alexion/dev
1
0 Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

chore(vendor): import pi 1.3.0

Browse Source 
Imported as a linear vendor commit so forge rebases do not replay the upstream root tree.

git-subtree-dir: third_party/pi

git-subtree-split: 71023aec1546d6d677a65669c5665c0927ce0dd0
This commit is contained in:
alexion committed 2026-10-15 17:58:48 -04:00
1 parent 719370c8f7
commit 6081e3e325
1743 files changed
+205459 -131704
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

third_party/pi/packages/ai/CHANGELOG.md
Vendored
+277
View File
@@ -1,5 +1,281 @@
# Changelog
## [1.3.0] - 2026-10-15
### Breaking Changes
- Removed `AssistantMessageFrame`, `AssistantMessageFrameEncoder`, and `reduceAssistantMessageFrames()`. They were only used by the removed experimental agent harness.
### Added
- Added cache-preserving thinking level changes for GPT-6 models on the openai provider: with an API key, the level is sent as configuration_update input items instead of changing reasoning.effort, controlled by compat.supportsMidConvoEffort ([#10853](https://github.com/earendil-works/pi/issues/10853))
### Changed
- Changed the `kimi-coding` provider to use OpenAI Chat Completions at `https://api.kimi.com/coding/v1`, the protocol used by the official Kimi Code client. K3, K3-256K, and Kimi For Coding now accept tools added mid-conversation without invalidating the prompt cache. `models.json` overrides of the `kimi-coding` base URL must now include `/v1` ([#10720](https://github.com/earendil-works/pi/issues/10720))
- Changed model switching to drop thinking blocks from earlier assistant messages produced by a different model, instead of sending them as plain text. The new model previously saw the old model's reasoning as part of its replies and could start writing its own reasoning into the visible answer ([#10995](https://github.com/earendil-works/pi/issues/10995))
### Fixed
- 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))
- Fixed Anthropic and Bedrock Claude requests failing with a 400 when an error tool result contains 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))
- Fixed OpenCode Go Gemini models failing with a 404 by moving OpenCode Zen and OpenCode Go to the `opencode.ai/inference` endpoints, and added Google Generative AI support to the OpenCode Go provider
- Fixed OpenCode Go MiniMax M2.7 failing with "Model does not support this protocol" by using the Anthropic Messages API
- Fixed automatic retry not triggering for gateway `Unable to forward upstream stream` errors ([#10920](https://github.com/earendil-works/pi/issues/10920))
- 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))
- 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))
- 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))
- 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))
- Fixed `openai-codex/gpt-6.1-sol` using a 272K context window instead of the 500K window Codex now provides ([#10840](https://github.com/earendil-works/pi/issues/10840))
- Fixed tool calls whose final arguments are malformed JSON silently running with only the fields parsed before the error; such calls now carry `argumentsError` and fail validation with the raw arguments text, so the model sees what it sent ([#10897](https://github.com/earendil-works/pi/issues/10897))
- Fixed Claude Haiku 5.5 and Claude Opus 5 not supporting the `off` thinking level; `off` now sends `thinking: {type: "disabled"}` without effort settings ([#10753](https://github.com/earendil-works/pi/issues/10753))
## [1.2.0] - 2026-10-12
### Added
- Added `personalizeOpenRouterCatalog()` (`@earendil-works/pi-ai/providers/openrouter`), which narrows an OpenRouter catalog to the chat models a key may use, with that key's context size, max output, and prices from `GET /models/user` ([#10353](https://github.com/earendil-works/pi/issues/10353), [#10672](https://github.com/earendil-works/pi/pull/10672) by [@davidbrai](https://github.com/davidbrai))
- Added `RefreshModelsContext.baseUrl`, the registered provider's base URL including configuration overrides applied by wrappers, so dynamic catalogs can follow a base URL set in `models.json` ([#10353](https://github.com/earendil-works/pi/issues/10353), [#10672](https://github.com/earendil-works/pi/pull/10672) by [@davidbrai](https://github.com/davidbrai))
### Fixed
- 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 ([#10609](https://github.com/earendil-works/pi/issues/10609))
