release aidev SIG 4/5

Pi v0.81.0: llama.cpp model management, full provider extensions, Qwen Token Plan providers

Pi adds native local model management via llama.cpp (HF search, download, load/unload with live progress), lets extensions register complete pi-ai providers with auth and streaming, ships built-in Qwen Token Plan providers for international and China regions, and expands usage accounting to cover tools, compaction, and branch summaries. A large batch of fixes improves session startup latency, replay reliability, terminal cursor handling, and more.

PUBLISHED2026-07-21
OBSERVED2026-08-21
AGE1mo
SOURCES1
  • llama.cpp integration: Connect to a llama.cpp router, search/download Hugging Face models, explicitly load or unload models with live progress
  • Full provider extensions: Extensions can now register complete pi-ai providers with authentication, model refresh, filtering, and custom streaming
  • Qwen Token Plan providers: Built-in international and China subscription providers with regional endpoints and API-key auth
  • Expanded usage accounting: Tool, compaction, and branch-summary usage persisted in sessions and included in footer totals
  • RPC: Added get_available_thinking_levels command and RpcClient.getAvailableThinkingLevels() method
  • Exported types: Message and tool execution lifecycle event types exported from package root
  • Fixes include: persisted remote model catalogs no longer override newer bundled catalogs after upgrade; API-key credentials apply provider-scoped env values; OpenAI-compatible cross-provider replay keeps tool call IDs unique; Kimi K3 thinking levels exposed (low, high, max); GPT-5.6 Codex models default to 272K context window; compaction queued messages preserve steering/follow-up delivery; read tool errors no longer syntax-highlighted as file contents; model catalog refresh moved out of startup into running interactive/RPC modes; persisted sessions no longer read twice on open; prompt-template defaults for all arguments; obsolete extension docs examples removed
featurecapability

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.