v2.27.1

Features

• Refresh the picker's models from inside the router
• Refresh stale model lists in the background
• Build the picker from the cached model lists
• Cache each source's model list on disk
• Read a source's live model list
• Pick the newest model per family from a listing
• Offer a DeepSeek API key as a subscription
• Offer the models the GLM Coding Plan actually serves

Bug Fixes

• Never follow a model-list redirect with the key
• Write the profile through a unique temp file
• Stop an unchanged model list from rewriting the profile
• Give the router the default login's tag file
• Stop probing ps with a pid list
• Refresh a borrowed login's OAuth token instead of 401ing
• Keep every subagent on the row the session picked
• Keep Explore subagents on the row the session picked
• Stop sending the one tool schema DeepSeek refuses
• Stamp the tool denial only on a provider the profile really names
• Stop sending the one tool schema z.ai refuses

Other Changes

• Perf(allin): stop fetching every subscription's quota at launch