- Fixed OpenRouter router models such as `openrouter/auto` showing a cost of -$1,000,000 per 1M tokens; OpenRouter's negative placeholder prices are now 0 ([#10672](https://github.com/earendil-works/pi/pull/10672) by [@davidbrai](https://github.com/davidbrai))
- Fixed Anthropic `strict: "prefer"` tools failing when the provider rejects the strict tool set for grammar size, tool count, or union limits; requests now retry once with preferred tools non-strict while keeping `strict: "require"` tools strict ([#10409](https://github.com/earendil-works/pi/issues/10409))
- Fixed "The upstream provider returned an error" not being retried as a transient provider error ([#10842](https://github.com/earendil-works/pi/issues/10842))
## [1.1.0] - 2026-10-07
### Breaking Changes
- A stream function must return an `AssistantMessageEventStream`, for example from `createAssistantMessageEventStream()`; a hand-written `EventStream<AssistantMessageEvent, AssistantMessage>` subclass no longer type-checks in its place
### Added
- Added `durationMs` to `AssistantMessage`: `AssistantMessageEventStream` measures each response with a monotonic clock from the start of the request to its final message, for every API implementation, including direct calls. Deferred results fetched later stay untimed ([#10549](https://github.com/earendil-works/pi/issues/10549))
- Added an optional `durationMs` to `ToolResultMessage` for the execution time of the tool ([#10549](https://github.com/earendil-works/pi/issues/10549))
- Added the `openai-decisions` classifier API for OpenAI's Decisions API, with `gpt-6-luna` as a classifier model of the `openai` provider. It needs an API key, so it is not listed as available while `openai` uses Sign in with ChatGPT
- Added Claude Haiku 5.5 (`claude-haiku-5-5`) to the `anthropic` provider with its prompt-length pricing tier, adaptive thinking with `xhigh`/`max` effort, per-message effort, and mid-conversation system messages and tool changes. Bedrock requests for Haiku 5.5 use adaptive thinking, native `xhigh`, and prompt caching
- Added `LoginOptions.agentName`, which replaces pi's name in the Sign in with ChatGPT agent name hint and the Codex browser login originator ([#10433](https://github.com/earendil-works/pi/pull/10433) by [@lucasmeijer](https://github.com/lucasmeijer))
- Added optional `images` to `ClassifierContext`. Models whose `input` includes `"image"` judge them with the state; `classify()` returns an error result for other models and for APIs that cannot send images
### Changed
- The faux provider's prompt-cache usage estimate compares the previous and current prompt message by message and compares characters only from the first differing message; the usage numbers are unchanged.
### Fixed
- Fixed `server_busy` and `servers are currently busy` provider errors ending the turn instead of being retried ([#10543](https://github.com/earendil-works/pi/issues/10543))
- Fixed the error message of a failed lazy API setup, such as a module load or auth failure, using its failure time as `timestamp` instead of the request start
- Reduced context-limit request failures by estimating input at 3.5 characters per token instead of 4 when calculating output limits ([#10497](https://github.com/earendil-works/pi/issues/10497))
- Fixed Radius models disabled by an organization owner still being listed: a fetched or cached Radius catalog now replaces the shipped default catalog instead of being merged into it
- Fixed OpenAI provider type-checking with cached catalogs that contain no classifier models
- Fixed Anthropic browser login failing with "localhost refused to connect" when port 53692 is reserved or in use, for example by Hyper-V/WSL port exclusions on Windows: login now falls back to a free loopback port ([#10571](https://github.com/earendil-works/pi/issues/10571))
- Fixed built-in model costs missing prompt-length pricing tiers for OpenCode, OpenCode Go, OpenRouter, Vercel AI Gateway, Google, MiniMax, and other models.dev providers, which undercounted the cost of long prompts on models such as Claude Haiku 5.5, Gemini 3.1 Pro, and GPT-5.4. OpenRouter time-of-day pricing is not modeled
- Fixed Mistral responses that end with `finish_reason: "error"` not being retried ([#10487](https://github.com/earendil-works/pi/issues/10487))
- Fixed Bedrock Converse not sending the reasoning effort to OpenAI models: gpt-oss gets `reasoning_effort` clamped to low/medium/high, other GPT models get `reasoning.effort` with `minimal` sent as `low` ([#9331](https://github.com/earendil-works/pi/issues/9331), [#10142](https://github.com/earendil-works/pi/pull/10142) by [@jsanter27](https://github.com/jsanter27))
