OpenClaw Main Stops Silently Dropping Unsupported Attachments

OpenClaw has merged a closed outcome model for inbound attachments into main. Documents, images, audio and video that are not processed should now leave a bounded marker the model can explain to the user instead of disappearing from the turn.

This fixes a particularly nasty failure class: the user attaches evidence, the agent answers as if no attachment existed, and neither side knows intake discarded it.

Documents now have an exhaustive outcome

PR #122025 replaces silent document continue paths with a typed result for each file. Unsupported Office formats, disabled URL access, read or extraction failure, and empty documents produce model-visible markers.

A DOCX that OpenClaw cannot read, for example, should no longer vanish. The model receives a concise explanation that the format is unsupported and can ask for PDF or plain text.

The important engineering change is not the exact wording. Every non-claimed result passes through one exhaustiveness-checked renderer. Adding a future gate requires adding an outcome the renderer must handle.

Images, audio and video record the reason at the owner

The first document fix deliberately did not guess whether media had been handled elsewhere. PR #122098 completes that custody chain with per-attachment dispositions such as:

• handled;
• handed to native vision;
• not selected by first-only policy;
• capability disabled;
• no configured model;
• denied by scope; or
• processing failed.

A late renderer consumes those recorded facts. It does not infer delivery from a missing output, which avoids telling the model an image was skipped while a native-vision route is actually looking at it.

Markers share one budget: at most five detailed document/media notices plus one overflow summary. That limits prompt amplification when a user sends a large batch of unsupported files.

What this does—and does not—prove

The change proves better intake accounting. It does not make an unsupported format readable, configure a missing vision model, grant denied media scope or guarantee that every provider understands every modality.

There is also a documented residual around partial native image resolution: the merged code suppresses conservatively when delivery facts are incomplete rather than risk a false “not analyzed” claim beside an image the model received. Operators should still inspect provider payloads and durable attachment state when a multimodal incident is ambiguous.

Rollout test

Use synthetic attachments with no sensitive contents:

1. send one unsupported Office document and require an explicit marker plus a useful user-facing explanation;
2. disable URL attachment access and test a URL-only file;
3. force a document read failure and confirm the turn does not silently omit it;
4. send four images with first-only selection and require dispositions for the unselected images;
5. test with no image-understanding model and with the capability disabled;
6. send an image through a verified native-vision route and require no false skip marker;
7. combine more than five failed document/media outcomes and verify one bounded overflow summary; and
8. compare model input, transcript and attachment store so a marker cannot hide a custody bug.

Do not use customer documents to test error reporting. The point is to prove classification, not provider quality.

Bottom line

An attachment should be handled or explicitly accounted for. OpenClaw’s new disposition model makes silent disappearance much harder and gives the agent enough bounded context to tell the user what went wrong. Treat that as observability, not as new modality support.