A vision editor action sends the image

Context

set_file_alternative_text and update_fal_asset_meta are editor actions on sys_file. The run they start says "table sys_file, uid 7", and nothing else reaches the model: it has no tool that shows it the image and no message that carries one. The action can only guess from nothing, and an alternative text written that way is worse than none.

VisionService analyses an image, but it is a separate call without tools. The run has to end in the tool call, because a declared write suspends for approval there (ADR-134), so the image has to be in the run.

Decision

An action declares the fields it fills from an image. EditorAction gains visionFields. A non-empty list makes the action a vision action; the two tools name their fields.

The image is a message of its own. The run holds the prompt as an ordinary string message and a second user message whose content is a part list with the image as a data URI. The providers convert such a part list: OpenAI, OpenRouter, Mistral and Groq pass it on, Claude and Gemini translate it, and Ollama, whose message content is a string, now takes it apart. A single part-list message would be shorter and would be wrong: InputGuardrailScreener reads string content only, so the editor's note would pass it unscreened.

The image is scaled first, and the result is measured. A vision provider needs no 6000-pixel photograph, and a large one would cost more on every send. FalImageDownscaler uses TYPO3's file processing, which caches the result in the storage's processed folder. In a backend request core defers that processing until a browser asks for the file, so the downscaler switches the deferral off around the call, as core's own ImageProcessingService does. Processing that fails hands back the original without an error, so the bytes are measured: an image whose longest edge is over vision.maxImageEdge, and a format no provider takes, are not sent. The original is never the fallback. A GIF becomes a PNG of its first frame, because Gemini takes no GIF and an animation can be large at any size. A file over 512 KiB that is no JPEG becomes one: the bound is on the edge, and a PNG within it would stay the original at any weight, in every waiting run.

Who may send which file. The file is resolved in the catalogue, before the run, which ADR-158 otherwise forbids. It has to be: the image is in the run from the first send. The read goes through FalStorageGate, the gate the tool itself uses, as the acting user. A file that does not exist, is not permitted, lies in an excluded folder or cannot be scaled gives one answer, no run, and the editor sees the message for an action that is not available. So the reason a catalogue had for not reading still holds: no uid is confirmed.

An action that cannot send the image does not run. Falling back to a run without it would be the defect this record removes. On a model that declares capabilities without vision the action is not offered; a model that declares none passes, as in model selection (ADR-138). A configuration that selects its model by criteria is offered the action only when its criteria require vision, because the model is chosen when the run sends.

The operator configures per field, in the extension configuration: vision.<field>.enabled, .maxLength and .prompt for alternative, title and description, plus vision.maxImageEdge and vision.excludedFolders. There is no page context for a file, so TSconfig is no place for them.

• enabled governs the editor action. An action with no field on is not offered, and a field that is off is not named in the prompt. It does not stop the assistant from using the tool: that is what the tool's own availability is for.

• maxLength is a rule about the field and applies to every writer of it, because the tool enforces it. The shipped value is 0, no limit of its own, so a fresh installation behaves as before and the shipped prompts ask for a length in words. A limit is named to the model as nine tenths of itself: models count characters badly, and an over-long value is refused on the approval card.

• .prompt replaces the shipped text for that field. It is the operator's text and is not fenced like the editor's note.

The prompt asks for the language of the acting backend user. A file has no page and so no site language to take it from.

Consequences

• An image leaves the installation, to the provider of the default configuration. That was never the case for these actions, and an operator who must keep a folder out names it in vision.excludedFolders.

• Input guardrails do not read images. Text in an image reaches the model unscreened. The single-tool allow-list, the acting user's rights and the approval card are the limits on what follows from it.

• The image is part of the transcript. It is in suspended_state, sealed, until the run settles. In the event stream it is verbatim at privacy.level = full and cut to the first 2000 characters of its base64 at redacted, and absent otherwise. It is sent again with the send after approval, which costs its tokens a second time.

• TranscriptEstimator counts string content, so the cost estimate on the batch page of ADR-162 does not count the image. It is an underestimate of the kind that record lists.

• An editor whose image cannot be sent reads that the action is not available, and not why. The price of confirming no uid.

• set_file_reference_alternative_text has no editor action and is unchanged.

• Core stores a failed processing attempt, so an image tried while the processor was broken stays unavailable until the processed files are cleared.

• FalStorageGate reaches storage 1 unless an integrator changes its arguments, so a file on another storage is not sent.

• A field that is off stays in the schema of update_fal_asset_meta. The prompt does not ask for it, and a model that fills it anyway shows that on the card.

Alternatives considered

A tool that returns the image to the model. Tool results are strings, on every provider path here, and a read tool in a one-tool run would have to be offered and paid for as a round trip of its own.

Generate with :php:`VisionService`, then start the run with the text. Two model calls, two configurations to keep apart, and a generated text in a prompt that tells a model to repeat it.

Revisit when

• Providers return image input cost separately: the estimate can count it then.

• A site language can be derived for a file, for example from the page that uses it.