@bastani/atomic-ai 0.0.0-placeholder.0 → 0.9.88-alpha.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +1601 -0
- package/LICENSE +22 -0
- package/NOTICE.md +16 -0
- package/README.md +166 -0
- package/dist/api/anthropic-assistant-content.d.ts +24 -0
- package/dist/api/anthropic-assistant-content.d.ts.map +1 -0
- package/dist/api/anthropic-assistant-content.js +99 -0
- package/dist/api/anthropic-assistant-content.js.map +1 -0
- package/dist/api/anthropic-messages.d.ts +71 -0
- package/dist/api/anthropic-messages.d.ts.map +1 -0
- package/dist/api/anthropic-messages.js +1814 -0
- package/dist/api/anthropic-messages.js.map +1 -0
- package/dist/api/anthropic-messages.lazy.d.ts +3 -0
- package/dist/api/anthropic-messages.lazy.d.ts.map +1 -0
- package/dist/api/anthropic-messages.lazy.js +3 -0
- package/dist/api/anthropic-messages.lazy.js.map +1 -0
- package/dist/api/azure-openai-config.d.ts +14 -0
- package/dist/api/azure-openai-config.d.ts.map +1 -0
- package/dist/api/azure-openai-config.js +61 -0
- package/dist/api/azure-openai-config.js.map +1 -0
- package/dist/api/azure-openai-responses.d.ts +14 -0
- package/dist/api/azure-openai-responses.d.ts.map +1 -0
- package/dist/api/azure-openai-responses.js +213 -0
- package/dist/api/azure-openai-responses.js.map +1 -0
- package/dist/api/azure-openai-responses.lazy.d.ts +3 -0
- package/dist/api/azure-openai-responses.lazy.d.ts.map +1 -0
- package/dist/api/azure-openai-responses.lazy.js +3 -0
- package/dist/api/azure-openai-responses.lazy.js.map +1 -0
- package/dist/api/bedrock-converse-stream.d.ts +38 -0
- package/dist/api/bedrock-converse-stream.d.ts.map +1 -0
- package/dist/api/bedrock-converse-stream.js +1301 -0
- package/dist/api/bedrock-converse-stream.js.map +1 -0
- package/dist/api/bedrock-converse-stream.lazy.d.ts +16 -0
- package/dist/api/bedrock-converse-stream.lazy.d.ts.map +1 -0
- package/dist/api/bedrock-converse-stream.lazy.js +43 -0
- package/dist/api/bedrock-converse-stream.lazy.js.map +1 -0
- package/dist/api/classifier-shared.d.ts +11 -0
- package/dist/api/classifier-shared.d.ts.map +1 -0
- package/dist/api/classifier-shared.js +90 -0
- package/dist/api/classifier-shared.js.map +1 -0
- package/dist/api/cloudflare-gateway-binding.d.ts +58 -0
- package/dist/api/cloudflare-gateway-binding.d.ts.map +1 -0
- package/dist/api/cloudflare-gateway-binding.js +33 -0
- package/dist/api/cloudflare-gateway-binding.js.map +1 -0
- package/dist/api/cloudflare-workers-ai-system-one.d.ts +4 -0
- package/dist/api/cloudflare-workers-ai-system-one.d.ts.map +1 -0
- package/dist/api/cloudflare-workers-ai-system-one.js +57 -0
- package/dist/api/cloudflare-workers-ai-system-one.js.map +1 -0
- package/dist/api/cloudflare-workers-ai-system-one.lazy.d.ts +3 -0
- package/dist/api/cloudflare-workers-ai-system-one.lazy.d.ts.map +1 -0
- package/dist/api/cloudflare-workers-ai-system-one.lazy.js +4 -0
- package/dist/api/cloudflare-workers-ai-system-one.lazy.js.map +1 -0
- package/dist/api/cloudflare.d.ts +11 -0
- package/dist/api/cloudflare.d.ts.map +1 -0
- package/dist/api/cloudflare.js +11 -0
- package/dist/api/cloudflare.js.map +1 -0
- package/dist/api/constrained-sampling.d.ts +38 -0
- package/dist/api/constrained-sampling.d.ts.map +1 -0
- package/dist/api/constrained-sampling.js +271 -0
- package/dist/api/constrained-sampling.js.map +1 -0
- package/dist/api/github-copilot-headers.d.ts +40 -0
- package/dist/api/github-copilot-headers.d.ts.map +1 -0
- package/dist/api/github-copilot-headers.js +76 -0
- package/dist/api/github-copilot-headers.js.map +1 -0
- package/dist/api/google-generative-ai.d.ts +15 -0
- package/dist/api/google-generative-ai.d.ts.map +1 -0
- package/dist/api/google-generative-ai.js +376 -0
- package/dist/api/google-generative-ai.js.map +1 -0
- package/dist/api/google-generative-ai.lazy.d.ts +3 -0
- package/dist/api/google-generative-ai.lazy.d.ts.map +1 -0
- package/dist/api/google-generative-ai.lazy.js +3 -0
- package/dist/api/google-generative-ai.lazy.js.map +1 -0
- package/dist/api/google-shared.d.ts +103 -0
- package/dist/api/google-shared.d.ts.map +1 -0
- package/dist/api/google-shared.js +536 -0
- package/dist/api/google-shared.js.map +1 -0
- package/dist/api/google-vertex.d.ts +17 -0
- package/dist/api/google-vertex.d.ts.map +1 -0
- package/dist/api/google-vertex.js +440 -0
- package/dist/api/google-vertex.js.map +1 -0
- package/dist/api/google-vertex.lazy.d.ts +3 -0
- package/dist/api/google-vertex.lazy.d.ts.map +1 -0
- package/dist/api/google-vertex.lazy.js +3 -0
- package/dist/api/google-vertex.lazy.js.map +1 -0
- package/dist/api/lazy.d.ts +19 -0
- package/dist/api/lazy.d.ts.map +1 -0
- package/dist/api/lazy.js +72 -0
- package/dist/api/lazy.js.map +1 -0
- package/dist/api/llama-cpp-classify.d.ts +33 -0
- package/dist/api/llama-cpp-classify.d.ts.map +1 -0
- package/dist/api/llama-cpp-classify.js +368 -0
- package/dist/api/llama-cpp-classify.js.map +1 -0
- package/dist/api/llama-cpp-classify.lazy.d.ts +3 -0
- package/dist/api/llama-cpp-classify.lazy.d.ts.map +1 -0
- package/dist/api/llama-cpp-classify.lazy.js +4 -0
- package/dist/api/llama-cpp-classify.lazy.js.map +1 -0
- package/dist/api/mistral-conversations.d.ts +25 -0
- package/dist/api/mistral-conversations.d.ts.map +1 -0
- package/dist/api/mistral-conversations.js +779 -0
- package/dist/api/mistral-conversations.js.map +1 -0
- package/dist/api/mistral-conversations.lazy.d.ts +3 -0
- package/dist/api/mistral-conversations.lazy.d.ts.map +1 -0
- package/dist/api/mistral-conversations.lazy.js +3 -0
- package/dist/api/mistral-conversations.lazy.js.map +1 -0
- package/dist/api/openai-codex-responses.d.ts +32 -0
- package/dist/api/openai-codex-responses.d.ts.map +1 -0
- package/dist/api/openai-codex-responses.js +1357 -0
- package/dist/api/openai-codex-responses.js.map +1 -0
- package/dist/api/openai-codex-responses.lazy.d.ts +3 -0
- package/dist/api/openai-codex-responses.lazy.d.ts.map +1 -0
- package/dist/api/openai-codex-responses.lazy.js +3 -0
- package/dist/api/openai-codex-responses.lazy.js.map +1 -0
- package/dist/api/openai-completions.d.ts +30 -0
- package/dist/api/openai-completions.d.ts.map +1 -0
- package/dist/api/openai-completions.js +1529 -0
- package/dist/api/openai-completions.js.map +1 -0
- package/dist/api/openai-completions.lazy.d.ts +3 -0
- package/dist/api/openai-completions.lazy.d.ts.map +1 -0
- package/dist/api/openai-completions.lazy.js +3 -0
- package/dist/api/openai-completions.lazy.js.map +1 -0
- package/dist/api/openai-decisions.d.ts +3 -0
- package/dist/api/openai-decisions.d.ts.map +1 -0
- package/dist/api/openai-decisions.js +157 -0
- package/dist/api/openai-decisions.js.map +1 -0
- package/dist/api/openai-decisions.lazy.d.ts +3 -0
- package/dist/api/openai-decisions.lazy.d.ts.map +1 -0
- package/dist/api/openai-decisions.lazy.js +4 -0
- package/dist/api/openai-decisions.lazy.js.map +1 -0
- package/dist/api/openai-images-options.d.ts +120 -0
- package/dist/api/openai-images-options.d.ts.map +1 -0
- package/dist/api/openai-images-options.js +112 -0
- package/dist/api/openai-images-options.js.map +1 -0
- package/dist/api/openai-images.d.ts +9 -0
- package/dist/api/openai-images.d.ts.map +1 -0
- package/dist/api/openai-images.js +317 -0
- package/dist/api/openai-images.js.map +1 -0
- package/dist/api/openai-images.lazy.d.ts +3 -0
- package/dist/api/openai-images.lazy.d.ts.map +1 -0
- package/dist/api/openai-images.lazy.js +4 -0
- package/dist/api/openai-images.lazy.js.map +1 -0
- package/dist/api/openai-prompt-cache.d.ts +3 -0
- package/dist/api/openai-prompt-cache.d.ts.map +1 -0
- package/dist/api/openai-prompt-cache.js +10 -0
- package/dist/api/openai-prompt-cache.js.map +1 -0
- package/dist/api/openai-responses-shared.d.ts +133 -0
- package/dist/api/openai-responses-shared.d.ts.map +1 -0
- package/dist/api/openai-responses-shared.js +970 -0
- package/dist/api/openai-responses-shared.js.map +1 -0
- package/dist/api/openai-responses.d.ts +15 -0
- package/dist/api/openai-responses.d.ts.map +1 -0
- package/dist/api/openai-responses.js +444 -0
- package/dist/api/openai-responses.js.map +1 -0
- package/dist/api/openai-responses.lazy.d.ts +3 -0
- package/dist/api/openai-responses.lazy.d.ts.map +1 -0
- package/dist/api/openai-responses.lazy.js +3 -0
- package/dist/api/openai-responses.lazy.js.map +1 -0
- package/dist/api/openai-retry-hint-client.d.ts +10 -0
- package/dist/api/openai-retry-hint-client.d.ts.map +1 -0
- package/dist/api/openai-retry-hint-client.js +17 -0
- package/dist/api/openai-retry-hint-client.js.map +1 -0
- package/dist/api/openrouter-images.d.ts +4 -0
- package/dist/api/openrouter-images.d.ts.map +1 -0
- package/dist/api/openrouter-images.js +135 -0
- package/dist/api/openrouter-images.js.map +1 -0
- package/dist/api/openrouter-images.lazy.d.ts +3 -0
- package/dist/api/openrouter-images.lazy.d.ts.map +1 -0
- package/dist/api/openrouter-images.lazy.js +4 -0
- package/dist/api/openrouter-images.lazy.js.map +1 -0
- package/dist/api/pi-messages.d.ts +101 -0
- package/dist/api/pi-messages.d.ts.map +1 -0
- package/dist/api/pi-messages.js +338 -0
- package/dist/api/pi-messages.js.map +1 -0
- package/dist/api/pi-messages.lazy.d.ts +3 -0
- package/dist/api/pi-messages.lazy.d.ts.map +1 -0
- package/dist/api/pi-messages.lazy.js +3 -0
- package/dist/api/pi-messages.lazy.js.map +1 -0
- package/dist/api/simple-options.d.ts +17 -0
- package/dist/api/simple-options.d.ts.map +1 -0
- package/dist/api/simple-options.js +89 -0
- package/dist/api/simple-options.js.map +1 -0
- package/dist/api/strict-tool-limits.d.ts +13 -0
- package/dist/api/strict-tool-limits.d.ts.map +1 -0
- package/dist/api/strict-tool-limits.js +50 -0
- package/dist/api/strict-tool-limits.js.map +1 -0
- package/dist/api/system-one-shared.d.ts +48 -0
- package/dist/api/system-one-shared.d.ts.map +1 -0
- package/dist/api/system-one-shared.js +175 -0
- package/dist/api/system-one-shared.js.map +1 -0
- package/dist/api/transform-messages.d.ts +27 -0
- package/dist/api/transform-messages.d.ts.map +1 -0
- package/dist/api/transform-messages.js +353 -0
- package/dist/api/transform-messages.js.map +1 -0
- package/dist/api/typesafe-system-one.d.ts +4 -0
- package/dist/api/typesafe-system-one.d.ts.map +1 -0
- package/dist/api/typesafe-system-one.js +29 -0
- package/dist/api/typesafe-system-one.js.map +1 -0
- package/dist/api/typesafe-system-one.lazy.d.ts +3 -0
- package/dist/api/typesafe-system-one.lazy.d.ts.map +1 -0
- package/dist/api/typesafe-system-one.lazy.js +4 -0
- package/dist/api/typesafe-system-one.lazy.js.map +1 -0
- package/dist/auth/context.d.ts +7 -0
- package/dist/auth/context.d.ts.map +1 -0
- package/dist/auth/context.js +42 -0
- package/dist/auth/context.js.map +1 -0
- package/dist/auth/credential-store.d.ts +17 -0
- package/dist/auth/credential-store.d.ts.map +1 -0
- package/dist/auth/credential-store.js +52 -0
- package/dist/auth/credential-store.js.map +1 -0
- package/dist/auth/helpers.d.ts +30 -0
- package/dist/auth/helpers.d.ts.map +1 -0
- package/dist/auth/helpers.js +66 -0
- package/dist/auth/helpers.js.map +1 -0
- package/dist/auth/oauth/anthropic.d.ts +9 -0
- package/dist/auth/oauth/anthropic.d.ts.map +1 -0
- package/dist/auth/oauth/anthropic.js +266 -0
- package/dist/auth/oauth/anthropic.js.map +1 -0
- package/dist/auth/oauth/callback-server.d.ts +31 -0
- package/dist/auth/oauth/callback-server.d.ts.map +1 -0
- package/dist/auth/oauth/callback-server.js +138 -0
- package/dist/auth/oauth/callback-server.js.map +1 -0
- package/dist/auth/oauth/device-code.d.ts +24 -0
- package/dist/auth/oauth/device-code.d.ts.map +1 -0
- package/dist/auth/oauth/device-code.js +69 -0
- package/dist/auth/oauth/device-code.js.map +1 -0
- package/dist/auth/oauth/github-copilot.d.ts +6 -0
- package/dist/auth/oauth/github-copilot.d.ts.map +1 -0
- package/dist/auth/oauth/github-copilot.js +352 -0
- package/dist/auth/oauth/github-copilot.js.map +1 -0
- package/dist/auth/oauth/kimi-coding.d.ts +10 -0
- package/dist/auth/oauth/kimi-coding.d.ts.map +1 -0
- package/dist/auth/oauth/kimi-coding.js +238 -0
- package/dist/auth/oauth/kimi-coding.js.map +1 -0
- package/dist/auth/oauth/load.d.ts +23 -0
- package/dist/auth/oauth/load.d.ts.map +1 -0
- package/dist/auth/oauth/load.js +67 -0
- package/dist/auth/oauth/load.js.map +1 -0
- package/dist/auth/oauth/meta.d.ts +17 -0
- package/dist/auth/oauth/meta.d.ts.map +1 -0
- package/dist/auth/oauth/meta.js +182 -0
- package/dist/auth/oauth/meta.js.map +1 -0
- package/dist/auth/oauth/oauth-page.d.ts +3 -0
- package/dist/auth/oauth/oauth-page.d.ts.map +1 -0
- package/dist/auth/oauth/oauth-page.js +105 -0
- package/dist/auth/oauth/oauth-page.js.map +1 -0
- package/dist/auth/oauth/openai-chatgpt.d.ts +3 -0
- package/dist/auth/oauth/openai-chatgpt.d.ts.map +1 -0
- package/dist/auth/oauth/openai-chatgpt.js +271 -0
- package/dist/auth/oauth/openai-chatgpt.js.map +1 -0
- package/dist/auth/oauth/openai-codex.d.ts +9 -0
- package/dist/auth/oauth/openai-codex.d.ts.map +1 -0
- package/dist/auth/oauth/openai-codex.js +360 -0
- package/dist/auth/oauth/openai-codex.js.map +1 -0
- package/dist/auth/oauth/openrouter.d.ts +15 -0
- package/dist/auth/oauth/openrouter.d.ts.map +1 -0
- package/dist/auth/oauth/openrouter.js +159 -0
- package/dist/auth/oauth/openrouter.js.map +1 -0
- package/dist/auth/oauth/pkce.d.ts +13 -0
- package/dist/auth/oauth/pkce.d.ts.map +1 -0
- package/dist/auth/oauth/pkce.js +31 -0
- package/dist/auth/oauth/pkce.js.map +1 -0
- package/dist/auth/oauth/token-response.d.ts +25 -0
- package/dist/auth/oauth/token-response.d.ts.map +1 -0
- package/dist/auth/oauth/token-response.js +43 -0
- package/dist/auth/oauth/token-response.js.map +1 -0
- package/dist/auth/oauth/trusted-url.d.ts +7 -0
- package/dist/auth/oauth/trusted-url.d.ts.map +1 -0
- package/dist/auth/oauth/trusted-url.js +12 -0
- package/dist/auth/oauth/trusted-url.js.map +1 -0
- package/dist/auth/oauth/xai.d.ts +6 -0
- package/dist/auth/oauth/xai.d.ts.map +1 -0
- package/dist/auth/oauth/xai.js +193 -0
- package/dist/auth/oauth/xai.js.map +1 -0
- package/dist/auth/oauth-refresh-errors.d.ts +46 -0
- package/dist/auth/oauth-refresh-errors.d.ts.map +1 -0
- package/dist/auth/oauth-refresh-errors.js +150 -0
- package/dist/auth/oauth-refresh-errors.js.map +1 -0
- package/dist/auth/resolve.d.ts +62 -0
- package/dist/auth/resolve.d.ts.map +1 -0
- package/dist/auth/resolve.js +357 -0
- package/dist/auth/resolve.js.map +1 -0
- package/dist/auth/types.d.ts +267 -0
- package/dist/auth/types.d.ts.map +1 -0
- package/dist/auth/types.js +2 -0
- package/dist/auth/types.js.map +1 -0
- package/dist/bedrock-provider.d.ts +5 -0
- package/dist/bedrock-provider.d.ts.map +1 -0
- package/dist/bedrock-provider.js +6 -0
- package/dist/bedrock-provider.js.map +1 -0
- package/dist/bun-oauth.d.ts +3 -0
- package/dist/bun-oauth.d.ts.map +1 -0
- package/dist/bun-oauth.js +23 -0
- package/dist/bun-oauth.js.map +1 -0
- package/dist/cache-warming.d.ts +53 -0
- package/dist/cache-warming.d.ts.map +1 -0
- package/dist/cache-warming.js +82 -0
- package/dist/cache-warming.js.map +1 -0
- package/dist/cli.d.ts +3 -0
- package/dist/cli.d.ts.map +1 -0
- package/dist/cli.js +118 -0
- package/dist/cli.js.map +1 -0
- package/dist/compat/extension-oauth-types.d.ts +43 -0
- package/dist/compat/extension-oauth-types.d.ts.map +1 -0
- package/dist/compat/extension-oauth-types.js +2 -0
- package/dist/compat/extension-oauth-types.js.map +1 -0
- package/dist/compat.d.ts +67 -0
- package/dist/compat.d.ts.map +1 -0
- package/dist/compat.js +202 -0
- package/dist/compat.js.map +1 -0
- package/dist/decision-models.d.ts +21 -0
- package/dist/decision-models.d.ts.map +1 -0
- package/dist/decision-models.generated.d.ts +3 -0
- package/dist/decision-models.generated.d.ts.map +1 -0
- package/dist/decision-models.generated.js +47 -0
- package/dist/decision-models.generated.js.map +1 -0
- package/dist/decision-models.js +8 -0
- package/dist/decision-models.js.map +1 -0
- package/dist/env-api-keys.d.ts +28 -0
- package/dist/env-api-keys.d.ts.map +1 -0
- package/dist/env-api-keys.js +185 -0
- package/dist/env-api-keys.js.map +1 -0
- package/dist/image-generation-error.d.ts +210 -0
- package/dist/image-generation-error.d.ts.map +1 -0
- package/dist/image-generation-error.js +442 -0
- package/dist/image-generation-error.js.map +1 -0
- package/dist/image-models.d.ts +21 -0
- package/dist/image-models.d.ts.map +1 -0
- package/dist/image-models.js +24 -0
- package/dist/image-models.js.map +1 -0
- package/dist/images-api-registry.d.ts +14 -0
- package/dist/images-api-registry.d.ts.map +1 -0
- package/dist/images-api-registry.js +22 -0
- package/dist/images-api-registry.js.map +1 -0
- package/dist/images.d.ts +9 -0
- package/dist/images.d.ts.map +1 -0
- package/dist/images.js +19 -0
- package/dist/images.js.map +1 -0
- package/dist/index.d.ts +51 -0
- package/dist/index.d.ts.map +1 -0
- package/dist/index.js +38 -0
- package/dist/index.js.map +1 -0
- package/dist/lazy-bedrock-provider.d.ts +10 -0
- package/dist/lazy-bedrock-provider.d.ts.map +1 -0
- package/dist/lazy-bedrock-provider.js +13 -0
- package/dist/lazy-bedrock-provider.js.map +1 -0
- package/dist/lazy-oauth.d.ts +10 -0
- package/dist/lazy-oauth.d.ts.map +1 -0
- package/dist/lazy-oauth.js +22 -0
- package/dist/lazy-oauth.js.map +1 -0
- package/dist/legacy-api-aliases.d.ts +42 -0
- package/dist/legacy-api-aliases.d.ts.map +1 -0
- package/dist/legacy-api-aliases.js +49 -0
- package/dist/legacy-api-aliases.js.map +1 -0
- package/dist/model-catalog.d.ts +33 -0
- package/dist/model-catalog.d.ts.map +1 -0
- package/dist/model-catalog.js +46 -0
- package/dist/model-catalog.js.map +1 -0
- package/dist/models-store.d.ts +30 -0
- package/dist/models-store.d.ts.map +1 -0
- package/dist/models-store.js +17 -0
- package/dist/models-store.js.map +1 -0
- package/dist/models.d.ts +297 -0
- package/dist/models.d.ts.map +1 -0
- package/dist/models.generated.d.ts +183 -0
- package/dist/models.generated.d.ts.map +1 -0
- package/dist/models.generated.js +185 -0
- package/dist/models.generated.js.map +1 -0
- package/dist/models.js +817 -0
- package/dist/models.js.map +1 -0
- package/dist/oauth.d.ts +3 -0
- package/dist/oauth.d.ts.map +1 -0
- package/dist/oauth.js +2 -0
- package/dist/oauth.js.map +1 -0
- package/dist/provider-id.d.ts +2 -0
- package/dist/provider-id.d.ts.map +1 -0
- package/dist/provider-id.js +4 -0
- package/dist/provider-id.js.map +1 -0
- package/dist/providers/all.d.ts +41 -0
- package/dist/providers/all.d.ts.map +1 -0
- package/dist/providers/all.js +145 -0
- package/dist/providers/all.js.map +1 -0
- package/dist/providers/amazon-bedrock.d.ts +3 -0
- package/dist/providers/amazon-bedrock.d.ts.map +1 -0
- package/dist/providers/amazon-bedrock.js +93 -0
- package/dist/providers/amazon-bedrock.js.map +1 -0
- package/dist/providers/amazon-bedrock.models.d.ts +5 -0
- package/dist/providers/amazon-bedrock.models.d.ts.map +1 -0
- package/dist/providers/amazon-bedrock.models.js +8 -0
- package/dist/providers/amazon-bedrock.models.js.map +1 -0
- package/dist/providers/ant-ling.d.ts +3 -0
- package/dist/providers/ant-ling.d.ts.map +1 -0
- package/dist/providers/ant-ling.js +15 -0
- package/dist/providers/ant-ling.js.map +1 -0
- package/dist/providers/ant-ling.models.d.ts +5 -0
- package/dist/providers/ant-ling.models.d.ts.map +1 -0
- package/dist/providers/ant-ling.models.js +8 -0
- package/dist/providers/ant-ling.models.js.map +1 -0
- package/dist/providers/anthropic.d.ts +4 -0
- package/dist/providers/anthropic.d.ts.map +1 -0
- package/dist/providers/anthropic.js +87 -0
- package/dist/providers/anthropic.js.map +1 -0
- package/dist/providers/anthropic.models.d.ts +5 -0
- package/dist/providers/anthropic.models.d.ts.map +1 -0
- package/dist/providers/anthropic.models.js +8 -0
- package/dist/providers/anthropic.models.js.map +1 -0
- package/dist/providers/azure-openai-responses.d.ts +3 -0
- package/dist/providers/azure-openai-responses.d.ts.map +1 -0
- package/dist/providers/azure-openai-responses.js +3 -0
- package/dist/providers/azure-openai-responses.js.map +1 -0
- package/dist/providers/azure-openai-responses.models.d.ts +3 -0
- package/dist/providers/azure-openai-responses.models.d.ts.map +1 -0
- package/dist/providers/azure-openai-responses.models.js +5 -0
- package/dist/providers/azure-openai-responses.models.js.map +1 -0
- package/dist/providers/azure.d.ts +3 -0
- package/dist/providers/azure.d.ts.map +1 -0
- package/dist/providers/azure.js +41 -0
- package/dist/providers/azure.js.map +1 -0
- package/dist/providers/azure.models.d.ts +5 -0
- package/dist/providers/azure.models.d.ts.map +1 -0
- package/dist/providers/azure.models.js +8 -0
- package/dist/providers/azure.models.js.map +1 -0
- package/dist/providers/baseten.d.ts +3 -0
- package/dist/providers/baseten.d.ts.map +1 -0
- package/dist/providers/baseten.js +15 -0
- package/dist/providers/baseten.js.map +1 -0
- package/dist/providers/baseten.models.d.ts +5 -0
- package/dist/providers/baseten.models.d.ts.map +1 -0
- package/dist/providers/baseten.models.js +8 -0
- package/dist/providers/baseten.models.js.map +1 -0
- package/dist/providers/cerebras.d.ts +3 -0
- package/dist/providers/cerebras.d.ts.map +1 -0
- package/dist/providers/cerebras.js +15 -0
- package/dist/providers/cerebras.js.map +1 -0
- package/dist/providers/cerebras.models.d.ts +5 -0
- package/dist/providers/cerebras.models.d.ts.map +1 -0
- package/dist/providers/cerebras.models.js +8 -0
- package/dist/providers/cerebras.models.js.map +1 -0
- package/dist/providers/cloudflare-ai-gateway.d.ts +5 -0
- package/dist/providers/cloudflare-ai-gateway.d.ts.map +1 -0
- package/dist/providers/cloudflare-ai-gateway.js +21 -0
- package/dist/providers/cloudflare-ai-gateway.js.map +1 -0
- package/dist/providers/cloudflare-ai-gateway.models.d.ts +5 -0
- package/dist/providers/cloudflare-ai-gateway.models.d.ts.map +1 -0
- package/dist/providers/cloudflare-ai-gateway.models.js +8 -0
- package/dist/providers/cloudflare-ai-gateway.models.js.map +1 -0
- package/dist/providers/cloudflare-auth.d.ts +4 -0
- package/dist/providers/cloudflare-auth.d.ts.map +1 -0
- package/dist/providers/cloudflare-auth.js +86 -0
- package/dist/providers/cloudflare-auth.js.map +1 -0
- package/dist/providers/cloudflare-stream.d.ts +12 -0
- package/dist/providers/cloudflare-stream.d.ts.map +1 -0
- package/dist/providers/cloudflare-stream.js +27 -0
- package/dist/providers/cloudflare-stream.js.map +1 -0
- package/dist/providers/cloudflare-workers-ai.d.ts +3 -0
- package/dist/providers/cloudflare-workers-ai.d.ts.map +1 -0
- package/dist/providers/cloudflare-workers-ai.js +22 -0
- package/dist/providers/cloudflare-workers-ai.js.map +1 -0
- package/dist/providers/cloudflare-workers-ai.models.d.ts +5 -0
- package/dist/providers/cloudflare-workers-ai.models.d.ts.map +1 -0
- package/dist/providers/cloudflare-workers-ai.models.js +8 -0
- package/dist/providers/cloudflare-workers-ai.models.js.map +1 -0
- package/dist/providers/compat-schema.d.ts +549 -0
- package/dist/providers/compat-schema.d.ts.map +1 -0
- package/dist/providers/compat-schema.js +452 -0
- package/dist/providers/compat-schema.js.map +1 -0
- package/dist/providers/data/.manifest.json +1 -0
- package/dist/providers/data/amazon-bedrock.json +1 -0
- package/dist/providers/data/ant-ling.json +1 -0
- package/dist/providers/data/anthropic.json +1 -0
- package/dist/providers/data/azure.json +1 -0
- package/dist/providers/data/baseten.json +1 -0
- package/dist/providers/data/cerebras.json +1 -0
- package/dist/providers/data/cloudflare-ai-gateway.json +1 -0
- package/dist/providers/data/cloudflare-workers-ai.json +1 -0
- package/dist/providers/data/deepseek.json +1 -0
- package/dist/providers/data/fireworks.json +1 -0
- package/dist/providers/data/github-copilot.json +1 -0
- package/dist/providers/data/google-vertex.json +1 -0
- package/dist/providers/data/google.json +1 -0
- package/dist/providers/data/groq.json +1 -0
- package/dist/providers/data/huggingface.json +1 -0
- package/dist/providers/data/kimi-coding.json +1 -0
- package/dist/providers/data/meta.json +1 -0
- package/dist/providers/data/minimax-cn.json +1 -0
- package/dist/providers/data/minimax.json +1 -0
- package/dist/providers/data/mistral.json +1 -0
- package/dist/providers/data/moonshotai-cn.json +1 -0
- package/dist/providers/data/moonshotai.json +1 -0
- package/dist/providers/data/nvidia.json +1 -0
- package/dist/providers/data/openai-codex.json +1 -0
- package/dist/providers/data/openai-decisions.json +1 -0
- package/dist/providers/data/openai-images.json +1 -0
- package/dist/providers/data/openai.json +1 -0
- package/dist/providers/data/opencode-go.json +1 -0
- package/dist/providers/data/opencode.json +1 -0
- package/dist/providers/data/openrouter.json +1 -0
- package/dist/providers/data/qwen-token-plan-cn.json +1 -0
- package/dist/providers/data/qwen-token-plan-individual-cn.json +1 -0
- package/dist/providers/data/qwen-token-plan-individual.json +1 -0
- package/dist/providers/data/qwen-token-plan.json +1 -0
- package/dist/providers/data/together.json +1 -0
- package/dist/providers/data/typesafe.json +1 -0
- package/dist/providers/data/vercel-ai-gateway.json +1 -0
- package/dist/providers/data/xai.json +1 -0
- package/dist/providers/data/xiaomi-token-plan-ams.json +1 -0
- package/dist/providers/data/xiaomi-token-plan-cn.json +1 -0
- package/dist/providers/data/xiaomi-token-plan-sgp.json +1 -0
- package/dist/providers/data/xiaomi.json +1 -0
- package/dist/providers/data/zai-coding-cn.json +1 -0
- package/dist/providers/data/zai.json +1 -0
- package/dist/providers/deepseek.d.ts +3 -0
- package/dist/providers/deepseek.d.ts.map +1 -0
- package/dist/providers/deepseek.js +15 -0
- package/dist/providers/deepseek.js.map +1 -0
- package/dist/providers/deepseek.models.d.ts +5 -0
- package/dist/providers/deepseek.models.d.ts.map +1 -0
- package/dist/providers/deepseek.models.js +8 -0
- package/dist/providers/deepseek.models.js.map +1 -0
- package/dist/providers/faux.d.ts +103 -0
- package/dist/providers/faux.d.ts.map +1 -0
- package/dist/providers/faux.js +521 -0
- package/dist/providers/faux.js.map +1 -0
- package/dist/providers/fireworks.d.ts +3 -0
- package/dist/providers/fireworks.d.ts.map +1 -0
- package/dist/providers/fireworks.js +19 -0
- package/dist/providers/fireworks.js.map +1 -0
- package/dist/providers/fireworks.models.d.ts +5 -0
- package/dist/providers/fireworks.models.d.ts.map +1 -0
- package/dist/providers/fireworks.models.js +8 -0
- package/dist/providers/fireworks.models.js.map +1 -0
- package/dist/providers/github-copilot-account.d.ts +18 -0
- package/dist/providers/github-copilot-account.d.ts.map +1 -0
- package/dist/providers/github-copilot-account.js +57 -0
- package/dist/providers/github-copilot-account.js.map +1 -0
- package/dist/providers/github-copilot-capabilities.d.ts +67 -0
- package/dist/providers/github-copilot-capabilities.d.ts.map +1 -0
- package/dist/providers/github-copilot-capabilities.js +81 -0
- package/dist/providers/github-copilot-capabilities.js.map +1 -0
- package/dist/providers/github-copilot-env.d.ts +41 -0
- package/dist/providers/github-copilot-env.d.ts.map +1 -0
- package/dist/providers/github-copilot-env.js +108 -0
- package/dist/providers/github-copilot-env.js.map +1 -0
- package/dist/providers/github-copilot-errors.d.ts +255 -0
- package/dist/providers/github-copilot-errors.d.ts.map +1 -0
- package/dist/providers/github-copilot-errors.js +764 -0
- package/dist/providers/github-copilot-errors.js.map +1 -0
- package/dist/providers/github-copilot-stream-attempt.d.ts +27 -0
- package/dist/providers/github-copilot-stream-attempt.d.ts.map +1 -0
- package/dist/providers/github-copilot-stream-attempt.js +45 -0
- package/dist/providers/github-copilot-stream-attempt.js.map +1 -0
- package/dist/providers/github-copilot.d.ts +40 -0
- package/dist/providers/github-copilot.d.ts.map +1 -0
- package/dist/providers/github-copilot.js +363 -0
- package/dist/providers/github-copilot.js.map +1 -0
- package/dist/providers/github-copilot.models.d.ts +5 -0
- package/dist/providers/github-copilot.models.d.ts.map +1 -0
- package/dist/providers/github-copilot.models.js +8 -0
- package/dist/providers/github-copilot.models.js.map +1 -0
- package/dist/providers/google-vertex.d.ts +3 -0
- package/dist/providers/google-vertex.d.ts.map +1 -0
- package/dist/providers/google-vertex.js +94 -0
- package/dist/providers/google-vertex.js.map +1 -0
- package/dist/providers/google-vertex.models.d.ts +5 -0
- package/dist/providers/google-vertex.models.d.ts.map +1 -0
- package/dist/providers/google-vertex.models.js +8 -0
- package/dist/providers/google-vertex.models.js.map +1 -0
- package/dist/providers/google.d.ts +3 -0
- package/dist/providers/google.d.ts.map +1 -0
- package/dist/providers/google.js +15 -0
- package/dist/providers/google.js.map +1 -0
- package/dist/providers/google.models.d.ts +5 -0
- package/dist/providers/google.models.d.ts.map +1 -0
- package/dist/providers/google.models.js +8 -0
- package/dist/providers/google.models.js.map +1 -0
- package/dist/providers/groq.d.ts +3 -0
- package/dist/providers/groq.d.ts.map +1 -0
- package/dist/providers/groq.js +15 -0
- package/dist/providers/groq.js.map +1 -0
- package/dist/providers/groq.models.d.ts +5 -0
- package/dist/providers/groq.models.d.ts.map +1 -0
- package/dist/providers/groq.models.js +8 -0
- package/dist/providers/groq.models.js.map +1 -0
- package/dist/providers/huggingface.d.ts +3 -0
- package/dist/providers/huggingface.d.ts.map +1 -0
- package/dist/providers/huggingface.js +15 -0
- package/dist/providers/huggingface.js.map +1 -0
- package/dist/providers/huggingface.models.d.ts +5 -0
- package/dist/providers/huggingface.models.d.ts.map +1 -0
- package/dist/providers/huggingface.models.js +8 -0
- package/dist/providers/huggingface.models.js.map +1 -0
- package/dist/providers/images/register-builtins.d.ts +5 -0
- package/dist/providers/images/register-builtins.d.ts.map +1 -0
- package/dist/providers/images/register-builtins.js +52 -0
- package/dist/providers/images/register-builtins.js.map +1 -0
- package/dist/providers/kimi-coding.d.ts +3 -0
- package/dist/providers/kimi-coding.d.ts.map +1 -0
- package/dist/providers/kimi-coding.js +24 -0
- package/dist/providers/kimi-coding.js.map +1 -0
- package/dist/providers/kimi-coding.models.d.ts +5 -0
- package/dist/providers/kimi-coding.models.d.ts.map +1 -0
- package/dist/providers/kimi-coding.models.js +8 -0
- package/dist/providers/kimi-coding.models.js.map +1 -0
- package/dist/providers/meta.d.ts +3 -0
- package/dist/providers/meta.d.ts.map +1 -0
- package/dist/providers/meta.js +24 -0
- package/dist/providers/meta.js.map +1 -0
- package/dist/providers/meta.models.d.ts +5 -0
- package/dist/providers/meta.models.d.ts.map +1 -0
- package/dist/providers/meta.models.js +8 -0
- package/dist/providers/meta.models.js.map +1 -0
- package/dist/providers/minimax-cn.d.ts +3 -0
- package/dist/providers/minimax-cn.d.ts.map +1 -0
- package/dist/providers/minimax-cn.js +15 -0
- package/dist/providers/minimax-cn.js.map +1 -0
- package/dist/providers/minimax-cn.models.d.ts +5 -0
- package/dist/providers/minimax-cn.models.d.ts.map +1 -0
- package/dist/providers/minimax-cn.models.js +8 -0
- package/dist/providers/minimax-cn.models.js.map +1 -0
- package/dist/providers/minimax.d.ts +3 -0
- package/dist/providers/minimax.d.ts.map +1 -0
- package/dist/providers/minimax.js +15 -0
- package/dist/providers/minimax.js.map +1 -0
- package/dist/providers/minimax.models.d.ts +5 -0
- package/dist/providers/minimax.models.d.ts.map +1 -0
- package/dist/providers/minimax.models.js +8 -0
- package/dist/providers/minimax.models.js.map +1 -0
- package/dist/providers/mistral.d.ts +3 -0
- package/dist/providers/mistral.d.ts.map +1 -0
- package/dist/providers/mistral.js +15 -0
- package/dist/providers/mistral.js.map +1 -0
- package/dist/providers/mistral.models.d.ts +5 -0
- package/dist/providers/mistral.models.d.ts.map +1 -0
- package/dist/providers/mistral.models.js +8 -0
- package/dist/providers/mistral.models.js.map +1 -0
- package/dist/providers/model-schema.d.ts +91 -0
- package/dist/providers/model-schema.d.ts.map +1 -0
- package/dist/providers/model-schema.js +57 -0
- package/dist/providers/model-schema.js.map +1 -0
- package/dist/providers/moonshotai-cn.d.ts +3 -0
- package/dist/providers/moonshotai-cn.d.ts.map +1 -0
- package/dist/providers/moonshotai-cn.js +15 -0
- package/dist/providers/moonshotai-cn.js.map +1 -0
- package/dist/providers/moonshotai-cn.models.d.ts +5 -0
- package/dist/providers/moonshotai-cn.models.d.ts.map +1 -0
- package/dist/providers/moonshotai-cn.models.js +8 -0
- package/dist/providers/moonshotai-cn.models.js.map +1 -0
- package/dist/providers/moonshotai.d.ts +3 -0
- package/dist/providers/moonshotai.d.ts.map +1 -0
- package/dist/providers/moonshotai.js +15 -0
- package/dist/providers/moonshotai.js.map +1 -0
- package/dist/providers/moonshotai.models.d.ts +5 -0
- package/dist/providers/moonshotai.models.d.ts.map +1 -0
- package/dist/providers/moonshotai.models.js +8 -0
- package/dist/providers/moonshotai.models.js.map +1 -0
- package/dist/providers/morph.d.ts +3 -0
- package/dist/providers/morph.d.ts.map +1 -0
- package/dist/providers/morph.js +22 -0
- package/dist/providers/morph.js.map +1 -0
- package/dist/providers/nvidia.d.ts +3 -0
- package/dist/providers/nvidia.d.ts.map +1 -0
- package/dist/providers/nvidia.js +15 -0
- package/dist/providers/nvidia.js.map +1 -0
- package/dist/providers/nvidia.models.d.ts +5 -0
- package/dist/providers/nvidia.models.d.ts.map +1 -0
- package/dist/providers/nvidia.models.js +8 -0
- package/dist/providers/nvidia.models.js.map +1 -0
- package/dist/providers/openai-codex.d.ts +3 -0
- package/dist/providers/openai-codex.d.ts.map +1 -0
- package/dist/providers/openai-codex.js +25 -0
- package/dist/providers/openai-codex.js.map +1 -0
- package/dist/providers/openai-codex.models.d.ts +5 -0
- package/dist/providers/openai-codex.models.d.ts.map +1 -0
- package/dist/providers/openai-codex.models.js +8 -0
- package/dist/providers/openai-codex.models.js.map +1 -0
