{"content":"fix(model-context): read opencode limit metadata + gemini family fallback (commit 4c2e28384, branch fix/model-context-studio-stop). Session cw-636ddce2 died mid-turn: gemini-4-argon fetched from the opencode gateway synced with NO window metadata (the parser only read flat context_window/context_length fields), resolved to the 128K fallback, and the ~105K-token cowork baseline plus one turn overflowed it. Now parsing the gateway's nested limit:{context,output} record (gemini-4-argon verified live at 1048576/65536), carrying maxOutputTokens through mergeSyncedProviderModels, and adding a gemini-* family fallback (1M window / 64K output / vision=true) so uncatalogued SKUs and already-stored rows resolve the real window even before a re-sync. Modalities parsing now understands opencode flat-MIME and {input,output} record shapes; Zhipu input_modalities stays authoritative both ways. Tests: providerModels 13/13, coworkModelLimits 37/37.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}