generate to populate .goslide-cache/

# Export to static HTML (reads cache only, zero network):
goslide build talk.md

# Or refresh cache during build (the one place we call LLM non-interactively):
goslide build talk.md --llm-refresh

Cache miss during goslide build is a hard error by default, listing
every affected slide / component / render-item. Pass --llm-refresh
to opt in to filling the cache during the build.

Offline build with a fixture file — fixture: ./sales.json on the
api component lets goslide build read static data instead of calling
a live endpoint. Great for committing a reproducible snapshot.

Reuses existing generate: config — no new YAML keys; the same
OpenAI-compatible endpoint that powers goslide generate powers the
LLM transformer. Works with any compatible provider (OpenAI,
OpenRouter, Ollama, vllm, sglang, etc.).

✅ Validation

The IR validator rejects llm render items missing prompt with code
llm-missing-prompt — goslide build / goslide serve won't start
against a broken deck.

📦 Validated example

examples/ai-generated/api-llm-sales/
is a fully self-contained directory. Clone this repo, cd into the
directory, run goslide build demo.md, and you'll see an LLM-written
analyst summary render in the output HTML without ever touching an LLM
— the committed .goslide-cache/ entry satisfies the bake. Real-LLM
regeneration requires OPENROUTER_API_KEY and --llm-refresh.

Compatibility

No breaking changes. v1.3.0 decks build unchanged; the new llm render
type is additive and only exercises the new code path when present.
The internal change to builder.Build (moving config.Load earlier so
the bake can mutate the IR before render) is invisible to callers.

Out of scope (future work)

• Streaming LLM responses (SSE). Current MVP buffers the full reply.
• JSONPath-style {{field.x}} expressions in the prompt. Current MVP
  injects the whole response as {{data}}.
• goslide generate emitting llm render items. Manual-author-only in
  v1.4.0.

Full Changelog

See v1.3.0...v1.4.0 for all changes.