- package/dist/providers/openai-decisions.d.ts +3 -0
- package/dist/providers/openai-decisions.d.ts.map +1 -0
- package/dist/providers/openai-decisions.js +19 -0
- package/dist/providers/openai-decisions.js.map +1 -0
- package/dist/providers/openai-decisions.models.d.ts +5 -0
- package/dist/providers/openai-decisions.models.d.ts.map +1 -0
- package/dist/providers/openai-decisions.models.js +8 -0
- package/dist/providers/openai-decisions.models.js.map +1 -0
- package/dist/providers/openai-images.d.ts +3 -0
- package/dist/providers/openai-images.d.ts.map +1 -0
- package/dist/providers/openai-images.js +19 -0
- package/dist/providers/openai-images.js.map +1 -0
- package/dist/providers/openai-images.models.d.ts +5 -0
- package/dist/providers/openai-images.models.d.ts.map +1 -0
- package/dist/providers/openai-images.models.js +8 -0
- package/dist/providers/openai-images.models.js.map +1 -0
- package/dist/providers/openai-mid-convo-effort.d.ts +2 -0
- package/dist/providers/openai-mid-convo-effort.d.ts.map +1 -0
- package/dist/providers/openai-mid-convo-effort.js +8 -0
- package/dist/providers/openai-mid-convo-effort.js.map +1 -0
- package/dist/providers/openai-subscription-models.d.ts +3 -0
- package/dist/providers/openai-subscription-models.d.ts.map +1 -0
- package/dist/providers/openai-subscription-models.js +19 -0
- package/dist/providers/openai-subscription-models.js.map +1 -0
- package/dist/providers/openai.d.ts +4 -0
- package/dist/providers/openai.d.ts.map +1 -0
- package/dist/providers/openai.js +48 -0
- package/dist/providers/openai.js.map +1 -0
- package/dist/providers/openai.models.d.ts +5 -0
- package/dist/providers/openai.models.d.ts.map +1 -0
- package/dist/providers/openai.models.js +8 -0
- package/dist/providers/openai.models.js.map +1 -0
- package/dist/providers/opencode-go.d.ts +5 -0
- package/dist/providers/opencode-go.d.ts.map +1 -0
- package/dist/providers/opencode-go.js +22 -0
- package/dist/providers/opencode-go.js.map +1 -0
- package/dist/providers/opencode-go.models.d.ts +5 -0
- package/dist/providers/opencode-go.models.d.ts.map +1 -0
- package/dist/providers/opencode-go.models.js +8 -0
- package/dist/providers/opencode-go.models.js.map +1 -0
- package/dist/providers/opencode.d.ts +5 -0
- package/dist/providers/opencode.d.ts.map +1 -0
- package/dist/providers/opencode.js +24 -0
- package/dist/providers/opencode.js.map +1 -0
- package/dist/providers/opencode.models.d.ts +5 -0
- package/dist/providers/opencode.models.d.ts.map +1 -0
- package/dist/providers/opencode.models.js +8 -0
- package/dist/providers/opencode.models.js.map +1 -0
- package/dist/providers/openrouter-api.d.ts +40 -0
- package/dist/providers/openrouter-api.d.ts.map +1 -0
- package/dist/providers/openrouter-api.js +41 -0
- package/dist/providers/openrouter-api.js.map +1 -0
- package/dist/providers/openrouter.d.ts +13 -0
- package/dist/providers/openrouter.d.ts.map +1 -0
- package/dist/providers/openrouter.js +88 -0
- package/dist/providers/openrouter.js.map +1 -0
- package/dist/providers/openrouter.models.d.ts +5 -0
- package/dist/providers/openrouter.models.d.ts.map +1 -0
- package/dist/providers/openrouter.models.js +8 -0
- package/dist/providers/openrouter.models.js.map +1 -0
- package/dist/providers/qwen-token-plan-cn.d.ts +3 -0
- package/dist/providers/qwen-token-plan-cn.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-cn.js +20 -0
- package/dist/providers/qwen-token-plan-cn.js.map +1 -0
- package/dist/providers/qwen-token-plan-cn.models.d.ts +5 -0
- package/dist/providers/qwen-token-plan-cn.models.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-cn.models.js +8 -0
- package/dist/providers/qwen-token-plan-cn.models.js.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.d.ts +3 -0
- package/dist/providers/qwen-token-plan-individual-cn.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.js +20 -0
- package/dist/providers/qwen-token-plan-individual-cn.js.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.d.ts +5 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.js +8 -0
- package/dist/providers/qwen-token-plan-individual-cn.models.js.map +1 -0
- package/dist/providers/qwen-token-plan-individual.d.ts +3 -0
- package/dist/providers/qwen-token-plan-individual.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-individual.js +20 -0
- package/dist/providers/qwen-token-plan-individual.js.map +1 -0
- package/dist/providers/qwen-token-plan-individual.models.d.ts +5 -0
- package/dist/providers/qwen-token-plan-individual.models.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan-individual.models.js +8 -0
- package/dist/providers/qwen-token-plan-individual.models.js.map +1 -0
- package/dist/providers/qwen-token-plan.d.ts +3 -0
- package/dist/providers/qwen-token-plan.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan.js +20 -0
- package/dist/providers/qwen-token-plan.js.map +1 -0
- package/dist/providers/qwen-token-plan.models.d.ts +5 -0
- package/dist/providers/qwen-token-plan.models.d.ts.map +1 -0
- package/dist/providers/qwen-token-plan.models.js +8 -0
- package/dist/providers/qwen-token-plan.models.js.map +1 -0
- package/dist/providers/together.d.ts +3 -0
- package/dist/providers/together.d.ts.map +1 -0
- package/dist/providers/together.js +15 -0
- package/dist/providers/together.js.map +1 -0
- package/dist/providers/together.models.d.ts +5 -0
- package/dist/providers/together.models.d.ts.map +1 -0
- package/dist/providers/together.models.js +8 -0
- package/dist/providers/together.models.js.map +1 -0
- package/dist/providers/typesafe.d.ts +3 -0
- package/dist/providers/typesafe.d.ts.map +1 -0
- package/dist/providers/typesafe.js +18 -0
- package/dist/providers/typesafe.js.map +1 -0
- package/dist/providers/typesafe.models.d.ts +5 -0
- package/dist/providers/typesafe.models.d.ts.map +1 -0
- package/dist/providers/typesafe.models.js +8 -0
- package/dist/providers/typesafe.models.js.map +1 -0
- package/dist/providers/vercel-ai-gateway.d.ts +3 -0
- package/dist/providers/vercel-ai-gateway.d.ts.map +1 -0
- package/dist/providers/vercel-ai-gateway.js +17 -0
- package/dist/providers/vercel-ai-gateway.js.map +1 -0
- package/dist/providers/vercel-ai-gateway.models.d.ts +5 -0
- package/dist/providers/vercel-ai-gateway.models.d.ts.map +1 -0
- package/dist/providers/vercel-ai-gateway.models.js +8 -0
- package/dist/providers/vercel-ai-gateway.models.js.map +1 -0
- package/dist/providers/xai.d.ts +3 -0
- package/dist/providers/xai.d.ts.map +1 -0
- package/dist/providers/xai.js +24 -0
- package/dist/providers/xai.js.map +1 -0
- package/dist/providers/xai.models.d.ts +5 -0
- package/dist/providers/xai.models.d.ts.map +1 -0
- package/dist/providers/xai.models.js +8 -0
- package/dist/providers/xai.models.js.map +1 -0
- package/dist/providers/xiaomi-token-plan-ams.d.ts +3 -0
- package/dist/providers/xiaomi-token-plan-ams.d.ts.map +1 -0
- package/dist/providers/xiaomi-token-plan-ams.js +15 -0
- package/dist/providers/xiaomi-token-plan-ams.js.map +1 -0
- package/dist/providers/xiaomi-token-plan-ams.models.d.ts +5 -0
- package/dist/providers/xiaomi-token-plan-ams.models.d.ts.map +1 -0
- package/dist/providers/xiaomi-token-plan-ams.models.js +8 -0
- package/dist/providers/xiaomi-token-plan-ams.models.js.map +1 -0
- package/dist/providers/xiaomi-token-plan-cn.d.ts +3 -0
- package/dist/providers/xiaomi-token-plan-cn.d.ts.map +1 -0
- package/dist/providers/xiaomi-token-plan-cn.js +15 -0
- package/dist/providers/xiaomi-token-plan-cn.js.map +1 -0
- package/dist/providers/xiaomi-token-plan-cn.models.d.ts +5 -0
- package/dist/providers/xiaomi-token-plan-cn.models.d.ts.map +1 -0
- package/dist/providers/xiaomi-token-plan-cn.models.js +8 -0
- package/dist/providers/xiaomi-token-plan-cn.models.js.map +1 -0
- package/dist/providers/xiaomi-token-plan-sgp.d.ts +3 -0
- package/dist/providers/xiaomi-token-plan-sgp.d.ts.map +1 -0
- package/dist/providers/xiaomi-token-plan-sgp.js +15 -0
- package/dist/providers/xiaomi-token-plan-sgp.js.map +1 -0
- package/dist/providers/xiaomi-token-plan-sgp.models.d.ts +5 -0
- package/dist/providers/xiaomi-token-plan-sgp.models.d.ts.map +1 -0
- package/dist/providers/xiaomi-token-plan-sgp.models.js +8 -0
- package/dist/providers/xiaomi-token-plan-sgp.models.js.map +1 -0
- package/dist/providers/xiaomi.d.ts +3 -0
- package/dist/providers/xiaomi.d.ts.map +1 -0
- package/dist/providers/xiaomi.js +15 -0
- package/dist/providers/xiaomi.js.map +1 -0
- package/dist/providers/xiaomi.models.d.ts +5 -0
- package/dist/providers/xiaomi.models.d.ts.map +1 -0
- package/dist/providers/xiaomi.models.js +8 -0
- package/dist/providers/xiaomi.models.js.map +1 -0
- package/dist/providers/zai-coding-cn.d.ts +3 -0
- package/dist/providers/zai-coding-cn.d.ts.map +1 -0
- package/dist/providers/zai-coding-cn.js +15 -0
- package/dist/providers/zai-coding-cn.js.map +1 -0
- package/dist/providers/zai-coding-cn.models.d.ts +5 -0
- package/dist/providers/zai-coding-cn.models.d.ts.map +1 -0
- package/dist/providers/zai-coding-cn.models.js +8 -0
- package/dist/providers/zai-coding-cn.models.js.map +1 -0
- package/dist/providers/zai.d.ts +3 -0
- package/dist/providers/zai.d.ts.map +1 -0
- package/dist/providers/zai.js +15 -0
- package/dist/providers/zai.js.map +1 -0
- package/dist/providers/zai.models.d.ts +5 -0
- package/dist/providers/zai.models.d.ts.map +1 -0
- package/dist/providers/zai.models.js +8 -0
- package/dist/providers/zai.models.js.map +1 -0
- package/dist/session-resources.d.ts +4 -0
- package/dist/session-resources.d.ts.map +1 -0
- package/dist/session-resources.js +22 -0
- package/dist/session-resources.js.map +1 -0
- package/dist/types.d.ts +845 -0
- package/dist/types.d.ts.map +1 -0
- package/dist/types.js +2 -0
- package/dist/types.js.map +1 -0
- package/dist/utils/abort-signals.d.ts +6 -0
- package/dist/utils/abort-signals.d.ts.map +1 -0
- package/dist/utils/abort-signals.js +34 -0
- package/dist/utils/abort-signals.js.map +1 -0
- package/dist/utils/abort.d.ts +8 -0
- package/dist/utils/abort.d.ts.map +1 -0
- package/dist/utils/abort.js +49 -0
- package/dist/utils/abort.js.map +1 -0
- package/dist/utils/azure-openai-thinking-level-map.d.ts +19 -0
- package/dist/utils/azure-openai-thinking-level-map.d.ts.map +1 -0
- package/dist/utils/azure-openai-thinking-level-map.js +45 -0
- package/dist/utils/azure-openai-thinking-level-map.js.map +1 -0
- package/dist/utils/diagnostics.d.ts +20 -0
- package/dist/utils/diagnostics.d.ts.map +1 -0
- package/dist/utils/diagnostics.js +25 -0
- package/dist/utils/diagnostics.js.map +1 -0
- package/dist/utils/document-input.d.ts +21 -0
- package/dist/utils/document-input.d.ts.map +1 -0
- package/dist/utils/document-input.js +28 -0
- package/dist/utils/document-input.js.map +1 -0
- package/dist/utils/error-body.d.ts +40 -0
- package/dist/utils/error-body.d.ts.map +1 -0
- package/dist/utils/error-body.js +177 -0
- package/dist/utils/error-body.js.map +1 -0
- package/dist/utils/estimate.d.ts +17 -0
- package/dist/utils/estimate.d.ts.map +1 -0
- package/dist/utils/estimate.js +107 -0
- package/dist/utils/estimate.js.map +1 -0
- package/dist/utils/event-stream.d.ts +24 -0
- package/dist/utils/event-stream.d.ts.map +1 -0
- package/dist/utils/event-stream.js +118 -0
- package/dist/utils/event-stream.js.map +1 -0
- package/dist/utils/hash.d.ts +3 -0
- package/dist/utils/hash.d.ts.map +1 -0
- package/dist/utils/hash.js +14 -0
- package/dist/utils/hash.js.map +1 -0
- package/dist/utils/headers.d.ts +4 -0
- package/dist/utils/headers.d.ts.map +1 -0
- package/dist/utils/headers.js +20 -0
- package/dist/utils/headers.js.map +1 -0
- package/dist/utils/identity.d.ts +67 -0
- package/dist/utils/identity.d.ts.map +1 -0
- package/dist/utils/identity.js +148 -0
- package/dist/utils/identity.js.map +1 -0
- package/dist/utils/inline-images.d.ts +24 -0
- package/dist/utils/inline-images.d.ts.map +1 -0
- package/dist/utils/inline-images.js +55 -0
- package/dist/utils/inline-images.js.map +1 -0
- package/dist/utils/json-parse.d.ts +18 -0
- package/dist/utils/json-parse.d.ts.map +1 -0
- package/dist/utils/json-parse.js +167 -0
- package/dist/utils/json-parse.js.map +1 -0
- package/dist/utils/model-operations.d.ts +12 -0
- package/dist/utils/model-operations.d.ts.map +1 -0
- package/dist/utils/model-operations.js +58 -0
- package/dist/utils/model-operations.js.map +1 -0
- package/dist/utils/models-dev-catalog.d.ts +324 -0
- package/dist/utils/models-dev-catalog.d.ts.map +1 -0
- package/dist/utils/models-dev-catalog.js +3166 -0
- package/dist/utils/models-dev-catalog.js.map +1 -0
- package/dist/utils/models-dev-reasoning-options.d.ts +18 -0
- package/dist/utils/models-dev-reasoning-options.d.ts.map +1 -0
- package/dist/utils/models-dev-reasoning-options.js +20 -0
- package/dist/utils/models-dev-reasoning-options.js.map +1 -0
- package/dist/utils/models-error.d.ts +8 -0
- package/dist/utils/models-error.d.ts.map +1 -0
- package/dist/utils/models-error.js +19 -0
- package/dist/utils/models-error.js.map +1 -0
- package/dist/utils/node-http-proxy.d.ts +4 -0
- package/dist/utils/node-http-proxy.d.ts.map +1 -0
- package/dist/utils/node-http-proxy.js +141 -0
- package/dist/utils/node-http-proxy.js.map +1 -0
- package/dist/utils/oauth-page.d.ts +2 -0
- package/dist/utils/oauth-page.d.ts.map +1 -0
- package/dist/utils/oauth-page.js +2 -0
- package/dist/utils/oauth-page.js.map +1 -0
- package/dist/utils/openrouter-cost.d.ts +24 -0
- package/dist/utils/openrouter-cost.d.ts.map +1 -0
- package/dist/utils/openrouter-cost.js +34 -0
- package/dist/utils/openrouter-cost.js.map +1 -0
- package/dist/utils/overflow.d.ts +73 -0
- package/dist/utils/overflow.d.ts.map +1 -0
- package/dist/utils/overflow.js +204 -0
- package/dist/utils/overflow.js.map +1 -0
- package/dist/utils/provider-env.d.ts +12 -0
- package/dist/utils/provider-env.d.ts.map +1 -0
- package/dist/utils/provider-env.js +51 -0
- package/dist/utils/provider-env.js.map +1 -0
- package/dist/utils/provider-retry.d.ts +21 -0
- package/dist/utils/provider-retry.d.ts.map +1 -0
- package/dist/utils/provider-retry.js +100 -0
- package/dist/utils/provider-retry.js.map +1 -0
- package/dist/utils/request-admission.d.ts +12 -0
- package/dist/utils/request-admission.d.ts.map +1 -0
- package/dist/utils/request-admission.js +37 -0
- package/dist/utils/request-admission.js.map +1 -0
- package/dist/utils/retry-hints.d.ts +15 -0
- package/dist/utils/retry-hints.d.ts.map +1 -0
- package/dist/utils/retry-hints.js +130 -0
- package/dist/utils/retry-hints.js.map +1 -0
- package/dist/utils/retry.d.ts +78 -0
- package/dist/utils/retry.d.ts.map +1 -0
- package/dist/utils/retry.js +251 -0
- package/dist/utils/retry.js.map +1 -0
- package/dist/utils/sanitize-unicode.d.ts +22 -0
- package/dist/utils/sanitize-unicode.d.ts.map +1 -0
- package/dist/utils/sanitize-unicode.js +26 -0
- package/dist/utils/sanitize-unicode.js.map +1 -0
- package/dist/utils/sleep.d.ts +2 -0
- package/dist/utils/sleep.d.ts.map +1 -0
- package/dist/utils/sleep.js +15 -0
- package/dist/utils/sleep.js.map +1 -0
- package/dist/utils/sse.d.ts +29 -0
- package/dist/utils/sse.d.ts.map +1 -0
- package/dist/utils/sse.js +73 -0
- package/dist/utils/sse.js.map +1 -0
- package/dist/utils/stream-deadline.d.ts +78 -0
- package/dist/utils/stream-deadline.d.ts.map +1 -0
- package/dist/utils/stream-deadline.js +205 -0
- package/dist/utils/stream-deadline.js.map +1 -0
- package/dist/utils/streaming-json.d.ts +84 -0
- package/dist/utils/streaming-json.d.ts.map +1 -0
- package/dist/utils/streaming-json.js +513 -0
- package/dist/utils/streaming-json.js.map +1 -0
- package/dist/utils/text.d.ts +21 -0
- package/dist/utils/text.d.ts.map +1 -0
- package/dist/utils/text.js +45 -0
- package/dist/utils/text.js.map +1 -0
- package/dist/utils/thinking-off.d.ts +23 -0
- package/dist/utils/thinking-off.d.ts.map +1 -0
- package/dist/utils/thinking-off.js +51 -0
- package/dist/utils/thinking-off.js.map +1 -0
- package/dist/utils/transcript.d.ts +101 -0
- package/dist/utils/transcript.d.ts.map +1 -0
- package/dist/utils/transcript.js +218 -0
- package/dist/utils/transcript.js.map +1 -0
- package/dist/utils/typebox-helpers.d.ts +17 -0
- package/dist/utils/typebox-helpers.d.ts.map +1 -0
- package/dist/utils/typebox-helpers.js +21 -0
- package/dist/utils/typebox-helpers.js.map +1 -0
- package/dist/utils/uuid.d.ts +3 -0
- package/dist/utils/uuid.d.ts.map +1 -0
- package/dist/utils/uuid.js +46 -0
- package/dist/utils/uuid.js.map +1 -0
- package/dist/utils/validation.d.ts +24 -0
- package/dist/utils/validation.d.ts.map +1 -0
- package/dist/utils/validation.js +363 -0
- package/dist/utils/validation.js.map +1 -0
- package/dist/utils/version.d.ts +2 -0
- package/dist/utils/version.d.ts.map +1 -0
- package/dist/utils/version.js +3 -0
- package/dist/utils/version.js.map +1 -0
- package/package.json +135 -8
package/CHANGELOG.md
ADDED
|
@@ -0,0 +1,1601 @@
|
|
|
1
|
+
# Changelog
|
|
2
|
+
|
|
3
|
+
This package is Atomic's fork of `@earendil-works/pi-ai`, published as `@bastani/atomic-ai` (formerly `@bastani/pi-ai`). Upstream history at the audited Pi `main` sync point (`cb7969d212836b8939001dce159fbd2ed6ad395f`) lives in [earendil-works/pi](https://github.com/earendil-works/pi/blob/cb7969d212836b8939001dce159fbd2ed6ad395f/packages/ai/CHANGELOG.md).
|
|
4
|
+
|
|
5
|
+
## [Unreleased]
|
|
6
|
+
|
|
7
|
+
## [0.9.88-alpha.1] - 2026-11-16
|
|
8
|
+
|
|
9
|
+
### Changed
|
|
10
|
+
|
|
11
|
+
- Renamed the package from `@bastani/pi-ai` to `@bastani/atomic-ai`. It now depends on Atomic's forks `@bastani/atomic-schema` and `@bastani/atomic-telemetry` instead of pi's packages. Atomic's extension loader still resolves `@bastani/pi-ai`, `@earendil-works/pi-ai` and `@mariozechner/pi-ai` imports to it ([#5174](https://github.com/bastani-inc/atomic/issues/5174))
|
|
12
|
+
- Requests now identify as Atomic by default: the default `User-Agent` is `atomic/<version>` (short form `atomic`), and the new `ATOMIC_ATTRIBUTION` attributes requests to Atomic for OpenRouter, NVIDIA and OpenCode. `PI_ATTRIBUTION` is deprecated and now equals `ATOMIC_ATTRIBUTION`. `ATOMIC_CACHE_RETENTION` and `ATOMIC_OAUTH_CALLBACK_HOST` replace `PI_CACHE_RETENTION` and `PI_OAUTH_CALLBACK_HOST`, which keep working as legacy names ([#5174](https://github.com/bastani-inc/atomic/issues/5174))
|
|
13
|
+
- OpenAI Codex requests and the Codex browser sign-in send `originator: atomic`, xAI OAuth sends `referrer: atomic`, and Bedrock's user-agent pair is `atomic/<version>`, all instead of `pi`. The OAuth client IDs are unchanged; if Codex or xAI sign-in or requests start failing, these identifiers are the change to revert ([#5174](https://github.com/bastani-inc/atomic/issues/5174))
|
|
14
|
+
|
|
15
|
+
### Fixed
|
|
16
|
+
|
|
17
|
+
- Google Gemini, Google Vertex, Anthropic, OpenAI Completions, OpenAI Responses, OpenAI Codex, Azure OpenAI Responses and Mistral failures now record the HTTP status, and Google's provider status such as `UNAVAILABLE`, in a `provider_response_failure` diagnostic on the failed message, so retry and fallback can decide on the status rather than the message text ([#4994](https://github.com/bastani-inc/atomic/issues/4994))
|
|
18
|
+
- Fixed requests failing when the conversation holds an image in a format providers do not accept inline, such as `image/svg+xml`: every provider now sends a text placeholder naming the type instead of the image, and Bedrock no longer throws "Unknown image type" ([#4908](https://github.com/bastani-inc/atomic/issues/4908)).
|
|
19
|
+
- Fixed Google Gemini, Google Vertex, Amazon Bedrock and Azure OpenAI Responses streams ignoring `streamDeadlineMs`: a stream that stalls after its first event is now cut at the deadline, closes its connection, and fails with the retryable "stream deadline exceeded" error so retry and model fallback run instead of the turn hanging ([#4904](https://github.com/bastani-inc/atomic/issues/4904))
|
|
20
|
+
- Fixed Anthropic, Mistral, OpenAI Codex, and pi-messages streams failing or losing events when a CRLF line ending, a multibyte character, or a `data:` prefix is split across network chunks. Every stream now parses the same way however it is chunked, and the OpenAI Codex and pi-messages streams accept CRLF and CR line endings ([#4899](https://github.com/bastani-inc/atomic/issues/4899))
|
|
21
|
+
- Fixed OpenAI-compatible streams merging parallel tool calls that a server sends with the same `index` into one call with malformed arguments; a named tool call delta with a new `id` now starts its own call ([#4899](https://github.com/bastani-inc/atomic/issues/4899))
|
|
22
|
+
- Fixed Anthropic tool calls losing their input when a server sends it complete in `content_block_start` with no `input_json_delta` events; the call now keeps that input ([#4899](https://github.com/bastani-inc/atomic/issues/4899))
|
|
23
|
+
- Fixed Anthropic, OpenAI Codex, OpenAI ChatGPT, Kimi Code, and Radius sign-ins asking you to run `/login` again after a token refresh whose response carries no new refresh token, which servers that don't rotate refresh tokens may send. The refresh now keeps the existing refresh token, and a response without `expires_in` takes its expiry from the access token's JWT `exp` claim or assumes one hour, instead of refreshing on every request ([#4959](https://github.com/bastani-inc/atomic/issues/4959))
|
|
24
|
+
- Fixed Gemini 2.x on Google and Vertex rejecting a parallel tool batch when a result other than the last one contains an image: every function response of the batch now stays in one turn, in call order, and the image turns follow it ([#5048](https://github.com/bastani-inc/atomic/issues/5048))
|
|
25
|
+
- Fixed `retryAssistantCall()` and `failureRetryDelayMs()` capping a server retry hint at `RetryPolicy.maxAgentDelayMs` when `maxRetryDelayMs` is not set. An unset `maxRetryDelayMs` now caps the hint at the documented 60 seconds, so a lower agent backoff cap no longer cuts a server's requested wait short; `maxRetryDelayMs: 0` still caps it at `maxAgentDelayMs`
|
|
26
|
+
|
|
27
|
+
### Removed
|
|
28
|
+
|
|
29
|
+
- Removed the `radius` provider, its OAuth flow and models, and the build-time fetch of its catalog from radius.pi.dev. The generic `pi-messages` API stays available for custom providers ([#5174](https://github.com/bastani-inc/atomic/issues/5174))
|
|
30
|
+
|
|
31
|
+
### Added
|
|
32
|
+
|
|
33
|
+
- Added `buildModelsDevChatModels()` and `finalizeChatModels()` in `@bastani/atomic-ai/utils/models-dev-catalog`, the transform from models.dev's `api.json` to the built-in chat catalog ([#5174](https://github.com/bastani-inc/atomic/issues/5174))
|
|
34
|
+
|
|
35
|
+
## [0.9.85] - 2026-11-13
|
|
36
|
+
|
|
37
|
+
Dropped connections are retried automatically.
|
|
38
|
+
|
|
39
|
+
- **Retries on bare socket errors.** Provider errors that report only a socket code such as `ECONNRESET`, `EPIPE`, or `ETIMEDOUT` now count as retryable, so a dropped connection no longer ends the request (upstream [#12128](https://github.com/earendil-works/pi/issues/12128), [#4740](https://github.com/bastani-inc/atomic/pull/4740)).
|
|
40
|
+
|
|
41
|
+
### Fixed
|
|
42
|
+
|
|
43
|
+
- Fixed provider errors that report only a Node socket error code, such as `read ECONNRESET`, `write EPIPE`, `connect ETIMEDOUT`, `connect ECONNREFUSED`, or `ECONNABORTED` (or `"Connection reset by peer"`), not being classified as retryable, without matching those codes inside longer identifiers (upstream [#12128](https://github.com/earendil-works/pi/issues/12128), [#12136](https://github.com/earendil-works/pi/pull/12136))
|
|
44
|
+
|
|
45
|
+
## [0.9.85-alpha.1] - 2026-11-13
|
|
46
|
+
|
|
47
|
+
### Fixed
|
|
48
|
+
|
|
49
|
+
- Fixed provider errors that report only a Node socket error code, such as `read ECONNRESET`, `write EPIPE`, `connect ETIMEDOUT`, `connect ECONNREFUSED`, or `ECONNABORTED` (or `"Connection reset by peer"`), not being classified as retryable, without matching those codes inside longer identifiers (upstream [#12128](https://github.com/earendil-works/pi/issues/12128), [#12136](https://github.com/earendil-works/pi/pull/12136))
|
|
50
|
+
|
|
51
|
+
## [0.9.81] - 2026-11-13
|
|
52
|
+
|
|
53
|
+
Qwen Token Plan moves to its current gateways, and tool schemas and Google requests fail less often.
|
|
54
|
+
|
|
55
|
+
- **Current Qwen Token Plan gateways.** The Team providers use the current hosts, the Individual providers gain `deepseek-v4.1-flash` models, and each plan can have its own API key variable (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112)).
|
|
56
|
+
- **Fewer rejected tool calls.** Tools whose schemas cannot be strict fall back to non-strict mode instead of drawing a 400 from OpenAI, and non-strict Anthropic schemas keep their root `additionalProperties` (upstream [#12049](https://github.com/earendil-works/pi/issues/12049), [#12100](https://github.com/earendil-works/pi/pull/12100)).
|
|
57
|
+
- **Steadier Google and retry handling.** Google adapters accept a custom `fetch` and end `CONTINUATION` turns cleanly, and `"Connection closed."` errors count as retryable (upstream [#12103](https://github.com/earendil-works/pi/pull/12103), [#12108](https://github.com/earendil-works/pi/pull/12108), [#12093](https://github.com/earendil-works/pi/pull/12093)).
|
|
58
|
+
|
|
59
|
+
### Added
|
|
60
|
+
|
|
61
|
+
- Added `deepseek-v4.1-flash` and `deepseek-v4.1-flash-1017` to the Qwen Token Plan Individual providers, and per-plan key variables `QWEN_TOKEN_PLAN_TEAM_API_KEY`, `QWEN_TOKEN_PLAN_TEAM_CN_API_KEY`, `QWEN_TOKEN_PLAN_INDIVIDUAL_API_KEY`, and `QWEN_TOKEN_PLAN_INDIVIDUAL_CN_API_KEY`; the shared `QWEN_TOKEN_PLAN_API_KEY` and `QWEN_TOKEN_PLAN_CN_API_KEY` still work as fallbacks (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112))
|
|
62
|
+
|
|
63
|
+
### Changed
|
|
64
|
+
|
|
65
|
+
- Updated `@google/genai` to 2.32.0 with automatic continuation turned off, so a `CONTINUATION` finish reason ends the turn with stop reason `length` (upstream [#12103](https://github.com/earendil-works/pi/pull/12103), [#12108](https://github.com/earendil-works/pi/pull/12108))
|
|
66
|
+
- Qwen Token Plan providers now use the current gateways (`token-plan.maas.qwencloudapi.com` international, `token-plan.maas.qianwenaiapi.com` China), and `qwen-token-plan` and `qwen-token-plan-cn` are named "Qwen Token Plan Team" and "Qwen Token Plan Team CN". Provider IDs are unchanged, and the old `aliyuncs.com` hosts still work (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112))
|
|
67
|
+
|
|
68
|
+
### Fixed
|
|
69
|
+
|
|
70
|
+
- Fixed tools with `strict: "prefer"` constrained sampling being sent in strict mode when their schema contains a node without a `type`, such as `items: {}`, `Type.Unknown()`, or a bare `const`, which OpenAI rejects with a 400 error; these tools now fall back to non-strict mode, `strict: "require"` fails locally, and nodes with only an `enum` stay strict. Non-strict Anthropic tool schemas, including those that fell back from `strict: "prefer"`, also keep their root `additionalProperties` keyword instead of dropping it (upstream [#12049](https://github.com/earendil-works/pi/issues/12049), [#12100](https://github.com/earendil-works/pi/pull/12100))
|
|
71
|
+
- Fixed provider errors reading `"Connection closed."` not being classified as retryable (upstream [#12093](https://github.com/earendil-works/pi/pull/12093))
|
|
72
|
+
- Fixed the `google-generative-ai` and `google-vertex` adapters rejecting a custom `fetch` option instead of forwarding it through `httpOptions.fetch`, and kept the global undici dispatcher's idle timeouts active when no custom `fetch` is supplied; Vertex ADC token requests still use `google-auth-library`'s own HTTP client (upstream [#12103](https://github.com/earendil-works/pi/pull/12103), [#12108](https://github.com/earendil-works/pi/pull/12108))
|
|
73
|
+
|
|
74
|
+
### Removed
|
|
75
|
+
|
|
76
|
+
- Removed the unused `mapStopReasonString` export from `@bastani/pi-ai/api/google-shared` (upstream [#12103](https://github.com/earendil-works/pi/pull/12103))
|
|
77
|
+
- Removed the retired `deepseek-v3.2` and `deepseek-v4-flash` from the Qwen Token Plan Team providers; use `deepseek-v4-flash-0731` or `deepseek-v4.1-flash` (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112))
|
|
78
|
+
|
|
79
|
+
## [0.9.81-alpha.1] - 2026-11-13
|
|
80
|
+
|
|
81
|
+
### Added
|
|
82
|
+
|
|
83
|
+
- Added `deepseek-v4.1-flash` and `deepseek-v4.1-flash-1017` to the Qwen Token Plan Individual providers, and per-plan key variables `QWEN_TOKEN_PLAN_TEAM_API_KEY`, `QWEN_TOKEN_PLAN_TEAM_CN_API_KEY`, `QWEN_TOKEN_PLAN_INDIVIDUAL_API_KEY`, and `QWEN_TOKEN_PLAN_INDIVIDUAL_CN_API_KEY`; the shared `QWEN_TOKEN_PLAN_API_KEY` and `QWEN_TOKEN_PLAN_CN_API_KEY` still work as fallbacks (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112))
|
|
84
|
+
|
|
85
|
+
### Changed
|
|
86
|
+
|
|
87
|
+
- Updated `@google/genai` to 2.32.0 with automatic continuation turned off, so a `CONTINUATION` finish reason ends the turn with stop reason `length` (upstream [#12103](https://github.com/earendil-works/pi/pull/12103), [#12108](https://github.com/earendil-works/pi/pull/12108))
|
|
88
|
+
- Qwen Token Plan providers now use the current gateways (`token-plan.maas.qwencloudapi.com` international, `token-plan.maas.qianwenaiapi.com` China), and `qwen-token-plan` and `qwen-token-plan-cn` are named "Qwen Token Plan Team" and "Qwen Token Plan Team CN". Provider IDs are unchanged, and the old `aliyuncs.com` hosts still work (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112))
|
|
89
|
+
|
|
90
|
+
### Fixed
|
|
91
|
+
|
|
92
|
+
- Fixed tools with `strict: "prefer"` constrained sampling being sent in strict mode when their schema contains a node without a `type`, such as `items: {}`, `Type.Unknown()`, or a bare `const`, which OpenAI rejects with a 400 error; these tools now fall back to non-strict mode, `strict: "require"` fails locally, and nodes with only an `enum` stay strict (upstream [#12049](https://github.com/earendil-works/pi/issues/12049))
|
|
93
|
+
- Fixed provider errors reading `"Connection closed."` not being classified as retryable (upstream [#12093](https://github.com/earendil-works/pi/pull/12093))
|
|
94
|
+
- Fixed non-strict Anthropic tool schemas dropping the root `additionalProperties` keyword, including when `strict: "prefer"` falls back to non-strict mode (upstream [#12100](https://github.com/earendil-works/pi/pull/12100))
|
|
95
|
+
- Fixed the `google-generative-ai` and `google-vertex` adapters rejecting a custom `fetch` option instead of forwarding it through `httpOptions.fetch`, and kept the global undici dispatcher's idle timeouts active when no custom `fetch` is supplied; Vertex ADC token requests still use `google-auth-library`'s own HTTP client (upstream [#12103](https://github.com/earendil-works/pi/pull/12103), [#12108](https://github.com/earendil-works/pi/pull/12108))
|
|
96
|
+
|
|
97
|
+
### Removed
|
|
98
|
+
|
|
99
|
+
- Removed the unused `mapStopReasonString` export from `@bastani/pi-ai/api/google-shared` (upstream [#12103](https://github.com/earendil-works/pi/pull/12103))
|
|
100
|
+
- Removed the retired `deepseek-v3.2` and `deepseek-v4-flash` from the Qwen Token Plan Team providers; use `deepseek-v4-flash-0731` or `deepseek-v4.1-flash` (upstream [#11773](https://github.com/earendil-works/pi/issues/11773), [#12112](https://github.com/earendil-works/pi/pull/12112))
|
|
101
|
+
|
|
102
|
+
## [0.9.79] - 2026-11-12
|
|
103
|
+
|
|
104
|
+
Anthropic message threads follow the endpoint, not the provider name.
|
|
105
|
+
|
|
106
|
+
- **Anthropic threads for custom providers.** A provider whose `baseUrl` is on `api.anthropic.com`, such as a second Anthropic API provider, gets message threads by default (upstream [#11987](https://github.com/earendil-works/pi/issues/11987)).
|
|
107
|
+
|
|
108
|
+
### Fixed
|
|
109
|
+
|
|
110
|
+
- Fixed Anthropic message threads staying off for providers other than `anthropic` whose `baseUrl` is on `api.anthropic.com`, such as a second Anthropic API provider; the default now follows the endpoint, not the provider ID (upstream [#11987](https://github.com/earendil-works/pi/issues/11987))
|
|
111
|
+
|
|
112
|
+
## [0.9.79-alpha.1] - 2026-11-11
|
|
113
|
+
|
|
114
|
+
### Fixed
|
|
115
|
+
|
|
116
|
+
- Fixed Anthropic message threads staying off for providers other than `anthropic` whose `baseUrl` is on `api.anthropic.com`, such as a second Anthropic API provider; the default now follows the endpoint, not the provider ID (upstream [#11987](https://github.com/earendil-works/pi/issues/11987))
|
|
117
|
+
|
|
118
|
+
## [0.9.78] - 2026-11-11
|
|
119
|
+
|
|
120
|
+
GitHub Copilot's newest Gemini and Claude models are always in the catalog.
|
|
121
|
+
|
|
122
|
+
- **New Copilot models.** `github-copilot/gemini-4-carbon` and `github-copilot/claude-fable-5.2` are available even from a catalog built before models.dev listed them, and Fable 5.2 offers only the thinking levels it accepts ([#4653](https://github.com/bastani-inc/atomic/pull/4653)).
|
|
123
|
+
|
|
124
|
+
### Added
|
|
125
|
+
|
|
126
|
+
- GitHub Copilot: `github-copilot/gemini-4-carbon` and `github-copilot/claude-fable-5.2` are always in the built-in catalog, also when the catalog was built before models.dev listed them ([#4653](https://github.com/bastani-inc/atomic/pull/4653))
|
|
127
|
+
|
|
128
|
+
### Fixed
|
|
129
|
+
|
|
130
|
+
- Claude Fable 5.2 no longer offers a `minimal` thinking level; like Fable 5.1 it accepts `low`, `medium`, `high`, `xhigh` and `max` ([#4653](https://github.com/bastani-inc/atomic/pull/4653))
|
|
131
|
+
|
|
132
|
+
## [0.9.78-alpha.1] - 2026-11-11
|
|
133
|
+
|
|
134
|
+
GitHub Copilot's newest Gemini and Claude models are always in the catalog.
|
|
135
|
+
|
|
136
|
+
- **New Copilot models.** `github-copilot/gemini-4-carbon` and `github-copilot/claude-fable-5.2` are available even from a catalog built before models.dev listed them, and Fable 5.2 offers only the thinking levels it accepts ([#4653](https://github.com/bastani-inc/atomic/pull/4653)).
