openclaw - 💡(How to fix) Fix Persist image understanding summaries for agent attachments

Problem

A local image-context probe showed that OpenClaw can send a native image to the model and openclaw infer image describe can produce a good durable summary/OCR-style description, but the agent/LCM path did not persist any attachment metadata or media-understanding output for reuse. Follow-up turns can only recover image content if the assistant happened to mention it later.

Desired behavior

• Agent image attachments should expose/persist durable media memory alongside the raw image reference: filename/source label, MIME type, byte size, dimensions/hash when known, attachment provenance, and image.description text/OCR output.
• Reuse existing media-understanding output when OpenClaw already produced it.
• If no prior summary exists, run image.describe once at ingest or maintenance time, not repeatedly during hot context assembly.
• Default summary budget should be about 512 tokens, with an OCR-heavy flyer/screenshot cap around 2,000 tokens.
• Fresh turns can still send the native image to capable models, but old/compacted/tool-stubbed context should receive a metadata+summary stub by default; raw image replay should be explicit.
• If summary generation fails, persist a metadata-only stub and raw reference so the agent still knows media existed.

Why now

Martian-Engineering/lossless-claw#775 is being hardened so LCM can store image files and assemble bounded image-memory stubs instead of replaying raw base64. For best results, OpenClaw should hand lossless-claw the media-understanding summary/metadata directly instead of requiring LCM to infer it later from assistant prose.

Acceptance criteria

• A deterministic image with visible text is sent through a local agent session.
• The first model call can use the native image.
• The durable context path stores/reuses image metadata and summary without retaining base64 in normal assembled context.
• A follow-up turn can answer from cached image memory after raw image replay is disabled.
• OCR-heavy flyer text is preserved under the 2,000-token cap.