v0.9.2

Changed

• Vendored inference library upgraded: @earendil-works/pi-ai 1.1.0 → 1.3.0 (both packages, lockstep). Upstream brings automatic retry for more transient provider failures (service is temporarily unavailable, gateway upstream errors, unable to forward upstream), an Anthropic retry when the provider rejects a strict tool set, Mistral requests no longer aborting mid-generation when only the response headers were slow, OpenRouter placeholder prices no longer reporting −$1,000,000 per 1M tokens, and model-specific thinking fixes (Gemini 4 thinking levels on Vertex and the Gemini API, Claude Haiku 5.5 / Opus 5 off level now honored upstream, GPT-6 cache-preserving effort changes, Xiaomi MiMo no longer offering unsupported thinking levels). Upstream-side: Seepient's own thinking-level mapping does not yet surface the Claude off level (mapping gap tracked for a follow-up).
• 42 upstream catalog entries were removed in pi-ai 1.3.0. Assignments pointing at a removed id stop resolving (unknown_model); reassign with seepient models set. Removed, grouped by provider — openrouter: the qwen3 family (qwen/qwen3-max, qwen/qwen3-max-thinking, qwen/qwen3-235b-a22b, qwen/qwen3-30b-a3b-thinking-2507, qwen/qwen3-8b, qwen/qwen3-coder-plus, qwen/qwen3-vl-8b-thinking, qwen/qwen3-vl-32b-instruct, qwen/qwen3-vl-235b-a22b-thinking, qwen/qwen3.6-max-preview, qwen/qwen-plus-2025-07-28), anthropic/claude-sonnet-4, the inclusionai ling-3.0-flash variants, apodex/apodex-1.1-mini:free, sao10k/l3.1-euryale-70b, the unbiased/pareto pair, and both image models inclusionai/ming-image-0.1-design and -design-layer; vercel-ai-gateway: alibaba/qwen3-max, alibaba/qwen3-max-preview, alibaba/qwen-3.6-max-preview, alibaba/qwen3-235b-a22b-thinking, alibaba/qwen3-coder-plus, alibaba/qwen3-vl-235b-a22b-instruct, alibaba/qwen3-vl-thinking, anthropic/claude-3-haiku, anthropic/claude-opus-4, and inclusionai/ling-3.1-flash-free (the other ~25 vercel qwen entries remain); opencode-go: all four chat models (deepseek-v4-flash, deepseek-v4-flash-vision-exp, grok-4.6, minimax-m2.7); opencode: the four free tiers; google-vertex: gemini-flash-latest and gemini-flash-lite-latest; nvidia: moonshotai/kimi-k3 and nvidia/nemotron-3-super-120b-a12b.

Fixed

• OpenCode Zen and OpenCode Go moved upstream to the opencode.ai/inference endpoints, which adds Google Generative AI (Gemini) support to OpenCode Go — a capability it did not offer before — and unblocks the provider's other models behind the retired endpoints; model resolution picks the new endpoints up from the catalog automatically.
• kimi-coding now dials the official Kimi Code endpoint (https://api.kimi.com/coding/v1) over the OpenAI chat-completions wire (previously Anthropic Messages), so tools added mid-conversation no longer invalidate its prompt cache. If you override the base URL of a kimi-coding account, it must now point at the /v1 root.

Known limitation (cross-model replay)

• Thinking blocks from earlier turns still reach a newly selected model. Upstream 1.3.0 drops (rather than inlines) thinking produced by a different model at provider switches, but Seepient's conversation replay stamps every prior assistant message with the current target's identity, so the upstream drop never fires here and prior-model reasoning can still enter the new model's context. Provenance threading through the replay path is tracked for a follow-up release.