|
|
137
|
+
|
|
138
|
+
### Added
|
|
139
|
+
|
|
140
|
+
- GitHub Copilot: `github-copilot/gemini-4-carbon` and `github-copilot/claude-fable-5.2` are always in the built-in catalog, also when the catalog was built before models.dev listed them ([#4653](https://github.com/bastani-inc/atomic/pull/4653))
|
|
141
|
+
|
|
142
|
+
### Fixed
|
|
143
|
+
|
|
144
|
+
- Claude Fable 5.2 no longer offers a `minimal` thinking level; like Fable 5.1 it accepts `low`, `medium`, `high`, `xhigh` and `max` ([#4653](https://github.com/bastani-inc/atomic/pull/4653))
|
|
145
|
+
|
|
146
|
+
## [0.9.77] - 2026-11-11
|
|
147
|
+
|
|
148
|
+
Providers can release their resources when a session ends.
|
|
149
|
+
|
|
150
|
+
- **Session resource cleanup.** Import `registerSessionResourceCleanup` and `cleanupSessionResources` from `@bastani/pi-ai/session-resources` to free provider resources when a session ends ([#4638](https://github.com/bastani-inc/atomic/pull/4638), upstream [#11961](https://github.com/earendil-works/pi/pull/11961)).
|
|
151
|
+
|
|
152
|
+
### Added
|
|
153
|
+
|
|
154
|
+
- Added the `@bastani/pi-ai/session-resources` subpath export, so `registerSessionResourceCleanup` and `cleanupSessionResources` can be imported to release provider resources when a session ends ([#4638](https://github.com/bastani-inc/atomic/pull/4638), upstream [#11961](https://github.com/earendil-works/pi/pull/11961))
|
|
155
|
+
|
|
156
|
+
## [0.9.77-alpha.1] - 2026-11-11
|
|
157
|
+
|
|
158
|
+
Providers can release their resources when a session ends.
|
|
159
|
+
|
|
160
|
+
- **Session resource cleanup.** Import `registerSessionResourceCleanup` and `cleanupSessionResources` from `@bastani/pi-ai/session-resources` to free provider resources when a session ends ([#4638](https://github.com/bastani-inc/atomic/pull/4638), upstream [#11961](https://github.com/earendil-works/pi/pull/11961)).
|
|
161
|
+
|
|
162
|
+
### Added
|
|
163
|
+
|
|
164
|
+
- Added the `@bastani/pi-ai/session-resources` subpath export, so `registerSessionResourceCleanup` and `cleanupSessionResources` can be imported to release provider resources when a session ends (upstream [#11961](https://github.com/earendil-works/pi/pull/11961))
|
|
165
|
+
|
|
166
|
+
## [0.9.76] - 2026-11-10
|
|
167
|
+
|
|
168
|
+
Claude Fable 5.2 works with an Anthropic subscription login.
|
|
169
|
+
|
|
170
|
+
- **Fable 5.2 over Claude subscriptions.** `claude-fable-5-2` no longer fails with `claude_code_version_too_old` when you sign in with an Anthropic subscription, because OAuth requests now identify as the latest Claude Code release (`2.1.322`) ([#4634](https://github.com/bastani-inc/atomic/pull/4634), upstream [#11969](https://github.com/earendil-works/pi/issues/11969)).
|
|
171
|
+
|
|
172
|
+
### Fixed
|
|
173
|
+
|
|
174
|
+
- Claude Fable 5.2 (`claude-fable-5-2`) works with an Anthropic subscription login. The Claude Code version sent on OAuth requests is now `2.1.322`, the latest Claude Code release, so Anthropic no longer rejects them with `claude_code_version_too_old` ([#4634](https://github.com/bastani-inc/atomic/pull/4634), upstream [#11969](https://github.com/earendil-works/pi/issues/11969))
|
|
175
|
+
|
|
176
|
+
## [0.9.76-alpha.1] - 2026-11-10
|
|
177
|
+
|
|
178
|
+
Claude Fable 5.2 works with an Anthropic subscription login.
|
|
179
|
+
|
|
180
|
+
- **Fable 5.2 over Claude subscriptions.** `claude-fable-5-2` no longer fails with `claude_code_version_too_old` when you sign in with an Anthropic subscription, because OAuth requests now identify as the latest Claude Code release (`2.1.322`) ([#4634](https://github.com/bastani-inc/atomic/pull/4634), upstream [#11969](https://github.com/earendil-works/pi/issues/11969)).
|
|
181
|
+
|
|
182
|
+
### Fixed
|
|
183
|
+
|
|
184
|
+
- Claude Fable 5.2 (`claude-fable-5-2`) works with an Anthropic subscription login. The Claude Code version sent on OAuth requests is now `2.1.322`, the latest Claude Code release, so Anthropic no longer rejects them with `claude_code_version_too_old` (upstream [#11969](https://github.com/earendil-works/pi/issues/11969))
|
|
185
|
+
|
|
186
|
+
## [0.9.75] - 2026-11-10
|
|
187
|
+
|
|
188
|
+
Claude Fable 5.2, steadier GitHub Copilot streams and SVG output that works without extra options.
|
|
189
|
+
|
|
190
|
+
- **Claude Fable 5.2.** `claude-fable-5-2` joins the `anthropic` provider with per-message effort and server-side fallback to Claude Opus 5, and Fable and Mythos 5.2 on other providers get the same effort support.
|
|
191
|
+
- **Diagnosable, steadier Copilot streams.** A Copilot stream that ends early reports how long it ran, whether the response had started and the request ID, Anthropic `ping` events keep a silently thinking model inside `streamDeadlineMs`, and a stored API key is filtered by the account's own model list ([#4626](https://github.com/bastani-inc/atomic/issues/4626), [#4627](https://github.com/bastani-inc/atomic/issues/4627)).
|
|
192
|
+
- **SVG images on GPT Image 2.6.** `outputFormat: "svg"` retries with a default authoring model when the API needs one, reports it in `svgModel`, and fails with a typed `invalid-option` error instead of `api-error` ([#4615](https://github.com/bastani-inc/atomic/issues/4615)).
|
|
193
|
+
- **Shared cache warming and correct costs.** Prompt cache warming decisions are shared helpers every host can use, Azure OpenAI keeps its long-context price tier, Anthropic and Bedrock stop resending rejected strict tools, and credential-free catalogs refresh even when a login is stale (upstream [#11857](https://github.com/earendil-works/pi/pull/11857), [#11948](https://github.com/earendil-works/pi/issues/11948), [#11626](https://github.com/earendil-works/pi/issues/11626), [#4610](https://github.com/bastani-inc/atomic/issues/4610)).
|
|
194
|
+
|
|
195
|
+
### Added
|
|
196
|
+
|
|
197
|
+
- Added the prompt cache warming policy as shared helpers, so every host that warms a prompt cache decides the same way: `getPromptCacheTtlMs()` (the lifetime of the entry a request writes, from the model's `promptCache` tier for the request's `cacheRetention`, then `PI_CACHE_RETENTION`, defaulting to `short`), `getCacheWarmingDelayMs()`, `isCacheWarmingReplayable()`, `getCacheWarmingRequestOptions()`, `evaluateCacheWarming()`, the `CACHE_WARMING_*` constants and the `CacheWarmingAction`, `CacheWarmingPhase` and `CacheWarmingDecision` types. They moved from the coding agent's cache warmer, which keeps resolving the model-specific default retention before it asks for the lifetime (upstream [#11857](https://github.com/earendil-works/pi/pull/11857))
|
|
198
|
+
- Added Claude Fable 5.2 (`claude-fable-5-2`) to the `anthropic` provider ($8/$40 per million input/output tokens, 1M context, 128K output). Like Fable 5.1 it uses per-message effort and mid-conversation system messages; requests its safety classifiers refuse before any output are answered by Claude Opus 5 through server-side fallback and billed at Opus 5's price. Claude Fable 5.2 and Mythos 5.2 on other providers, such as OpenRouter's `anthropic/claude-fable-5.2`, get the same per-message effort support
|
|
199
|
+
- Added `GitHubCopilotStreamAttempt`, which records the request ID, duration and progress of one Copilot streaming request, and its `GitHubCopilotStreamSnapshot` result type, which `classifyGitHubCopilotFailure()` takes as `stream` ([#4626](https://github.com/bastani-inc/atomic/issues/4626))
|
|
200
|
+
|
|
201
|
+
### Fixed
|
|
202
|
+
|
|
203
|
+
- GitHub Copilot now filters a stored API key by its account's model list as it does an OAuth login. `ApiKeyCredential` gains the optional `availableModelIds` and `fastModelIds` that a model refresh saves from the account's list, `RefreshModelsContext.updateCredential` is offered for a stored API key too and now takes and returns a `Credential` (it writes only while the store still holds the key or token the refresh used), and `refreshGitHubCopilotAccountModelLists()` renews just those lists for a provider alias. Nothing is hidden for a key whose list was not read or names no usable model ([#4627](https://github.com/bastani-inc/atomic/issues/4627))
|
|
204
|
+
- Catalog providers can set `Provider.refreshModelsAuth: "configured"` to refresh without renewing OAuth or receiving a provider credential, while still skipping unconfigured providers. Stale logins no longer block credential-free catalogs ([#4610](https://github.com/bastani-inc/atomic/issues/4610))
|
|
205
|
+
- Azure OpenAI models now keep OpenAI's long-context price tier, so requests above 272K input tokens are costed at Azure's long-context rates instead of the standard rates (upstream [#11948](https://github.com/earendil-works/pi/issues/11948))
|
|
206
|
+
- Fixed Anthropic and Bedrock sending `strict: "prefer"` tools strict again on every request after Claude rejected them for exceeding its strict tool limits, which cost a rejected request (400) per turn. Later requests with the same `sessionId` now send those tools non-strict right away, until `cleanupSessionResources(sessionId)` (upstream [#11626](https://github.com/earendil-works/pi/issues/11626))
|
|
207
|
+
- Fixed `outputFormat: "svg"` on GPT Image 2.6 failing with a generic `api-error` (HTTP 400, `missing_required_parameter` for `svg.model`) on accounts where the API requires the SVG authoring model. A request without `svg` now retries once with a default authoring model, `DEFAULT_SVG_AUTHOR_MODEL` (`gpt-6.1-sol`, a value that worked in testing: OpenAI documents no default), and retries once more without `svg` if the API rejects it as unsupported, so a call sends at most three requests. An `svg` you pass is always sent as given and never retried or replaced. `AssistantImages.svgModel` reports the model that wrote the SVG. A rejected `svg` or `svg.model` now fails with `ImageGenerationConfigError` (`code: "invalid-option"`, `field` `svg.model` or `svg`, and the new `reason` values `required-by-api`, `rejected-by-api` and `unsupported-by-api`) instead of `api-error`. Image requests pass their 10-minute default timeout to the SDK explicitly, so SVG authoring that takes tens of seconds never times out early ([#4615](https://github.com/bastani-inc/atomic/issues/4615))
|
|
208
|
+
- Fixed GitHub Copilot `stream-ended-early` failures giving almost nothing to diagnose them with. `GitHubCopilotStreamEndedEarlyError` now carries `durationMs` (how long the attempt ran) and `startedContent` (whether any stream event had arrived), and takes its `requestId` from the response's `x-request-id`, on Anthropic Messages, Chat Completions and Responses streams. The fields appear in `fields()`, survive serialization, and show in the message (`Attempt ran 312.4 s, response started, request ID …`). The hint no longer says Atomic retries "when retries are enabled", which was wrong once the retries were used up; it now suggests splitting the work, lowering the thinking effort or adding a fallback model. `GitHubCopilotFailureContext.streamStarted` is replaced by `stream`, and `recordGitHubCopilotFailure()` takes a `GitHubCopilotStreamSnapshot` (or `undefined` when no response arrived) instead of a boolean. `GitHubCopilotErrorFieldValue` now includes `boolean` ([#4626](https://github.com/bastani-inc/atomic/issues/4626))
|
|
209
|
+
- Fixed Anthropic `ping` events not counting as activity for the stream idle deadline (`streamDeadlineMs`), so a model thinking silently while the server kept the connection alive was cut at the deadline as if the stream had stalled. `withStreamDeadline()` takes an optional keep-alive predicate: a match restarts the idle timer and is not yielded, so pings never reach `onProviderStreamEvent` ([#4626](https://github.com/bastani-inc/atomic/issues/4626))
|
|
210
|
+
|
|
211
|
+
## [0.9.75-alpha.1] - 2026-11-10
|
|
212
|
+
|
|
213
|
+
### Added
|
|
214
|
+
|
|
215
|
+
- Added the prompt cache warming policy as shared helpers, so every host that warms a prompt cache decides the same way: `getPromptCacheTtlMs()` (the lifetime of the entry a request writes, from the model's `promptCache` tier for the request's `cacheRetention`, then `PI_CACHE_RETENTION`, defaulting to `short`), `getCacheWarmingDelayMs()`, `isCacheWarmingReplayable()`, `getCacheWarmingRequestOptions()`, `evaluateCacheWarming()`, the `CACHE_WARMING_*` constants and the `CacheWarmingAction`, `CacheWarmingPhase` and `CacheWarmingDecision` types. They moved from the coding agent's cache warmer, which keeps resolving the model-specific default retention before it asks for the lifetime (upstream [#11857](https://github.com/earendil-works/pi/pull/11857))
|
|
216
|
+
- Added Claude Fable 5.2 (`claude-fable-5-2`) to the `anthropic` provider ($8/$40 per million input/output tokens, 1M context, 128K output). Like Fable 5.1 it uses per-message effort and mid-conversation system messages; requests its safety classifiers refuse before any output are answered by Claude Opus 5 through server-side fallback and billed at Opus 5's price. Claude Fable 5.2 and Mythos 5.2 on other providers, such as OpenRouter's `anthropic/claude-fable-5.2`, get the same per-message effort support
|
|
217
|
+
- Added `GitHubCopilotStreamAttempt`, which records the request ID, duration and progress of one Copilot streaming request, and its `GitHubCopilotStreamSnapshot` result type, which `classifyGitHubCopilotFailure()` takes as `stream` ([#4626](https://github.com/bastani-inc/atomic/issues/4626))
|
|
218
|
+
|
|
219
|
+
### Fixed
|
|
220
|
+
|
|
221
|
+
- GitHub Copilot now filters a stored API key by its account's model list as it does an OAuth login. `ApiKeyCredential` gains the optional `availableModelIds` and `fastModelIds` that a model refresh saves from the account's list, `RefreshModelsContext.updateCredential` is offered for a stored API key too and now takes and returns a `Credential` (it writes only while the store still holds the key or token the refresh used), and `refreshGitHubCopilotAccountModelLists()` renews just those lists for a provider alias. Nothing is hidden for a key whose list was not read or names no usable model ([#4627](https://github.com/bastani-inc/atomic/issues/4627))
|
|
222
|
+
- Catalog providers can set `Provider.refreshModelsAuth: "configured"` to refresh without renewing OAuth or receiving a provider credential, while still skipping unconfigured providers. Stale logins no longer block credential-free catalogs ([#4610](https://github.com/bastani-inc/atomic/issues/4610))
|
|
223
|
+
- Azure OpenAI models now keep OpenAI's long-context price tier, so requests above 272K input tokens are costed at Azure's long-context rates instead of the standard rates (upstream [#11948](https://github.com/earendil-works/pi/issues/11948))
|
|
224
|
+
- Fixed Anthropic and Bedrock sending `strict: "prefer"` tools strict again on every request after Claude rejected them for exceeding its strict tool limits, which cost a rejected request (400) per turn. Later requests with the same `sessionId` now send those tools non-strict right away, until `cleanupSessionResources(sessionId)` (upstream [#11626](https://github.com/earendil-works/pi/issues/11626))
|
|
225
|
+
- Fixed `outputFormat: "svg"` on GPT Image 2.6 failing with a generic `api-error` (HTTP 400, `missing_required_parameter` for `svg.model`) on accounts where the API requires the SVG authoring model. A request without `svg` now retries once with a default authoring model, `DEFAULT_SVG_AUTHOR_MODEL` (`gpt-6.1-sol`, a value that worked in testing: OpenAI documents no default), and retries once more without `svg` if the API rejects it as unsupported, so a call sends at most three requests. An `svg` you pass is always sent as given and never retried or replaced. `AssistantImages.svgModel` reports the model that wrote the SVG. A rejected `svg` or `svg.model` now fails with `ImageGenerationConfigError` (`code: "invalid-option"`, `field` `svg.model` or `svg`, and the new `reason` values `required-by-api`, `rejected-by-api` and `unsupported-by-api`) instead of `api-error`. Image requests pass their 10-minute default timeout to the SDK explicitly, so SVG authoring that takes tens of seconds never times out early ([#4615](https://github.com/bastani-inc/atomic/issues/4615))
|
|
226
|
+
- Fixed GitHub Copilot `stream-ended-early` failures giving almost nothing to diagnose them with. `GitHubCopilotStreamEndedEarlyError` now carries `durationMs` (how long the attempt ran) and `startedContent` (whether any stream event had arrived), and takes its `requestId` from the response's `x-request-id`, on Anthropic Messages, Chat Completions and Responses streams. The fields appear in `fields()`, survive serialization, and show in the message (`Attempt ran 312.4 s, response started, request ID …`). The hint no longer says Atomic retries "when retries are enabled", which was wrong once the retries were used up; it now suggests splitting the work, lowering the thinking effort or adding a fallback model. `GitHubCopilotFailureContext.streamStarted` is replaced by `stream`, and `recordGitHubCopilotFailure()` takes a `GitHubCopilotStreamSnapshot` (or `undefined` when no response arrived) instead of a boolean. `GitHubCopilotErrorFieldValue` now includes `boolean` ([#4626](https://github.com/bastani-inc/atomic/issues/4626))
|
|
227
|
+
- Fixed Anthropic `ping` events not counting as activity for the stream idle deadline (`streamDeadlineMs`), so a model thinking silently while the server kept the connection alive was cut at the deadline as if the stream had stalled. `withStreamDeadline()` takes an optional keep-alive predicate: a match restarts the idle timer and is not yielded, so pings never reach `onProviderStreamEvent` ([#4626](https://github.com/bastani-inc/atomic/issues/4626))
|
|
228
|
+
|
|
229
|
+
## [0.9.72] - 2026-11-08
|
|
230
|
+
|
|
231
|
+
GPT Image 2.6, typed Copilot failures, and thinking `off` that every provider honors.
|
|
232
|
+
|
|
233
|
+
- **GPT Image 2.6.** `gpt-image-2.6-sunburst` and `gpt-image-2.6-flare` generate animated GIF and SVG output and markup-guided edits, and invalid requests fail locally with a typed error naming the field ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
234
|
+
- **Typed GitHub Copilot failures.** Every failed Copilot request carries a typed error with a stable code, so retries, fallback and compaction no longer depend on message text; Gemini on Copilot keeps its reasoning and counts its reasoning tokens ([#4324](https://github.com/bastani-inc/atomic/issues/4324)).
|
|
235
|
+
- **Thinking off at the lowest level.** Models that cannot turn thinking off run at their lowest supported level, with a `thinking_off_fallback` diagnostic, and Azure and Gemini models offer only levels they accept ([#4545](https://github.com/bastani-inc/atomic/issues/4545), [#4552](https://github.com/bastani-inc/atomic/issues/4552), [#4556](https://github.com/bastani-inc/atomic/issues/4556)).
|
|
236
|
+
- **Upstream pi fixes.** Connection errors keep their cause, Responses requests send `session-id`, Azure GPT-6 models get their 1,050,000-token windows, and new compat flags cover stricter proxies (upstream [#11783](https://github.com/earendil-works/pi/pull/11783), [#11793](https://github.com/earendil-works/pi/pull/11793), [#11781](https://github.com/earendil-works/pi/pull/11781), [#9442](https://github.com/earendil-works/pi/pull/9442)).
|
|
237
|
+
|
|
238
|
+
**Breaking change:** If a proxy requires `tools` once the history has tool calls, such as older LiteLLM in front of Anthropic, set `compat.requiresToolsWithToolHistory: true` on that model.
|
|
239
|
+
|
|
240
|
+
### Breaking Changes
|
|
241
|
+
|
|
242
|
+
- `openai-completions` no longer sends `"tools": []` when a request has no tools but the conversation contains tool calls or tool results; the field is omitted. This fixes 400 errors (`tools must not be an empty array`) from vLLM 0.20+, also behind LiteLLM. Proxies that require `tools` once the history has tool calls, such as older LiteLLM versions in front of Anthropic, now reject these requests; set `compat.requiresToolsWithToolHistory: true` on those models to send the empty array again (upstream [#11891](https://github.com/earendil-works/pi/pull/11891))
|
|
243
|
+
|
|
244
|
+
### Added
|
|
245
|
+
|
|
246
|
+
- Added the GPT Image 2.6 models `gpt-image-2.6-sunburst` and `gpt-image-2.6-flare` to the `openai-images` provider, priced at $5 text input, $1.25 cached input and $30 image output per million tokens ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
247
|
+
- `OpenAIImagesOptions` supports GPT Image 2.6: `outputFormat` (`png`, `jpeg`, `webp`, `svg`, `gif`) with `outputCompression`, animated GIFs with `frameCount` (2-16) and `frameDurationMs`, SVG output with the authoring `svg` model, `complexity`, `moderation` and `safetyIdentifier`. Outputs carry `image/gif` and `image/svg+xml` MIME types. `OPENAI_IMAGES_LIMITS` exports the documented bounds. A live request with 17 GIF frames was rejected, so the maximum is 16, as the API reference says, not the guide's 64 ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
248
|
+
- Image blocks in an image request accept `role: "markup"` (a transparent annotation layer for markup-guided edits) or `role: "mask"`. The first image without a role is the base image and the others are references ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
249
|
+
- `ImageModel.capabilities` declares the optional request features a model supports (`complexity`, `markup`, `svg`, `gif`). The GPT Image 2.6 models declare all four and the 2.5 models declare `markup`.
|
|
250
|
+
- Requests that cannot succeed fail before anything is sent with an `ImageGenerationConfigError` naming the offending `field`: values out of range, GIF or SVG on an edit, frame options without GIF, SVG with `n` above 1, compression with `png`, `gif` or `svg`, markup with a mask, and features the model does not declare. The checks also cover the request an `onPayload` hook returns or changes in place. Each failure has a `reason` (`out-of-range`, `requires-option`, `conflicts-with-option`, `generation-only`, `unsupported-by-model`, `requires-base-image`, `duplicate-role`), documented in the README ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
251
|
+
- Typed image-generation errors: `ImageGenerationError` and its subclasses for configuration, API, quota, rate-limit, server, user (`image_generation_user_error`), moderation (`moderation_blocked`, with `moderationStage` and `categories`) and timeout failures, each with a stable `code`, `retryable`, a `hint`, guards that check the `name` and `code` of the value itself, `findImageGenerationError(error, guard)` to search a `cause` chain, and `imageGenerationErrorFromData()` to rebuild one, with its typed cause chain, from `AssistantImages.error` after it crossed a process boundary. Typed errors carry the HTTP status and the provider's type, code and request ID, never the provider's message, which can quote the prompt. Cause chains retain up to eight links, including foreign wrappers and known errno codes; foreign messages are replaced with `[foreign error message omitted]` ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
252
|
+
- `github-copilot` request failures are typed. Every failed Copilot request on Chat Completions, Responses or Messages carries a `GitHubCopilotError` subclass in its assistant message's `diagnostics`, read back with `getGitHubCopilotError(message)`. Each has a stable `code`, `retryable`, a `hint`, structured context and bounded causes. Guards, `findGitHubCopilotError()`, `classifyGitHubCopilotFailure()` and serialization helpers are exported. Retry, fallback and compaction decisions use the typed code ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
253
|
+
- `describeGitHubCopilotModel(model)` returns a Copilot model's capabilities from its metadata: API, thinking levels and reasoning replay, tool calls, usage, images and limits ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
254
|
+
- Added the `requiresNonNullAssistantContent` OpenAI Completions compat flag for providers that reject `null` assistant content: replayed assistant messages with no text, such as tool-call-only messages, then send `""` (upstream [#8118](https://github.com/earendil-works/pi/pull/8118))
|
|
255
|
+
- Added the `promptCacheKeyMode` compat option (`auto`, `enabled`, `disabled`) for OpenAI Completions and Responses models. `enabled` sends `prompt_cache_key` to compatible proxies whenever caching is active, independently of long cache retention; `disabled` never sends it, for proxies that reject it. `auto` keeps each API's previous behavior, and `cacheRetention: "none"` still sends no key (upstream [#9442](https://github.com/earendil-works/pi/pull/9442))
|
|
256
|
+
- Added the `qwen-token-plan-individual-cn` provider: the Qwen Token Plan Individual catalog on the China endpoint, using `QWEN_TOKEN_PLAN_CN_API_KEY` (upstream [#7989](https://github.com/earendil-works/pi/pull/7989))
|
|
257
|
+
- Added `@bastani/pi-ai/lazy-oauth` (`registerLazyOAuthFlows()`) and `@bastani/pi-ai/lazy-bedrock-provider` (`registerLazyBedrockProvider()`, built on the new `setBedrockProviderLoader()`) for Node bundles that split at dynamic imports: the OAuth login flows and Bedrock load through literal dynamic imports, so the bundler keeps them inside the bundle as chunks that load on first use, instead of through the variable-specifier imports that load pi-ai's own files at runtime. Nothing changes without calling them (upstream [`da95cd79`](https://github.com/earendil-works/pi/commit/da95cd79c6adc4f273dee84b26f2255b4f215658))
|
|
258
|
+
|
|
259
|
+
### Changed
|
|
260
|
+
|
|
261
|
+
- `openai-images` no longer retries a quota-exhausted HTTP 429 when `maxRetries` is set; rate-limit and server errors are still retried ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
262
|
+
- The bundled `github-copilot` catalog picks each model's API from its models.dev SDK metadata (Anthropic SDK: Messages, OpenAI SDK: Responses, otherwise Chat Completions) instead of from model ID prefixes. The generated catalog is unchanged ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
263
|
+
- A discovered `github-copilot` model sends strict tool schemas only when its entry lists `structured_outputs`, and takes its images per request and a resize cap from the entry's `limits.vision` ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
264
|
+
- A `github-copilot` Chat Completions or Responses turn that completes with no text and no tool call now fails with a retryable `empty-response` error instead of finishing empty ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
265
|
+
- The `openai-codex` provider is named "OpenAI Codex" again instead of "OpenAI Codex (legacy)" (upstream [`36142c1c`](https://github.com/earendil-works/pi/commit/36142c1c7eab1d756e977b5f39149a634cc472ac))
|
|
266
|
+
|
|
267
|
+
### Fixed
|
|
268
|
+
|
|
269
|
+
- Image requests reject transparent JPEG, GIF and SVG output locally, and an `onPayload` hook cannot leave a markup or mask image without a base image. Rebuilt API errors preserve explicit `retryable` overrides ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
270
|
+
- `github-copilot` Gemini models keep the opaque `reasoning_opaque` value Copilot streams as a redacted thinking block and send it back on the same model's next request, so multi-turn tool loops keep their reasoning state ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
271
|
+
- Chat Completions usage counts a top-level `reasoning_tokens` as output when the response's `total_tokens` shows it sits outside `completion_tokens`, as on `github-copilot` Gemini models, so their output tokens, totals and cost are no longer understated ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
272
|
+
- Responses stream `error` and `response.failed` events keep the provider's error code on the thrown error ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
273
|
+
- Messages stream errors keep their structured provider type, request ID and retry hints, so a rejected Copilot request is not retried as a cut-off stream. Error guards reject malformed serialized payloads instead of rebuilding them without their context ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
274
|
+
- Fixed Azure GPT-6 Luna, Sol, GPT-6.1 Sol and GPT-6.1 Terra using a 272K context window instead of the 1,050,000 tokens Azure Foundry deploys them with, which made sessions compact far earlier than needed (upstream [#11781](https://github.com/earendil-works/pi/pull/11781))
|
|
275
|
+
- Fixed `qwen-token-plan-individual` missing `glm-5.3`, which `qwen-token-plan` and `qwen-token-plan-cn` already list (upstream [#11785](https://github.com/earendil-works/pi/pull/11785))
|
|
276
|
+
- Fixed Claude Sonnet 5 requests failing with a 400 when a temperature was set: Sonnet 5 rejects non-default `temperature`, like Sonnet 5.5, so it is no longer sent (upstream [#7961](https://github.com/earendil-works/pi/pull/7961))
|
|
277
|
+
- Connection errors keep their underlying cause: a request that fails before the provider answers (connection refused or reset, DNS, TLS or certificate, proxy) now reports, for example, `Connection error.: connect ECONNREFUSED 127.0.0.1:8080` instead of only `Connection error.`. The message gets the deepest error of the `cause` chain (up to 5 levels deep) and its code (upstream [#11783](https://github.com/earendil-works/pi/pull/11783))
|
|
278
|
+
- OpenAI Responses requests send the session-affinity header as `session-id` instead of `session_id`. Proxies such as nginx with `underscores_in_headers off` silently dropped the underscored header, and Envoy can be set to reject it. `x-client-request-id` and `prompt_cache_key` are unchanged, and OpenAI Completions still sends `session_id` (upstream [#11793](https://github.com/earendil-works/pi/pull/11793))
|
|
279
|
+
- A request aborted while its setup runs (credential resolution, lazy module loading) now ends with stop reason `aborted` instead of `error`. `lazyStream(model, setup, signal?)` takes the request's abort signal, and `lazyApi`, providers and `Models` pass it through (upstream [#8635](https://github.com/earendil-works/pi/pull/8635))
|
|
280
|
+
- Fixed auto-retry not firing when a gateway reports a stream that dropped mid-reply as `... was cut short after the reply had begun`; that wording is now a retryable provider error (upstream [#11885](https://github.com/earendil-works/pi/pull/11885))
|
|
281
|
+
- Thinking `off` no longer leaves a thinking model at its provider's default effort. On a model whose `thinkingLevelMap.off` is `null`, Anthropic Messages, OpenAI Responses, Azure OpenAI Responses and OpenAI Codex Responses requests run at the model's lowest supported level, the same request as selecting that level, where Claude Opus 5.5 and Sonnet 5.5 used to get a final `high` effort marker and Claude Fable 5, OpenAI o-series and GPT-5 models, Grok and `github-copilot` models omitted the setting. On the Responses APIs the fallback needs that level's effort listed in the model's thinking map. The generated `azure` catalog now copies the efforts the matching `openai` model lists, except `xhigh` and `max`, so Azure GPT-5.1 and later send `none` for `off` and Azure GPT-5, GPT-5 mini and GPT-5 nano fall back to `minimal`. Azure GPT-5.1 and later now offer `off`; their other levels are unchanged. The reply carries a `thinking_off_fallback` diagnostic with the level and provider value sent, as do Gemini requests that already used the lowest level. `github-copilot` Responses models whose thinking map lists an off effort, such as GPT-6 Luna, are now sent it ([#4545](https://github.com/bastani-inc/atomic/issues/4545))
|
|
282
|
+
- A request without a reasoning level now runs a model that cannot turn thinking off at its lowest supported level on Amazon Bedrock, OpenAI Chat Completions and Mistral, with the same `thinking_off_fallback` diagnostic, instead of leaving the effort to the provider's default or sending a disabled toggle the model does not support. This covers Claude Fable 5 and 5.1 and GPT-6 Astra and GPT-6.1 Sol on Bedrock, and Chat Completions and Mistral models such as OpenRouter `openai/gpt-5.1-codex`. On OpenAI Responses, Chat Completions and Mistral the fallback is the lowest supported level whose effort the thinking map lists, so Azure GPT-5 Pro, GPT-5.2 Pro, GPT-5.4 Pro and GPT-5.3 Codex Spark send their lowest listed effort instead of omitting `reasoning`. OpenRouter translates an effort for the routed model, so OpenRouter models whose map lists no effort, such as `openai/o3`, are sent their lowest supported level. A Chat Completions model whose map denies `off` and lists no effort is no longer sent a disabled thinking toggle. The generated `azure` catalog copies the matching `openai` model's `off` entry, so Azure o-series models and `gpt-realtime-2.1` no longer offer `off`, and a request without a reasoning level falls back to `low` or `minimal` instead of sending `none`. Gemini 2.5 Pro no longer offers `off` on `google` and `google-vertex`, and a budget-thinking Gemini model that cannot turn thinking off is sent its lowest level's thinking budget instead of `0`. Levels other than `off` send the same requests as before ([#4552](https://github.com/bastani-inc/atomic/issues/4552))
|
|
283
|
+
- Azure GPT-5 Pro, GPT-5.2 Pro, GPT-5.4 Pro, GPT-5.3 Codex Spark and o-series (`o1`, `o1-pro`, `o3`, `o3-mini`, `o3-pro`, `o4-mini`) deployments no longer offer thinking levels the model rejects, such as `minimal`. Where the matching `openai` model cannot turn thinking off, the generated `azure` catalog copies its `minimal`, `low` and `medium` masks, so thinking `off`, including `streamSimple` with `reasoning: "off"`, runs at the lowest accepted effort (`high` on GPT-5 Pro, `medium` on GPT-5.2 Pro and GPT-5.4 Pro, `low` on the others) instead of sending `minimal`, and an explicit level below that effort runs at it too. The Azure copy of OpenAI's efforts now takes only `off` and efforts listed under their own level name, so a remapped OpenAI effort cannot change an explicit Azure level. `gemini-pro-latest` is sent a `thinkingLevel` (`LOW`, `MEDIUM` or `HIGH`) for every level instead of `thinkingBudget: -1`, and thinking `off` runs it at `LOW`. Thinking `off` on a budget-thinking Gemini model that cannot turn thinking off uses the request's `thinkingBudgets` for its lowest level instead of the default budget; `GoogleOptions` and `GoogleVertexOptions` take `thinkingBudgets` for this ([#4556](https://github.com/bastani-inc/atomic/issues/4556))
|
|
284
|
+
|
|
285
|
+
## [0.9.72-alpha.1] - 2026-11-08
|
|
286
|
+
|
|
287
|
+
GPT Image 2.6, typed Copilot failures, and thinking `off` that every provider honors.
|
|
288
|
+
|
|
289
|
+
- **GPT Image 2.6.** `gpt-image-2.6-sunburst` and `gpt-image-2.6-flare` generate animated GIF and SVG output and markup-guided edits, and invalid requests fail locally with a typed error naming the field ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
290
|
+
- **Typed GitHub Copilot failures.** Every failed Copilot request carries a typed error with a stable code, so retries, fallback and compaction no longer depend on message text; Gemini on Copilot keeps its reasoning and counts its reasoning tokens ([#4324](https://github.com/bastani-inc/atomic/issues/4324)).
|
|
291
|
+
- **Thinking off at the lowest level.** Models that cannot turn thinking off run at their lowest supported level, with a `thinking_off_fallback` diagnostic, and Azure and Gemini models offer only levels they accept ([#4545](https://github.com/bastani-inc/atomic/issues/4545), [#4552](https://github.com/bastani-inc/atomic/issues/4552), [#4556](https://github.com/bastani-inc/atomic/issues/4556)).
|
|
292
|
+
- **Upstream pi fixes.** Connection errors keep their cause, Responses requests send `session-id`, Azure GPT-6 models get their 1,050,000-token windows, and new compat flags cover stricter proxies (upstream [#11783](https://github.com/earendil-works/pi/pull/11783), [#11793](https://github.com/earendil-works/pi/pull/11793), [#11781](https://github.com/earendil-works/pi/pull/11781), [#9442](https://github.com/earendil-works/pi/pull/9442)).
|
|
293
|
+
|
|
294
|
+
**Breaking change:** If a proxy requires `tools` once the history has tool calls, such as older LiteLLM in front of Anthropic, set `compat.requiresToolsWithToolHistory: true` on that model.
|
|
295
|
+
|
|
296
|
+
### Breaking Changes
|
|
297
|
+
|
|
298
|
+
- `openai-completions` no longer sends `"tools": []` when a request has no tools but the conversation contains tool calls or tool results; the field is omitted. This fixes 400 errors (`tools must not be an empty array`) from vLLM 0.20+, also behind LiteLLM. Proxies that require `tools` once the history has tool calls, such as older LiteLLM versions in front of Anthropic, now reject these requests; set `compat.requiresToolsWithToolHistory: true` on those models to send the empty array again (upstream [#11891](https://github.com/earendil-works/pi/pull/11891))
|
|
299
|
+
|
|
300
|
+
### Added
|
|
301
|
+
|
|
302
|
+
- Added the GPT Image 2.6 models `gpt-image-2.6-sunburst` and `gpt-image-2.6-flare` to the `openai-images` provider, priced at $5 text input, $1.25 cached input and $30 image output per million tokens ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
303
|
+
- `OpenAIImagesOptions` supports GPT Image 2.6: `outputFormat` (`png`, `jpeg`, `webp`, `svg`, `gif`) with `outputCompression`, animated GIFs with `frameCount` (2-16) and `frameDurationMs`, SVG output with the authoring `svg` model, `complexity`, `moderation` and `safetyIdentifier`. Outputs carry `image/gif` and `image/svg+xml` MIME types. `OPENAI_IMAGES_LIMITS` exports the documented bounds. A live request with 17 GIF frames was rejected, so the maximum is 16, as the API reference says, not the guide's 64 ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
304
|
+
- Image blocks in an image request accept `role: "markup"` (a transparent annotation layer for markup-guided edits) or `role: "mask"`. The first image without a role is the base image and the others are references ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
305
|
+
- `ImageModel.capabilities` declares the optional request features a model supports (`complexity`, `markup`, `svg`, `gif`). The GPT Image 2.6 models declare all four and the 2.5 models declare `markup`.
