v0.5.0

   🚀 Features

• Add unified video generation providers  -  by @sheepbox8646 in #20 (a8230)
• Add Xiaomi MiMo speech providers  -  by @MasterYuan418 in #21 (4d62b)
• Add Alibaba Cloud DashScope image generation provider  -  by @akazwz in #25 (f7bcb)
• Support developer and fallback message roles  -  by @sheepbox8646 in #27 (c9803)
• Add step commit barrier  -  by @chen-ran in #30 (ceb51)
• anthropic:
  • Support output_config.effort and budget-free adaptive thinking  -  by @qqqqqf-q and @cursoragent in #17 (f4a48)
  • Cache_control on tool_result and assistant blocks  -  by @ChrAlpha in #22 (8bc4b)
  • Model thinking config as a union, retire NormalizeReasoningEffort  -  by @HoneyBBQ (f1e49)
• google:
  • ThinkingConfig request-side support (T1 of #41)  -  by @HoneyBBQ in #42 and #41 (70292)
• openai:
  • Add MiniMax chat completions compat  -  by @akazwz in #16 (eedff)
  • Support prompt_cache_key for Chat Completions and Responses API  -  by @ChrAlpha in #24 (79661)
• provider:
  • Translate FileParts into each provider's native file format  -  by @wishworldbetter in #31 (93241)
• speech:
  • Support Deepgram model discovery  -  by @StringKe in #15 (0ca2d)

   🐞 Bug Fixes

• anthropic:
  • Always send input for tool_use blocks  -  by @StringKe in #11 (fa5c3)
  • Put the empty thinking string on the wire  -  by @HoneyBBQ (1e567)
• google:
  • Classify embedding models by supportedGenerationMethods  -  by @qqqqqf-q in #14 (1a82f)
  • Send tool schemas via parametersJsonSchema  -  by @akazwz in #18 (94281)
  • Preserve tool call thought signatures  -  by @akazwz in #19 (9b1f5)
• openai:
  • Handle usage delivered in trailing stream chunk  -  by @akazwz in #12 (43c38)
  • Support DeepSeek thinking disable  -  by @akazwz in #13 (a9dad)
  • Add Kimi chat completions compat for anyOf schema  -  by @chen-ran (84827)
  • Preserve max effort for Codex  -  by @qqqqqf-q in #26 (c19f5)
  • Rewrite Kimi MFJS schema normalization, require explicit compat opt-in  -  by @akazwz in #28 (198e8)
  • Request reasoning summaries  -  by @alexma233 in #38 (63be1)
  • Accept reasoning text delta events  -  by @alexma233 in #39 (ea90a)
  • Capture encrypted_content on output_item.done, not added  -  by @HoneyBBQ (0b060)
  • Handle Responses failed events  -  by @chen-ran (d852d)
• provider:
  • Close three review findings on the reasoning round-trip  -  by @HoneyBBQ (72553)
• reasoning:
  • Replay every reasoning block with its own opaque token  -  by @HoneyBBQ (c630a)
• sdk:
  • Aggregate cache write ttl usage  -  by @ChrAlpha in #23 (c5df6)

   🏎 Performance

• provider: Use strings.Builder for streamed tool-arg accumulation  -  by @Fodesu in #32 (14e8a)

    View changes on GitHub