fix(storage-resize-images): a malformed model reply to CUSTOM_FILTER_PROMPT blocks the image (existing extension bug, also in the kit) #3269

Existing bug in the storage-resize-images extension (0.3.6), carried over to the kit. It is not a kit/extension parity difference, so it is not tracked under #3263. Priority: medium. With CUSTOM_FILTER_PROMPT set, an image that should pass can be replaced by the placeholder, and the log blames the safety filters.

Problem

When the model's reply fails genkit's output schema check (must have required property 'response'), the code treats it as a block. One malformed reply turns a "no" into a placeholder, and the log says Image content blocked by Vertex AI content filters., which points the user at CONTENT_FILTER_LEVEL rather than the prompt. By reading, with DELETE_ORIGINAL_FILE=true a blocked image counts as a success, so the original is also deleted (not run).

Reproduced

CONTENT_FILTER_LEVEL=BLOCK_MEDIUM_AND_ABOVE, CUSTOM_FILTER_PROMPT="Does this image contain a cat?" and a custom placeholder. An image with no cat was uploaded three times to each side:

• Kit: placeholder in 2 of 3 runs.
• Extension: resized in 3 of 3 runs.

Direct calls with the extension's genkit version reproduced the schema failure on some replies. The moderation code is the same on both sides (by reading), so the split is put down to model nondeterminism.

Code

• Extension: content-filter.ts#L84, #L200-L203.
• Kit: content-filter.ts#L51-L60, #L140-L146.

Resolution

With a custom prompt, treat a schema failure as a retryable error rather than a block, and log a message that names the custom prompt. Keep the safety-filter block (finishReason: "blocked") as it is.