release aidev SIG 2/5

Pi v0.79.9: chat-template thinking for DeepSeek/vLLM, GLM-5.2 routing fixes

OpenAI-compatible custom providers can now map Pi thinking levels into chat_template_kwargs, enabling vLLM/Hugging Face chat-template models such as DeepSeek to use provider-native thinking controls. GLM-5.2 gets corrected Fireworks OpenAI-compatible routing and OpenRouter xhigh thinking support. Multiple fixes include session extension reuse, deep branch quadratic-time fix, streaming code fence flicker, fuzzy edit file rewrite, and /model ranking improvements.

PUBLISHED2026-06-20
OBSERVED2026-08-21
AGE2mo
SOURCES1
  • Chat-template thinking: OpenAI-compatible providers can map Pi thinking levels into chat_template_kwargs for vLLM/HF models like DeepSeek
  • GLM-5.2: corrected Fireworks routing to OpenAI-compatible Chat Completions endpoint with reasoning_effort
  • GLM-5.2: OpenRouter xhigh reasoning effort exposed via native xhigh parameter
  • Fixed same-directory session switches reusing imported extension modules while preserving fresh instances
  • Fixed deep session branches taking quadratic time to build context/branch paths
  • Fixed Markdown streaming code fence rendering — partial closing fences no longer shrink/flicker code blocks
  • Fixed fuzzy edit matches preserving untouched line blocks instead of rewriting whole file
  • Fixed bash commands through legacy WSL bash.exe passing scripts over stdin for variable expansion
  • Fixed /model hiding GitHub Copilot models unavailable to authenticated account
  • Fixed /model search ranking exact provider-prefixed matches before proxy-provider model ID matches
  • Fixed RPC unknown-command errors including request id to prevent client hangs
featurefix

COMMUNITY

No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.