|
|
306
|
+
- Requests that cannot succeed fail before anything is sent with an `ImageGenerationConfigError` naming the offending `field`: values out of range, GIF or SVG on an edit, frame options without GIF, SVG with `n` above 1, compression with `png`, `gif` or `svg`, markup with a mask, and features the model does not declare. The checks also cover the request an `onPayload` hook returns or changes in place. Each failure has a `reason` (`out-of-range`, `requires-option`, `conflicts-with-option`, `generation-only`, `unsupported-by-model`, `requires-base-image`, `duplicate-role`), documented in the README ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
307
|
+
- Typed image-generation errors: `ImageGenerationError` and its subclasses for configuration, API, quota, rate-limit, server, user (`image_generation_user_error`), moderation (`moderation_blocked`, with `moderationStage` and `categories`) and timeout failures, each with a stable `code`, `retryable`, a `hint`, guards that check the `name` and `code` of the value itself, `findImageGenerationError(error, guard)` to search a `cause` chain, and `imageGenerationErrorFromData()` to rebuild one, with its typed cause chain, from `AssistantImages.error` after it crossed a process boundary. Typed errors carry the HTTP status and the provider's type, code and request ID, never the provider's message, which can quote the prompt. Cause chains retain up to eight links, including foreign wrappers and known errno codes; foreign messages are replaced with `[foreign error message omitted]` ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
308
|
+
- `github-copilot` request failures are typed. Every failed Copilot request on Chat Completions, Responses or Messages carries a `GitHubCopilotError` subclass in its assistant message's `diagnostics`, read back with `getGitHubCopilotError(message)`. Each has a stable `code`, `retryable`, a `hint`, structured context and bounded causes. Guards, `findGitHubCopilotError()`, `classifyGitHubCopilotFailure()` and serialization helpers are exported. Retry, fallback and compaction decisions use the typed code ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
309
|
+
- `describeGitHubCopilotModel(model)` returns a Copilot model's capabilities from its metadata: API, thinking levels and reasoning replay, tool calls, usage, images and limits ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
310
|
+
- Added the `requiresNonNullAssistantContent` OpenAI Completions compat flag for providers that reject `null` assistant content: replayed assistant messages with no text, such as tool-call-only messages, then send `""` (upstream [#8118](https://github.com/earendil-works/pi/pull/8118))
|
|
311
|
+
- Added the `promptCacheKeyMode` compat option (`auto`, `enabled`, `disabled`) for OpenAI Completions and Responses models. `enabled` sends `prompt_cache_key` to compatible proxies whenever caching is active, independently of long cache retention; `disabled` never sends it, for proxies that reject it. `auto` keeps each API's previous behavior, and `cacheRetention: "none"` still sends no key (upstream [#9442](https://github.com/earendil-works/pi/pull/9442))
|
|
312
|
+
- Added the `qwen-token-plan-individual-cn` provider: the Qwen Token Plan Individual catalog on the China endpoint, using `QWEN_TOKEN_PLAN_CN_API_KEY` (upstream [#7989](https://github.com/earendil-works/pi/pull/7989))
|
|
313
|
+
- Added `@bastani/pi-ai/lazy-oauth` (`registerLazyOAuthFlows()`) and `@bastani/pi-ai/lazy-bedrock-provider` (`registerLazyBedrockProvider()`, built on the new `setBedrockProviderLoader()`) for Node bundles that split at dynamic imports: the OAuth login flows and Bedrock load through literal dynamic imports, so the bundler keeps them inside the bundle as chunks that load on first use, instead of through the variable-specifier imports that load pi-ai's own files at runtime. Nothing changes without calling them (upstream [`da95cd79`](https://github.com/earendil-works/pi/commit/da95cd79c6adc4f273dee84b26f2255b4f215658))
|
|
314
|
+
|
|
315
|
+
### Changed
|
|
316
|
+
|
|
317
|
+
- `openai-images` no longer retries a quota-exhausted HTTP 429 when `maxRetries` is set; rate-limit and server errors are still retried ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
318
|
+
- The bundled `github-copilot` catalog picks each model's API from its models.dev SDK metadata (Anthropic SDK: Messages, OpenAI SDK: Responses, otherwise Chat Completions) instead of from model ID prefixes. The generated catalog is unchanged ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
319
|
+
- A discovered `github-copilot` model sends strict tool schemas only when its entry lists `structured_outputs`, and takes its images per request and a resize cap from the entry's `limits.vision` ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
320
|
+
- A `github-copilot` Chat Completions or Responses turn that completes with no text and no tool call now fails with a retryable `empty-response` error instead of finishing empty ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
321
|
+
- The `openai-codex` provider is named "OpenAI Codex" again instead of "OpenAI Codex (legacy)" (upstream [`36142c1c`](https://github.com/earendil-works/pi/commit/36142c1c7eab1d756e977b5f39149a634cc472ac))
|
|
322
|
+
|
|
323
|
+
### Fixed
|
|
324
|
+
|
|
325
|
+
- Image requests reject transparent JPEG, GIF and SVG output locally, and an `onPayload` hook cannot leave a markup or mask image without a base image. Rebuilt API errors preserve explicit `retryable` overrides ([#4332](https://github.com/bastani-inc/atomic/issues/4332)).
|
|
326
|
+
- `github-copilot` Gemini models keep the opaque `reasoning_opaque` value Copilot streams as a redacted thinking block and send it back on the same model's next request, so multi-turn tool loops keep their reasoning state ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
327
|
+
- Chat Completions usage counts a top-level `reasoning_tokens` as output when the response's `total_tokens` shows it sits outside `completion_tokens`, as on `github-copilot` Gemini models, so their output tokens, totals and cost are no longer understated ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
328
|
+
- Responses stream `error` and `response.failed` events keep the provider's error code on the thrown error ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
329
|
+
- Messages stream errors keep their structured provider type, request ID and retry hints, so a rejected Copilot request is not retried as a cut-off stream. Error guards reject malformed serialized payloads instead of rebuilding them without their context ([#4324](https://github.com/bastani-inc/atomic/issues/4324))
|
|
330
|
+
- Fixed Azure GPT-6 Luna, Sol, GPT-6.1 Sol and GPT-6.1 Terra using a 272K context window instead of the 1,050,000 tokens Azure Foundry deploys them with, which made sessions compact far earlier than needed (upstream [#11781](https://github.com/earendil-works/pi/pull/11781))
|
|
331
|
+
- Fixed `qwen-token-plan-individual` missing `glm-5.3`, which `qwen-token-plan` and `qwen-token-plan-cn` already list (upstream [#11785](https://github.com/earendil-works/pi/pull/11785))
|
|
332
|
+
- Fixed Claude Sonnet 5 requests failing with a 400 when a temperature was set: Sonnet 5 rejects non-default `temperature`, like Sonnet 5.5, so it is no longer sent (upstream [#7961](https://github.com/earendil-works/pi/pull/7961))
|
|
333
|
+
- Connection errors keep their underlying cause: a request that fails before the provider answers (connection refused or reset, DNS, TLS or certificate, proxy) now reports, for example, `Connection error.: connect ECONNREFUSED 127.0.0.1:8080` instead of only `Connection error.`. The message gets the deepest error of the `cause` chain (up to 5 levels deep) and its code (upstream [#11783](https://github.com/earendil-works/pi/pull/11783))
|
|
334
|
+
- OpenAI Responses requests send the session-affinity header as `session-id` instead of `session_id`. Proxies such as nginx with `underscores_in_headers off` silently dropped the underscored header, and Envoy can be set to reject it. `x-client-request-id` and `prompt_cache_key` are unchanged, and OpenAI Completions still sends `session_id` (upstream [#11793](https://github.com/earendil-works/pi/pull/11793))
|
|
335
|
+
- A request aborted while its setup runs (credential resolution, lazy module loading) now ends with stop reason `aborted` instead of `error`. `lazyStream(model, setup, signal?)` takes the request's abort signal, and `lazyApi`, providers and `Models` pass it through (upstream [#8635](https://github.com/earendil-works/pi/pull/8635))
|
|
336
|
+
- Fixed auto-retry not firing when a gateway reports a stream that dropped mid-reply as `... was cut short after the reply had begun`; that wording is now a retryable provider error (upstream [#11885](https://github.com/earendil-works/pi/pull/11885))
|
|
337
|
+
- Thinking `off` no longer leaves a thinking model at its provider's default effort. On a model whose `thinkingLevelMap.off` is `null`, Anthropic Messages, OpenAI Responses, Azure OpenAI Responses and OpenAI Codex Responses requests run at the model's lowest supported level, the same request as selecting that level, where Claude Opus 5.5 and Sonnet 5.5 used to get a final `high` effort marker and Claude Fable 5, OpenAI o-series and GPT-5 models, Grok and `github-copilot` models omitted the setting. On the Responses APIs the fallback needs that level's effort listed in the model's thinking map. The generated `azure` catalog now copies the efforts the matching `openai` model lists, except `xhigh` and `max`, so Azure GPT-5.1 and later send `none` for `off` and Azure GPT-5, GPT-5 mini and GPT-5 nano fall back to `minimal`. Azure GPT-5.1 and later now offer `off`; their other levels are unchanged. The reply carries a `thinking_off_fallback` diagnostic with the level and provider value sent, as do Gemini requests that already used the lowest level. `github-copilot` Responses models whose thinking map lists an off effort, such as GPT-6 Luna, are now sent it ([#4545](https://github.com/bastani-inc/atomic/issues/4545))
|
|
338
|
+
- A request without a reasoning level now runs a model that cannot turn thinking off at its lowest supported level on Amazon Bedrock, OpenAI Chat Completions and Mistral, with the same `thinking_off_fallback` diagnostic, instead of leaving the effort to the provider's default or sending a disabled toggle the model does not support. This covers Claude Fable 5 and 5.1 and GPT-6 Astra and GPT-6.1 Sol on Bedrock, and Chat Completions and Mistral models such as OpenRouter `openai/gpt-5.1-codex`. On OpenAI Responses, Chat Completions and Mistral the fallback is the lowest supported level whose effort the thinking map lists, so Azure GPT-5 Pro, GPT-5.2 Pro, GPT-5.4 Pro and GPT-5.3 Codex Spark send their lowest listed effort instead of omitting `reasoning`. OpenRouter translates an effort for the routed model, so OpenRouter models whose map lists no effort, such as `openai/o3`, are sent their lowest supported level. A Chat Completions model whose map denies `off` and lists no effort is no longer sent a disabled thinking toggle. The generated `azure` catalog copies the matching `openai` model's `off` entry, so Azure o-series models and `gpt-realtime-2.1` no longer offer `off`, and a request without a reasoning level falls back to `low` or `minimal` instead of sending `none`. Gemini 2.5 Pro no longer offers `off` on `google` and `google-vertex`, and a budget-thinking Gemini model that cannot turn thinking off is sent its lowest level's thinking budget instead of `0`. Levels other than `off` send the same requests as before ([#4552](https://github.com/bastani-inc/atomic/issues/4552))
|
|
339
|
+
- Azure GPT-5 Pro, GPT-5.2 Pro, GPT-5.4 Pro, GPT-5.3 Codex Spark and o-series (`o1`, `o1-pro`, `o3`, `o3-mini`, `o3-pro`, `o4-mini`) deployments no longer offer thinking levels the model rejects, such as `minimal`. Where the matching `openai` model cannot turn thinking off, the generated `azure` catalog copies its `minimal`, `low` and `medium` masks, so thinking `off`, including `streamSimple` with `reasoning: "off"`, runs at the lowest accepted effort (`high` on GPT-5 Pro, `medium` on GPT-5.2 Pro and GPT-5.4 Pro, `low` on the others) instead of sending `minimal`, and an explicit level below that effort runs at it too. The Azure copy of OpenAI's efforts now takes only `off` and efforts listed under their own level name, so a remapped OpenAI effort cannot change an explicit Azure level. `gemini-pro-latest` is sent a `thinkingLevel` (`LOW`, `MEDIUM` or `HIGH`) for every level instead of `thinkingBudget: -1`, and thinking `off` runs it at `LOW`. Thinking `off` on a budget-thinking Gemini model that cannot turn thinking off uses the request's `thinkingBudgets` for its lowest level instead of the default budget; `GoogleOptions` and `GoogleVertexOptions` take `thinkingBudgets` for this ([#4556](https://github.com/bastani-inc/atomic/issues/4556))
|
|
340
|
+
|
|
341
|
+
## [0.9.71] - 2026-11-03
|
|
342
|
+
|
|
343
|
+
Copilot discovery now preserves readable model names alongside exact request IDs.
|
|
344
|
+
|
|
345
|
+
- **Account model discovery.** OAuth and API-key users get the chat models their account offers, with advertised names and exact IDs, without replacing curated catalog metadata ([#4279](https://github.com/bastani-inc/atomic/pull/4279), [#4303](https://github.com/bastani-inc/atomic/pull/4303)).
|
|
346
|
+
- **Anthropic conversation reuse.** Message threads send only new messages when the server can continue a session, with full-request fallback when its context changes ([#4293](https://github.com/bastani-inc/atomic/pull/4293)).
|
|
347
|
+
- **Tool compatibility.** Plain JSON Schema tools gain coercion and validation through pi-schema while TypeBox tools keep their existing checks ([#4300](https://github.com/bastani-inc/atomic/pull/4300), [#4301](https://github.com/bastani-inc/atomic/pull/4301)).
|
|
348
|
+
|
|
349
|
+
### Added
|
|
350
|
+
|
|
351
|
+
- Added the text-only `github-copilot/microsoft-decision-1` classifier with OAuth and API-key support (upstream [#11679](https://github.com/earendil-works/pi/pull/11679)).
|
|
352
|
+
- `github-copilot` discovers models from the account's `/models` list. `discoverGitHubCopilotModels(catalog, context)` reads the list with the effective OAuth or API-key credential and returns the chat models the catalog lacks, each under its exact advertised ID, with its advertised display name or the ID when no name is provided, and marked `discovered: true`; `parseGitHubCopilotDiscoveredModels(response, catalog)` applies the same rules to a response. It returns `undefined`, not an empty list, when the list cannot be read (no usable credential, or a 400, 401, 403 or 404 response), so a caller keeps the models it found before. Entries with a disabled policy, hidden from the model picker, without tool-call support, or already in the catalog are skipped, so curated entries keep their metadata. Every entry, fast entries included, takes its API, limits, image input and reasoning efforts from its own row, and a Messages entry that advertises adaptive thinking gets `compat.forceAdaptiveThinking`, so the selected effort is sent. A fast entry carries fast route metadata only when its `family` names another model.
|
|
353
|
+
- `Model.discovered` marks a model found in the account's own model list rather than in the built-in catalog.
|
|
354
|
+
- `RefreshModelsContext.updateCredential` lets a provider's network refresh save account data it read, such as the model lists an availability filter checks, to the OAuth login the refresh used. Nothing is written once the store holds a different token, so a renewed login or another account is never changed.
|
|
355
|
+
- Anthropic requests with a `sessionId` (and `cacheRetention` not `"none"`) now use message threads (`message-threads-2026-08-12` beta): the server continues from the previous reply and only the new messages are sent instead of the whole conversation. When anything earlier in the request changed, or the server can't continue the thread, the full request is sent. Token counts and prompt caching are unchanged. On by default only for the `anthropic` provider on `api.anthropic.com`; set `compat.supportsMessageThreads` to change that, for example for gateways that forward the `thread` field. Fast-mode requests keep their `speed`, and changing it starts a new thread; requests made with a caller-supplied `client` never use threads. `cleanupSessionResources(sessionId)` frees a session's thread state (upstream [earendil-works/pi#11545](https://github.com/earendil-works/pi/pull/11545))
|
|
356
|
+
- `StreamOptions.disableContinuation` keeps a request from using or replacing server-side continuation state such as Anthropic message threads. Set it on requests that repeat an earlier one, such as cache warming (upstream [earendil-works/pi#11545](https://github.com/earendil-works/pi/pull/11545))
|
|
357
|
+
|
|
358
|
+
### Changed
|
|
359
|
+
|
|
360
|
+
- `validateToolArguments()` and `validateToolCall()` validate a plain JSON Schema, one TypeBox did not build (an MCP tool's, a JSON round-tripped declaration, a schema from `@earendil-works/pi-schema`'s builders), with `@earendil-works/pi-schema` (`compile(schema, { coerce: true }).parse()`), a new dependency. Such tools get pi-schema's coercion, the superset of TypeBox's `Value.Convert` rules and pi-ai's own: `"1"`/`"0"` to `true`/`false`, `"null"` to `null`, a fractional numeric string to an integer (`"42.1"` to 42), a single value wrapped into an array, and `null` for a required string is now `"null"` instead of `""`. Their validation errors list pi-schema's messages (`- value: missing`, `- limit: expected number, got string`), and the `url` and `json-pointer-uri-fragment` formats are checked. TypeBox-built schemas keep TypeBox's conversion, checks (refinements and `Format.Set()` formats included) and messages, and `Type`, `Static`, `TSchema` and the TypeBox `StringEnum()` stay exported (upstream [`6536f801`](https://github.com/earendil-works/pi/commit/6536f80129a266c3619c044f8487c0279fc39cd9), adapted to keep TypeBox)
|
|
361
|
+
- `Tool.parameters` accepts any JSON Schema object: `Tool<TParameters extends object = TSchema>`. The default is still `TSchema`, so existing `Tool` and `Tool<typeof schema>` types are unchanged (upstream [`6536f801`](https://github.com/earendil-works/pi/commit/6536f80129a266c3619c044f8487c0279fc39cd9))
|
|
362
|
+
- `github-copilot` availability keeps fast-route models for a credential without a fast list, such as an API key, because only the account's own model list adds them. A login that lists its fast IDs offers each listed ID even when it is not a picker entry. A network refresh that reads the account's list also saves its model lists to the OAuth login that read it, so a model the refresh discovers is offered at once, while the login's lists still decide for every model, so models another login discovered are not offered.
|
|
363
|
+
|
|
364
|
+
### Fixed
|
|
365
|
+
|
|
366
|
+
- `github-copilot` models discovered from the account's `/models` list are named by the display name their list entry advertises (for example `Claude Sonnet 4.5`) instead of repeating their ID, and fall back to the ID when the entry has no name. IDs are unchanged, so requests still use the exact advertised ID ([#4303](https://github.com/bastani-inc/atomic/pull/4303)).
|
|
367
|
+
- A plain JSON Schema tool whose schema has a `$ref` that names nothing (a missing definition, an external document), as some MCP servers send, now rejects only arguments at that place (`- extra: unresolved reference #/definitions/Missing`) instead of failing every call of the tool (upstream [`6536f801`](https://github.com/earendil-works/pi/commit/6536f80129a266c3619c044f8487c0279fc39cd9))
|
|
368
|
+
- OpenAI Codex SSE retries no longer leave an `abort` listener on the caller's signal after each completed backoff (upstream [#11708](https://github.com/earendil-works/pi/issues/11708))
|
|
369
|
+
- The default output limit now leaves 15% of the estimated input unused (at least 4096 tokens), so OpenRouter, which counts the input with its own tokenizer and rejects input plus `max_tokens` above the window, no longer rejects models whose output limit is close to the window with a 400 that was then treated as a context overflow. When that margin would leave less than 32768 tokens of output, only the fixed 4096 margin is kept and output is capped at 32768 (upstream [#10497](https://github.com/earendil-works/pi/issues/10497))
|
|
370
|
+
- Anthropic non-strict tool schemas keep the root `$defs` and `definitions`, so property `$ref`s resolve. Before, Claude could send a JSON string where an object was expected, and some Anthropic-compatible gateways rejected the request. Root `anyOf`/`oneOf`/`allOf` are still left out because Anthropic rejects them (upstream [#9557](https://github.com/earendil-works/pi/issues/9557))
|
|
371
|
+
- OpenCode Go and Zen Claude 5.5 models, Opus 5 and Fable 5.1 now use per-message effort and send `thinking.block_binding: drop_block`, like the Anthropic provider. OpenCode Go Claude Haiku 5.5 failed with a 400 (`Invalid signature in thinking block ... bound to a different conversation`) after the tool list, system prompt or earlier history changed, and kept failing on every retry (upstream [#10954](https://github.com/earendil-works/pi/issues/10954))
|
|
372
|
+
|
|
373
|
+
## [0.9.71-alpha.1] - 2026-11-03
|
|
374
|
+
|
|
375
|
+
Copilot discovery now preserves readable model names alongside exact request IDs.
|
|
376
|
+
|
|
377
|
+
- **Account model discovery.** OAuth and API-key users get the chat models their account offers, with advertised names and exact IDs, without replacing curated catalog metadata ([#4279](https://github.com/bastani-inc/atomic/pull/4279), [#4303](https://github.com/bastani-inc/atomic/pull/4303)).
|
|
378
|
+
- **Anthropic conversation reuse.** Message threads send only new messages when the server can continue a session, with full-request fallback when its context changes ([#4293](https://github.com/bastani-inc/atomic/pull/4293)).
|
|
379
|
+
- **Tool compatibility.** Plain JSON Schema tools gain coercion and validation through pi-schema while TypeBox tools keep their existing checks ([#4300](https://github.com/bastani-inc/atomic/pull/4300), [#4301](https://github.com/bastani-inc/atomic/pull/4301)).
|
|
380
|
+
|
|
381
|
+
### Added
|
|
382
|
+
|
|
383
|
+
- Added the text-only `github-copilot/microsoft-decision-1` classifier with OAuth and API-key support (upstream [#11679](https://github.com/earendil-works/pi/pull/11679)).
|
|
384
|
+
- `github-copilot` discovers models from the account's `/models` list. `discoverGitHubCopilotModels(catalog, context)` reads the list with the effective OAuth or API-key credential and returns the chat models the catalog lacks, each under its exact advertised ID, with its advertised display name or the ID when no name is provided, and marked `discovered: true`; `parseGitHubCopilotDiscoveredModels(response, catalog)` applies the same rules to a response. It returns `undefined`, not an empty list, when the list cannot be read (no usable credential, or a 400, 401, 403 or 404 response), so a caller keeps the models it found before. Entries with a disabled policy, hidden from the model picker, without tool-call support, or already in the catalog are skipped, so curated entries keep their metadata. Every entry, fast entries included, takes its API, limits, image input and reasoning efforts from its own row, and a Messages entry that advertises adaptive thinking gets `compat.forceAdaptiveThinking`, so the selected effort is sent. A fast entry carries fast route metadata only when its `family` names another model.
|
|
385
|
+
- `Model.discovered` marks a model found in the account's own model list rather than in the built-in catalog.
|
|
386
|
+
- `RefreshModelsContext.updateCredential` lets a provider's network refresh save account data it read, such as the model lists an availability filter checks, to the OAuth login the refresh used. Nothing is written once the store holds a different token, so a renewed login or another account is never changed.
|
|
387
|
+
- Anthropic requests with a `sessionId` (and `cacheRetention` not `"none"`) now use message threads (`message-threads-2026-08-12` beta): the server continues from the previous reply and only the new messages are sent instead of the whole conversation. When anything earlier in the request changed, or the server can't continue the thread, the full request is sent. Token counts and prompt caching are unchanged. On by default only for the `anthropic` provider on `api.anthropic.com`; set `compat.supportsMessageThreads` to change that, for example for gateways that forward the `thread` field. Fast-mode requests keep their `speed`, and changing it starts a new thread; requests made with a caller-supplied `client` never use threads. `cleanupSessionResources(sessionId)` frees a session's thread state (upstream [earendil-works/pi#11545](https://github.com/earendil-works/pi/pull/11545))
|
|
388
|
+
- `StreamOptions.disableContinuation` keeps a request from using or replacing server-side continuation state such as Anthropic message threads. Set it on requests that repeat an earlier one, such as cache warming (upstream [earendil-works/pi#11545](https://github.com/earendil-works/pi/pull/11545))
|
|
389
|
+
|
|
390
|
+
### Changed
|
|
391
|
+
|
|
392
|
+
- `validateToolArguments()` and `validateToolCall()` validate a plain JSON Schema, one TypeBox did not build (an MCP tool's, a JSON round-tripped declaration, a schema from `@earendil-works/pi-schema`'s builders), with `@earendil-works/pi-schema` (`compile(schema, { coerce: true }).parse()`), a new dependency. Such tools get pi-schema's coercion, the superset of TypeBox's `Value.Convert` rules and pi-ai's own: `"1"`/`"0"` to `true`/`false`, `"null"` to `null`, a fractional numeric string to an integer (`"42.1"` to 42), a single value wrapped into an array, and `null` for a required string is now `"null"` instead of `""`. Their validation errors list pi-schema's messages (`- value: missing`, `- limit: expected number, got string`), and the `url` and `json-pointer-uri-fragment` formats are checked. TypeBox-built schemas keep TypeBox's conversion, checks (refinements and `Format.Set()` formats included) and messages, and `Type`, `Static`, `TSchema` and the TypeBox `StringEnum()` stay exported (upstream [`6536f801`](https://github.com/earendil-works/pi/commit/6536f80129a266c3619c044f8487c0279fc39cd9), adapted to keep TypeBox)
|
|
393
|
+
- `Tool.parameters` accepts any JSON Schema object: `Tool<TParameters extends object = TSchema>`. The default is still `TSchema`, so existing `Tool` and `Tool<typeof schema>` types are unchanged (upstream [`6536f801`](https://github.com/earendil-works/pi/commit/6536f80129a266c3619c044f8487c0279fc39cd9))
|
|
394
|
+
- `github-copilot` availability keeps fast-route models for a credential without a fast list, such as an API key, because only the account's own model list adds them. A login that lists its fast IDs offers each listed ID even when it is not a picker entry. A network refresh that reads the account's list also saves its model lists to the OAuth login that read it, so a model the refresh discovers is offered at once, while the login's lists still decide for every model, so models another login discovered are not offered.
|
|
395
|
+
|
|
396
|
+
### Fixed
|
|
397
|
+
|
|
398
|
+
- `github-copilot` models discovered from the account's `/models` list are named by the display name their list entry advertises (for example `Claude Sonnet 4.5`) instead of repeating their ID, and fall back to the ID when the entry has no name. IDs are unchanged, so requests still use the exact advertised ID ([#4303](https://github.com/bastani-inc/atomic/pull/4303)).
|
|
399
|
+
- A plain JSON Schema tool whose schema has a `$ref` that names nothing (a missing definition, an external document), as some MCP servers send, now rejects only arguments at that place (`- extra: unresolved reference #/definitions/Missing`) instead of failing every call of the tool (upstream [`6536f801`](https://github.com/earendil-works/pi/commit/6536f80129a266c3619c044f8487c0279fc39cd9))
|
|
400
|
+
- OpenAI Codex SSE retries no longer leave an `abort` listener on the caller's signal after each completed backoff (upstream [#11708](https://github.com/earendil-works/pi/issues/11708))
|
|
401
|
+
- The default output limit now leaves 15% of the estimated input unused (at least 4096 tokens), so OpenRouter, which counts the input with its own tokenizer and rejects input plus `max_tokens` above the window, no longer rejects models whose output limit is close to the window with a 400 that was then treated as a context overflow. When that margin would leave less than 32768 tokens of output, only the fixed 4096 margin is kept and output is capped at 32768 (upstream [#10497](https://github.com/earendil-works/pi/issues/10497))
|
|
402
|
+
- Anthropic non-strict tool schemas keep the root `$defs` and `definitions`, so property `$ref`s resolve. Before, Claude could send a JSON string where an object was expected, and some Anthropic-compatible gateways rejected the request. Root `anyOf`/`oneOf`/`allOf` are still left out because Anthropic rejects them (upstream [#9557](https://github.com/earendil-works/pi/issues/9557))
|
|
403
|
+
- OpenCode Go and Zen Claude 5.5 models, Opus 5 and Fable 5.1 now use per-message effort and send `thinking.block_binding: drop_block`, like the Anthropic provider. OpenCode Go Claude Haiku 5.5 failed with a 400 (`Invalid signature in thinking block ... bound to a different conversation`) after the tool list, system prompt or earlier history changed, and kept failing on every retry (upstream [#10954](https://github.com/earendil-works/pi/issues/10954))
|
|
404
|
+
|
|
405
|
+
## [0.9.64] - 2026-11-02
|
|
406
|
+
|
|
407
|
+
### Added
|
|
408
|
+
|
|
409
|
+
- `api/anthropic-assistant-content` exports `anthropicAssistantReplayContent(message, model)`, which returns the content blocks an Anthropic Messages request for `model` carries for one transcript turn, or none when the request omits the turn, using the same transformation and conversion as the request. Callers that repair provider payloads can match transcript turns to payload messages without copying the converter's rules ([#4185](https://github.com/bastani-inc/atomic/issues/4185))
|
|
410
|
+
|
|
411
|
+
### Fixed
|
|
412
|
+
|
|
413
|
+
- The Anthropic Messages converter sends signed `thinking` text back exactly as Anthropic returned it instead of stripping unpaired surrogates from it, which changed the latest assistant message and failed the request with `thinking or redacted_thinking blocks in the latest assistant message cannot be modified` ([#4185](https://github.com/bastani-inc/atomic/issues/4185))
|
|
414
|
+
|
|
415
|
+
## [0.9.64-alpha.1] - 2026-11-02
|
|
416
|
+
|
|
417
|
+
### Added
|
|
418
|
+
|
|
419
|
+
- `api/anthropic-assistant-content` exports `anthropicAssistantReplayContent(message, model)`, which returns the content blocks an Anthropic Messages request for `model` carries for one transcript turn, or none when the request omits the turn, using the same transformation and conversion as the request. Callers that repair provider payloads can match transcript turns to payload messages without copying the converter's rules ([#4185](https://github.com/bastani-inc/atomic/issues/4185))
|
|
420
|
+
|
|
421
|
+
### Fixed
|
|
422
|
+
|
|
423
|
+
- The Anthropic Messages converter sends signed `thinking` text back exactly as Anthropic returned it instead of stripping unpaired surrogates from it, which changed the latest assistant message and failed the request with `thinking or redacted_thinking blocks in the latest assistant message cannot be modified` ([#4185](https://github.com/bastani-inc/atomic/issues/4185))
|
|
424
|
+
|
|
425
|
+
## [0.9.59] - 2026-11-01
|
|
426
|
+
|
|
427
|
+
### Fixed
|
|
428
|
+
|
|
429
|
+
- `withCopilotEnvBaseUrl` accepts an optional `getStoredApiKey` callback so a Copilot token saved as an API key is routed like `COPILOT_GITHUB_TOKEN`, with the saved key taking precedence. It now also routes `getAllModels`, and reads the token on each access so login and logout reroute without rebuilding the provider. ([#4151](https://github.com/bastani-inc/atomic/pull/4151))
|
|
430
|
+
|
|
431
|
+
## [0.9.59-alpha.1] - 2026-11-01
|
|
432
|
+
|
|
433
|
+
### Fixed
|
|
434
|
+
|
|
435
|
+
- `withCopilotEnvBaseUrl` accepts an optional `getStoredApiKey` callback so a Copilot token saved as an API key is routed like `COPILOT_GITHUB_TOKEN`, with the saved key taking precedence. It now also routes `getAllModels`, and reads the token on each access so login and logout reroute without rebuilding the provider. ([#4151](https://github.com/bastani-inc/atomic/pull/4151))
|
|
436
|
+
|
|
437
|
+
## [0.9.54] - 2026-10-29
|
|
438
|
+
|
|
439
|
+
### Fixed
|
|
440
|
+
|
|
441
|
+
- Fixed Zhipu/Z.ai usage-limit errors (code 1308, sent as HTTP 429) being retried, although the limit resets after hours ([#11537](https://github.com/earendil-works/pi/issues/11537)).
|
|
442
|
+
- Fixed OpenRouter's `~anthropic/claude-*-latest` aliases using the Chat Completions API. They now use Anthropic Messages with adaptive thinking ([pi 44f6851b](https://github.com/earendil-works/pi/commit/44f6851bf)).
|
|
443
|
+
|
|
444
|
+
## [0.9.54-alpha.1] - 2026-10-29
|
|
445
|
+
|
|
446
|
+
### Fixed
|
|
447
|
+
|
|
448
|
+
- Fixed Zhipu/Z.ai usage-limit errors (code 1308, sent as HTTP 429) being retried, although the limit resets after hours ([#11537](https://github.com/earendil-works/pi/issues/11537)).
|
|
449
|
+
- Fixed OpenRouter's `~anthropic/claude-*-latest` aliases using the Chat Completions API. They now use Anthropic Messages with adaptive thinking ([pi 44f6851b](https://github.com/earendil-works/pi/commit/44f6851bf)).
|
|
450
|
+
|
|
451
|
+
## [0.9.47] - 2026-10-27
|
|
452
|
+
|
|
453
|
+
### Fixed
|
|
454
|
+
|
|
455
|
+
- LiteLLM's pre-call `ContextWindowExceededError` ("Context Window exceeded for given call") is now recognized as a context overflow, so the session compacts and retries instead of failing the turn with the 400 ([pi #11492](https://github.com/earendil-works/pi/issues/11492))
|
|
456
|
+
|
|
457
|
+
## [0.9.47-alpha.1] - 2026-10-27
|
|
458
|
+
|
|
459
|
+
### Fixed
|
|
460
|
+
|
|
461
|
+
- LiteLLM's pre-call `ContextWindowExceededError` ("Context Window exceeded for given call") is now recognized as a context overflow, so the session compacts and retries instead of failing the turn with the 400 ([pi #11492](https://github.com/earendil-works/pi/issues/11492))
|
|
462
|
+
|
|
463
|
+
## [0.9.46] - 2026-10-27
|
|
464
|
+
|
|
465
|
+
### Added
|
|
466
|
+
|
|
467
|
+
- Added GPT-6.1 Terra to the OpenAI, Azure OpenAI Responses, and OpenAI Codex providers with official pricing ($1 input, $0.05 cached input, $1.25 cache writes, $5 output, long-context tier above 272k), Fast at twice those rates on the OpenAI API, tool search, and mid-conversation effort changes. Unlike GPT-6.1 Sol it supports `off` (reasoning effort `none`); it rejects `minimal`. Codex serves it with a 500K context window ([pi 363ad78b](https://github.com/earendil-works/pi/commit/363ad78bf0625e1dfbe848429782f5e07bb1584e))
|
|
468
|
+
- `OAuthAuth.needsRefresh()` (optional) reports that an unexpired stored credential lacks data only a refresh restores; `Models.refresh()` with network access then refreshes it once for providers without a model source ([#3998](https://github.com/bastani-inc/atomic/issues/3998))
|
|
469
|
+
|
|
470
|
+
### Fixed
|
|
471
|
+
|
|
472
|
+
- GitHub Copilot offers fewer models its account can't use: a model enabled during login is offered only if the model list re-read afterwards marks it usable, and an OAuth login without a stored model list re-reads the list on the next `Models.refresh()` with network access (until then it keeps offering the catalog) ([#3998](https://github.com/bastani-inc/atomic/issues/3998))
|
|
473
|
+
- Anthropic Messages, Azure OpenAI Responses, Mistral and pi-messages failures now report the server retry hint (`Retry-After-Ms`, `Retry-After` in seconds or as an HTTP date, or a body `retry_after`) as `AssistantMessage.retryAfterMs`, as the OpenAI providers already did. A hint longer than `maxRetryDelayMs` stays on the failure that rejects it, and HTTP 529 overload responses are retryable like 503 ([#3984](https://github.com/bastani-inc/atomic/issues/3984))
|
|
474
|
+
- OAuth logins accept only `http:` and `https:` URLs from servers. Radius now also checks its browser authorization endpoint and device verification URI, and GitHub Copilot, Kimi Coding and Meta share one check ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
475
|
+
- pi-messages streams now fail on a `contentIndex` that is not the next or an existing content block, such as `__proto__` or a huge index ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
476
|
+
- The `pi-ai` CLI writes its `auth.json` with mode `0600`, so other users can no longer read the stored tokens; an existing file is tightened before new tokens are written ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
477
|
+
- OpenCode DeepSeek V4.1 Flash no longer sends `prompt_cache_retention` with long cache retention, which the OpenCode gateway rejects ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
478
|
+
- Google Generative AI, Google Vertex and Amazon Bedrock failures now report the server's retry hint as `retryAfterMs`: Google's `RetryInfo.retryDelay` from the error body, and Bedrock's `Retry-After` / `Retry-After-Ms` response headers. A Google request retried by the shared retry policy now waits that delay instead of a guessed backoff, and a delay above the retry-delay cap fails at once with the hint attached.
|
|
479
|
+
- ChatGPT (Codex) HTTP failures other than usage limits now keep the HTTP status in their message (`402: …`), so an exhausted quota or server error is still recognizable from the message.
|
|
480
|
+
- Amazon Bedrock GPT-6.1 Sol (`openai.gpt-6.1-sol`, `us.openai.gpt-6.1-sol`, `global.openai.gpt-6.1-sol`) now offers its published reasoning efforts `low` through `max`, including `xhigh` and `max`, and no longer offers `off` or `minimal`, which the model does not support ([OpenAI](https://developers.openai.com/api/docs/models/gpt-6.1-sol)). Bedrock GPT-6 Sol and GPT-6 Luna routes now offer `xhigh` and `max` and no longer offer `minimal`, matching their AWS model cards ([GPT-6 Sol](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-sol.html), [GPT-6 Luna](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-luna.html))
|
|
481
|
+
- Amazon Bedrock OpenAI GPT models with reasoning `off` now send the model's mapped `off` effort, as the OpenAI Responses provider does, so Bedrock GPT-6 Sol and GPT-6 Luna run with `reasoning.effort: "none"` instead of silently using their default `medium`. Models whose map gives `off` no value, such as GPT-6 Astra and GPT-6.1 Sol, still send no reasoning field.
|
|
482
|
+
- OpenCode Zen and OpenCode Go requests use the documented protocol-specific endpoints (`/openai/v1`, `/anthropic`, `/google/v1beta`, and `/systemone/v1` for Jev classifiers) instead of the shared `/inference/v1` base ([pi #11132](https://github.com/earendil-works/pi/issues/11132))
|
|
483
|
+
- GPT-5.4 and GPT-5.5 no longer fail with `Missing namespace for function_call` after switching from another model that called tools. On the OpenAI, OpenCode and GitHub Copilot providers they now declare their tools in the top-level tool list instead of loading them through `additional_tools` or tool search ([pi #11415](https://github.com/earendil-works/pi/issues/11415))
|
|
484
|
+
- Bedrock requests no longer fail when `strict: "prefer"` tools exceed Claude's strict tool limits, such as more than 16 optional parameters across all strict tools. They retry once with preferred tools non-strict and keep `strict: "require"` tools strict, as the Anthropic provider already does ([pi #11416](https://github.com/earendil-works/pi/issues/11416))
|
|
485
|
+
|
|
486
|
+
## [0.9.46-alpha.1] - 2026-10-27
|
|
487
|
+
|
|
488
|
+
### Added
|
|
489
|
+
|
|
490
|
+
- Added GPT-6.1 Terra to the OpenAI, Azure OpenAI Responses, and OpenAI Codex providers with official pricing ($1 input, $0.05 cached input, $1.25 cache writes, $5 output, long-context tier above 272k), Fast at twice those rates on the OpenAI API, tool search, and mid-conversation effort changes. Unlike GPT-6.1 Sol it supports `off` (reasoning effort `none`); it rejects `minimal`. Codex serves it with a 500K context window ([pi 363ad78b](https://github.com/earendil-works/pi/commit/363ad78bf0625e1dfbe848429782f5e07bb1584e))
|
|
491
|
+
- `OAuthAuth.needsRefresh()` (optional) reports that an unexpired stored credential lacks data only a refresh restores; `Models.refresh()` with network access then refreshes it once for providers without a model source ([#3998](https://github.com/bastani-inc/atomic/issues/3998))
|
|
492
|
+
|
|
493
|
+
### Fixed
|
|
494
|
+
|
|
495
|
+
- GitHub Copilot offers fewer models its account can't use: a model enabled during login is offered only if the model list re-read afterwards marks it usable, and an OAuth login without a stored model list re-reads the list on the next `Models.refresh()` with network access (until then it keeps offering the catalog) ([#3998](https://github.com/bastani-inc/atomic/issues/3998))
|
|
496
|
+
- Anthropic Messages, Azure OpenAI Responses, Mistral and pi-messages failures now report the server retry hint (`Retry-After-Ms`, `Retry-After` in seconds or as an HTTP date, or a body `retry_after`) as `AssistantMessage.retryAfterMs`, as the OpenAI providers already did. A hint longer than `maxRetryDelayMs` stays on the failure that rejects it, and HTTP 529 overload responses are retryable like 503 ([#3984](https://github.com/bastani-inc/atomic/issues/3984))
|
|
497
|
+
- OAuth logins accept only `http:` and `https:` URLs from servers. Radius now also checks its browser authorization endpoint and device verification URI, and GitHub Copilot, Kimi Coding and Meta share one check ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
498
|
+
- pi-messages streams now fail on a `contentIndex` that is not the next or an existing content block, such as `__proto__` or a huge index ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
499
|
+
- The `pi-ai` CLI writes its `auth.json` with mode `0600`, so other users can no longer read the stored tokens; an existing file is tightened before new tokens are written ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
500
|
+
- OpenCode DeepSeek V4.1 Flash no longer sends `prompt_cache_retention` with long cache retention, which the OpenCode gateway rejects ([pi 89c921d8](https://github.com/earendil-works/pi/commit/89c921d8b56b59e768b26e7f48b6832978119310))
|
|
501
|
+
- Google Generative AI, Google Vertex and Amazon Bedrock failures now report the server's retry hint as `retryAfterMs`: Google's `RetryInfo.retryDelay` from the error body, and Bedrock's `Retry-After` / `Retry-After-Ms` response headers. A Google request retried by the shared retry policy now waits that delay instead of a guessed backoff, and a delay above the retry-delay cap fails at once with the hint attached.
|
|
502
|
+
- ChatGPT (Codex) HTTP failures other than usage limits now keep the HTTP status in their message (`402: …`), so an exhausted quota or server error is still recognizable from the message.
|
|
503
|
+
- Amazon Bedrock GPT-6.1 Sol (`openai.gpt-6.1-sol`, `us.openai.gpt-6.1-sol`, `global.openai.gpt-6.1-sol`) now offers its published reasoning efforts `low` through `max`, including `xhigh` and `max`, and no longer offers `off` or `minimal`, which the model does not support ([OpenAI](https://developers.openai.com/api/docs/models/gpt-6.1-sol)). Bedrock GPT-6 Sol and GPT-6 Luna routes now offer `xhigh` and `max` and no longer offer `minimal`, matching their AWS model cards ([GPT-6 Sol](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-sol.html), [GPT-6 Luna](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-luna.html))
|
|
504
|
+
- Amazon Bedrock OpenAI GPT models with reasoning `off` now send the model's mapped `off` effort, as the OpenAI Responses provider does, so Bedrock GPT-6 Sol and GPT-6 Luna run with `reasoning.effort: "none"` instead of silently using their default `medium`. Models whose map gives `off` no value, such as GPT-6 Astra and GPT-6.1 Sol, still send no reasoning field.
