Vision was the story this week. Three separate toolkits shipped releases within days of each other, and the pattern is now clear: the ecosystem has settled on structured evidence rather than prose description as the right output for a model that cannot see. Prose is a demo format. Coordinates, region labels and crop requests are an API.
Modlens leads the vision pack
Modlens updated with tighter layout segmentation, and its JSON evidence format is now the reference implementation other vision plugins are converging on. Text with coordinates, region labels, and a crop tool that lets the model ask for a closer look is a fundamentally better contract than “describe this image”.
The pin on install matters for the same reason: parsers downstream of the bridge break if the schema shifts under them. Prefer a bridge that versions its evidence shape.
Vision Router and Vision Toolkit both shipped
Vision Router added pixel diffing to its tool set, which makes visual regression checking a normal tool call rather than a bespoke workflow. Vision Toolkit improved long screenshot tiling, the failure mode that made OCR useless on full-page captures.
Zero-key setups matter here: both ship a free vision chain, which explains the install numbers. Free chains are fine for experiments and wrong for ongoing screenshots of customer UIs sent to an anonymous endpoint. Configure a keyed path before the first sensitive capture.
Context tooling quietly became essential
DSH Context keeps climbing. The reason is practical: once sessions run past an hour, knowing what is in the window stops being curiosity and becomes necessary maintenance. Its composition view and compaction event log are the two features people mention most.
Seeing the window is still not the same as budgeting it. If you need a hard stop, pair inspection with an orchestration plugin that has an explicit budget — or accept that visibility alone will not save an unbounded run.
Memory release worth pinning
Memento shipped a schema refinement. If you depend on its snapshot format, pin the version, as the early releases are still settling. The approval gate on writes remains the feature that distinguishes it from the rest of the memory category.
Also worth a look
Background Agents released a durability fix so child agents survive a page reload, and MisakaNet continued to fill its shared debugging-lesson library.
What we are watching
Two things for next week: whether the interface plugins keep pace with the latest preview release, and whether any of the vision plugins adopt a shared evidence schema instead of three slightly different JSON shapes.
Convergence would let prompts and parsers move between plugins. Until then, pick one vision contract per workflow and do not stack bridges that all claim “image understanding”.
What this means if you install
If screenshot OCR is already in your workflow, verify long-capture tiling before you compare star counts. For pixel regression, Vision Router’s diff tools are closer to the job; for layout-aware text coordinates, Modlens’s evidence format is the safer contract. A free vision chain is fine for experiments and wrong for ongoing screenshots of customer UIs sent to an anonymous endpoint. If Memento is already in your pipeline, pin the version rather than chasing every patch.
Install order if you are starting fresh: one vision bridge with a stable schema, then a context dashboard if sessions run long. Do not add a second vision toolkit until the first one fails a concrete task you care about.