v0.10.9
What's Changed
LLM Retry System — Now With 15 Minutes of Patience
Ever had an agent get stuck retrying while you're staring at the screen wondering "is it dead yet?" We've completely overhauled the retry logic to be smarter, more predictable, and way more patient. The old system had a convoluted timeout calculation that was hard to follow and varied based on retry counts. Now? A clean, unified 15-minute timeout across all LLM requests — plenty of time for those slow reasoning models to do their thing. We also extracted the retry schedule into a dedicated utility (makeLLMRetrySchedule) so the logic is centralized and easier to maintain. No more guessing games — if it takes more than 15 minutes, it fails fast and cleanly.
AI SDK Compatibility — No More Warning Noise
The AI SDK was complaining about how we handle system messages — specifically a "system-in-messages" warning that kept popping up in logs. We silenced that beast by adding allowSystemInMessages: true to both streaming and non-streaming LLM calls. Your logs are now cleaner, and the SDK is happier. Everyone wins.
CI Agent Model Upgrades
Our automated agents (ci-reviewer, pr-assistant, and release-notes) got a model upgrade — they're now running on minimax/minimax-m2.5:free instead of the previous openai/gpt-oss-120b:free. Faster, better, ready to review your PRs with renewed vigor.
Commits
5aabc714fix(llm): clean up retry design — scoped fallback, unified schedule, 15 min timeout (#237)e28bd614fix(llm): suppress AI SDK system-in-messages warning (#235)734dedf70.10.9