|
|
505
|
+
|
|
506
|
+
## [0.9.44] - 2026-10-25
|
|
507
|
+
|
|
508
|
+
### Breaking Changes
|
|
509
|
+
|
|
510
|
+
- `appendGrammarToolInputJsonDelta()` (`api/constrained-sampling`) takes the input delta instead of the whole input so far and no longer closes the input; close it with the new `closeGrammarToolInputJson()` ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
511
|
+
|
|
512
|
+
### Added
|
|
513
|
+
|
|
514
|
+
- Assistant messages from the `pi-messages` API now carry `responseModel`, the concrete model the server reports on its terminal `done` or `error` event, so clients of virtual models can see which model answered ([pi #11265](https://github.com/earendil-works/pi/pull/11265))
|
|
515
|
+
- Added `StreamingJsonParser`, an incremental, tolerant parser for JSON text that arrives in pieces (each piece is read once; snapshots of the partial value never change once returned), and `ToolCallArgumentsStream`, which every provider now uses for streamed tool-call arguments; `StreamingJsonParser.append()` also returns the changes each piece made (`JsonChange`), for writers that keep their own copy ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
516
|
+
|
|
517
|
+
### Changed
|
|
518
|
+
|
|
519
|
+
- The faux provider streams partial tool-call arguments, so `toolcall_delta` partials show growing arguments as real providers do ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
520
|
+
|
|
521
|
+
### Fixed
|
|
522
|
+
|
|
523
|
+
- Claude models with mid-conversation effort no longer fail with `Invalid effort level` on accounts whose organization caps effort below `high`, even at `low` or `medium` thinking: the fixed top-level effort is now `low`, and the per-message effort markers still set the level for each reply ([pi #10063](https://github.com/earendil-works/pi/issues/10063))
|
|
524
|
+
- The faux provider no longer counts newly cached prompt tokens both as `input` and as `cacheWrite` when it simulates prompt caching for a `sessionId`. Every prompt token now goes to either `cacheRead` or `cacheWrite`, and `input` is 0 ([pi f3fa8bd9](https://github.com/earendil-works/pi/commit/f3fa8bd92430a2bf443681027a01c58586f35464))
|
|
525
|
+
- Gemini tool schemas sent through `parametersJsonSchema` now express string and number `const` values as single-value `enum`s, which Gemini honors (it ignores `const`). Only schema-bearing keywords are rewritten, so literal data under `default`, `examples`, `enum`, `const` and custom keywords, and property names such as `__proto__`, are sent unchanged ([pi 7314819f](https://github.com/earendil-works/pi/commit/7314819f0fe30fa616f32fa3a1fc8d1a856a1446), [pi #11308](https://github.com/earendil-works/pi/issues/11308))
|
|
526
|
+
- Kimi K3.1 Pro and K3.1 Pro Highspeed on Kimi Coding, Moonshot, and OpenCode Go no longer use Kimi K2 request settings. They now send `reasoning_effort` (low/high/max), return earlier reasoning, accept tools added mid-conversation, and report Moonshot rates on Kimi Coding; OpenRouter and Vercel AI Gateway clamp their output limit like Kimi K3. Kimi Coding `k3-256k` now reports Kimi K3 rates ([pi 1de2d7ad](https://github.com/earendil-works/pi/commit/1de2d7adeacfd1b065d9df091997d44c43120ade))
|
|
527
|
+
- Streamed tool-call arguments are no longer re-parsed in full on every delta, which made CPU time and memory grow quadratically with the size of a tool call (a 90 KB `write` on Anthropic peaked at about 140 MB of heap; a 500 KB one took minutes). Every provider and the pi-messages client now read each delta once, and keep their stream state (`partialJson`, `partialArgs`, `index`, `customInput`, `streamIndex`, Bedrock's encrypted reasoning bytes) off the content blocks, so it no longer reaches partial messages or transcripts of failed streams. Grammar (custom) tool input streaming no longer copies the whole input on every delta, and truncated provider error bodies and pi-messages diagnostics no longer keep the whole text they were cut from alive ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
528
|
+
|
|
529
|
+
## [0.9.44-alpha.1] - 2026-10-24
|
|
530
|
+
|
|
531
|
+
### Breaking Changes
|
|
532
|
+
|
|
533
|
+
- `appendGrammarToolInputJsonDelta()` (`api/constrained-sampling`) takes the input delta instead of the whole input so far and no longer closes the input; close it with the new `closeGrammarToolInputJson()` ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
534
|
+
|
|
535
|
+
### Added
|
|
536
|
+
|
|
537
|
+
- Assistant messages from the `pi-messages` API now carry `responseModel`, the concrete model the server reports on its terminal `done` or `error` event, so clients of virtual models can see which model answered ([pi #11265](https://github.com/earendil-works/pi/pull/11265))
|
|
538
|
+
- Added `StreamingJsonParser`, an incremental, tolerant parser for JSON text that arrives in pieces (each piece is read once; snapshots of the partial value never change once returned), and `ToolCallArgumentsStream`, which every provider now uses for streamed tool-call arguments; `StreamingJsonParser.append()` also returns the changes each piece made (`JsonChange`), for writers that keep their own copy ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
539
|
+
|
|
540
|
+
### Changed
|
|
541
|
+
|
|
542
|
+
- The faux provider streams partial tool-call arguments, so `toolcall_delta` partials show growing arguments as real providers do ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
543
|
+
|
|
544
|
+
### Fixed
|
|
545
|
+
|
|
546
|
+
- Claude models with mid-conversation effort no longer fail with `Invalid effort level` on accounts whose organization caps effort below `high`, even at `low` or `medium` thinking: the fixed top-level effort is now `low`, and the per-message effort markers still set the level for each reply ([pi #10063](https://github.com/earendil-works/pi/issues/10063))
|
|
547
|
+
- The faux provider no longer counts newly cached prompt tokens both as `input` and as `cacheWrite` when it simulates prompt caching for a `sessionId`. Every prompt token now goes to either `cacheRead` or `cacheWrite`, and `input` is 0 ([pi f3fa8bd9](https://github.com/earendil-works/pi/commit/f3fa8bd92430a2bf443681027a01c58586f35464))
|
|
548
|
+
- Gemini tool schemas sent through `parametersJsonSchema` now express string and number `const` values as single-value `enum`s, which Gemini honors (it ignores `const`). Only schema-bearing keywords are rewritten, so literal data under `default`, `examples`, `enum`, `const` and custom keywords, and property names such as `__proto__`, are sent unchanged ([pi 7314819f](https://github.com/earendil-works/pi/commit/7314819f0fe30fa616f32fa3a1fc8d1a856a1446), [pi #11308](https://github.com/earendil-works/pi/issues/11308))
|
|
549
|
+
- Kimi K3.1 Pro and K3.1 Pro Highspeed on Kimi Coding, Moonshot, and OpenCode Go no longer use Kimi K2 request settings. They now send `reasoning_effort` (low/high/max), return earlier reasoning, accept tools added mid-conversation, and report Moonshot rates on Kimi Coding; OpenRouter and Vercel AI Gateway clamp their output limit like Kimi K3. Kimi Coding `k3-256k` now reports Kimi K3 rates ([pi 1de2d7ad](https://github.com/earendil-works/pi/commit/1de2d7adeacfd1b065d9df091997d44c43120ade))
|
|
550
|
+
- Streamed tool-call arguments are no longer re-parsed in full on every delta, which made CPU time and memory grow quadratically with the size of a tool call (a 90 KB `write` on Anthropic peaked at about 140 MB of heap; a 500 KB one took minutes). Every provider and the pi-messages client now read each delta once, and keep their stream state (`partialJson`, `partialArgs`, `index`, `customInput`, `streamIndex`, Bedrock's encrypted reasoning bytes) off the content blocks, so it no longer reaches partial messages or transcripts of failed streams. Grammar (custom) tool input streaming no longer copies the whole input on every delta, and truncated provider error bodies and pi-messages diagnostics no longer keep the whole text they were cut from alive ([pi #9265](https://github.com/earendil-works/pi/issues/9265))
|
|
551
|
+
|
|
552
|
+
## [0.9.42] - 2026-10-23
|
|
553
|
+
|
|
554
|
+
### Added
|
|
555
|
+
|
|
556
|
+
- Added `openBrowserOnConfirm` and `copyCodeOnConfirm` to `OAuthDeviceCodeInfo` and the `device_code` `AuthEvent`, so a provider can tell clients whether to offer opening the verification URL and copying the code when the user confirms. Both default to `true` ([pi #9282](https://github.com/earendil-works/pi/issues/9282))
|
|
557
|
+
|
|
558
|
+
### Fixed
|
|
559
|
+
|
|
560
|
+
- OpenRouter requests through Chat Completions and Anthropic Messages now set `usage.cost.total` to the cost OpenRouter reports, including the upstream inference cost for BYOK requests, and mark it `totalFromProvider: true`. When the response has no valid accounting, the catalog estimate stays. The `input`, `output`, `cacheRead` and `cacheWrite` components remain catalog estimates and may not add up to the total ([pi #9980](https://github.com/earendil-works/pi/issues/9980))
|
|
561
|
+
- Claude models that accept mid-conversation system messages no longer reject a request with a 400 (`role 'system' must precede an 'assistant' message`) when a system update is followed by an assistant reply with nothing to send to Anthropic, such as an empty reply or another model's reasoning-only reply ([pi #11091](https://github.com/earendil-works/pi/issues/11091))
|
|
562
|
+
- Cancelling an Anthropic request while it waits for federation authentication now stops that request promptly without cancelling the token exchange shared by other requests ([#3864](https://github.com/bastani-inc/atomic/issues/3864)).
|
|
563
|
+
- Provider request-budget waits no longer consume the Mistral or explicit Codex SSE response-header timeout; the timer now starts once the request is admitted, and cancellation still interrupts the wait ([#3864](https://github.com/bastani-inc/atomic/issues/3864)).
|
|
564
|
+
|
|
565
|
+
## [0.9.42-alpha.1] - 2026-10-23
|
|
566
|
+
|
|
567
|
+
### Added
|
|
568
|
+
|
|
569
|
+
- Added `openBrowserOnConfirm` and `copyCodeOnConfirm` to `OAuthDeviceCodeInfo` and the `device_code` `AuthEvent`, so a provider can tell clients whether to offer opening the verification URL and copying the code when the user confirms. Both default to `true` ([pi #9282](https://github.com/earendil-works/pi/issues/9282))
|
|
570
|
+
|
|
571
|
+
### Fixed
|
|
572
|
+
|
|
573
|
+
- OpenRouter requests through Chat Completions and Anthropic Messages now set `usage.cost.total` to the cost OpenRouter reports, including the upstream inference cost for BYOK requests, and mark it `totalFromProvider: true`. When the response has no valid accounting, the catalog estimate stays. The `input`, `output`, `cacheRead` and `cacheWrite` components remain catalog estimates and may not add up to the total ([pi #9980](https://github.com/earendil-works/pi/issues/9980))
|
|
574
|
+
- Claude models that accept mid-conversation system messages no longer reject a request with a 400 (`role 'system' must precede an 'assistant' message`) when a system update is followed by an assistant reply with nothing to send to Anthropic, such as an empty reply or another model's reasoning-only reply ([pi #11091](https://github.com/earendil-works/pi/issues/11091))
|
|
575
|
+
- Cancelling an Anthropic request while it waits for federation authentication now stops that request promptly without cancelling the token exchange shared by other requests ([#3864](https://github.com/bastani-inc/atomic/issues/3864)).
|
|
576
|
+
- Provider request-budget waits no longer consume the Mistral or explicit Codex SSE response-header timeout; the timer now starts once the request is admitted, and cancellation still interrupts the wait ([#3864](https://github.com/bastani-inc/atomic/issues/3864)).
|
|
577
|
+
|
|
578
|
+
## [0.9.41] - 2026-10-22
|
|
579
|
+
|
|
580
|
+
### Breaking Changes
|
|
581
|
+
|
|
582
|
+
- The default `User-Agent` is now `pi/<version> (<platform>; <runtime>; <arch>)` instead of `pi (<platform> <release>; <arch>)`, and `@bastani/pi-ai/utils/pi-user-agent` (`getPiUserAgent()`) is removed; use `formatUserAgent()`.
|
|
583
|
+
|
|
584
|
+
### Added
|
|
585
|
+
|
|
586
|
+
- Added `Usage.cost.totalFromProvider`. It is `true` when `cost.total` is the billed cost reported by the provider and `false` when the total is estimated from the model's catalog rates, as `calculateCost()` and the image generators now record. A missing value, as in sessions recorded before this field existed, means `false`.
|
|
587
|
+
- Added image input to more System One classifiers: `ClassifierContext.images` now also works with Jev and Clef on `cloudflare-workers-ai` and with OpenRouter decision models whose catalog lists image input, next to Jev on `typesafe`. Each service gets its own wire form: TypeSafe `assets` (PNG and JPEG) with the `TypeSafe-Danger-Zone-Images` opt-in header, Clef `images` data URLs (PNG, JPEG, and WebP, at most four), or OpenRouter `image_url` parts in the state. A text-only model's error now names the model: `System One API does not support image input for <model>`.
|
|
588
|
+
- Added `compat` to System One classifier models (`SystemOneCompat`, with `imageInput` and `sendImageOptInHeader`). Built-in catalog entries set it explicitly; models without it detect the form from the provider, `baseUrl`, and model ID.
|
|
589
|
+
- Added the `@cf/cloudflare/clef-omni` and `typesafe/jev-1.14.0` classifier models to `cloudflare-workers-ai`.
|
|
590
|
+
- Added client identity configuration: an `identity` request option and `createModels({ identity })` default choose the `User-Agent` (`"detailed"` with a version, `"short"`, or custom) and provider attribution headers for OpenRouter, NVIDIA NIM, and OpenCode from `attribution: { name, url, categories? }`. `withClientIdentity()` applies an identity to a `Models` or `MutableModels` a library receives from its host. `formatUserAgent()`, `getIdentityHeaders()`, `getOpenCodeSessionHeaders()`, `mergeClientIdentity()`, `PI_ATTRIBUTION`, and `PI_AI_VERSION` are exported. Attribution matches OpenRouter by its exact `openrouter.ai` host, so a custom endpoint that merely contains that name is not attributed.
|
|
591
|
+
- Bedrock, OpenRouter image generation, pi-messages, and classifier requests now send pi's `User-Agent` too.
|
|
592
|
+
- `Models` requests with a `sessionId` to custom providers pointing at OpenCode (`opencode.ai`) now send the `x-opencode-session` header too, like the built-in OpenCode providers.
|
|
593
|
+
|
|
594
|
+
### Fixed
|
|
595
|
+
|
|
596
|
+
- Fixed Cloudflare Workers AI classifier catalog data: Clef Flash now costs $0.038 per million input tokens with a 24,576-token context window, and Jev costs $0.042 per million input tokens instead of being listed as free.
|
|
597
|
+
- `refreshStoredOAuthCredential()` no longer aborts a token exchange when its caller stops waiting. A token rotation that completes after the 60-second refresh timeout is now saved instead of leaving the replaced refresh token stored. The refresh keeps the credential lock until the exchange settles, for up to `OAUTH_REFRESH_LOCK_HOLD_MS`, so no other refresh presents the token it may already have rotated. When the provider rejects a refresh, it returns a newer credential that another process has saved. A refresh rejected with `invalid_grant` reports only the provider's error code and description, never request URLs, bodies, or tokens, followed by an instruction to run `/login <provider>`. `OAUTH_REFRESH_TIMEOUT_MS` and `OAUTH_REFRESH_LOCK_HOLD_MS` are exported.
|
|
598
|
+
- Anthropic OAuth no longer cuts off a token refresh request after 30 seconds, because the provider may already have rotated the refresh token. The 30-second limit still applies to the login code exchange.
|
|
599
|
+
- Kimi Code OAuth no longer cuts off a token refresh request after 30 seconds. It also no longer sends the same refresh token again after an attempt that may have reached the server (a timeout, an abort, a connection lost mid-request, or an HTTP 429 or 5xx response), because the server may already have rotated that token. A refresh is retried only when the connection was never made. The 30-second limit still applies to device login.
|
|
600
|
+
- A refresh rejected because the stored login is revoked or expired now tells you to run `/login <provider>` for more cases than a literal `invalid_grant`. This includes OpenAI's `refresh_token_expired`, `refresh_token_invalidated` and `refresh_token_reused` codes (also in a nested `error.code`), any HTTP 401 from a token endpoint, Kimi Code's 401 and 403 responses, Meta's expired session, and a revoked GitHub Copilot login. Built-in refreshers throw the new exported `OAuthRefreshRejectedError`, which carries only the HTTP status and the provider's error code and description. Custom refreshers can throw it too, and `refreshStoredOAuthCredential()` maps it to the `/login <provider>` message. The description is dropped when it contains the presented token or a URL. Kimi Code's own refresh errors no longer include the provider's raw error description. `isOAuthRefreshRejectedError`, `isPermanentRefreshRejection`, `parseOAuthErrorBody` and the `OAuthRefreshRejection` and `OAuthErrorDetail` types are exported alongside it.
|
|
601
|
+
- A token rotation that completes after its caller stopped waiting is no longer lost when the process exits. The new `drainPendingOAuthRotations(timeoutMs = OAUTH_SHUTDOWN_DRAIN_MS)` waits, for at most `OAUTH_SHUTDOWN_DRAIN_MS` (3 seconds) by default, until every token exchange that has started settles and its result is saved. It resolves `true` when none is left and `false` when the bound passes first, and never rejects. Hosts call it before a graceful exit. A process killed or crashing mid-exchange still cannot save a rotation it never received; `/login <provider>` signs in again. If a late rotation succeeds but saving it fails, the next refresh failure for that provider adds that an earlier refresh succeeded but its new login could not be saved, with only the save error's code or name (for example `ELOCKED`).
|
|
602
|
+
- Sign in with ChatGPT: a ChatGPT Business workspace usage-limit error (`subscription_sharing_workplace_individual_usage_limit_exceeded`) is no longer classified as a transient error to retry, and its message now tells you to ask your workspace admin to increase the limit.
|
|
603
|
+
- Fixed automatic retry not triggering for `stream disconnected before completion` errors relayed by proxies that wrap the Codex CLI ([#11057](https://github.com/earendil-works/pi/issues/11057))
|
|
604
|
+
- Fixed automatic retry treating digits inside request ids or other numbers (for example `req_011CU503xyz` or `5242880 bytes`) as retryable HTTP status codes. Status codes 429, 500, 502-504 and 520-524 now match only as standalone numbers ([#11057](https://github.com/earendil-works/pi/issues/11057))
|
|
605
|
+
|
|
606
|
+
## [0.9.41-alpha.3] - 2026-10-22
|
|
607
|
+
|
|
608
|
+
### Added
|
|
609
|
+
|
|
610
|
+
- Added `Usage.cost.totalFromProvider`. It is `true` when `cost.total` is the billed cost reported by the provider and `false` when the total is estimated from the model's catalog rates, as `calculateCost()` and the image generators now record. A missing value, as in sessions recorded before this field existed, means `false`.
|
|
611
|
+
|
|
612
|
+
## [0.9.41-alpha.2] - 2026-10-22
|
|
613
|
+
|
|
614
|
+
### Added
|
|
615
|
+
|
|
616
|
+
- Added image input to more System One classifiers: `ClassifierContext.images` now also works with Jev and Clef on `cloudflare-workers-ai` and with OpenRouter decision models whose catalog lists image input, next to Jev on `typesafe`. Each service gets its own wire form: TypeSafe `assets` (PNG and JPEG) with the `TypeSafe-Danger-Zone-Images` opt-in header, Clef `images` data URLs (PNG, JPEG, and WebP, at most four), or OpenRouter `image_url` parts in the state. A text-only model's error now names the model: `System One API does not support image input for <model>`.
|
|
617
|
+
- Added `compat` to System One classifier models (`SystemOneCompat`, with `imageInput` and `sendImageOptInHeader`). Built-in catalog entries set it explicitly; models without it detect the form from the provider, `baseUrl`, and model ID.
|
|
618
|
+
- Added the `@cf/cloudflare/clef-omni` and `typesafe/jev-1.14.0` classifier models to `cloudflare-workers-ai`.
|
|
619
|
+
|
|
620
|
+
### Fixed
|
|
621
|
+
|
|
622
|
+
- Fixed Cloudflare Workers AI classifier catalog data: Clef Flash now costs $0.038 per million input tokens with a 24,576-token context window, and Jev costs $0.042 per million input tokens instead of being listed as free.
|
|
623
|
+
|
|
624
|
+
## [0.9.41-alpha.1] - 2026-10-21
|
|
625
|
+
|
|
626
|
+
### Breaking Changes
|
|
627
|
+
|
|
628
|
+
- The default `User-Agent` is now `pi/<version> (<platform>; <runtime>; <arch>)` instead of `pi (<platform> <release>; <arch>)`, and `@bastani/pi-ai/utils/pi-user-agent` (`getPiUserAgent()`) is removed; use `formatUserAgent()`.
|
|
629
|
+
|
|
630
|
+
### Added
|
|
631
|
+
|
|
632
|
+
- Added client identity configuration: an `identity` request option and `createModels({ identity })` default choose the `User-Agent` (`"detailed"` with a version, `"short"`, or custom) and provider attribution headers for OpenRouter, NVIDIA NIM, and OpenCode from `attribution: { name, url, categories? }`. `withClientIdentity()` applies an identity to a `Models` or `MutableModels` a library receives from its host. `formatUserAgent()`, `getIdentityHeaders()`, `getOpenCodeSessionHeaders()`, `mergeClientIdentity()`, `PI_ATTRIBUTION`, and `PI_AI_VERSION` are exported. Attribution matches OpenRouter by its exact `openrouter.ai` host, so a custom endpoint that merely contains that name is not attributed.
|
|
633
|
+
- Bedrock, OpenRouter image generation, pi-messages, and classifier requests now send pi's `User-Agent` too.
|
|
634
|
+
- `Models` requests with a `sessionId` to custom providers pointing at OpenCode (`opencode.ai`) now send the `x-opencode-session` header too, like the built-in OpenCode providers.
|
|
635
|
+
|
|
636
|
+
### Fixed
|
|
637
|
+
|
|
638
|
+
- `refreshStoredOAuthCredential()` no longer aborts a token exchange when its caller stops waiting. A token rotation that completes after the 60-second refresh timeout is now saved instead of leaving the replaced refresh token stored. The refresh keeps the credential lock until the exchange settles, for up to `OAUTH_REFRESH_LOCK_HOLD_MS`, so no other refresh presents the token it may already have rotated. When the provider rejects a refresh, it returns a newer credential that another process has saved. A refresh rejected with `invalid_grant` reports only the provider's error code and description, never request URLs, bodies, or tokens, followed by an instruction to run `/login <provider>`. `OAUTH_REFRESH_TIMEOUT_MS` and `OAUTH_REFRESH_LOCK_HOLD_MS` are exported.
|
|
639
|
+
- Anthropic OAuth no longer cuts off a token refresh request after 30 seconds, because the provider may already have rotated the refresh token. The 30-second limit still applies to the login code exchange.
|
|
640
|
+
- Kimi Code OAuth no longer cuts off a token refresh request after 30 seconds. It also no longer sends the same refresh token again after an attempt that may have reached the server (a timeout, an abort, a connection lost mid-request, or an HTTP 429 or 5xx response), because the server may already have rotated that token. A refresh is retried only when the connection was never made. The 30-second limit still applies to device login.
|
|
641
|
+
- A refresh rejected because the stored login is revoked or expired now tells you to run `/login <provider>` for more cases than a literal `invalid_grant`. This includes OpenAI's `refresh_token_expired`, `refresh_token_invalidated` and `refresh_token_reused` codes (also in a nested `error.code`), any HTTP 401 from a token endpoint, Kimi Code's 401 and 403 responses, Meta's expired session, and a revoked GitHub Copilot login. Built-in refreshers throw the new exported `OAuthRefreshRejectedError`, which carries only the HTTP status and the provider's error code and description. Custom refreshers can throw it too, and `refreshStoredOAuthCredential()` maps it to the `/login <provider>` message. The description is dropped when it contains the presented token or a URL. Kimi Code's own refresh errors no longer include the provider's raw error description. `isOAuthRefreshRejectedError`, `isPermanentRefreshRejection`, `parseOAuthErrorBody` and the `OAuthRefreshRejection` and `OAuthErrorDetail` types are exported alongside it.
|
|
642
|
+
- A token rotation that completes after its caller stopped waiting is no longer lost when the process exits. The new `drainPendingOAuthRotations(timeoutMs = OAUTH_SHUTDOWN_DRAIN_MS)` waits, for at most `OAUTH_SHUTDOWN_DRAIN_MS` (3 seconds) by default, until every token exchange that has started settles and its result is saved. It resolves `true` when none is left and `false` when the bound passes first, and never rejects. Hosts call it before a graceful exit. A process killed or crashing mid-exchange still cannot save a rotation it never received; `/login <provider>` signs in again. If a late rotation succeeds but saving it fails, the next refresh failure for that provider adds that an earlier refresh succeeded but its new login could not be saved, with only the save error's code or name (for example `ELOCKED`).
|
|
643
|
+
- Sign in with ChatGPT: a ChatGPT Business workspace usage-limit error (`subscription_sharing_workplace_individual_usage_limit_exceeded`) is no longer classified as a transient error to retry, and its message now tells you to ask your workspace admin to increase the limit.
|
|
644
|
+
- Fixed automatic retry not triggering for `stream disconnected before completion` errors relayed by proxies that wrap the Codex CLI ([#11057](https://github.com/earendil-works/pi/issues/11057))
|
|
645
|
+
- Fixed automatic retry treating digits inside request ids or other numbers (for example `req_011CU503xyz` or `5242880 bytes`) as retryable HTTP status codes. Status codes 429, 500, 502-504 and 520-524 now match only as standalone numbers ([#11057](https://github.com/earendil-works/pi/issues/11057))
|
|
646
|
+
|
|
647
|
+
## [0.9.40] - 2026-10-20
|
|
648
|
+
|
|
649
|
+
### Breaking Changes
|
|
650
|
+
|
|
651
|
+
- `resolveTranscriptTools()` now takes a `TranscriptToolSupport` (`"none"`, `"additions"`, or `"changes"`) instead of a boolean, and `TranscriptTools.anchorsAdditions` is replaced by `foldedSystemMessages`, the number of leading system messages folded into `requestTools` ([#10940](https://github.com/earendil-works/pi/issues/10940))
|
|
652
|
+
|
|
653
|
+
### Added
|
|
654
|
+
|
|
655
|
+
- Added `MiniMax-M3.1-Flash` to the `minimax` and `minimax-cn` providers, with thinking levels low through max sent as `output_config.effort`. Thinking cannot be turned off for this model. New MiniMax models from models.dev now appear without a manual allowlist entry; only the older M2, M2.1, M2.5, and M2.5-highspeed models stay hidden ([#11121](https://github.com/earendil-works/pi/issues/11121))
|
|
656
|
+
- Added the `typesafe/jev-1.14.0` classifier for Jev 1.14 alongside `typesafe/jev-latest`.
|
|
657
|
+
- Direct TypeSafe Jev classifiers (`typesafe/jev-latest`, `typesafe/jev-1.14.0`) now send TypeSafe's long-context and image preview headers on every request, accept up to 262K tokens, and accept PNG or JPEG `images`. Images are sent as referenced assets and listed in order under the state's `images` field; other image types, or a state that already has an `images` field, return an error result. Jev on Vercel AI Gateway, OpenCode Zen, OpenRouter, and Cloudflare Workers AI still rejects images.
|
|
658
|
+
|
|
659
|
+
### Changed
|
|
660
|
+
|
|
661
|
+
- Grok 4.3 and Grok 4.5 on the `xai` provider now offer the `xhigh` thinking level, matching the reasoning efforts xAI supports for these models.
|
|
662
|
+
|
|
663
|
+
### Fixed
|
|
664
|
+
|
|
665
|
+
- Treat any HTTP 413 from a provider as request-size overflow, so a GitHub Copilot request over its byte-size limit (`413 ... failed to parse request`) now compacts and retries instead of surfacing a raw API error ([pi 73f8de80](https://github.com/earendil-works/pi/commit/73f8de80bc87021375f8b525292ac07492253002))
|
|
666
|
+
- Fixed OpenAI Responses deferred tool loads rewriting the top-level tool list, and losing the prompt cache, on every load after a tool removal. Models with `additional_tools` now declare all tools in `additional_tools` items, starting with one at the head of the input, so removals (announced as text) and redefinitions no longer rewrite the prefix. Tool search and Kimi tool messages rewrite the top-level tools once per removal or redeclaration and keep later loads anchored ([#10940](https://github.com/earendil-works/pi/issues/10940))
|
|
667
|
+
|
|
668
|
+
### Removed
|
|
669
|
+
|
|
670
|
+
- Removed `gpt-5.5` and `gpt-5.3-codex-spark` from the `openai-codex` provider. The Codex backend rejects both for ChatGPT accounts ("not supported when using Codex with a ChatGPT account"), so selecting them always failed ([#11173](https://github.com/earendil-works/pi/pull/11173))
|
|
671
|
+
|
|
672
|
+
## [0.9.40-alpha.2] - 2026-10-20
|
|
673
|
+
|
|
674
|
+
### Breaking Changes
|
|
675
|
+
|
|
676
|
+
- `resolveTranscriptTools()` now takes a `TranscriptToolSupport` (`"none"`, `"additions"`, or `"changes"`) instead of a boolean, and `TranscriptTools.anchorsAdditions` is replaced by `foldedSystemMessages`, the number of leading system messages folded into `requestTools` ([#10940](https://github.com/earendil-works/pi/issues/10940))
|
|
677
|
+
|
|
678
|
+
### Changed
|
|
679
|
+
|
|
680
|
+
- Grok 4.3 and Grok 4.5 on the `xai` provider now offer the `xhigh` thinking level, matching the reasoning efforts xAI supports for these models.
|
|
681
|
+
|
|
682
|
+
### Fixed
|
|
683
|
+
|
|
684
|
+
- Fixed OpenAI Responses deferred tool loads rewriting the top-level tool list, and losing the prompt cache, on every load after a tool removal. Models with `additional_tools` now declare all tools in `additional_tools` items, starting with one at the head of the input, so removals (announced as text) and redefinitions no longer rewrite the prefix. Tool search and Kimi tool messages rewrite the top-level tools once per removal or redeclaration and keep later loads anchored ([#10940](https://github.com/earendil-works/pi/issues/10940))
|
|
685
|
+
|
|
686
|
+
### Removed
|
|
687
|
+
|
|
688
|
+
- Removed `gpt-5.5` and `gpt-5.3-codex-spark` from the `openai-codex` provider. The Codex backend rejects both for ChatGPT accounts ("not supported when using Codex with a ChatGPT account"), so selecting them always failed ([#11173](https://github.com/earendil-works/pi/pull/11173))
|
|
689
|
+
|
|
690
|
+
## [0.9.40-alpha.1] - 2026-10-20
|
|
691
|
+
|
|
692
|
+
### Added
|
|
693
|
+
|
|
694
|
+
- Added `MiniMax-M3.1-Flash` to the `minimax` and `minimax-cn` providers, with thinking levels low through max sent as `output_config.effort`. Thinking cannot be turned off for this model. New MiniMax models from models.dev now appear without a manual allowlist entry; only the older M2, M2.1, M2.5, and M2.5-highspeed models stay hidden ([#11121](https://github.com/earendil-works/pi/issues/11121))
|
|
695
|
+
- Added the `typesafe/jev-1.14.0` classifier for Jev 1.14 alongside `typesafe/jev-latest`.
|
|
696
|
+
- Direct TypeSafe Jev classifiers (`typesafe/jev-latest`, `typesafe/jev-1.14.0`) now send TypeSafe's long-context and image preview headers on every request, accept up to 262K tokens, and accept PNG or JPEG `images`. Images are sent as referenced assets and listed in order under the state's `images` field; other image types, or a state that already has an `images` field, return an error result. Jev on Vercel AI Gateway, OpenCode Zen, OpenRouter, and Cloudflare Workers AI still rejects images.
|
|
697
|
+
|
|
698
|
+
### Fixed
|
|
699
|
+
|
|
700
|
+
- Treat any HTTP 413 from a provider as request-size overflow, so a GitHub Copilot request over its byte-size limit (`413 ... failed to parse request`) now compacts and retries instead of surfacing a raw API error ([pi 73f8de80](https://github.com/earendil-works/pi/commit/73f8de80bc87021375f8b525292ac07492253002))
|
|
701
|
+
|
|
702
|
+
|
|
703
|
+
## [0.9.38] - 2026-10-18
|
|
704
|
+
|
|
705
|
+
### Breaking Changes
|
|
706
|
+
|
|
707
|
+
- Removed `AssistantMessageFrame`, `AssistantMessageFrameEncoder`, and `reduceAssistantMessageFrames()`. Use `AssistantMessageEventStream` events for streaming progress and `result()` for the final message instead.
|
|
708
|
+
|
|
709
|
+
### Added
|
|
710
|
+
|
|
711
|
+
- Added cache-preserving thinking-level changes for GPT-6 models on `openai` and `openai-api` with API-key authentication, controlled by `compat.supportsMidConvoEffort` ([#10853](https://github.com/earendil-works/pi/issues/10853)).
|
|
712
|
+
|
|
713
|
+
### Changed
|
|
714
|
+
|
|
715
|
+
- Changed `kimi-coding` to OpenAI Chat Completions at `https://api.kimi.com/coding/v1`, matching the official Kimi Code client. K3, K3-256K, and Kimi For Coding support tools added mid-conversation without invalidating the prompt cache. Custom `models.json` base URLs for this provider must include `/v1` ([#10720](https://github.com/earendil-works/pi/issues/10720)).
|
|
716
|
+
|
|
717
|
+
### Fixed
|
|
718
|
+
|
|
719
|
+
- Fixed OpenCode Go Gemini models failing with a 404, and builds failing when the OpenCode Go catalog lists Gemini models, by moving OpenCode Zen and OpenCode Go to the `opencode.ai/inference` endpoints and registering the Google Generative AI API on the `opencode-go` provider, matching OpenCode Zen.
|
|
720
|
+
- Fixed OpenCode Go MiniMax M2.7 failing with "Model does not support this protocol" by using the Anthropic Messages API.
|
|
721
|
+
- Fixed tool-call argument parsing dropping fields when a string contains an invalid `\u` escape such as `\usepackage` ([#10868](https://github.com/earendil-works/pi/issues/10868)).
|
|
722
|
+
- Fixed Anthropic requests failing with a 400 when an error tool result contains images; images are now replaced with a text note ([#10794](https://github.com/earendil-works/pi/issues/10794)).
|
|
723
|
+
- Fixed automatic retry not triggering for gateway `Unable to forward upstream stream` errors ([#10920](https://github.com/earendil-works/pi/issues/10920)).
|
|
724
|
+
- Fixed OpenAI-compatible responses showing thinking after the answer when a proxy such as LiteLLM sends reasoning and content in the same stream chunk ([#10926](https://github.com/earendil-works/pi/issues/10926)).
|
|
725
|
+
- Fixed Gemini 4 Argon requests on Google Vertex failing with "Thinking budget is not supported", and its thinking level being ignored on the Gemini API, by sending `thinkingLevel` for Gemini 4 and later models ([#10890](https://github.com/earendil-works/pi/issues/10890)).
|
|
726
|
+
- Fixed Xiaomi MiMo models failing with a 400 at the `minimal` thinking level by sending only the documented thinking on/off toggle and exposing only the `off` and `high` levels ([#10818](https://github.com/earendil-works/pi/issues/10818)).
|
|
727
|
+
- Fixed tool calls with malformed final JSON arguments silently running with only a parsed prefix. These calls now fail validation with the original arguments so the model can correct them ([#10897](https://github.com/earendil-works/pi/issues/10897)).
|
|
728
|
+
- Fixed `retryAssistantCall()` ignoring server retry hints. A failed response's `Retry-After-Ms`, `Retry-After` (seconds or HTTP date) or Cloudflare body `retry_after` is reported as `AssistantMessage.retryAfterMs` by the OpenAI Responses, OpenAI Completions and ChatGPT (Codex) providers, and replaces the exponential backoff for that retry, capped by the new `RetryPolicy.maxRetryDelayMs` (60 seconds by default). Cloudflare 521-523 responses and connection resets are now retryable ([#3623](https://github.com/bastani-inc/atomic/issues/3623))
|
|
729
|
+
- Fixed streaming tool arguments losing `__proto__` fields and changing object prototypes when JSON is incomplete, including nested objects and arrays ([#3679](https://github.com/bastani-inc/atomic/issues/3679)).
|
|
730
|
+
- Fixed `closeOpenAICodexWebSocketSessions("")` closing other sessions' cached sockets. Empty keys now do nothing; omitting the key still closes all cached sockets.
|
|
731
|
+
- Fixed the root `AssistantMessageEventStream` export being visible only as a type to TypeScript consumers. It can now be imported and constructed as a value, matching the runtime API.
|
|
732
|
+
- Fixed automatic retry not triggering for `The service is temporarily unavailable` errors, which also made compaction fail on the first attempt ([#10965](https://github.com/earendil-works/pi/issues/10965)).
|
|
733
|
+
- Fixed Claude Haiku 5.5 and Claude Opus 5 not offering the `off` thinking level on Anthropic. `off` now sends `thinking: {type: "disabled"}` without effort settings ([#10753](https://github.com/earendil-works/pi/issues/10753)).
|
|
734
|
+
- Fixed images in error tool results being replaced with a note on Anthropic, and Claude on Amazon Bedrock rejecting error tool results that contain images. The images are now sent after the tool results of the turn, so the model still sees them ([#10794](https://github.com/earendil-works/pi/issues/10794)).
|
|
735
|
+
- Fixed model switching exposing another model's thinking as visible reply text. Foreign thinking blocks are now dropped; Anthropic models that adjudicate thinking-block model binding still receive compatible blocks from the same provider ([#10995](https://github.com/earendil-works/pi/issues/10995)).
|
|
736
|
+
- Fixed automatic retry not triggering for node-fetch request failures during Google Vertex authentication token refresh ([#10996](https://github.com/earendil-works/pi/issues/10996)).
|
|
737
|
+
- Fixed whitespace-only tool output (for example a bare `\r\n` from bash) causing OpenAI Completions, OpenAI Responses, and Google providers to reject requests. It now uses the same placeholder as empty output ([#8720](https://github.com/earendil-works/pi/issues/8720)).
|
|
738
|
+
- Fixed automatic retry not triggering for gateway upstream failures such as `upstream_error: Upstream request failed`, `upstream_stream_read_error`, and `Upstream HTTP/2 stream failed` ([#10920](https://github.com/earendil-works/pi/issues/10920)).
|
|
739
|
+
|
|
740
|
+
## [0.9.38-alpha.7] - 2026-10-17
|
|
741
|
+
|
|
742
|
+
### Fixed
|
|
743
|
+
|
|
744
|
+
- Fixed `closeOpenAICodexWebSocketSessions("")` closing other sessions' cached sockets. Empty keys now do nothing; omitting the key still closes all cached sockets.
|
|
745
|
+
- Fixed the root `AssistantMessageEventStream` export being visible only as a type to TypeScript consumers. It can now be imported and constructed as a value, matching the runtime API.
|
|
746
|
+
- Fixed automatic retry not triggering for `The service is temporarily unavailable` errors, which also made compaction fail on the first attempt ([#10965](https://github.com/earendil-works/pi/issues/10965)).
|
|
747
|
+
- Fixed Claude Haiku 5.5 and Claude Opus 5 not offering the `off` thinking level on Anthropic. `off` now sends `thinking: {type: "disabled"}` without effort settings ([#10753](https://github.com/earendil-works/pi/issues/10753)).
|
|
748
|
+
- Fixed images in error tool results being replaced with a note on Anthropic, and Claude on Amazon Bedrock rejecting error tool results that contain images. The images are now sent after the tool results of the turn, so the model still sees them ([#10794](https://github.com/earendil-works/pi/issues/10794)).
