[google-drive] Require local PNG inspection for URL thumbnails

plugins/google-drive/skills/google-drive/references/slides/chart-refresh-workflows.md

@@ -51,7 +51,9 @@ Use this guidance when the user asks to refresh, update, replace, or repair a ch
 
 - A chart refresh or replacement is only complete when the chart slide was re-read after the write and the visual result was checked with a thumbnail when the image itself matters.
 - Do not claim success just because the write call succeeded or because adjacent chart labels changed.
-- When thumbnail responses include an `image_asset_pointer`, image content part, or other rendered image artifact, inspect that returned image directly instead of downloading the `contentUrl` just to look at the same slide.
+- When thumbnail responses include an `image_asset_pointer`, image content part, or other rendered image artifact, inspect that returned image directly.
+- If the thumbnail response only exposes a thumbnail URL or `contentUrl`, curl the rendered image to a local PNG, for example `curl -L "$contentUrl" -o /tmp/slides-thumb-<slide-id>.png`, and inspect that local image before claiming the chart visual was verified.
+- After chart or layout edits, fetch a fresh thumbnail and re-curl any URL-backed thumbnail to a new local PNG before final success.
 - When the refresh or replacement changed visible layout, placeholder cleanup, or nearby styling, follow [visual-change-loop](./visual-change-loop.md) and keep the review loop going through the third fresh visual pass.
 - Verify both of these things before declaring success:
   - the chart graphic itself now matches the intended source chart

plugins/google-drive/skills/google-drive/references/slides/visual-change-loop.md

@@ -20,6 +20,8 @@ Use this recipe whenever a Slides write can change anything the user will see, e
 2. Start with a thumbnail.
 - Fetch a `LARGE` thumbnail when spacing, clipping, or shape alignment matters.
 - Inspect the image returned by `get_slide_thumbnail` directly. That may appear as an `image_asset_pointer`, an image content part, or another rendered image artifact in the tool response; do not require base64 bytes before starting visual review.
+- If the thumbnail response only provides a thumbnail URL or `contentUrl`, save the rendered image locally with `curl -L "$contentUrl" -o /tmp/slides-thumb-<slide-id>.png` before visual review.
+- Use the returned image artifact or the saved local PNG as the visual source of truth for overlap, clipping, padding, footer collisions, text rendering, and shape alignment.
 - Use that thumbnail as the primary visual signal, but not as the only signal for overflow or collision checks.
 - Write down the 2-4 concrete visible issues you are fixing in the next pass.
 
@@ -40,6 +42,7 @@ Use this recipe whenever a Slides write can change anything the user will see, e
 
 5. Verify immediately.
 - Fetch another thumbnail right after the write.
+- If the fresh thumbnail is URL-backed, curl it to a new local PNG and inspect that new file before declaring the pass verified.
 - Confirm both text and non-text visual targets actually changed.
 - Re-read the live slide structure after the write when text boxes, wrapping, or neighboring layout relationships were in scope.
 - If the write fixed the main text but left stale bars, arrows, borders, wrapping, or collisions, the slide is not done.

plugins/google-drive/skills/google-slides-visual-iteration/SKILL.md

@@ -59,7 +59,8 @@ If a dedicated visual-iteration tool exists in the runtime, use it. Otherwise, e
 - When the tool returns an `image_asset_pointer`, image content part, thumbnail artifact, or other rendered image in `content`, treat that as analyzable visual input for this workflow and inspect it directly.
 - Do not wait for image bytes, a data URL, or base64 image content before reviewing the slide. A returned image artifact from `get_slide_thumbnail` is the slide screenshot for this loop.
 - If `get_slide_thumbnail` succeeds, treat that as the visual verification path for this workflow even if the transcript view looks metadata-shaped. Do not abandon the thumbnail loop just because the runtime shows a thumbnail artifact, asset pointer, URL, or metadata wrapper instead of inline pixels in the message body.
-- The response may also include `contentUrl` metadata, but prefer the returned image content instead of downloading the URL or switching to another image-analysis path.
+- If the thumbnail response only exposes a thumbnail URL or `contentUrl`, curl the rendered image to a local PNG before visual review, for example `curl -L "$contentUrl" -o /tmp/slides-thumb-<slide-id>.png`.
+- Inspect the saved local PNG with the available image-viewing path before diagnosing overlap, clipping, padding, footer collisions, text rendering, or before claiming visual verification.
 - Do not switch to deck export, PDF rendering, or other fallback rendering paths when the thumbnail tool already succeeded. Only use a fallback path if the thumbnail action itself failed or is unavailable.
 
 4. Diagnose concrete visual problems.
@@ -100,6 +101,7 @@ If a dedicated visual-iteration tool exists in the runtime, use it. Otherwise, e
 
 6. Verify immediately.
 - Call `get_slide_thumbnail` again after every batch update.
+- If the fresh thumbnail is URL-backed, curl the new URL to a local PNG and inspect that new file. Do not reuse a PNG from before the write.
 - State which issues are now fixed, which issues remain, and whether the pass introduced any new regressions.
 - Confirm the targeted issue cluster is actually fixed before moving on.
 - Verify that small target or benchmark text changed along with the main headline value when both were in scope.

plugins/google-drive/skills/google-slides/SKILL.md

@@ -52,7 +52,9 @@ Confirm the runtime exposes the relevant Google Slides actions before editing:
 - For slide summaries or inspection, do not rely on text extraction alone when a slide contains charts, graphs, screenshots, diagrams, or image-heavy content.
 - Use `get_slide_thumbnail` alongside text/structure reads when visual evidence matters so the summary reflects both what the slide says and what the slide shows.
 - If a candidate slide has little useful extracted text but may still contain relevant evidence in a chart, screenshot, or diagram, inspect its thumbnail before ruling it out.
-- If the thumbnail response includes an `image_asset_pointer`, image content part, or other rendered image artifact, inspect that returned image directly as the slide image input. The response may also include `contentUrl` metadata; prefer the returned image content for visual understanding instead of downloading that URL or relying only on metadata.
+- If the thumbnail response includes an `image_asset_pointer`, image content part, or other rendered image artifact, inspect that returned image directly as the slide image input.
+- If the thumbnail response only exposes a thumbnail URL or `contentUrl`, curl the rendered image to a local PNG, for example `curl -L "$contentUrl" -o /tmp/slides-thumb-<slide-id>.png`, then inspect that local image before making or reporting visual judgments.
+- Do not claim visual inspection from slide JSON, text extraction, geometry, or thumbnail metadata alone.
 - Treat the slide page size as a hard boundary for every shape, text box, image, and color band you create.
 - If the write can change anything visible on the slide, such as text wrapping, shape styling, arrow direction, accent bars, chart placement, or connector styling, follow [visual-change-loop](./visual-change-loop.md) before the first `batch_update`.