Critical Fixes
Remove channel timeout that killed autonomous tasks
The channel adapter had a hardcoded 600s (10 min) timeout on wait_for(). Long-running autonomous tasks (browser registration flows, multi-step promotion, etc.) were killed after 10 minutes even though the agent was still working fine. The agent kept running in the background but the user saw "Request timed out." Removed the timeout entirely — the agent loop has its own safety limits (step caps, budget limits, cost caps).
Fix embeddings wiped on every startup
create_vec_table() was doing DROP TABLE IF EXISTS + CREATE on every startup, deleting all embeddings. This forced a 20-26s full re-index on every launch. Now only recreates the table if it doesn't exist or embedding dimensions changed.
Smart timeout fallback for LLM providers
On LLM call timeout, immediately fall through to the next provider instead of retrying the same one 3x. Z.ai timeout bumped 120s → 180s.
Upgrade from v2026.02.24.1
All three fixes are critical for autonomous operation. The channel timeout bug made it impossible to run tasks longer than 10 minutes. The embeddings bug wasted 20+ seconds on every startup.