|
|
749
|
+
- Fixed model switching exposing another model's thinking as visible reply text. Foreign thinking blocks are now dropped; Anthropic models that adjudicate thinking-block model binding still receive compatible blocks from the same provider ([#10995](https://github.com/earendil-works/pi/issues/10995)).
|
|
750
|
+
- Fixed automatic retry not triggering for node-fetch request failures during Google Vertex authentication token refresh ([#10996](https://github.com/earendil-works/pi/issues/10996)).
|
|
751
|
+
- Fixed whitespace-only tool output (for example a bare `\r\n` from bash) causing OpenAI Completions, OpenAI Responses, and Google providers to reject requests. It now uses the same placeholder as empty output ([#8720](https://github.com/earendil-works/pi/issues/8720)).
|
|
752
|
+
- Fixed automatic retry not triggering for gateway upstream failures such as `upstream_error: Upstream request failed`, `upstream_stream_read_error`, and `Upstream HTTP/2 stream failed` ([#10920](https://github.com/earendil-works/pi/issues/10920)).
|
|
753
|
+
|
|
754
|
+
## [0.9.38-alpha.6] - 2026-10-16
|
|
755
|
+
|
|
756
|
+
### Fixed
|
|
757
|
+
|
|
758
|
+
- Fixed streaming tool arguments losing `__proto__` fields and changing object prototypes when JSON is incomplete, including nested objects and arrays ([#3679](https://github.com/bastani-inc/atomic/issues/3679)).
|
|
759
|
+
|
|
760
|
+
## [0.9.38-alpha.1] - 2026-10-15
|
|
761
|
+
|
|
762
|
+
### Breaking Changes
|
|
763
|
+
|
|
764
|
+
- Removed `AssistantMessageFrame`, `AssistantMessageFrameEncoder`, and `reduceAssistantMessageFrames()`. Use `AssistantMessageEventStream` events for streaming progress and `result()` for the final message instead.
|
|
765
|
+
|
|
766
|
+
### Added
|
|
767
|
+
|
|
768
|
+
- Added cache-preserving thinking-level changes for GPT-6 models on `openai` and `openai-api` with API-key authentication, controlled by `compat.supportsMidConvoEffort` ([#10853](https://github.com/earendil-works/pi/issues/10853)).
|
|
769
|
+
|
|
770
|
+
### Changed
|
|
771
|
+
|
|
772
|
+
- Changed `kimi-coding` to OpenAI Chat Completions at `https://api.kimi.com/coding/v1`, matching the official Kimi Code client. K3, K3-256K, and Kimi For Coding support tools added mid-conversation without invalidating the prompt cache. Custom `models.json` base URLs for this provider must include `/v1` ([#10720](https://github.com/earendil-works/pi/issues/10720)).
|
|
773
|
+
|
|
774
|
+
### Fixed
|
|
775
|
+
|
|
776
|
+
- Fixed OpenCode Go Gemini models failing with a 404, and builds failing when the OpenCode Go catalog lists Gemini models, by moving OpenCode Zen and OpenCode Go to the `opencode.ai/inference` endpoints and registering the Google Generative AI API on the `opencode-go` provider, matching OpenCode Zen.
|
|
777
|
+
- Fixed OpenCode Go MiniMax M2.7 failing with "Model does not support this protocol" by using the Anthropic Messages API.
|
|
778
|
+
- Fixed tool-call argument parsing dropping fields when a string contains an invalid `\u` escape such as `\usepackage` ([#10868](https://github.com/earendil-works/pi/issues/10868)).
|
|
779
|
+
- Fixed Anthropic requests failing with a 400 when an error tool result contains images; images are now replaced with a text note ([#10794](https://github.com/earendil-works/pi/issues/10794)).
|
|
780
|
+
- Fixed automatic retry not triggering for gateway `Unable to forward upstream stream` errors ([#10920](https://github.com/earendil-works/pi/issues/10920)).
|
|
781
|
+
- Fixed OpenAI-compatible responses showing thinking after the answer when a proxy such as LiteLLM sends reasoning and content in the same stream chunk ([#10926](https://github.com/earendil-works/pi/issues/10926)).
|
|
782
|
+
- Fixed Gemini 4 Argon requests on Google Vertex failing with "Thinking budget is not supported", and its thinking level being ignored on the Gemini API, by sending `thinkingLevel` for Gemini 4 and later models ([#10890](https://github.com/earendil-works/pi/issues/10890)).
|
|
783
|
+
- Fixed Xiaomi MiMo models failing with a 400 at the `minimal` thinking level by sending only the documented thinking on/off toggle and exposing only the `off` and `high` levels ([#10818](https://github.com/earendil-works/pi/issues/10818)).
|
|
784
|
+
- Fixed tool calls with malformed final JSON arguments silently running with only a parsed prefix. These calls now fail validation with the original arguments so the model can correct them ([#10897](https://github.com/earendil-works/pi/issues/10897)).
|
|
785
|
+
- Fixed `retryAssistantCall()` ignoring server retry hints. A failed response's `Retry-After-Ms`, `Retry-After` (seconds or HTTP date) or Cloudflare body `retry_after` is reported as `AssistantMessage.retryAfterMs` by the OpenAI Responses, OpenAI Completions and ChatGPT (Codex) providers, and replaces the exponential backoff for that retry, capped by the new `RetryPolicy.maxRetryDelayMs` (60 seconds by default). Cloudflare 521-523 responses and connection resets are now retryable ([#3623](https://github.com/bastani-inc/atomic/issues/3623))
|
|
786
|
+
|
|
787
|
+
## [0.9.36] - 2026-10-13
|
|
788
|
+
|
|
789
|
+
### Fixed
|
|
790
|
+
|
|
791
|
+
- Fixed GPT-6.1 Sol using API context limits with ChatGPT subscription sign-in through `openai`. Subscription model availability now uses a 500,000-token context with no long-context price tier on Standard, Fast, and Ultrafast, including `openai-codex` catalog refreshes. API-key authentication keeps its API limits and pricing.
|
|
792
|
+
|
|
793
|
+
## [0.9.36-alpha.1] - 2026-10-13
|
|
794
|
+
|
|
795
|
+
### Fixed
|
|
796
|
+
|
|
797
|
+
- Fixed GPT-6.1 Sol using API context limits with ChatGPT subscription sign-in through `openai`. Subscription model availability now uses a 500,000-token context with no long-context price tier on Standard, Fast, and Ultrafast, including `openai-codex` catalog refreshes. API-key authentication keeps its API limits and pricing.
|
|
798
|
+
|
|
799
|
+
## [0.9.35] - 2026-10-13
|
|
800
|
+
|
|
801
|
+
### Added
|
|
802
|
+
|
|
803
|
+
- Added `personalizeOpenRouterCatalog()` to restrict OpenRouter chat catalogs to the models a key may use, with its context limits, output limits and prices from `/models/user`. Image and classifier models are preserved.
|
|
804
|
+
- Added `RefreshModelsContext.baseUrl` so catalog refreshes can use the registered provider's configured endpoint.
|
|
805
|
+
|
|
806
|
+
### Changed
|
|
807
|
+
|
|
808
|
+
- Changed the `openai-codex/gpt-6.1-sol` catalog entry to a 500,000-token context window (up from 272,000) priced without a long-context tier on its base cost and its Fast and Ultrafast service tiers, following OpenAI's Codex & Work change for ChatGPT subscriptions. The `openai` GPT-6.1 Sol entry keeps its 272,000-token cap and long-context tier, and the other Codex models are unchanged.
|
|
809
|
+
|
|
810
|
+
### Fixed
|
|
811
|
+
|
|
812
|
+
- Fixed Anthropic requests rejected by strict-tool capacity limits by retrying once with ordinary schemas for tools marked `strict: "prefer"`. Required strict tools remain unchanged.
|
|
813
|
+
- Fixed "The upstream provider returned an error" not being retried as a transient provider error ([#10842](https://github.com/earendil-works/pi/issues/10842)).
|
|
814
|
+
- Personalized OpenRouter catalogs now treat negative placeholder prices, such as those returned for `openrouter/auto`, as zero instead of negative token costs.
|
|
815
|
+
|
|
816
|
+
## [0.9.35-alpha.4] - 2026-10-13
|
|
817
|
+
|
|
818
|
+
### Added
|
|
819
|
+
|
|
820
|
+
- Added `personalizeOpenRouterCatalog()` to restrict OpenRouter chat catalogs to the models a key may use, with its context limits, output limits and prices from `/models/user`. Image and classifier models are preserved.
|
|
821
|
+
- Added `RefreshModelsContext.baseUrl` so catalog refreshes can use the registered provider's configured endpoint.
|
|
822
|
+
|
|
823
|
+
### Fixed
|
|
824
|
+
|
|
825
|
+
- Personalized OpenRouter catalogs now treat negative placeholder prices, such as those returned for `openrouter/auto`, as zero instead of negative token costs.
|
|
826
|
+
|
|
827
|
+
## [0.9.35-alpha.3] - 2026-10-12
|
|
828
|
+
|
|
829
|
+
### Fixed
|
|
830
|
+
|
|
831
|
+
- Fixed Anthropic requests rejected by strict-tool capacity limits by retrying once with ordinary schemas for tools marked `strict: "prefer"`. Required strict tools remain unchanged.
|
|
832
|
+
- Fixed "The upstream provider returned an error" not being retried as a transient provider error ([#10842](https://github.com/earendil-works/pi/issues/10842)).
|
|
833
|
+
|
|
834
|
+
## [0.9.35-alpha.1] - 2026-10-12
|
|
835
|
+
|
|
836
|
+
### Changed
|
|
837
|
+
|
|
838
|
+
- Changed the `openai-codex/gpt-6.1-sol` catalog entry to a 500,000-token context window (up from 272,000) priced without a long-context tier on its base cost and its Fast and Ultrafast service tiers, following OpenAI's Codex & Work change for ChatGPT subscriptions. The `openai` GPT-6.1 Sol entry keeps its 272,000-token cap and long-context tier, and the other Codex models are unchanged.
|
|
839
|
+
|
|
840
|
+
## [0.9.30] - 2026-10-09
|
|
841
|
+
|
|
842
|
+
### Added
|
|
843
|
+
|
|
844
|
+
- Added the `providers/compat-schema` and `providers/model-schema` modules, which export the TypeBox schemas behind the compatibility, cost, input-limit, prompt-cache, and thinking-level types. The existing types, such as `OpenAICompletionsCompat` and `ModelCost`, are now derived from these schemas and remain exported from the package entry points, and `ProviderCompatSchema` combines every API's compatibility fields.
|
|
845
|
+
|
|
846
|
+
### Fixed
|
|
847
|
+
|
|
848
|
+
- Fixed Mistral streams being aborted mid-generation (e.g. during long thinking) after the request timeout; the timeout now applies only to waiting for response headers and to reading an error response, and cancelling a request also stops a stalled error response ([#10609](https://github.com/earendil-works/pi/issues/10609))
|
|
849
|
+
|
|
850
|
+
## [0.9.30-alpha.1] - 2026-10-09
|
|
851
|
+
|
|
852
|
+
### Added
|
|
853
|
+
|
|
854
|
+
- Added the `providers/compat-schema` and `providers/model-schema` modules, which export the TypeBox schemas behind the compatibility, cost, input-limit, prompt-cache, and thinking-level types. The existing types, such as `OpenAICompletionsCompat` and `ModelCost`, are now derived from these schemas and remain exported from the package entry points, and `ProviderCompatSchema` combines every API's compatibility fields.
|
|
855
|
+
|
|
856
|
+
### Fixed
|
|
857
|
+
|
|
858
|
+
- Fixed Mistral streams being aborted mid-generation (e.g. during long thinking) after the request timeout; the timeout now applies only to waiting for response headers and to reading an error response, and cancelling a request also stops a stalled error response ([#10609](https://github.com/earendil-works/pi/issues/10609))
|
|
859
|
+
|
|
860
|
+
## [0.9.29] - 2026-10-09
|
|
861
|
+
|
|
862
|
+
### Added
|
|
863
|
+
|
|
864
|
+
- Added the `ultrafast` service tier to `gpt-6.1-sol` and `gpt-5.6-sol` on `openai` and `openai-api`, and to `gpt-6.1-sol` on `openai-codex`. GPT-6.1 Sol uses OpenAI's published Ultrafast rates ($12 input, $60 output per million tokens), with cache rates derived at the same six-times-Standard ratio; GPT-5.6 Sol Ultrafast is a preview without published rates and is priced at six times its Standard rates ([#3529](https://github.com/bastani-inc/atomic/issues/3529)).
|
|
865
|
+
- Added the `openai-images` provider (`openaiImagesProvider()`) and the `openai-images` image API for OpenAI's Images API, with `gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, `gpt-image-2`, `gpt-image-1.5`, `gpt-image-1`, and `gpt-image-1-mini`. Text input becomes the prompt; image inputs send the request to `/images/edits`. `OpenAIImagesOptions` adds `size`, `quality`, `n`, and `background`, and `Models.generateImages()` options are now typed per image API through `ImageApiOptionsMap`. Like `openai-decisions`, the provider has no login of its own: it reuses the API key stored for `openai-api`, then the one stored for `openai`, then `OPENAI_API_KEY`, and skips a ChatGPT sign-in on `openai` ([#3527](https://github.com/bastani-inc/atomic/issues/3527)).
|
|
866
|
+
|
|
867
|
+
### Changed
|
|
868
|
+
|
|
869
|
+
- OpenAI Responses and OpenAI Codex Responses requests now retry once at the default tier when the server rejects the requested `service_tier` before producing output, and record a `service_tier_unavailable` warning diagnostic on the assistant message once the default-tier request is accepted. On OpenAI Responses with an API key, a response that reports a slower tier than requested records the same warning. Requests made with a ChatGPT sign-in token, on `openai` or `openai-codex`, do not warn on a reported `default` and are priced at the requested tier, because the ChatGPT backend reports `default` for every tier it applies. Otherwise pricing follows the tier the server reports ([#3529](https://github.com/bastani-inc/atomic/issues/3529)).
|
|
870
|
+
|
|
871
|
+
## [0.9.29-alpha.1] - 2026-10-08
|
|
872
|
+
|
|
873
|
+
### Added
|
|
874
|
+
|
|
875
|
+
- Added the `ultrafast` service tier to `gpt-6.1-sol` and `gpt-5.6-sol` on `openai` and `openai-api`, and to `gpt-6.1-sol` on `openai-codex`. GPT-6.1 Sol uses OpenAI's published Ultrafast rates ($12 input, $60 output per million tokens), with cache rates derived at the same six-times-Standard ratio; GPT-5.6 Sol Ultrafast is a preview without published rates and is priced at six times its Standard rates ([#3529](https://github.com/bastani-inc/atomic/issues/3529)).
|
|
876
|
+
- Added the `openai-images` provider (`openaiImagesProvider()`) and the `openai-images` image API for OpenAI's Images API, with `gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, `gpt-image-2`, `gpt-image-1.5`, `gpt-image-1`, and `gpt-image-1-mini`. Text input becomes the prompt; image inputs send the request to `/images/edits`. `OpenAIImagesOptions` adds `size`, `quality`, `n`, and `background`, and `Models.generateImages()` options are now typed per image API through `ImageApiOptionsMap`. Like `openai-decisions`, the provider has no login of its own: it reuses the API key stored for `openai-api`, then the one stored for `openai`, then `OPENAI_API_KEY`, and skips a ChatGPT sign-in on `openai` ([#3527](https://github.com/bastani-inc/atomic/issues/3527)).
|
|
877
|
+
|
|
878
|
+
### Changed
|
|
879
|
+
|
|
880
|
+
- OpenAI Responses and OpenAI Codex Responses requests now retry once at the default tier when the server rejects the requested `service_tier` before producing output, and record a `service_tier_unavailable` warning diagnostic on the assistant message once the default-tier request is accepted. On OpenAI Responses with an API key, a response that reports a slower tier than requested records the same warning. Requests made with a ChatGPT sign-in token, on `openai` or `openai-codex`, do not warn on a reported `default` and are priced at the requested tier, because the ChatGPT backend reports `default` for every tier it applies. Otherwise pricing follows the tier the server reports ([#3529](https://github.com/bastani-inc/atomic/issues/3529)).
|
|
881
|
+
|
|
882
|
+
## [0.9.28] - 2026-10-08
|
|
883
|
+
|
|
884
|
+
### Breaking Changes
|
|
885
|
+
|
|
886
|
+
- Stream functions must return `AssistantMessageEventStream`, for example from `createAssistantMessageEventStream()`. Plain `EventStream` subclasses are no longer assignable to this response stream type.
|
|
887
|
+
|
|
888
|
+
### Added
|
|
889
|
+
|
|
890
|
+
- Added `ApiKeyAuth.borrowCredentialsFrom` and `sharedApiKeyAuth()`, which let a provider without a login of its own reuse the stored API key of the first listed provider that has one, then fall back to its environment variables. A borrowing provider has no stored credential of its own, and stored OAuth credentials of the listed providers are skipped.
|
|
891
|
+
- Added `openaiApiProvider()` (`openai-api`) and `anthropicApiProvider()` (`anthropic-api`): API-key-only twins of `openai` and `anthropic` that list the same chat models under their own provider ID, so a subscription login on the source provider no longer displaces the API key.
|
|
892
|
+
- Added Claude Haiku 5.5 with adaptive thinking, `xhigh`/`max` effort, prompt-length pricing tiers, and mid-conversation effort, system-message, and tool changes. Bedrock supports adaptive thinking, native `xhigh`, thinking block binding, and prompt caching for Haiku 5.5.
|
|
893
|
+
- Added optional `durationMs` to assistant messages and tool results. Assistant response streams measure elapsed time with a monotonic clock, preserving existing measurements and leaving deferred results untimed.
|
|
894
|
+
- Added `LoginOptions.agentName` for apps to identify themselves during OpenAI ChatGPT and Codex browser login.
|
|
895
|
+
- Added OpenAI's Decisions classifier API and the `openai/gpt-6-luna` classifier, available with an API key but not Sign in with ChatGPT credentials.
|
|
896
|
+
- Added optional `images` to classifier contexts. Unsupported models and classifier APIs return an error result instead of ignoring images.
|
|
897
|
+
- Added Azure Foundry Chat Completions and `azure/deepseek-v4-pro`, with Azure endpoint resolution, deployment-name mapping, and supported reasoning efforts.
|
|
898
|
+
|
|
899
|
+
### Changed
|
|
900
|
+
|
|
901
|
+
- `openai-decisions` no longer has a login or stored credential of its own. It reuses the API key stored for `openai-api`, then the one stored for `openai`, then `OPENAI_API_KEY`; a ChatGPT sign-in on `openai` is skipped.
|
|
902
|
+
- Moved the GPT-6 Luna Decisions classifier to the `openai-decisions` provider (`openaiDecisionsProvider()`), with its own API-key credential. The `openai` provider no longer lists classifiers.
|
|
903
|
+
- The faux provider compares prompt-cache prefixes message by message without changing simulated usage counts.
|
|
904
|
+
- Renamed the Azure provider to `azure`, retaining the `azure-openai-responses` API ID and compatibility with the former provider name and exports.
|
|
905
|
+
|
|
906
|
+
### Fixed
|
|
907
|
+
|
|
908
|
+
- Fixed Kimi K3 cost estimates on Moonshot providers to use official pricing without cache-write charges.
|
|
909
|
+
- Fixed Magistral models retaining Mistral's `prompt_mode` reasoning controls when catalog metadata advertises effort levels.
|
|
910
|
+
- Fixed failed lazy API setup messages using the failure time instead of the request start as their timestamp.
|
|
911
|
+
- Fixed prompt-length pricing tiers being lost when generating model catalogs from AI Gateway, OpenRouter, and models.dev.
|
|
912
|
+
- Corrected the Claude Sonnet 5.5 fallback catalog's cache-read rate to the official $0.10 per million tokens.
|
|
913
|
+
- Fixed Bedrock Claude Haiku 5.5 requests sending unsupported caller temperature settings.
|
|
914
|
+
- Reduced context-limit request failures by estimating new input at 3.5 characters per token when calculating output limits.
|
|
915
|
+
- Fixed transient `server_busy`, `servers are currently busy`, and Mistral `finish_reason: "error"` responses ending the turn instead of being retried.
|
|
916
|
+
- Fixed Anthropic browser login when port 53692 is reserved or in use by falling back to a free loopback port. Copy-code login remains available for headless use.
|
|
917
|
+
- Fixed Codex requests ignoring caller and model overrides for `originator` and `User-Agent` headers; authentication headers remain authoritative.
|
|
918
|
+
- Fixed the selected reasoning effort not reaching OpenAI GPT models on Bedrock Converse. GPT-oss effort is clamped to `low`, `medium`, or `high`; Astra keeps its existing payload.
|
|
919
|
+
- Fixed Radius models disabled by an organization still appearing in the catalog. Fetched or cached gateway catalogs now replace the bundled defaults, including empty catalogs.
|
|
920
|
+
- Fixed OpenAI Decisions accepting unknown choices or out-of-range probabilities, confidence, and scores as successful answers, and failing to return error results when request hooks or fetch throw null or undefined.
|
|
921
|
+
- Retry transient HTTP/2 errors containing "pending stream has been canceled".
|
|
922
|
+
|
|
923
|
+
## [0.9.28-alpha.5] - 2026-10-08
|
|
924
|
+
|
|
925
|
+
### Added
|
|
926
|
+
|
|
927
|
+
- Added `ApiKeyAuth.borrowCredentialsFrom` and `sharedApiKeyAuth()`, which let a provider without a login of its own reuse the stored API key of the first listed provider that has one, then fall back to its environment variables. A borrowing provider has no stored credential of its own, and stored OAuth credentials of the listed providers are skipped.
|
|
928
|
+
|
|
929
|
+
### Changed
|
|
930
|
+
|
|
931
|
+
- `openai-decisions` no longer has a login or stored credential of its own. It reuses the API key stored for `openai-api`, then the one stored for `openai`, then `OPENAI_API_KEY`; a ChatGPT sign-in on `openai` is skipped.
|
|
932
|
+
|
|
933
|
+
## [0.9.28-alpha.4] - 2026-10-07
|
|
934
|
+
|
|
935
|
+
### Breaking Changes
|
|
936
|
+
|
|
937
|
+
- Stream functions must return `AssistantMessageEventStream`, for example from `createAssistantMessageEventStream()`. Plain `EventStream` subclasses are no longer assignable to this response stream type.
|
|
938
|
+
|
|
939
|
+
### Added
|
|
940
|
+
|
|
941
|
+
- Added `openaiApiProvider()` (`openai-api`) and `anthropicApiProvider()` (`anthropic-api`): API-key-only twins of `openai` and `anthropic` that list the same chat models under their own provider ID, so a subscription login on the source provider no longer displaces the API key.
|
|
942
|
+
- Added Claude Haiku 5.5 with adaptive thinking, `xhigh`/`max` effort, prompt-length pricing tiers, and mid-conversation effort, system-message, and tool changes. Bedrock supports adaptive thinking, native `xhigh`, thinking block binding, and prompt caching for Haiku 5.5.
|
|
943
|
+
- Added optional `durationMs` to assistant messages and tool results. Assistant response streams measure elapsed time with a monotonic clock, preserving existing measurements and leaving deferred results untimed.
|
|
944
|
+
|
|
945
|
+
### Changed
|
|
946
|
+
|
|
947
|
+
- Moved the GPT-6 Luna Decisions classifier to the `openai-decisions` provider (`openaiDecisionsProvider()`), with its own API-key credential. The `openai` provider no longer lists classifiers.
|
|
948
|
+
|
|
949
|
+
### Fixed
|
|
950
|
+
|
|
951
|
+
- Fixed Kimi K3 cost estimates on Moonshot providers to use official pricing without cache-write charges.
|
|
952
|
+
- Fixed Magistral models retaining Mistral's `prompt_mode` reasoning controls when catalog metadata advertises effort levels.
|
|
953
|
+
- Fixed failed lazy API setup messages using the failure time instead of the request start as their timestamp.
|
|
954
|
+
- Fixed prompt-length pricing tiers being lost when generating model catalogs from AI Gateway, OpenRouter, and models.dev.
|
|
955
|
+
- Corrected the Claude Sonnet 5.5 fallback catalog's cache-read rate to the official $0.10 per million tokens.
|
|
956
|
+
- Fixed Bedrock Claude Haiku 5.5 requests sending unsupported caller temperature settings.
|
|
957
|
+
|
|
958
|
+
## [0.9.28-alpha.3] - 2026-10-07
|
|
959
|
+
|
|
960
|
+
### Added
|
|
961
|
+
|
|
962
|
+
- Added `LoginOptions.agentName` for apps to identify themselves during OpenAI ChatGPT and Codex browser login.
|
|
963
|
+
- Added OpenAI's Decisions classifier API and the `openai/gpt-6-luna` classifier, available with an API key but not Sign in with ChatGPT credentials.
|
|
964
|
+
- Added optional `images` to classifier contexts. Unsupported models and classifier APIs return an error result instead of ignoring images.
|
|
965
|
+
|
|
966
|
+
### Changed
|
|
967
|
+
|
|
968
|
+
- The faux provider compares prompt-cache prefixes message by message without changing simulated usage counts.
|
|
969
|
+
|
|
970
|
+
### Fixed
|
|
971
|
+
|
|
972
|
+
- Reduced context-limit request failures by estimating new input at 3.5 characters per token when calculating output limits.
|
|
973
|
+
- Fixed transient `server_busy`, `servers are currently busy`, and Mistral `finish_reason: "error"` responses ending the turn instead of being retried.
|
|
974
|
+
- Fixed Anthropic browser login when port 53692 is reserved or in use by falling back to a free loopback port. Copy-code login remains available for headless use.
|
|
975
|
+
- Fixed Codex requests ignoring caller and model overrides for `originator` and `User-Agent` headers; authentication headers remain authoritative.
|
|
976
|
+
- Fixed the selected reasoning effort not reaching OpenAI GPT models on Bedrock Converse. GPT-oss effort is clamped to `low`, `medium`, or `high`; Astra keeps its existing payload.
|
|
977
|
+
- Fixed Radius models disabled by an organization still appearing in the catalog. Fetched or cached gateway catalogs now replace the bundled defaults, including empty catalogs.
|
|
978
|
+
- Fixed OpenAI Decisions accepting unknown choices or out-of-range probabilities, confidence, and scores as successful answers, and failing to return error results when request hooks or fetch throw null or undefined.
|
|
979
|
+
|
|
980
|
+
## [0.9.28-alpha.1] - 2026-10-05
|
|
981
|
+
|
|
982
|
+
### Added
|
|
983
|
+
|
|
984
|
+
- Added Azure Foundry Chat Completions and `azure/deepseek-v4-pro`, with Azure endpoint resolution, deployment-name mapping, and supported reasoning efforts.
|
|
985
|
+
|
|
986
|
+
### Changed
|
|
987
|
+
|
|
988
|
+
- Renamed the Azure provider to `azure`, retaining the `azure-openai-responses` API ID and compatibility with the former provider name and exports.
|
|
989
|
+
|
|
990
|
+
### Fixed
|
|
991
|
+
|
|
992
|
+
- Retry transient HTTP/2 errors containing "pending stream has been canceled".
|
|
993
|
+
|
|
994
|
+
## [0.9.27] - 2026-10-05
|
|
995
|
+
|
|
996
|
+
### Fixed
|
|
997
|
+
|
|
998
|
+
- Fixed OAuth credentials being invalidated when a request or model refresh was cancelled during token rotation. A started refresh now completes and persists its rotated token ([#3429](https://github.com/bastani-inc/atomic/issues/3429)).
|
|
999
|
+
- Fixed slow OAuth token rotation preventing model refresh from reaching the network phase. Rotation now has its own bounded timeout while request authentication keeps its shorter waiting deadline ([#3429](https://github.com/bastani-inc/atomic/issues/3429)).
|
|
1000
|
+
|
|
1001
|
+
## [0.9.27-alpha.1] - 2026-10-04
|
|
1002
|
+
|
|
1003
|
+
### Fixed
|
|
1004
|
+
|
|
1005
|
+
- Fixed OAuth credentials being invalidated when a request or model refresh was cancelled during token rotation. A started refresh now completes and persists its rotated token ([#3429](https://github.com/bastani-inc/atomic/issues/3429)).
|
|
1006
|
+
- Fixed slow OAuth token rotation preventing model refresh from reaching the network phase. Rotation now has its own bounded timeout while request authentication keeps its shorter waiting deadline ([#3429](https://github.com/bastani-inc/atomic/issues/3429)).
|
|
1007
|
+
|
|
1008
|
+
## [0.9.26] - 2026-10-04
|
|
1009
|
+
|
|
1010
|
+
### Added
|
|
1011
|
+
|
|
1012
|
+
- Added per-thinking-level sampling parameter overrides (`Model.samplingParamsByThinkingLevel`) for `openai-completions`, `openai-responses`, and `azure-openai-responses` requests ([#9776](https://github.com/earendil-works/pi/pull/9776) by [@mrexodia](https://github.com/mrexodia))
|
|
1013
|
+
- Added Cloudflare's Clef and Clef Flash classifier models (`@cf/cloudflare/clef`, `@cf/cloudflare/clef-flash`) to the `cloudflare-workers-ai` provider ([#10316](https://github.com/earendil-works/pi/pull/10316) by [@ndisidore](https://github.com/ndisidore), [#10322](https://github.com/earendil-works/pi/pull/10322) by [@RealAlexandreAI](https://github.com/RealAlexandreAI)).
|
|
1014
|
+
- Added `OAuthAuth.usesCallbackServer` so native providers can declare callback-server login flows with a manual-paste fallback. Built-in Anthropic, OpenAI, OpenAI Codex, and OpenRouter OAuth definitions carry this metadata through lazy loading ([#3400](https://github.com/bastani-inc/atomic/issues/3400)).
|
|
1015
|
+
|
|
1016
|
+
### Changed
|
|
1017
|
+
|
|
1018
|
+
- Native Anthropic tool changes now define added and redefined tools inline, keeping the initial tool list stable for prompt caching. `hasToolRedefinitions()` is deprecated.
|
|
1019
|
+
- OAuth browser pages now show the Atomic logo in color instead of white.
|
|
1020
|
+
|
|
1021
|
+
### Fixed
|
|
1022
|
+
|
|
1023
|
+
- Provider errors containing "Selected model is at capacity" are now retried.
|
|
1024
|
+
- Cloudflare AI Gateway now uses dashed Claude model IDs accepted by its Anthropic endpoint.
|
|
1025
|
+
- Fixed Sign in with ChatGPT continuing when its callback port is taken by another login, which made the browser show "OAuth state mismatch"; it now fails with a port-in-use error before opening the browser ([#10265](https://github.com/earendil-works/pi/issues/10265)).
|
|
1026
|
+
- Amazon Bedrock OpenAI models are no longer costed at the short-context rate above 272k input tokens; Bedrock models now include the pricing tiers listed on models.dev ([#10326](https://github.com/earendil-works/pi/issues/10326)).
|
|
1027
|
+
- Amazon Bedrock Claude requests no longer fail with "Invalid `signature` in `thinking` block" after the system prompt or tools change; Claude Opus 4.7+, Sonnet 5+, and Fable 5 now drop stale thinking blocks like the Anthropic provider ([#10324](https://github.com/earendil-works/pi/issues/10324)).
|
|
1028
|
+
- Fixed OpenAI Responses requests failing with `Expected an ID that begins with 'ctc'` when replaying grammar tool calls, such as `codemode`, from another provider or a gateway.
|
|
1029
|
+
- Fixed Together's `deepseek-ai/DeepSeek-V4-Pro-0813` (the renamed DeepSeek V4 Pro) losing its `high` reasoning control after the model catalog refreshed.
|
|
1030
|
+
|
|
1031
|
+
## [0.9.26-alpha.6] - 2026-10-04
|
|
1032
|
+
|
|
1033
|
+
### Added
|
|
1034
|
+
|
|
1035
|
+
- Added per-thinking-level sampling parameter overrides (`Model.samplingParamsByThinkingLevel`) for `openai-completions`, `openai-responses`, and `azure-openai-responses` requests ([#9776](https://github.com/earendil-works/pi/pull/9776) by [@mrexodia](https://github.com/mrexodia))
|
|
1036
|
+
|
|
1037
|
+
## [0.9.26-alpha.3] - 2026-10-03
|
|
1038
|
+
|
|
1039
|
+
### Added
|
|
1040
|
+
|
|
1041
|
+
- Added Cloudflare's Clef and Clef Flash classifier models (`@cf/cloudflare/clef`, `@cf/cloudflare/clef-flash`) to the `cloudflare-workers-ai` provider ([#10316](https://github.com/earendil-works/pi/pull/10316) by [@ndisidore](https://github.com/ndisidore), [#10322](https://github.com/earendil-works/pi/pull/10322) by [@RealAlexandreAI](https://github.com/RealAlexandreAI)).
|
|
1042
|
+
|
|
1043
|
+
### Changed
|
|
1044
|
+
|
|
1045
|
+
- Native Anthropic tool changes now define added and redefined tools inline, keeping the initial tool list stable for prompt caching. `hasToolRedefinitions()` is deprecated.
|
|
1046
|
+
|
|
1047
|
+
### Fixed
|
|
1048
|
+
|
|
1049
|
+
- Provider errors containing "Selected model is at capacity" are now retried.
|
|
1050
|
+
- Cloudflare AI Gateway now uses dashed Claude model IDs accepted by its Anthropic endpoint.
|
|
1051
|
+
- Fixed Sign in with ChatGPT continuing when its callback port is taken by another login, which made the browser show "OAuth state mismatch"; it now fails with a port-in-use error before opening the browser ([#10265](https://github.com/earendil-works/pi/issues/10265)).
|
|
1052
|
+
- Amazon Bedrock OpenAI models are no longer costed at the short-context rate above 272k input tokens; Bedrock models now include the pricing tiers listed on models.dev ([#10326](https://github.com/earendil-works/pi/issues/10326)).
|
|
1053
|
+
- Amazon Bedrock Claude requests no longer fail with "Invalid `signature` in `thinking` block" after the system prompt or tools change; Claude Opus 4.7+, Sonnet 5+, and Fable 5 now drop stale thinking blocks like the Anthropic provider ([#10324](https://github.com/earendil-works/pi/issues/10324)).
|
|
1054
|
+
|
|
1055
|
+
## [0.9.26-alpha.2] - 2026-10-02
|
|
1056
|
+
|
|
1057
|
+
### Added
|
|
1058
|
+
|
|
1059
|
+
- Added `OAuthAuth.usesCallbackServer` so native providers can declare callback-server login flows with a manual-paste fallback. Built-in Anthropic, OpenAI, OpenAI Codex, and OpenRouter OAuth definitions carry this metadata through lazy loading ([#3400](https://github.com/bastani-inc/atomic/issues/3400)).
|
|
1060
|
+
|
|
1061
|
+
## [0.9.26-alpha.1] - 2026-10-01
|
|
1062
|
+
|
|
1063
|
+
### Changed
|
|
1064
|
+
|
|
1065
|
+
- OAuth browser pages now show the Atomic logo in color instead of white.
|
|
1066
|
+
|
|
1067
|
+
### Fixed
|
|
1068
|
+
|
|
1069
|
+
- Fixed OpenAI Responses requests failing with `Expected an ID that begins with 'ctc'` when replaying grammar tool calls, such as `codemode`, from another provider or a gateway.
|
|
1070
|
+
- Fixed Together's `deepseek-ai/DeepSeek-V4-Pro-0813` (the renamed DeepSeek V4 Pro) losing its `high` reasoning control after the model catalog refreshed.
|
|
1071
|
+
|
|
1072
|
+
## [0.9.25] - 2026-10-01
|
|
1073
|
+
|
|
1074
|
+
### Added
|
|
1075
|
+
|
|
1076
|
+
- Added Anthropic workload identity federation from the Anthropic SDK environment variables `ANTHROPIC_FEDERATION_RULE_ID`, `ANTHROPIC_ORGANIZATION_ID`, and `ANTHROPIC_IDENTITY_TOKEN_FILE` (plus optional `ANTHROPIC_SERVICE_ACCOUNT_ID` and `ANTHROPIC_WORKSPACE_ID`). API keys and `ANTHROPIC_AUTH_TOKEN` take precedence ([#10177](https://github.com/earendil-works/pi/issues/10177), [#10242](https://github.com/earendil-works/pi/pull/10242) by [@philfreo](https://github.com/philfreo)).
|
|
1077
|
+
- Anthropic subscription login now offers a copy-code method for remote or headless machines, alongside the default browser callback method ([#10194](https://github.com/earendil-works/pi/pull/10194) by [@lucasmeijer](https://github.com/lucasmeijer)).
|
|
1078
|
+
- Added GPT-6.1 Sol for OpenAI and Codex with supported reasoning efforts, prompt-cache and tool capabilities, and long-context pricing metadata.
|
|
1079
|
+
- Added GPT-6.1 Sol to the GitHub Copilot catalog with Copilot's Responses endpoint, supported reasoning efforts, and provider-published context and output limits. Availability remains subject to the account's model policy.
|
|
1080
|
+
- Added **Sign in with ChatGPT** for the OpenAI Responses API alongside API-key authentication, separate from Codex subscription login.
|
|
1081
|
+
- Added Jev classifiers on Vercel AI Gateway and OpenCode Zen, and provider-reported classifier usage and costs, including billed responses with invalid answers.
|
|
1082
|
+
- Added the lightweight `@bastani/pi-ai/models` entry point for model catalog consumers.
|
|
1083
|
+
- Added `Model.serviceTiers` and `getServiceTierCost()` for the Fast and Ultrafast tiers a model advertises and their published rates. The OpenAI catalog follows OpenAI's Fast and Ultrafast pricing tables. The Codex catalog follows Codex's per-model tiers: Fast for every model except GPT-5.3 Codex Spark, and Ultrafast for GPT-6 Astra.
|
|
1084
|
+
- The OpenAI Responses adapter can send GPT-6 Astra's `ultrafast` tier.
|
|
1085
|
+
|
|
1086
|
+
### Changed
|
|
1087
|
+
|
|
1088
|
+
- The Codex Responses adapter sends `service_tier` only when the model advertises it, following Codex: `flex` always passes through, `default` sends no tier, and an unadvertised tier is left out instead of failing.
|
|
1089
|
+
- The OpenAI Responses adapter leaves out a Fast or Ultrafast tier the model doesn't advertise. It sends other tiers as requested.
|
|
1090
|
+
- Fast and Ultrafast usage is priced at the served tier's published rates. A model without tier metadata keeps the previous Fast multiplier.
|
|
1091
|
+
|
|
1092
|
+
### Fixed
|
|
1093
|
+
|
|
1094
|
+
- Fixed context overflow detection for Z.AI CN endpoint `Prompt exceeds max length` errors ([#10208](https://github.com/earendil-works/pi/issues/10208)).
|
|
1095
|
+
- Anthropic `strict: "prefer"` tools now fall back to non-strict mode for unsupported string formats and additional array/object constraints instead of rejecting the request. Supported string length and pattern constraints remain strict ([#9953](https://github.com/earendil-works/pi/issues/9953)).
|
|
1096
|
+
- Provider retries now use exponential backoff when `Retry-After` or `retry-after-ms` cannot be parsed as a finite delay ([#9571](https://github.com/earendil-works/pi/issues/9571)).
|
|
1097
|
+
|
|
1098
|
+
## [0.9.25-alpha.3] - 2026-09-30
|
|
1099
|
+
|
|
1100
|
+
### Added
|
|
1101
|
+
|
|
1102
|
+
- Added Anthropic workload identity federation from the Anthropic SDK environment variables `ANTHROPIC_FEDERATION_RULE_ID`, `ANTHROPIC_ORGANIZATION_ID`, and `ANTHROPIC_IDENTITY_TOKEN_FILE` (plus optional `ANTHROPIC_SERVICE_ACCOUNT_ID` and `ANTHROPIC_WORKSPACE_ID`). API keys and `ANTHROPIC_AUTH_TOKEN` take precedence ([#10177](https://github.com/earendil-works/pi/issues/10177), [#10242](https://github.com/earendil-works/pi/pull/10242) by [@philfreo](https://github.com/philfreo)).
