v0.1.1

• Attachment guard at execute time: unsupported attachment types now fail locally and free, before any HTTP request. Works around llm core skipping attachment validation on the Conversation.prompt path (llm -c, llm chat), which forwarded unsupported attachments to the endpoint as a paid 400. Covers the PDF case on vision models too (the endpoint refuses OpenAI file content parts).

v0.1.0

Initial release.

• Live model catalog from api.tiyuvta.ai — new models appear without a plugin update
• Per-model capability detection: vision attachment types only where the model has eyes
• Streaming, tools, schemas via the OpenAI-compatible endpoint
• Resilient catalog cache: atomic writes, corrupt-cache repair, stale-cache fallback on network failure or interstitial HTML
• llm tiyuvta models / llm tiyuvta refresh commands