v2.4.7: Pin Your OpenRouter Provider, Resume Any Conversation, and a Fresh macOS App

Third release of the day, and this one is heavier than the last. v2.4.7 hands you real control over how your OpenRouter requests get routed, repairs a "Continue conversation" bug that quietly affected every engine, raises the file upload ceiling, and pushes the macOS desktop app to 1.0.9.

Here is the full breakdown.

OpenRouter provider routing, with your prompt cache in mind

This is the headline. When you add or edit an OpenRouter model, you can now control which provider actually serves the request, right from the model picker. That matters more than it sounds, because OpenRouter sits in front of many providers for the same model, and switching between them mid-conversation quietly resets your prompt cache, which costs you both money and latency.

So we gave you the controls:

• Pin a provider from the list of the ones that genuinely serve that model. Each option shows its real per-provider price (in, out, and cache) plus a cache or no-cache badge, so you can deliberately keep your prompt cache warm. Pinning turns fallbacks off by default. If you leave them on, you get a warning, because OpenRouter could otherwise swap providers mid-conversation and wipe the cache you were trying to protect.
• Advanced routing options for the people who want them: provider order, allow fallbacks, only or ignore lists, sort by price, throughput or latency, require parameters, max price, data-collection and zero-data-retention.
• Provider search shows up automatically the moment a model has more than one provider behind it.
• The model list now sorts by "Most used" by default, based on real token usage from OpenRouter's daily rankings, with Recent, Price, Context and Name options available too.
• Image, audio and video models are hidden from the picker now (things like nano-banana). Only text-generation models show up, so the list stays relevant.
• Editing an existing OpenRouter model expands inline under its card instead of jumping you back up to a form at the top of the list.

"Continue conversation" works again, on every engine

This was a real one. After we moved to the headless streaming bridges, resuming a past conversation from History stopped restoring context. The old mechanism typed a /resume command into the terminal, but the new bridges only read JSON on stdin, so the command was simply ignored. The agent would answer your follow-up with no memory of the conversation you were trying to continue.

Fixed across the board:

• Claude: the bridge now relaunches with --resume <id> to reload the full context when the conversation already exists, instead of --session-id, which was erroring out with "Session ID already in use."
• Codex: the bridge now reloads the saved thread id on start, so your first message continues the existing thread instead of quietly opening a fresh one.
• Continue itself now repoints the session and respawns the bridge on the conversation you chose, rather than firing off a command the bridge was never going to act on.

If you rely on picking a thread back up days later, this is the fix that makes that trustworthy again.

Upload files up to 256 MB

The file explorer upload ceiling moves from 100 MB to 256 MB. Anything larger than 50 MB now uploads in chunks, which gets it past the per-request size limits that proxies and tunnels tend to impose, and the pieces are reassembled server-side. You get a per-chunk progress bar so you can watch it move. Smaller files still go up in a single request, exactly as before.

macOS desktop app 1.0.9

The macOS desktop build is now at 1.0.9. The dashboard's macOS download serves the .dmg installer, and the auto-updater feed (the manifest plus the update archive) is served right alongside it, so existing installs can quietly update themselves rather than making you reinstall.

The short version

Pin your OpenRouter provider and keep your cache warm, resume any conversation and actually get your context back, upload files up to 256 MB, and let the Mac app update itself. If you are running Geta.Team, it is already waiting for you.

Curious what an AI employee with real memory and real provider control can do on your own infrastructure? Have a look at geta.team.