|
|
1103
|
+
- Anthropic subscription login now offers a copy-code method for remote or headless machines, alongside the default browser callback method ([#10194](https://github.com/earendil-works/pi/pull/10194) by [@lucasmeijer](https://github.com/lucasmeijer)).
|
|
1104
|
+
|
|
1105
|
+
### Fixed
|
|
1106
|
+
|
|
1107
|
+
- Fixed context overflow detection for Z.AI CN endpoint `Prompt exceeds max length` errors ([#10208](https://github.com/earendil-works/pi/issues/10208)).
|
|
1108
|
+
- Anthropic `strict: "prefer"` tools now fall back to non-strict mode for unsupported string formats and additional array/object constraints instead of rejecting the request. Supported string length and pattern constraints remain strict ([#9953](https://github.com/earendil-works/pi/issues/9953)).
|
|
1109
|
+
- Provider retries now use exponential backoff when `Retry-After` or `retry-after-ms` cannot be parsed as a finite delay ([#9571](https://github.com/earendil-works/pi/issues/9571)).
|
|
1110
|
+
|
|
1111
|
+
## [0.9.25-alpha.1] - 2026-09-29
|
|
1112
|
+
|
|
1113
|
+
### Added
|
|
1114
|
+
|
|
1115
|
+
- Added GPT-6.1 Sol for OpenAI and Codex with supported reasoning efforts, prompt-cache and tool capabilities, and long-context pricing metadata.
|
|
1116
|
+
- Added GPT-6.1 Sol to the GitHub Copilot catalog with Copilot's Responses endpoint, supported reasoning efforts, and provider-published context and output limits. Availability remains subject to the account's model policy.
|
|
1117
|
+
- Added **Sign in with ChatGPT** for the OpenAI Responses API alongside API-key authentication, separate from Codex subscription login.
|
|
1118
|
+
- Added Jev classifiers on Vercel AI Gateway and OpenCode Zen, and provider-reported classifier usage and costs, including billed responses with invalid answers.
|
|
1119
|
+
- Added the lightweight `@bastani/pi-ai/models` entry point for model catalog consumers.
|
|
1120
|
+
- Added `Model.serviceTiers` and `getServiceTierCost()` for the Fast and Ultrafast tiers a model advertises and their published rates. The OpenAI catalog follows OpenAI's Fast and Ultrafast pricing tables. The Codex catalog follows Codex's per-model tiers: Fast for every model except GPT-5.3 Codex Spark, and Ultrafast for GPT-6 Astra.
|
|
1121
|
+
- The OpenAI Responses adapter can send GPT-6 Astra's `ultrafast` tier.
|
|
1122
|
+
|
|
1123
|
+
### Changed
|
|
1124
|
+
|
|
1125
|
+
- The Codex Responses adapter sends `service_tier` only when the model advertises it, following Codex: `flex` always passes through, `default` sends no tier, and an unadvertised tier is left out instead of failing.
|
|
1126
|
+
- The OpenAI Responses adapter leaves out a Fast or Ultrafast tier the model doesn't advertise. It sends other tiers as requested.
|
|
1127
|
+
- Fast and Ultrafast usage is priced at the served tier's published rates. A model without tier metadata keeps the previous Fast multiplier.
|
|
1128
|
+
|
|
1129
|
+
## [0.9.24] - 2026-09-29
|
|
1130
|
+
|
|
1131
|
+
### Added
|
|
1132
|
+
|
|
1133
|
+
- Added Claude Sonnet 5.5 to the built-in Anthropic model catalog with adaptive thinking, mid-conversation effort, 1M context, and official pricing metadata.
|
|
1134
|
+
- Added the `llama-cpp-classify` classifier API, which turns a chat model served by llama.cpp's `llama-server` into a classifier by reading the next-token probabilities of single-token answer labels, and a `temperature` classifier option that softens or sharpens answer probabilities on APIs that can apply it ([#10119](https://github.com/earendil-works/pi/pull/10119)).
|
|
1135
|
+
|
|
1136
|
+
### Fixed
|
|
1137
|
+
|
|
1138
|
+
- Removed the OpenCode Go Kimi K2.6 model overrides because models.dev deprecated that model; OpenCode Zen Kimi K2.6 keeps its overrides.
|
|
1139
|
+
- Fixed Mistral reasoning models ignoring the requested thinking level: GLM 5.3 now uses `reasoning_effort` instead of `prompt_mode`, GLM 5.2 accepts `max`, and Mistral models only offer the effort levels the API supports ([#9678](https://github.com/earendil-works/pi/issues/9678)).
|
|
1140
|
+
- Fixed OpenCode Zen and OpenCode Go `qwen3.8-flash` thinking being replayed as plain text on later turns because the endpoint returns empty thinking signatures ([#10047](https://github.com/earendil-works/pi/issues/10047)).
|
|
1141
|
+
- Fixed OpenAI Responses streams returning unfinished tool calls as runnable, which made servers that omit `output_index` (such as llama.cpp) run mixed-up commands; such streams now end with an error ([#9974](https://github.com/earendil-works/pi/issues/9974)).
|
|
1142
|
+
|
|
1143
|
+
## [0.9.24-alpha.1] - 2026-09-28
|
|
1144
|
+
|
|
1145
|
+
### Added
|
|
1146
|
+
|
|
1147
|
+
- Added Claude Sonnet 5.5 to the built-in Anthropic model catalog with adaptive thinking, mid-conversation effort, 1M context, and official pricing metadata.
|
|
1148
|
+
- Added the `llama-cpp-classify` classifier API, which turns a chat model served by llama.cpp's `llama-server` into a classifier by reading the next-token probabilities of single-token answer labels, and a `temperature` classifier option that softens or sharpens answer probabilities on APIs that can apply it ([#10119](https://github.com/earendil-works/pi/pull/10119)).
|
|
1149
|
+
|
|
1150
|
+
### Fixed
|
|
1151
|
+
|
|
1152
|
+
- Removed the OpenCode Go Kimi K2.6 model overrides because models.dev deprecated that model; OpenCode Zen Kimi K2.6 keeps its overrides.
|
|
1153
|
+
- Fixed Mistral reasoning models ignoring the requested thinking level: GLM 5.3 now uses `reasoning_effort` instead of `prompt_mode`, GLM 5.2 accepts `max`, and Mistral models only offer the effort levels the API supports ([#9678](https://github.com/earendil-works/pi/issues/9678)).
|
|
1154
|
+
- Fixed OpenCode Zen and OpenCode Go `qwen3.8-flash` thinking being replayed as plain text on later turns because the endpoint returns empty thinking signatures ([#10047](https://github.com/earendil-works/pi/issues/10047)).
|
|
1155
|
+
- Fixed OpenAI Responses streams returning unfinished tool calls as runnable, which made servers that omit `output_index` (such as llama.cpp) run mixed-up commands; such streams now end with an error ([#9974](https://github.com/earendil-works/pi/issues/9974)).
|
|
1156
|
+
|
|
1157
|
+
## [0.9.21] - 2026-09-26
|
|
1158
|
+
|
|
1159
|
+
### Changed
|
|
1160
|
+
|
|
1161
|
+
- Upgraded the bundled OpenAI SDK to 7.19.0, so the OpenAI Responses and Codex `serviceTier` option types accept `"fast"` ([#10044](https://github.com/earendil-works/pi/pull/10044)).
|
|
1162
|
+
|
|
1163
|
+
### Fixed
|
|
1164
|
+
|
|
1165
|
+
- Fixed model-level `samplingParams` being dropped by direct `stream()`/`complete()` calls on OpenAI-compatible APIs ([#9506](https://github.com/earendil-works/pi/issues/9506)).
|
|
1166
|
+
- Fixed Mistral GLM models producing empty text blocks and split thinking blocks from empty content deltas, which could make later requests fail with "Expected at most one leading ThinkChunk" ([#9674](https://github.com/earendil-works/pi/issues/9674)).
|
|
1167
|
+
- Fixed OpenAI Responses requests that set `serviceTier: "fast"` being priced at the standard rate when the response does not report a service tier ([#10034](https://github.com/earendil-works/pi/issues/10034)).
|
|
1168
|
+
|
|
1169
|
+
## [0.9.21-alpha.2] - 2026-09-26
|
|
1170
|
+
|
|
1171
|
+
### Changed
|
|
1172
|
+
|
|
1173
|
+
- Upgraded the bundled OpenAI SDK to 7.19.0, so the OpenAI Responses and Codex `serviceTier` option types accept `"fast"` ([#10044](https://github.com/earendil-works/pi/pull/10044)).
|
|
1174
|
+
|
|
1175
|
+
### Fixed
|
|
1176
|
+
|
|
1177
|
+
- Fixed model-level `samplingParams` being dropped by direct `stream()`/`complete()` calls on OpenAI-compatible APIs ([#9506](https://github.com/earendil-works/pi/issues/9506)).
|
|
1178
|
+
- Fixed Mistral GLM models producing empty text blocks and split thinking blocks from empty content deltas, which could make later requests fail with "Expected at most one leading ThinkChunk" ([#9674](https://github.com/earendil-works/pi/issues/9674)).
|
|
1179
|
+
- Fixed OpenAI Responses requests that set `serviceTier: "fast"` being priced at the standard rate when the response does not report a service tier ([#10034](https://github.com/earendil-works/pi/issues/10034)).
|
|
1180
|
+
|
|
1181
|
+
## [0.9.20] - 2026-09-24
|
|
1182
|
+
|
|
1183
|
+
### Breaking Changes
|
|
1184
|
+
|
|
1185
|
+
- Image generation now uses `ImageModel` entries in the regular `Provider` and `Models` collection instead of the separate `ImagesModels`/`ImagesProvider` collection. Replace `createImagesModels()`, `createImagesProvider()`, `builtinImagesModels()`, and `openrouterImagesProvider()` with `createModels()`/`builtinModels()` and `models.getModelOfType("image", ...)`/`models.generateImages()`. The old plural image type names are removed; generated image catalog data now ships alongside chat and classifier entries.
|
|
1186
|
+
- Provider implementations and direct API modules now consume `TranscriptContext`; use `normalizeContext()` at direct-call boundaries and replay helpers to read prompt/tool state. Durable tool arguments and results use JSON-value types.
|
|
1187
|
+
|
|
1188
|
+
### Added
|
|
1189
|
+
|
|
1190
|
+
- Added `onProviderStreamEvent` to observe parsed provider stream events before normalization, including provider-specific fields not retained in assistant messages ([#9784](https://github.com/earendil-works/pi/issues/9784)).
|
|
1191
|
+
- Added operation-specific model accessors (`getModelsOfType()`, `getModelOfType()`, `getAvailableOfType()`, `getAllModels()`, `getAllAvailable()`), image and classifier dispatch on providers, and `classify()` for structured choice, score, and bool questions. The built-in TypeSafe `jev-latest` classifier uses `TYPESAFE_API_KEY`; OpenRouter image models share OpenRouter authentication. Chat-only reads and models with no `type` continue to mean chat.
|
|
1192
|
+
- Added typed JSON catalog variants (`models.all.json` and `providers/{id}.all.json`) alongside the existing chat-only variants for clients that request all operation types.
|
|
1193
|
+
- Added Jev classifier models on OpenRouter (`typesafe/jev-1.13`, `~typesafe/jev-latest`) through its TypeSafe-compatible System One endpoint, and on Cloudflare Workers AI (`typesafe/jev`) through the new `cloudflare-workers-ai-system-one` classifier API.
|
|
1194
|
+
- Added Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna to the GitHub Copilot catalog.
|
|
1195
|
+
- Added GPT-6 Sol and GPT-6 Luna for OpenAI API keys and OpenAI Codex subscriptions, with full reasoning-effort, prompt-caching, tool-search, long-context pricing, and official cost metadata.
|
|
1196
|
+
- Added `getDecisionModels()` and a generated models.dev decision-model catalog (`type: "decision"`, fetched from `https://models.dev/api.json?type=all` by `npm run generate-decision-models`) listing TypeSafe Jev on the gateways that resell it, with context limits and prices.
|
|
1197
|
+
- Added Claude Opus 5.5 to the built-in Anthropic model catalog with adaptive thinking, 1M context, and official pricing metadata.
|
|
1198
|
+
- Added Anthropic fast mode for models whose `fastRoute` declares `speed: "fast"`: the Anthropic Messages adapter sends the route's upstream model with `speed: "fast"` and the `fast-mode-2026-02-01` beta, prices responses reporting fast speed at the 2x fast-mode rates, and rejects payload hooks that change the route-owned model or speed.
|
|
1199
|
+
- Added Meta provider (Model API key and Muse subscription OAuth) with Muse Spark models.
|
|
1200
|
+
- Added model image-input limit and cache-safe resize metadata (`inputLimits`) to the `Model` type and the generated catalog ([#9631](https://github.com/earendil-works/pi/issues/9631)).
|
|
1201
|
+
- Added Grok 4.7 to the built-in xAI model catalog with long-context pricing metadata.
|
|
1202
|
+
- Chronological system messages with named prompt patches and tool additions/removals, including native provider transitions where supported and replay checkpoints elsewhere.
|
|
1203
|
+
- A static Radius catalog with authenticated dynamic refresh, refreshed image models, and prompt-cache lifetime metadata for direct Anthropic models.
|
|
1204
|
+
|
|
1205
|
+
### Fixed
|
|
1206
|
+
|
|
1207
|
+
- Fixed Claude Opus 5.5 on GitHub Copilot offering thinking levels other than low, medium, high, xhigh, and max when models.dev lists the model before its effort metadata is complete.
|
|
1208
|
+
- Fixed 1-hour Anthropic cache writes reported by Vercel AI Gateway in streaming deltas being priced at the 5-minute rate ([#9210](https://github.com/earendil-works/pi/issues/9210)).
|
|
1209
|
+
- Rejected malformed TypeSafe classifier answers when the returned choice, score, confidence, or probability falls outside the submitted question's bounds.
|
|
1210
|
+
- Fixed Anthropic and Bedrock Claude requests failing with HTTP 400 when a tool opted into `strict: "prefer"` JSON-schema constrained sampling with numeric, string-length, or `maxItems`/`minItems` constraints that Claude's strict mode rejects. These tools now fall back to non-strict tool use, and `strict: "require"` reports the unsupported keyword.
|
|
1211
|
+
- Fixed Anthropic OAuth requests for Claude Opus 5.5 rejected with `claude_code_version_too_old` by advertising Claude Code version `2.1.280`, the minimum the API requires.
|
|
1212
|
+
- Fixed Fast-mode usage costs for GPT-6 models being recorded at standard rates when OpenAI reports the tier as `service_tier: "fast"` instead of `priority`.
|
|
1213
|
+
- Fixed Claude Opus 5.5 offering thinking levels other than low, medium, high, xhigh, and max when models.dev lists the model before its effort metadata is complete.
|
|
1214
|
+
- Fixed unknown OpenAI-compatible Chat Completions endpoints receiving strict tool schemas unless they explicitly advertise support (`compat.supportsStrictMode` now defaults to `false`), while preserving strict tools for capable built-in models ([#9816](https://github.com/earendil-works/pi/issues/9816)).
|
|
1215
|
+
- Fixed image-only user messages being rejected by some OpenAI-compatible providers because they included an empty text part ([#9797](https://github.com/earendil-works/pi/issues/9797)).
|
|
1216
|
+
- Derive Gemini thinking levels from model metadata, preserve renamed Anthropic/Vercel unsigned thinking replay and DeepSeek V4 effort, retry Cloudflare 520 and Azure peak-load errors, and scope bodyless overflow detection to Cerebras.
|
|
1217
|
+
- Documented Fireworks deferred tool loading against chronological system-message `toolsAdded`/`toolsRemoved` instead of the removed tool-result `addedToolNames` field.
|
|
1218
|
+
- Request-auth preparation now times out after 15 seconds when OAuth refresh, credential-store reads, or auth derivation ignore cancellation, and late refresh results cannot overwrite stored credentials. The timeout diagnostic is source-neutral and does not instruct you to log in ([#3085](https://github.com/bastani-inc/atomic/issues/3085), [#3087](https://github.com/bastani-inc/atomic/pull/3087)).
|
|
1219
|
+
- Bedrock requests now honor explicit `maxRetries`, including zero for a single transport attempt, instead of silently using the AWS SDK retry default. Omitting the option preserves SDK/environment configuration ([#3089](https://github.com/bastani-inc/atomic/issues/3089), [#3090](https://github.com/bastani-inc/atomic/issues/3090)).
|
|
1220
|
+
- Kept Kimi Coding models available after the upstream catalog split into regional coding plans, preserving Atomic's existing kimi.com endpoint.
|
|
1221
|
+
- Model catalog declarations preserve JSON import attributes for strict NodeNext consumers without requiring `skipLibCheck` ([#3105](https://github.com/bastani-inc/atomic/issues/3105)).
|
|
1222
|
+
- Credential screening now recognizes `TYPESAFE_API_KEY` instead of `TYPESAFE_AI_API_KEY`, matching the renamed TypeSafe Jev environment variable.
|
|
1223
|
+
- Fixed z.ai `Prompt too long` errors not being recognized as context overflow ([earendil-works/pi#9805](https://github.com/earendil-works/pi/issues/9805)).
|
|
1224
|
+
- Fixed Cerebras models advertising unsupported strict tool schemas, which caused HTTP 400 errors when strict and non-strict tools were mixed ([earendil-works/pi#9804](https://github.com/earendil-works/pi/pull/9804) by [@EdenGottlieb](https://github.com/EdenGottlieb)).
|
|
1225
|
+
- Fixed OpenAI-compatible Responses errors to identify the actual provider instead of always labeling them as OpenAI errors ([#9298](https://github.com/earendil-works/pi/issues/9298)).
|
|
1226
|
+
- Fixed Amazon Bedrock one-hour cache writes being priced at the five-minute rate ([#9457](https://github.com/earendil-works/pi/issues/9457)).
|
|
1227
|
+
- Fixed Baseten requests to send session-affinity headers from `sessionId` for automatic prompt-cache routing ([#9629](https://github.com/earendil-works/pi/issues/9629)).
|
|
1228
|
+
|
|
1229
|
+
## [0.9.20-alpha.9] - 2026-09-24
|
|
1230
|
+
|
|
1231
|
+
### Breaking Changes
|
|
1232
|
+
|
|
1233
|
+
- Image generation now uses `ImageModel` entries in the regular `Provider` and `Models` collection instead of the separate `ImagesModels`/`ImagesProvider` collection. Replace `createImagesModels()`, `createImagesProvider()`, `builtinImagesModels()`, and `openrouterImagesProvider()` with `createModels()`/`builtinModels()` and `models.getModelOfType("image", ...)`/`models.generateImages()`. The old plural image type names are removed; generated image catalog data now ships alongside chat and classifier entries.
|
|
1234
|
+
|
|
1235
|
+
### Added
|
|
1236
|
+
|
|
1237
|
+
- Added `onProviderStreamEvent` to observe parsed provider stream events before normalization, including provider-specific fields not retained in assistant messages ([#9784](https://github.com/earendil-works/pi/issues/9784)).
|
|
1238
|
+
- Added operation-specific model accessors (`getModelsOfType()`, `getModelOfType()`, `getAvailableOfType()`, `getAllModels()`, `getAllAvailable()`), image and classifier dispatch on providers, and `classify()` for structured choice, score, and bool questions. The built-in TypeSafe `jev-latest` classifier uses `TYPESAFE_API_KEY`; OpenRouter image models share OpenRouter authentication. Chat-only reads and models with no `type` continue to mean chat.
|
|
1239
|
+
- Added typed JSON catalog variants (`models.all.json` and `providers/{id}.all.json`) alongside the existing chat-only variants for clients that request all operation types.
|
|
1240
|
+
- Added Jev classifier models on OpenRouter (`typesafe/jev-1.13`, `~typesafe/jev-latest`) through its TypeSafe-compatible System One endpoint, and on Cloudflare Workers AI (`typesafe/jev`) through the new `cloudflare-workers-ai-system-one` classifier API.
|
|
1241
|
+
|
|
1242
|
+
### Fixed
|
|
1243
|
+
|
|
1244
|
+
- Fixed Claude Opus 5.5 on GitHub Copilot offering thinking levels other than low, medium, high, xhigh, and max when models.dev lists the model before its effort metadata is complete.
|
|
1245
|
+
- Fixed 1-hour Anthropic cache writes reported by Vercel AI Gateway in streaming deltas being priced at the 5-minute rate ([#9210](https://github.com/earendil-works/pi/issues/9210)).
|
|
1246
|
+
- Rejected malformed TypeSafe classifier answers when the returned choice, score, confidence, or probability falls outside the submitted question's bounds.
|
|
1247
|
+
- Fixed Anthropic and Bedrock Claude requests failing with HTTP 400 when a tool opted into `strict: "prefer"` JSON-schema constrained sampling with numeric, string-length, or `maxItems`/`minItems` constraints that Claude's strict mode rejects. These tools now fall back to non-strict tool use, and `strict: "require"` reports the unsupported keyword.
|
|
1248
|
+
|
|
1249
|
+
## [0.9.20-alpha.7] - 2026-09-22
|
|
1250
|
+
|
|
1251
|
+
### Added
|
|
1252
|
+
|
|
1253
|
+
- Added Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna to the GitHub Copilot catalog.
|
|
1254
|
+
- Added GPT-6 Sol and GPT-6 Luna for OpenAI API keys and OpenAI Codex subscriptions, with full reasoning-effort, prompt-caching, tool-search, long-context pricing, and official cost metadata.
|
|
1255
|
+
- Added `getDecisionModels()` and a generated models.dev decision-model catalog (`type: "decision"`, fetched from `https://models.dev/api.json?type=all` by `npm run generate-decision-models`) listing TypeSafe Jev on the gateways that resell it, with context limits and prices.
|
|
1256
|
+
- Added Claude Opus 5.5 to the built-in Anthropic model catalog with adaptive thinking, 1M context, and official pricing metadata.
|
|
1257
|
+
- Added Anthropic fast mode for models whose `fastRoute` declares `speed: "fast"`: the Anthropic Messages adapter sends the route's upstream model with `speed: "fast"` and the `fast-mode-2026-02-01` beta, prices responses reporting fast speed at the 2x fast-mode rates, and rejects payload hooks that change the route-owned model or speed.
|
|
1258
|
+
|
|
1259
|
+
### Fixed
|
|
1260
|
+
|
|
1261
|
+
- Fixed Anthropic OAuth requests for Claude Opus 5.5 rejected with `claude_code_version_too_old` by advertising Claude Code version `2.1.280`, the minimum the API requires.
|
|
1262
|
+
- Fixed Fast-mode usage costs for GPT-6 models being recorded at standard rates when OpenAI reports the tier as `service_tier: "fast"` instead of `priority`.
|
|
1263
|
+
- Fixed Claude Opus 5.5 offering thinking levels other than low, medium, high, xhigh, and max when models.dev lists the model before its effort metadata is complete.
|
|
1264
|
+
|
|
1265
|
+
## [0.9.20-alpha.6] - 2026-09-22
|
|
1266
|
+
|
|
1267
|
+
### Added
|
|
1268
|
+
|
|
1269
|
+
- Added Meta provider (Model API key and Muse subscription OAuth) with Muse Spark models.
|
|
1270
|
+
- Added model image-input limit and cache-safe resize metadata (`inputLimits`) to the `Model` type and the generated catalog ([#9631](https://github.com/earendil-works/pi/issues/9631)).
|
|
1271
|
+
- Added Grok 4.7 to the built-in xAI model catalog with long-context pricing metadata.
|
|
1272
|
+
|
|
1273
|
+
### Fixed
|
|
1274
|
+
|
|
1275
|
+
- Fixed unknown OpenAI-compatible Chat Completions endpoints receiving strict tool schemas unless they explicitly advertise support (`compat.supportsStrictMode` now defaults to `false`), while preserving strict tools for capable built-in models ([#9816](https://github.com/earendil-works/pi/issues/9816)).
|
|
1276
|
+
- Fixed image-only user messages being rejected by some OpenAI-compatible providers because they included an empty text part ([#9797](https://github.com/earendil-works/pi/issues/9797)).
|
|
1277
|
+
|
|
1278
|
+
## [0.9.20-alpha.4] - 2026-09-20
|
|
1279
|
+
|
|
1280
|
+
### Breaking Changes
|
|
1281
|
+
|
|
1282
|
+
- Provider implementations and direct API modules now consume `TranscriptContext`; use `normalizeContext()` at direct-call boundaries and replay helpers to read prompt/tool state. Durable tool arguments and results use JSON-value types.
|
|
1283
|
+
|
|
1284
|
+
### Added
|
|
1285
|
+
|
|
1286
|
+
- Chronological system messages with named prompt patches and tool additions/removals, including native provider transitions where supported and replay checkpoints elsewhere.
|
|
1287
|
+
- A static Radius catalog with authenticated dynamic refresh, refreshed image models, and prompt-cache lifetime metadata for direct Anthropic models.
|
|
1288
|
+
|
|
1289
|
+
### Fixed
|
|
1290
|
+
|
|
1291
|
+
- Derive Gemini thinking levels from model metadata, preserve renamed Anthropic/Vercel unsigned thinking replay and DeepSeek V4 effort, retry Cloudflare 520 and Azure peak-load errors, and scope bodyless overflow detection to Cerebras.
|
|
1292
|
+
|
|
1293
|
+
- Documented Fireworks deferred tool loading against chronological system-message `toolsAdded`/`toolsRemoved` instead of the removed tool-result `addedToolNames` field.
|
|
1294
|
+
|
|
1295
|
+
- Request-auth preparation now times out after 15 seconds when OAuth refresh, credential-store reads, or auth derivation ignore cancellation, and late refresh results cannot overwrite stored credentials. The timeout diagnostic is source-neutral and does not instruct you to log in ([#3085](https://github.com/bastani-inc/atomic/issues/3085), [#3087](https://github.com/bastani-inc/atomic/pull/3087)).
|
|
1296
|
+
- Bedrock requests now honor explicit `maxRetries`, including zero for a single transport attempt, instead of silently using the AWS SDK retry default. Omitting the option preserves SDK/environment configuration ([#3089](https://github.com/bastani-inc/atomic/issues/3089), [#3090](https://github.com/bastani-inc/atomic/issues/3090)).
|
|
1297
|
+
- Kept Kimi Coding models available after the upstream catalog split into regional coding plans, preserving Atomic's existing kimi.com endpoint.
|
|
1298
|
+
- Model catalog declarations preserve JSON import attributes for strict NodeNext consumers without requiring `skipLibCheck` ([#3105](https://github.com/bastani-inc/atomic/issues/3105)).
|
|
1299
|
+
- Credential screening now recognizes `TYPESAFE_API_KEY` instead of `TYPESAFE_AI_API_KEY`, matching the renamed TypeSafe Jev environment variable.
|
|
1300
|
+
- Fixed z.ai `Prompt too long` errors not being recognized as context overflow ([earendil-works/pi#9805](https://github.com/earendil-works/pi/issues/9805)).
|
|
1301
|
+
- Fixed Cerebras models advertising unsupported strict tool schemas, which caused HTTP 400 errors when strict and non-strict tools were mixed ([earendil-works/pi#9804](https://github.com/earendil-works/pi/pull/9804) by [@EdenGottlieb](https://github.com/EdenGottlieb)).
|
|
1302
|
+
|
|
1303
|
+
## [0.9.20-alpha.3] - 2026-09-16
|
|
1304
|
+
|
|
1305
|
+
### Fixed
|
|
1306
|
+
|
|
1307
|
+
- Fixed OpenAI-compatible Responses errors to identify the actual provider instead of always labeling them as OpenAI errors ([#9298](https://github.com/earendil-works/pi/issues/9298)).
|
|
1308
|
+
- Fixed Amazon Bedrock one-hour cache writes being priced at the five-minute rate ([#9457](https://github.com/earendil-works/pi/issues/9457)).
|
|
1309
|
+
- Fixed Baseten requests to send session-affinity headers from `sessionId` for automatic prompt-cache routing ([#9629](https://github.com/earendil-works/pi/issues/9629)).
|
|
1310
|
+
|
|
1311
|
+
## [0.9.19] - 2026-09-13
|
|
1312
|
+
|
|
1313
|
+
Cumulative release of the `0.9.19-alpha.2` through `0.9.19-alpha.4` prereleases. Per-change details remain in the unchanged prerelease sections below.
|
|
1314
|
+
|
|
1315
|
+
### Added
|
|
1316
|
+
|
|
1317
|
+
- Enabled native deferred tool loading for Fireworks Messages models through `ToolSearch` or `tool_search` ([#9323](https://github.com/earendil-works/pi/issues/9323)).
|
|
1318
|
+
|
|
1319
|
+
### Fixed
|
|
1320
|
+
|
|
1321
|
+
- Fixed Anthropic object-union tool schemas, including own property names such as `__proto__` and explicitly nonobject branches, without changing authored schemas or local validation ([#2190](https://github.com/bastani-inc/atomic/pull/2190) by [@elefthei](https://github.com/elefthei), [#2189](https://github.com/bastani-inc/atomic/issues/2189)).
|
|
1322
|
+
- Capped shared assistant and summary retry backoff with `RetryPolicy.maxAgentDelayMs`, defaulting to 60 seconds ([#8826](https://github.com/earendil-works/pi/issues/8826)).
|
|
1323
|
+
- Updated Anthropic, Bedrock, Google and HTTP proxy SDKs; Google's consecutive tool-call limit now returns an error stop reason. GitHub Copilot GPT models use the Responses API ([#9209](https://github.com/earendil-works/pi/issues/9209)).
|
|
1324
|
+
- Fixed quadratic CPU usage when draining buffered `EventStream` events ([#9055](https://github.com/earendil-works/pi/issues/9055)).
|
|
1325
|
+
- Fixed OpenAI Codex Off reasoning requests and removed unavailable GPT-5.4 and GPT-5.4 mini catalog entries ([#9191](https://github.com/earendil-works/pi/issues/9191), [#9394](https://github.com/earendil-works/pi/issues/9394)).
|
|
1326
|
+
- Fixed Fireworks unsigned thinking replay and catalog-based reasoning selection, including DeepSeek V4 and Qwen3.8 fallbacks. Refreshed DeepSeek V4.1 Flash identity and pricing ([#9323](https://github.com/earendil-works/pi/issues/9323), [#9423](https://github.com/earendil-works/pi/issues/9423)).
|
|
1327
|
+
- Fixed OpenRouter prompt-cached requests sending `x-session-id` for Chat Completions and Anthropic Messages ([#9102](https://github.com/earendil-works/pi/issues/9102)).
|
|
1328
|
+
- Fixed Mistral Medium aliases and Mistral-hosted GLM-5.2 using `reasoning_effort` instead of unsupported or ignored `prompt_mode` ([#8700](https://github.com/earendil-works/pi/issues/8700), [#9375](https://github.com/earendil-works/pi/issues/9375)).
|
|
1329
|
+
|
|
1330
|
+
## [0.9.19-alpha.4] - 2026-09-10
|
|
1331
|
+
|
|
1332
|
+
### Added
|
|
1333
|
+
|
|
1334
|
+
- Enabled native deferred tool loading for Fireworks Messages models. Use `ToolSearch` or `tool_search` as the loader name for prompt-prefix deferral ([#9323](https://github.com/earendil-works/pi/issues/9323)).
|
|
1335
|
+
|
|
1336
|
+
### Fixed
|
|
1337
|
+
|
|
1338
|
+
- Fixed OpenAI Codex requests to send the model's Off reasoning effort instead of omitting it, while respecting unsupported Off mappings ([#9191](https://github.com/earendil-works/pi/issues/9191)).
|
|
1339
|
+
- Fixed Fireworks unsigned thinking replay and reasoning effort selection using catalog metadata, with verified DeepSeek V4 and Qwen3.8 fallbacks and removal of redundant GLM 5.2 and Kimi K3 effort aliases ([#9323](https://github.com/earendil-works/pi/issues/9323)).
|
|
1340
|
+
- Fixed OpenRouter requests to send `x-session-id` from `sessionId` for Chat Completions and Anthropic Messages models when prompt caching is enabled ([#9102](https://github.com/earendil-works/pi/issues/9102)).
|
|
1341
|
+
- Fixed the DeepSeek catalog to advertise `deepseek-flash` for DeepSeek V4.1 Flash instead of retired Flash aliases, and refreshed DeepSeek pricing metadata ([#9423](https://github.com/earendil-works/pi/issues/9423)).
|
|
1342
|
+
- Removed GPT-5.4 and GPT-5.4 mini from the OpenAI Codex catalog after they became unavailable to ChatGPT accounts ([#9394](https://github.com/earendil-works/pi/issues/9394)).
|
|
1343
|
+
- Fixed Mistral-hosted GLM-5.2 reasoning requests to use `reasoning_effort` instead of the ignored `prompt_mode` ([#9375](https://github.com/earendil-works/pi/issues/9375)).
|
|
1344
|
+
|
|
1345
|
+
## [0.9.19-alpha.3] - 2026-09-09
|
|
1346
|
+
|
|
1347
|
+
### Fixed
|
|
1348
|
+
|
|
1349
|
+
- Capped shared assistant retry backoff with `RetryPolicy.maxAgentDelayMs` (60 seconds by default), including summary calls ([#8826](https://github.com/earendil-works/pi/issues/8826)).
|
|
1350
|
+
|
|
1351
|
+
## [0.9.19-alpha.2] - 2026-09-08
|
|
1352
|
+
|
|
1353
|
+
### Fixed
|
|
1354
|
+
|
|
1355
|
+
- Fixed Anthropic Messages tool requests advertising root object-union parameter schemas without fields by projecting their branch fields into Anthropic-compatible object schemas ([#2190](https://github.com/bastani-inc/atomic/pull/2190) by [@elefthei](https://github.com/elefthei)).
|
|
1356
|
+
- Fixed Anthropic object-union tool projection dropping own property names such as `__proto__` and incorrectly projecting explicitly nonobject branches with `properties` metadata; authored schemas and runtime validation remain unchanged ([#2190](https://github.com/bastani-inc/atomic/pull/2190), [#2189](https://github.com/bastani-inc/atomic/issues/2189)).
|
|
1357
|
+
- Updated Anthropic, Bedrock, Google and HTTP proxy SDKs to upstream Pi's current versions, and mapped Google's consecutive tool-call limit to an error stop reason.
|
|
1358
|
+
- Routed all GitHub Copilot GPT model IDs through the Responses API ([#9209](https://github.com/earendil-works/pi/issues/9209)).
|
|
1359
|
+
- Fixed quadratic CPU usage when draining buffered `EventStream` events ([#9055](https://github.com/earendil-works/pi/issues/9055)).
|
|
1360
|
+
- Fixed reasoning-capable Mistral Medium aliases using unsupported `prompt_mode` instead of `reasoning_effort` ([#8700](https://github.com/earendil-works/pi/issues/8700)).
|
|
1361
|
+
|
|
1362
|
+
## [0.9.18] - 2026-09-05
|
|
1363
|
+
|
|
1364
|
+
Cumulative release of the `0.9.18-alpha.5` through `0.9.18-alpha.7` prereleases. Per-change details remain in the unchanged prerelease sections below.
|
|
1365
|
+
|
|
1366
|
+
### Breaking Changes
|
|
1367
|
+
|
|
1368
|
+
- `createGatewayBindingFetch` requires `binding.fetch()` and forwards requests verbatim. The `gateway(id).run(...)` fallback is removed, `baseUrl` and `gateway` options are ignored, and bindings without `fetch()` fail at construction. Set model base URLs to `https://workers-binding.ai/ai-gateway/gateways/{gateway}/{provider}` ([#8287](https://github.com/earendil-works/pi/pull/8287)).
|
|
1369
|
+
- Fast inference is explicit `Model.fastRoute` metadata, not a name suffix. Responses adapters accept the canonical model, send its route's upstream ID, enforce its declared tier even against caller options, and price priority using its base model. A route without a tier sends none. Payload hooks that replace the body with a non-object or change the route's model or tier now fail. Models without routes retain caller-controlled tiers and hooks.
|
|
1370
|
+
- GitHub Copilot exposes route-bearing fast models only when OAuth `fastModelIds` advertises the exact ID. Provider-owned names ending in `-fast` remain ordinary models governed by `availableModelIds`.
|
|
1371
|
+
|
|
1372
|
+
### Added
|
|
1373
|
+
|
|
1374
|
+
- Added Claude Fable 5.1 with provider-specific pricing, a 1M-token context window, 128K output, always-on adaptive thinking, and exactly `low`, `medium`, `high`, `xhigh`, and `max` efforts. Added its server-side fallback targets, Claude Opus 4.8 and Opus 5, and compatibility controls for thinking binding, forced tool choice, and temperature.
|
|
1375
|
+
- Added Gemini 3.8 Flash across supported provider catalogs, including GitHub Copilot. Google and Vertex use a 1,048,576-token context window and 65,536 output tokens with low, medium, and high thinking; other providers retain their own metadata. Added Copilot Claude Fable 5/5.1 and Baseten `zai-org/GLM-5.3-Fast`.
|
|
1376
|
+
- Added GPT-6-Astra on OpenAI, OpenAI Codex, Bedrock, OpenRouter, and Vercel AI Gateway, retaining provider-owned IDs, reasoning levels, pricing, and request-wide long-context tiers. Added a provisional Copilot entry with zero costs for unverified pricing; account availability still depends on Copilot's authenticated catalog.
|
|
1377
|
+
- Added OpenRouter `microsoft/mai-image-2.6` and `microsoft/mai-image-2.6-flash` image models.
|
|
1378
|
+
- Added public `DocumentContent` PDF input for Anthropic Messages and Bedrock Converse. Unsupported models receive a visible placeholder; non-PDF MIME types are rejected, and token estimates measure the encoded document payload.
|
|
1379
|
+
- Added public `FallbackContent`, per-turn Anthropic effort persistence, historical effort markers, signed-thinking recovery, and diagnostics for thinking blocks dropped at request start or after server-side fallback.
|
|
1380
|
+
- Added `ModelFastRoute`, `resolveRequestedServiceTier`, and `assertPayloadPreservesFastRoute`, plus `AssistantMessageFrameEncoder` and `reduceAssistantMessageFrames` for compact, replayable stream progress.
|
|
1381
|
+
- Added `compat.vllmPriority` for servers using priority scheduling, `supportsMaxOutputTokens` for Responses gateways that reject the output cap, and compatibility flags for unsupported temperature and forced-tool-choice fields.
|
|
1382
|
+
|
|
1383
|
+
### Changed
|
|
1384
|
+
|
|
1385
|
+
- Exported `./utils/*` subpaths for Pi 0.85 runtime packages.
|
|
1386
|
+
- Qwen3.8 Max and Flash thinking choices follow models.dev on Qwen Token Plan providers, offering low, medium, and xhigh without off.
|
|
1387
|
+
- Responses models supporting `prompt_cache_options`, including Astra, use `ttl: "30m"` for long retention. Earlier models keep the 24-hour legacy field; short and explicit no-cache modes retain capability checks.
|
|
1388
|
+
- Assistant-frame reduction reconstructs owned block values rather than deleting fields on externally supplied content, preserving output and field order.
|
|
1389
|
+
|
|
1390
|
+
### Fixed
|
|
1391
|
+
|
|
1392
|
+
- Fixed Claude Fable temperature rejection across Anthropic, Bedrock, OpenRouter, and Copilot, including sampling defaults that reintroduced `temperature`, `top_p`, or `top_k`. Forced tool choices on Fable 5.1 now fail explicitly across supported APIs, including all OpenAI forcing shapes and Bedrock requests without tools; auto and none remain unchanged.
|
|
1393
|
+
- Fixed first-party Anthropic model switches and Fable 5.1 prefix changes by preserving eligible signed reasoning and letting the API drop incompatible blocks. Disabled thinking no longer sends the interleaved-thinking beta, and recovery diagnostics survive empty later reports or failed streams.
|
|
1394
|
+
- Fixed mid-stream Anthropic fallback attribution, replay, and billing. Handoff markers survive, serving-model prices apply, earlier output-producing attempts are charged once, and pre-handoff reasoning and unexecuted calls/results are dropped together. Faux-provider replay keeps content indices aligned around fallback markers.
|
|
1395
|
+
- Fixed Fable 5.1 OAuth requests rejected as `claude_code_version_too_old` by advertising the verified minimum Claude Code version, `2.1.251`.
|
|
1396
|
+
- Fixed Copilot Claude Fable routing through Anthropic Messages and all Fireworks GLM models through OpenAI-compatible completions. Removed retired xAI `grok-build-0.1` and restored Qwen3.8 Flash on Qwen Token Plan Individual.
|
|
1397
|
+
- Fixed Codex SSE terminal events without a trailing blank line and assistant-frame start snapshots losing `providerThinkingLevel`.
|
|
1398
|
+
- Fixed `NO_PROXY` matching for root domains, subdomains, IPv6, port-scoped entries, and wildcard entries.
|
|
1399
|
+
|
|
1400
|
+
## [0.9.18-alpha.7] - 2026-09-05
|
|
1401
|
+
|
|
1402
|
+
### Added
|
|
1403
|
+
|
|
1404
|
+
- Added OpenRouter's `microsoft/mai-image-2.6` and `microsoft/mai-image-2.6-flash` image models from the Pi 0.85.1 catalog.
|
|
1405
|
+
|
|
1406
|
+
## [0.9.18-alpha.6] - 2026-09-04
|
|
1407
|
+
|
|
1408
|
+
### Breaking Changes
|
|
1409
|
+
|
|
1410
|
+
- **Breaking:** `createGatewayBindingFetch` now requires `binding.fetch()` and always forwards requests verbatim. The `gateway(id).run(...)` universal-endpoint fallback is removed. `baseUrl` and `gateway` options are ignored. Bindings without `fetch()` throw at construction. Point each model's `baseUrl` at `https://workers-binding.ai/ai-gateway/gateways/{gateway}/{provider}`.
|
|
1411
|
+
|
|
1412
|
+
### Added
|
|
1413
|
+
|
|
1414
|
+
- Added Anthropic per-turn effort persistence, deterministic historical effort markers, and signed-thinking mismatch recovery for supported Claude models across Anthropic Messages transports, including OpenRouter.
|
|
1415
|
+
- Added `compat.vllmPriority` to the OpenAI Completions compatibility options. When set, it is sent as the top-level `priority` request field; lower values are handled earlier and the vLLM server default is `0`, so it only takes effect when vLLM runs with `--scheduling-policy priority`. Setting it on a background or batch model defers that model's long prefills behind interactive sessions. Off by default and never set on the generated catalog ([#9004](https://github.com/earendil-works/pi/pull/9004)).
|
|
1416
|
+
- `Model.fastRoute` and the `ModelFastRoute` type. This optional field marks a model as the fast-inference variant of another model and carries the upstream routing it needs: the base model it pairs with, the model ID to send upstream, and an optional OpenAI-style `serviceTier`. Presence of this field — never a `-fast` name suffix — is what gives a model fast semantics.
|
|
1417
|
+
- Added `AssistantMessageFrameEncoder` and `reduceAssistantMessageFrames` so callers can persist compact, replayable stream progress without the full event stream ([upstream `8b5899dc`](https://github.com/earendil-works/pi/commit/8b5899dce26f9f6b8d313ee6a4b4a8dccbb9bfc2)).
|
|
1418
|
+
- Added Gemini 3.8 Flash to the generated provider catalogs. `google` and `google-vertex` publish `gemini-3.8-flash` with a 1,048,576-token context window, 65,536 maximum output tokens, text and image input, and $0.75 input / $3.75 output / $0.075 cache read per million tokens. Thinking is always on, so `off` is denied as on Gemini 3.6 and 3.7 Flash, and `minimal` is denied too because Google publishes only `LOW`, `MEDIUM`, and `HIGH` for this model and states that `MINIMAL` is unsupported. That denial is scoped to 3.8: Gemini 3.5 and 3.6 Flash do publish `MINIMAL` and keep offering it. The model also lands on `opencode`, `openrouter` (`google/gemini-3.8-flash` and `:batch`), and `vercel-ai-gateway` (`google/gemini-3.8-flash`) from those providers' own catalogs ([#9076](https://github.com/earendil-works/pi/issues/9076)).
|
|
1419
|
+
- Added `github-copilot/gemini-3.8-flash` from the models.dev Copilot catalog. Atomic consumes that row without supplementing or overriding it; the current metadata provides a 1,000,000-token context window, 64,000 maximum output tokens, text and image input, and `reasoning_effort` values of `low`, `medium`, and `high`. The model routes through Copilot's OpenAI-compatible completions endpoint with Copilot's static headers.
|
|
1420
|
+
- Added `github-copilot/claude-fable-5` and `github-copilot/claude-fable-5-1`. The generator uses each model's models.dev metadata, falling back to the models.dev Anthropic row when the Copilot catalog has not published the same model yet; the authenticated Copilot picker still controls which models an account may select.
|
|
1421
|
+
- Added `zai-org/GLM-5.3-Fast` to the Baseten catalog from models.dev.
|
|
1422
|
+
- Added GPT-6-Astra to the generated OpenAI and OpenAI Codex catalogs with text and image input, 272,000 default input context, 128,000 maximum output, exact `low` through `max` reasoning efforts, tool search, additional tools, and the published $10 input / $1 cache read / $12.50 cache write / $50 output rates per million tokens. Requests above 272,000 aggregate input carry the published $20 / $2 / $25 / $75 request-wide tier. Amazon Bedrock also gains the exact Codex-advertised `openai.gpt-6-astra`, `global.openai.gpt-6-astra`, and `us.openai.gpt-6-astra` IDs, sends the selected reasoning effort through Converse, and uses zero catalog costs until AWS publishes Astra prices. The current OpenRouter catalog contributes `openai/gpt-6-astra` and `openai/gpt-6-astra-pro`; the Vercel AI Gateway contributes `openai/gpt-6-astra` and its provider-owned `openai/gpt-6-astra-fast`. Both dynamic catalogs retain their request-wide long-context prices. Azure OpenAI remains omitted because its authoritative catalog did not advertise Astra.
|
|
1423
|
+
- Added a provisional `github-copilot/gpt-6-astra` catalog entry using the Responses API, known Astra capabilities, and zero costs to mark unverified Copilot pricing. Copilot's own models.dev metadata takes precedence when present; account availability and fast sibling entitlement remain controlled by Copilot.
|
|
1424
|
+
|
|
1425
|
+
### Changed
|
|
1426
|
+
|
|
1427
|
+
- Exported `./utils/*` subpaths required by Pi 0.85 runtime packages.
|
|
1428
|
+
- **Breaking:** the **presence** of `Model.fastRoute` — not just its declared tier value — decides the service tier. A route that declares no tier now means *no tier*, so a GitHub Copilot fast variant emits no `service_tier` field and takes no tier cost multiplier whatever the caller passes. Previously a caller's option leaked through on such a route, which contradicted the documented Copilot contract. A model with no route still honors an explicit `serviceTier`.
|
|
1429
|
+
- **Breaking:** the OpenAI Responses and ChatGPT Codex Responses adapters now reject a payload hook that changes the `model` or `service_tier` of a model carrying a `fastRoute`, or replaces its request body with a non-object that cannot carry those fields. The error names the model, the violation, and the normal sibling to select instead. A hook could previously send a different model, an ordinary-tier request, or a malformed replacement under the `-fast` identity the caller recorded and is billed for. Hooks keep unrestricted freedom over every other field and over models with no route. New export: `assertPayloadPreservesFastRoute` from `api/openai-responses-shared`.
|
|
1430
|
+
- **Breaking:** a fast model variant's declared `service_tier` now wins over a per-request `serviceTier` option. Fast versus normal is model identity, so a caller who wants another tier selects the normal sibling; letting the option win allowed a model that is still selected, recorded, persisted, and billed as `-fast` to route as an ordinary request. A model with no route honors an explicit `serviceTier` exactly as before. Payload hooks may rewrite other fields, but cannot override a fast route's model or tier.
|
|
1431
|
+
- **Breaking:** the OpenAI Responses and ChatGPT Codex Responses adapters now default `service_tier` from `model.fastRoute?.serviceTier` when a request supplies no explicit `serviceTier` option, and use that resolved tier for serialization, response-tier resolution, and cost multiplication alike. A fast model variant therefore routes correctly through `Models.stream`/`complete`, `streamSimple`/`completeSimple`, and a provider's stream functions taken directly — previously only a caller that injected the option got the tier, so every other caller was billed and served at normal tier with no warning. New export: `resolveRequestedServiceTier` from `api/openai-responses-shared`.
|
|
1432
|
+
- **Breaking:** the OpenAI Responses and ChatGPT Codex Responses adapters now serialize `model` as `model.fastRoute?.upstreamModelId ?? model.id`. A fast variant therefore routes to its base upstream model while the model object — and so the assistant message the adapter emits, its `model` field, and everything downstream that reads it — keeps the canonical `-fast` identity. Callers that previously handed these adapters a pre-substituted model should hand them the canonical model instead.
|
|
1433
|
+
- **Breaking:** the `service_tier` cost multiplier in both adapters now keys on `model.fastRoute?.baseModelId ?? model.id`. Without this, `gpt-5.5-fast` would have been priced at the generic 2x priority rate instead of gpt-5.5's 2.5x once the adapters started receiving the canonical model.
|
|
1434
|
+
- **Breaking:** `githubCopilotProvider().filterModels` no longer strips every model ID ending in `-fast` from the selectable list. It now exposes a model carrying `fastRoute` only when the OAuth credential's `fastModelIds` advertises that exact ID, and treats a Copilot-owned model that merely ends in `-fast` as an ordinary picker model gated by `availableModelIds`.
|
|
1435
|
+
- Qwen3.8 Max and Qwen3.8 Flash no longer offer the `off` thinking level on the `qwen-token-plan`, `qwen-token-plan-cn`, and `qwen-token-plan-individual` providers. Their selectable effort levels now follow models.dev metadata: `low`, `medium`, and `xhigh`.
|
|
1436
|
+
- `reduceAssistantMessageFrames` now rebuilds an ended text, thinking, or tool-call block as a value the reducer owns, assembled from the end frame and the rest of the replayed block's own data, instead of deleting optional fields on a value read out of externally supplied frame content. Reduced output is unchanged for every input, including field order and any property the frame carried that the declared block types do not name.
|
|
1437
|
+
- OpenAI Responses models that accept `prompt_cache_options`, including GPT-6 Astra, now send `ttl: "30m"` for long prompt-cache retention instead of combining that API with the legacy `prompt_cache_retention: "24h"` field. Earlier Responses models keep the 24-hour field, explicit no-cache mode remains limited to capable models, and short retention omits both controls.
|
|
1438
|
+
|
|
1439
|
+
### Fixed
|
|
1440
|
+
|
|
1441
|
+
- Removed the retired `grok-build-0.1` preview from the generated xAI catalog.
|
|
1442
|
+
- Fixed native Anthropic Messages requests sending the interleaved-thinking beta when thinking was disabled, and preserved request-start thinking-drop diagnostics when a later provider report is empty or the stream fails.
|
|
1443
|
+
- Fixed GitHub Copilot Claude Fable requests to use the Anthropic Messages adapter so selected reasoning levels are sent. The generated Copilot catalog now routes `claude-fable-*` alongside the other Claude 4.x/5.x entries ([#8961](https://github.com/earendil-works/pi/issues/8961)).
|
|
1444
|
+
- Fixed the generated Fireworks catalog to serve every GLM model through the OpenAI-compatible completions API. Previously only the `glm-5p2` family took that route and newer GLM entries such as `glm-5p3` were generated against the Anthropic-compatible endpoint ([#8978](https://github.com/earendil-works/pi/issues/8978)).
|
|
1445
|
+
- Fast-route payload enforcement now rejects `null`, array, and primitive hook replacements before they can reach the OpenAI Responses or ChatGPT Codex transport.
|
|
1446
|
+
- Fixed the Workers AI binding transport to use the binding's plain `fetch` passthrough for models migrated to the `workers-binding.ai` base URL, preserving request methods, headers, query strings, and streaming bodies instead of translating requests through the universal-endpoint shim. See **Breaking Changes** for the required base-URL and binding-type migration ([#8287](https://github.com/earendil-works/pi/pull/8287)).
|
|
1447
|
+
- Fixed OpenAI Codex SSE parsing to process terminal events that are not followed by a blank line ([#9047](https://github.com/earendil-works/pi/issues/9047)).
|
|
1448
|
+
- Fixed the Qwen Token Plan Individual catalog to include Qwen3.8 Flash ([#9021](https://github.com/earendil-works/pi/issues/9021)).
|
|
1449
|
+
- Fixed assistant-message frames dropping `providerThinkingLevel` from the start snapshot ([upstream `0fdec07b`](https://github.com/earendil-works/pi/commit/0fdec07ba397)).
|
|
1450
|
+
|
|
1451
|
+
## [0.9.18-alpha.5] - 2026-09-01
|
|
1452
|
+
|
|
1453
|
+
### Added
|
|
1454
|
+
|
|
1455
|
+
- Added first-class Claude Fable 5.1 (`claude-fable-5-1`) support to the generated catalogs, for the providers Atomic has a matching runtime integration for. Each entry carries the model's 1,000,000-token context window, 128,000-token maximum output, and that provider's own pricing, including the reduced $0.25 per million cache read (a quarter of Claude Fable 5's rate) and the US-only inference premium on the Amazon Bedrock `us.` profile. Adaptive thinking is always on, `off` is denied, and exactly the five efforts Anthropic publishes — `low`, `medium`, `high`, `xhigh`, `max` — are selectable, with the `high` default preserved. Provider catalogs move independently of releases, so the exact mirror set tracks them rather than being fixed here; a provider "latest" alias may also route to Fable 5.1 without naming it.
|
|
1456
|
+
- Added the published server-side fallback targets for Claude Fable 5.1. Requests now send `fallbacks` naming Claude Opus 4.8 and Claude Opus 5, so a classifier refusal can be retried server-side and can redeem the prompt-cache fallback credit.
|
|
1457
|
+
- Added `compat.enforcesPreservedThinkingBinding`, `compat.delegatesThinkingModelBinding`, and `compat.supportsForcedToolChoice` to the Anthropic Messages compatibility options, and `compat.supportsForcedToolChoice` plus `compat.supportsTemperature` to the Amazon Bedrock ones, so a provider can declare that its API adjudicates thinking-block signatures, that its model runs Anthropic's conversation check, and whether its model accepts forced tool use or a `temperature`.
|
|
1458
|
+
- Added visibility for thinking blocks the Anthropic API drops. When a response reports `input_transformations`, the assistant message now carries an `anthropic_input_transformations` diagnostic with the dropped-block count, the reasons (`prefix_binding_mismatch` or `model_binding_mismatch`), and the block paths, instead of the drop being silent. Both report sites are read: `message_start`, and the final `message_delta` after a mid-stream server-side fallback.
|
|
1459
|
+
- Added `compat.supportsTemperature` to the OpenAI-completions compatibility options, mirroring the field the Anthropic Messages compat already had, so a model that rejects the `temperature` request field can be marked as such on that API too.
|
|
1460
|
+
- Added a public `FallbackContent` block (`{ type: "fallback", fromModel, toModel }`) to `AssistantMessage.content`. Anthropic's server-side fallback marks a mid-stream handoff with a `fallback` content block, and clients must echo it back in place because the API validates the surrounding thinking blocks against its position. It is public content rather than a stream-only event for exactly that reason.
|
|
1461
|
+
- Added PDF document input. `DocumentContent` is a new public content block on `UserMessage.content`, and `Model.input` can now contain `"pdf"`. The Anthropic Messages path emits a base64 `document` block and the Amazon Bedrock Converse path emits a `DocumentBlock` with decoded bytes and a sanitized, neutral `name`, as AWS requires. PDF is a platform capability rather than a per-model one — "All active models support PDF processing", routed through the same vision path as images — so `"pdf"` is gated on the **API** that can serialize a document, not on the model: Claude entries on Anthropic and Amazon Bedrock gain it, and every other mirror keeps `["text", "image"]`. A document sent to a model without it is replaced by a visible placeholder, and token estimation measures the encoded payload rather than counting it as an image. `DocumentContent.mimeType` is the literal `"application/pdf"` rather than a free `string`, because that is the only media type either serializer implements — both hardcode PDF instead of reading the field — and a request carrying any other value is now rejected by name instead of being sent mislabelled. Widening that union later is non-breaking; adding a second format is not a media-type swap, since Anthropic needs a structurally different source variant for plain text and has no variant at all for Bedrock's office and markup formats.
|
|
1462
|
+
|
|
1463
|
+
### Fixed
|
|
1464
|
+
|
|
1465
|
+
- `NO_PROXY` now matches a root domain and its subdomains consistently, and understands bracketed IPv6 hosts, bare IPv6 hosts, port-scoped entries, and a bare `*` entry anywhere in the list. Previously an entry such as `example.com` did not exempt `api.example.com`, and `[2001:db8::1]` was parsed as a host/port pair ([#8737](https://github.com/earendil-works/pi/pull/8737)).
|
|
1466
|
+
- Added a `supportsMaxOutputTokens` compatibility flag to `OpenAIResponsesCompat` (default `true`). Setting it to `false` omits `max_output_tokens` from openai-responses requests, for Codex-protocol gateways that reject the parameter with a 400 ([#8941](https://github.com/earendil-works/pi/pull/8941)).
|
|
1467
|
+
- Fixed Claude Fable 5 and Claude Fable 5.1 sending `temperature` on the Anthropic Messages API. Anthropic rejects non-default `temperature`, `top_p`, and `top_k` on both models with a 400 on every request, whether or not thinking is on, so both are now generated with `supportsTemperature: false` and the field is omitted.
|
|
1468
|
+
- Fixed Claude Fable 5 and Claude Fable 5.1 still sending `temperature` on Amazon Bedrock, OpenRouter, and GitHub Copilot. The Bedrock Converse and OpenAI-completions request builders emitted the field unconditionally, so every Claude Fable mirror on those APIs sent a parameter the model rejects — including OpenRouter's `~anthropic/claude-fable-latest` alias, whose id names no version. All Claude Fable entries on those two APIs are now generated with `supportsTemperature: false`, matching what the Anthropic Messages path already did for the same family, and both request builders honour it. OpenRouter's own `supported_parameters` for those entries also omits `temperature`. Models outside the Claude Fable family are unchanged.
|
|
1469
|
+
- Fixed Claude Fable 5.1 sessions failing with a 400 `invalid_request_error` when a thinking block was replayed behind a changed conversation prefix. Anthropic enforces that check by default for organizations created on or after 2026-08-31, so a dynamic system prompt, a tool-set change, or a model switch could break an otherwise healthy session. Requests for that model now send the `thinking-binding-controls-2026-08-01` beta header with `thinking.block_binding.prefix_mismatch_behavior: "drop_block"`, so the API drops the affected thinking blocks and answers the turn. The field accompanies the header on every request for that model, including turns with no reasoning level, where the header alone would have left the behavior at its rejecting `"error"` default.
|
|
1470
|
+
- Fixed mid-conversation model switches discarding reasoning that the target model was allowed to read. First-party Anthropic models now replay another Claude model's signed `thinking` and `redacted_thinking` blocks unchanged and let the API decide which ones the target model may use, instead of rewriting them into visible assistant text. Switching up to Claude Fable 5.1 keeps the conversation's reasoning; switching down to an earlier Claude model has the API drop it server-side. Assistant text, tool calls, and tool results are preserved in both directions. The change is scoped to `provider: "anthropic"` on the `anthropic-messages` API; Amazon Bedrock, Google Vertex, and Anthropic-compatible proxies are unchanged.
|
|
1471
|
+
- Fixed forced tool use being sent to Claude Fable 5.1, which rejects it on every request, whichever platform serves it. The Anthropic, Amazon Bedrock, and OpenAI-completions entry points all accept a forced tool choice, so a caller could construct a request the model cannot honor. A model marked `supportsForcedToolChoice: false` now **rejects the request with an error** naming the model and Anthropic's documented remedy, rather than sending it. An earlier revision of this branch silently substituted `{"type": "auto"}`; that discarded an explicit caller instruction and is no longer done. On a gateway the old behavior was worse than a provider error, because OpenRouter drops parameters a model does not support, so the forced choice vanished and the response looked normal. `auto` and `none` are never altered, and every other model passes forced choices through unchanged. The guard is keyed on the versioned Claude Fable 5.1 model id across all three APIs, rather than on `provider: "anthropic"`, because the rejection is a property of the model rather than of Anthropic's endpoint; it therefore covers every mirror a provider currently publishes without naming a count, since those catalogs move. It is deliberately **not** applied to Claude Fable 5, which accepts forced tool use, nor to an unversioned "latest" alias whose target cannot be known from its id. On the OpenAI-completions path the check now covers all four forcing shapes in OpenAI's wider tool-choice union — `"required"`, a named `function`, a named `custom` tool, and `allowed_tools` with `mode: "required"` — where it previously tested only the first two. `allowed_tools` with `mode: "auto"` constrains the candidate set rather than forcing a call and still passes through untouched.
|
|
1472
|
+
- Fixed Claude Fable 5.1 offering a `minimal` thinking level that Anthropic does not publish for it. The generated thinking-level map was sparse on the Anthropic, opencode, Vercel AI Gateway, and Amazon Bedrock entries, and an unmapped level is treated as available, so six levels were selectable instead of five. `minimal` is now denied explicitly, leaving exactly `low`, `medium`, `high`, `xhigh`, and `max`. The OpenRouter entries already carried a full map and are unaffected. Claude Fable 5 has the same sparse map and is deliberately left alone here.
|
|
1473
|
+
- Fixed `samplingParams` reopening the sampling fields a model rejects. On the OpenAI-completions adapter that object is merged last so its keys override the named request fields, which meant a caller could reintroduce `temperature` on a model generated with `supportsTemperature: false` — and `top_p` and `top_k`, which are never named fields there, could only ever arrive that way. Those three keys are now stripped after the merge, so a model-level `samplingParams` default is covered too. Precedence is otherwise unchanged: every other custom key still overrides, and models without the restriction are untouched.
|
|
1474
|
+
- Fixed mid-stream server-side fallback being silently discarded. When a classifier declined partway through a response, the `fallback` content block marking the handoff was dropped on the floor by the streaming dispatch, the message stayed attributed to the requested model, and usage was therefore costed at that model's rates instead of the model that actually produced the answer. The marker is now recorded in place, the message is re-attributed to the serving model, and cost is recomputed from the fallback target's own rates. Replay follows Anthropic's documented rules: the marker is echoed exactly where it appeared, **including onto the wire**, and `thinking`, `redacted_thinking`, and unexecuted client-side tool calls before it are dropped, because a request that echoes thinking from both sides of the boundary is rejected. A tool result whose call was dropped at the boundary is dropped with it, since an unmatched `tool_result` is rejected in turn. A decline before any output and a sticky-routed later turn were already correct and are unchanged.
|
|
1475
|
+
- Fixed earlier fallback attempts being billed as free. Anthropic bills every attempt that produced output at the rates of the model that ran it, and reports them in `usage.iterations`, while the top-level usage describes only the attempt that produced the returned message. Those earlier attempts are now added to `usage.cost` at their own model's rates. An attempt that declined before producing any output is not billed, matching the documented rule, and the serving attempt is not counted twice. Token counts are deliberately not summed across models, which the documentation forbids.
|
|
1476
|
+
- Fixed the Amazon Bedrock forced-tool-choice guard running after the early returns in `convertToolConfig`, so a forced choice paired with an empty or absent tool list was discarded in silence instead of rejected. The validation now runs before both returns. `toolChoice: "none"` is unaffected, since `none` is never a forced choice.
|
|
1477
|
+
- Fixed Anthropic OAuth requests for Claude Fable 5.1 being rejected with a 400 `claude_code_version_too_old`. The OAuth path impersonates Claude Code, and Anthropic gates newer models on the `claude-cli/<version>` user agent alone; the pinned `2.1.75` predated Fable 5.1's floor, so every OAuth turn for that model failed with "version 2.1.251 or newer is required" regardless of the betas or other identity headers sent. The advertised version is now `2.1.251`, the exact published minimum the API accepts, bisected against the live endpoint rather than read off the error text. It is a strict capability superset of the old value: every Claude model this package ships answers normally at the new version. API-key requests are unaffected, since only the OAuth branch sends a Claude Code user agent.
|
|
1478
|
+
- Fixed a `fallback` content block corrupting the content indices that follow it when a response is replayed through the faux provider. The block was skipped without occupying its slot in the streamed `partial.content`, so every later block was addressed one position short: a `[text, fallback, text]` message aborted the stream with an error and lost its second text block, and a tool call after a fallback was announced at an index holding something else. The marker now occupies its slot and simply receives no stream event, matching how the real Anthropic path already records it, so `contentIndex` continues to address `partial.content` directly for consumers that index it.
|
|
1479
|
+
|
|
1480
|
+
|
|
1481
|
+
## [0.9.16] - 2026-08-29
|
|
1482
|
+
|
|
1483
|
+
Cumulative release of the `0.9.16-alpha.5` – `0.9.16-alpha.11` prereleases. The summary below covers the user-visible outcome of that work; the per-change detail remains in the prerelease sections below.
|
|
1484
|
+
|
|
1485
|
+
### Added
|
|
1486
|
+
|
|
1487
|
+
- Added the Meta Muse Image and Recraft V4 Styles, V4 Styles Pro, V4 Styles Pro Vector, and V4 Styles Vector image models to the generated OpenRouter catalog.
|
|
1488
|
+
- Added the DeepSeek V4 Flash Vision experimental model to generated catalogs.
|
|
1489
|
+
|
|
1490
|
+
### Fixed
|
|
1491
|
+
|
|
1492
|
+
- Fixed fragmented Mistral tool calls splitting when continuation chunks omit the tool-call ID ([#8387](https://github.com/earendil-works/pi/issues/8387)).
|
|
1493
|
+
- Fixed Cloudflare AI Gateway typing and mirrored supported Workers AI models through its compatibility endpoint.
|
|
1494
|
+
- Fixed OpenAI-compatible Chat Completions reasoning streams to concatenate incremental reasoning deltas instead of replacing earlier content ([#8605](https://github.com/earendil-works/pi/pull/8605)).
|
|
1495
|
+
- Fixed OpenRouter reasoning controls by deriving `off` support and available effort levels from model metadata, preventing reasoning-mandatory models from receiving `effort: "none"` ([#8454](https://github.com/earendil-works/pi/issues/8454)).
|
|
1496
|
+
- Fixed OpenAI-compatible Chat Completions requests sending `tool_choice` without tools, which gateways can reject during compaction ([#8607](https://github.com/earendil-works/pi/issues/8607)).
|
|
1497
|
+
- Fixed OpenAI-compatible Chat Completions ignoring an explicitly requested `toolChoice` when no tools are defined ([#8649](https://github.com/earendil-works/pi/issues/8649), [#8638](https://github.com/earendil-works/pi/issues/8638)).
|
|
1498
|
+
- Fixed OpenAI-compatible streaming rewriting `thinkingSignature` on every `reasoning_details` delta. Replay metadata is buffered during streaming and serialized once when the thinking block is finalized, including on the error path ([#8671](https://github.com/earendil-works/pi/issues/8671)).
|
|
1499
|
+
|
|
1500
|
+
## [0.9.16-alpha.11] - 2026-08-28
|
|
1501
|
+
|
|
1502
|
+
### Added
|
|
1503
|
+
|
|
1504
|
+
- Added the Meta Muse Image and Recraft V4 Styles, V4 Styles Pro, V4 Styles Pro Vector, and V4 Styles Vector image models to the generated OpenRouter catalog.
|
|
1505
|
+
|
|
1506
|
+
## [0.9.16-alpha.10] - 2026-08-28
|
|
1507
|
+
|
|
1508
|
+
### Fixed
|
|
1509
|
+
|
|
1510
|
+
- Fixed fragmented Mistral tool calls splitting when continuation chunks omit the tool-call ID ([#8387](https://github.com/earendil-works/pi/issues/8387)).
|
|
1511
|
+
|
|
1512
|
+
## [0.9.16-alpha.5] - 2026-08-26
|
|
1513
|
+
|
|
1514
|
+
### Added
|
|
1515
|
+
|
|
1516
|
+
- Added the DeepSeek V4 Flash Vision experimental model to generated catalogs.
|
|
1517
|
+
|
|
1518
|
+
### Fixed
|
|
1519
|
+
|
|
1520
|
+
- Fixed Cloudflare AI Gateway typing and mirrored supported Workers AI models through its compatibility endpoint.
|
|
1521
|
+
- Fixed OpenAI-compatible Chat Completions reasoning streams to concatenate incremental reasoning deltas instead of replacing earlier content ([#8605](https://github.com/earendil-works/pi/pull/8605)).
|
|
1522
|
+
- Fixed OpenRouter reasoning controls by deriving `off` support and available effort levels from model metadata, preventing reasoning-mandatory models from receiving `effort: "none"` ([#8454](https://github.com/earendil-works/pi/issues/8454)).
|
|
1523
|
+
- Fixed OpenAI-compatible Chat Completions requests sending `tool_choice` without tools, which gateways can reject during compaction ([#8607](https://github.com/earendil-works/pi/issues/8607)).
|
|
1524
|
+
- Fixed OpenAI-compatible Chat Completions ignoring an explicitly requested `toolChoice` when no tools are defined ([#8649](https://github.com/earendil-works/pi/issues/8649), [#8638](https://github.com/earendil-works/pi/issues/8638)).
|
|
1525
|
+
- Fixed OpenAI-compatible streaming rewriting `thinkingSignature` on every `reasoning_details` delta. Replay metadata is buffered during streaming and serialized once when the thinking block is finalized, including on the error path ([#8671](https://github.com/earendil-works/pi/issues/8671)).
|
|
1526
|
+
|
|
1527
|
+
## [0.9.15] - 2026-08-21
|
|
1528
|
+
|
|
1529
|
+
Cumulative release of the `0.9.15-alpha.1` prerelease. The summary below covers the user-visible outcome of that work; the per-change detail remains in the prerelease section below.
|
|
1530
|
+
|
|
1531
|
+
### Breaking Changes
|
|
1532
|
+
|
|
1533
|
+
- Built-in xAI models now use the OpenAI Responses API only; Completions is no longer registered on the xAI provider.
|
|
1534
|
+
- Renamed the exported `GoogleThinkingLevel` type to `GoogleApiThinkingLevel` and added `ResolvedGoogleThinkingLevel` for normalized adapter levels.
|
|
1535
|
+
|
|
1536
|
+
### Added
|
|
1537
|
+
|
|
1538
|
+
- Added provider-neutral `toolChoice` on simple stream requests.
|
|
1539
|
+
- Generalized OpenAI Completions thinking-token budget fields through `thinkingTokenBudgetField`, with `supportsThinkingTokenBudget` retained as the vLLM alias.
|
|
1540
|
+
- Added pi's runtime `User-Agent` to Anthropic, Azure Responses, Google, Vertex, Mistral, OpenAI Completions, and OpenAI Responses requests.
|
|
1541
|
+
|
|
1542
|
+
### Changed
|
|
1543
|
+
|
|
1544
|
+
- `@bastani/pi-ai` now generates its models.dev catalog during package builds and uses the generated catalog for offline compilation.
|
|
1545
|
+
- Vendored and rebranded the package as `@bastani/pi-ai` in the Atomic monorepo, with later tagged Atomic releases publishing it through trusted publishing.
|
|
1546
|
+
- Moved `COPILOT_GITHUB_TOKEN` env-token host routing into the exported `@bastani/pi-ai/providers/github-copilot-env` module ([#2522](https://github.com/bastani-inc/atomic/issues/2522)).
|
|
1547
|
+
|
|
1548
|
+
### Fixed
|
|
1549
|
+
|
|
1550
|
+
- Fixed raw `COPILOT_GITHUB_TOKEN` chat authentication by sending `Copilot-Integration-Id: copilot-developer-cli`, while preserving exchanged OAuth-token behavior ([#2522](https://github.com/bastani-inc/atomic/issues/2522)).
|
|
1551
|
+
- Fixed Amazon Bedrock Converse Stream and `pi-messages` dropping static custom-model headers; caller headers still override model headers and `null` still suppresses a static header.
|
|
1552
|
+
- Fixed stalled OpenAI Completions, OpenAI Responses, Anthropic Messages, `pi-messages`, Mistral, and Codex provider streams by applying idle deadlines that abort the request and surface retryable transport errors. Native Codex WebSocket streams now use `streamDeadlineMs` rather than the HTTP timeout ([#2553](https://github.com/bastani-inc/atomic/issues/2553)).
|
|
1553
|
+
- Ported upstream fixes for Copilot throttling, Kimi cache reads, DeepSeek thinking levels, Google thinking maps, Bedrock headers and reasoning, model catalogs, Anthropic fallback pricing, Azure tool choice, Z.AI effort metadata, and OpenAI reasoning replay. Cerebras now defaults to `gpt-oss-120b`, and the unused OpenTelemetry dependency was removed.
|
|
1554
|
+
|
|
1555
|
+
## [0.9.15-alpha.1] - 2026-08-21
|
|
1556
|
+
|
|
1557
|
+
### Breaking Changes
|
|
1558
|
+
|
|
1559
|
+
- Built-in xAI models now use the OpenAI Responses API only. Completions is no longer registered on the xAI provider.
|
|
1560
|
+
|
|
1561
|
+
- Renamed the exported `GoogleThinkingLevel` type to `GoogleApiThinkingLevel` and added `ResolvedGoogleThinkingLevel` for normalized adapter levels.
|
|
1562
|
+
|
|
1563
|
+
### Added
|
|
1564
|
+
|
|
1565
|
+
- Added provider-neutral `toolChoice` on simple stream requests.
|
|
1566
|
+
- Generalized OpenAI Completions thinking-token budget fields (`thinkingTokenBudgetField`, with `supportsThinkingTokenBudget` as the vLLM alias).
|
|
1567
|
+
- Send pi's runtime `User-Agent` on Anthropic, Azure Responses, Google, Vertex, Mistral, OpenAI Completions, and OpenAI Responses requests.
|
|
1568
|
+
|
|
1569
|
+
### Changed
|
|
1570
|
+
|
|
1571
|
+
- `@bastani/pi-ai` `build` now runs `generate-models` then `build:offline`, matching upstream. Catalog JSON under `src/providers/data/` is generated from models.dev at build time and is no longer committed.
|
|
1572
|
+
|
|
1573
|
+
- Vendored into the Atomic monorepo as `packages/ai` and rebranded the published package to `@bastani/pi-ai`. The first npm version must be published by hand so trusted publishing can be attached; later tagged Atomic releases publish it from `publish.yml`.
|
|
1574
|
+
|
|
1575
|
+
- Moved `COPILOT_GITHUB_TOKEN` env-token host routing into the exported `@bastani/pi-ai/providers/github-copilot-env` module ([#2522](https://github.com/bastani-inc/atomic/issues/2522)).
|
|
1576
|
+
|
|
1577
|
+
### Fixed
|
|
1578
|
+
|
|
1579
|
+
- Fixed raw `COPILOT_GITHUB_TOKEN` Copilot chat authentication by sending `Copilot-Integration-Id: copilot-developer-cli`; exchanged OAuth tokens containing a `tid=` segment retain their existing behavior ([#2522](https://github.com/bastani-inc/atomic/issues/2522)).
|
|
1580
|
+
- Fixed the Amazon Bedrock Converse Stream and `pi-messages` APIs dropping a custom model's static `headers`: both now merge `model.headers` beneath caller `options.headers`, matching every other API implementation, with `null` caller values still suppressing a static header.
|
|
1581
|
+
- Fixed GitHub Copilot requests on the default `transport: "auto"` hanging forever after a response-body decompression failure such as `Library error: zlib error: incorrect header check`. The stalled body never rejected the adapter's async iterator, so the attempt never settled and retry/model fallback never advanced. Provider streams for the `openai-completions`, `openai-responses`, and `anthropic-messages` APIs are now wrapped in an idle stream deadline that closes the source iterator and settles the attempt as a retryable transport error, and the transient-error classifier recognises zlib, `incorrect header check`, decompression, and `Library error:` wrapper text ([#2553](https://github.com/bastani-inc/atomic/issues/2553)).
|
|
1582
|
+
- Fixed the idle stream deadline leaving the stalled provider request open: expiry now aborts an attempt-local signal, combined with the caller's signal, so the underlying HTTP request and socket are torn down before retry or model fallback starts instead of accumulating abandoned connections. The caller's own signal semantics are unchanged, the deadline error still surfaces as the retryable transport failure, and deadlines above the platform's 32-bit timer limit no longer clamp to an immediate timeout ([#2553](https://github.com/bastani-inc/atomic/issues/2553)).
|
|
1583
|
+
- Fixed idle stream deadlines for the native `pi-messages`, Mistral conversations, and Codex SSE transports: each now threads the deadline-owned abort signal through the request and bounds the decoded event loop, so a stalled response settles and closes its body instead of remaining pending ([#2553](https://github.com/bastani-inc/atomic/issues/2553)).
|
|
1584
|
+
- Fixed the native Codex `transport: "auto"` WebSocket path using the HTTP `timeoutMs` as its idle limit: connected streams now use the effective `streamDeadlineMs`, so a configured stream deadline triggers idle recovery without waiting for the unrelated HTTP timeout ([#2553](https://github.com/bastani-inc/atomic/issues/2553)).
|
|
1585
|
+
- Ported unreleased `@earendil-works/pi-ai` `fix(ai)` commits from `earendil-works/pi` main after v0.84.2:
|
|
1586
|
+
- GitHub Copilot login now updates only known tool-capable models with unconfigured policies, runs those updates sequentially, and retries throttled `/models` and policy requests within a bounded delay ([#7850](https://github.com/earendil-works/pi/issues/7850), [#8254](https://github.com/earendil-works/pi/pull/8254)).
|
|
1587
|
+
- Kimi OpenAI-compatible usage now treats top-level `cached_tokens` as cache reads ([#8119](https://github.com/earendil-works/pi/pull/8119), [#8075](https://github.com/earendil-works/pi/issues/8075)).
|
|
1588
|
+
- DeepSeek V4 Flash on OpenCode and OpenCode Go exposes a `low` thinking level ([#8181](https://github.com/earendil-works/pi/pull/8181)).
|
|
1589
|
+
- Google Generative AI and Vertex AI honor `thinkingLevelMap` on custom models ([#8135](https://github.com/earendil-works/pi/issues/8135)).
|
|
1590
|
+
- Amazon Bedrock `onResponse` forwards raw Smithy response headers ([#8243](https://github.com/earendil-works/pi/pull/8243), [#8234](https://github.com/earendil-works/pi/issues/8234)).
|
|
1591
|
+
- Xiaomi catalog generation drops shut-down MiMo V2 model names ([#8187](https://github.com/earendil-works/pi/issues/8187)).
|
|
1592
|
+
- China ZAI Coding Plan uses the `zhipuai-coding-plan` catalog, including GLM-4.6V, and PAYG-equivalent usage estimates ([#8220](https://github.com/earendil-works/pi/issues/8220)).
|
|
1593
|
+
- Qwen Token Plan Individual includes `deepseek-v4-pro-0813` ([#8194](https://github.com/earendil-works/pi/issues/8194)).
|
|
1594
|
+
- Anthropic server-side refusal fallbacks are declared in model metadata and priced from the returned fallback model, not stream options ([#8258](https://github.com/earendil-works/pi/pull/8258), [#8319](https://github.com/earendil-works/pi/pull/8319), [#8352](https://github.com/earendil-works/pi/pull/8352), [#8285](https://github.com/earendil-works/pi/issues/8285)).
|
|
1595
|
+
- Azure OpenAI Responses forwards `toolChoice`.
|
|
1596
|
+
- Baseten GLM-5.2 declares image input.
|
|
1597
|
+
- Bedrock round-trips non-Anthropic redacted reasoning ([#8314](https://github.com/earendil-works/pi/pull/8314)).
|
|
1598
|
+
- Z.AI reasoning effort metadata is derived from models.dev options, preserving GLM-5.2 `none` and exposing GLM-5.3 low/high/max ([#8336](https://github.com/earendil-works/pi/issues/8336)).
|
|
1599
|
+
- OpenAI-compatible Chat Completions preserves and resends assistant-level `reasoning_details` via `thinkingSignature` ([#8246](https://github.com/earendil-works/pi/pull/8246), [#7994](https://github.com/earendil-works/pi/issues/7994)).
|
|
1600
|
+
- Removed the unused `@opentelemetry/api` dependency from `@bastani/pi-ai`.
|
|
1601
|
+
- Cerebras default model is now `gpt-oss-120b`.
|