- Fixed model and caller headers not overriding the `originator` and `User-Agent` headers of `openai-codex-responses` requests ([#10429](https://github.com/earendil-works/pi/pull/10429) by [@lucasmeijer](https://github.com/lucasmeijer))
## [1.0.4] - 2026-10-05
### Fixed
- Fixed Bedrock requests that fail with `The pending stream has been canceled` after a stalled HTTP/2 connection not being retried automatically ([#10379](https://github.com/earendil-works/pi/issues/10379))
## [1.0.3] - 2026-10-05
### Breaking Changes
- Renamed the Azure provider from `azure-openai-responses` to `azure`, since it now serves Chat Completions as well as the Responses API. Use `getModel("azure", ...)`, and import `azureProvider` and `AZURE_MODELS` from `@earendil-works/pi-ai/providers/azure` instead of `azureOpenAIResponsesProvider` and `AZURE_OPENAI_RESPONSES_MODELS` from `providers/azure-openai-responses`. The `azure-openai-responses` api id and the `AZURE_OPENAI_*` environment variables are unchanged ([#9714](https://github.com/earendil-works/pi/pull/9714) by [@jsanter27](https://github.com/jsanter27))
### Added
- Added Chat Completions support to the Azure provider for Foundry deployments, with DeepSeek V4 Pro in the built-in catalog. Other Foundry models can be added under the `azure` provider with `api: "openai-completions"`, and `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` and `azureDeploymentName` apply to both APIs ([#9645](https://github.com/earendil-works/pi/issues/9645), [#9714](https://github.com/earendil-works/pi/pull/9714) by [@jsanter27](https://github.com/jsanter27))
### Fixed
- Fixed OAuth credentials being invalidated when a request or model refresh was cancelled or superseded during a token refresh: a token refresh that has started now completes and persists the rotated refresh token
## [1.0.2] - 2026-10-04
### Added
- Added per-thinking-level sampling parameter overrides (`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))
## [1.0.1] - 2026-10-03
### Added
- 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))
### Changed
- Anthropic models with native mid-conversation tool changes now use the `inline-tools-2026-09-15` beta: later tools are defined by value in `tool_addition` blocks instead of being appended to the top-level tool list, and redefining a tool under the same name no longer falls back to resending the full tool list, so the prompt cache survives it. Upgraded `@anthropic-ai/sdk` to 0.129.0.
- Deprecated `hasToolRedefinitions()`; no built-in transport needs it anymore.
### Fixed
- Fixed "Selected model is at capacity" provider errors ending the turn instead of being retried ([#10278](https://github.com/earendil-works/pi/issues/10278))
- Fixed Cloudflare AI Gateway Claude models failing with a 404 by using dashed model IDs (`claude-opus-5-5` instead of `claude-opus-5.5`), which Anthropic requires
- 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 ([#10265](https://github.com/earendil-works/pi/issues/10265))
- Fixed Amazon Bedrock OpenAI models costing requests above 272k input tokens at the short-context rate; Bedrock models now include the pricing tiers listed on models.dev ([#10326](https://github.com/earendil-works/pi/issues/10326))
- Fixed Amazon Bedrock Claude requests failing with "Invalid `signature` in `thinking` block" after the system prompt or tools changed; 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))
- Fixed Together DeepSeek V4 Pro losing its thinking level controls after Together renamed it to `deepseek-ai/DeepSeek-V4-Pro-0813` ([#10336](https://github.com/earendil-works/pi/pull/10336) by [@cv](https://github.com/cv))
## [1.0.0] - 2026-10-01
### Added
- Added a copy code login method to Anthropic OAuth. Login asks for browser login (default) or copy code login, which shows the authorization code on Anthropic's page for pasting into pi and works when the browser runs on another machine ([#10194](https://github.com/earendil-works/pi/pull/10194) by [@lucasmeijer](https://github.com/lucasmeijer)).
### Changed
- Changed OAuth browser pages to use the color Pi logo.
### Fixed
- 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 like Radius.
## [0.99.2] - 2026-09-30
### Added
- Added the lightweight `@earendil-works/pi-ai/models` entry point for model collections and provider construction without loading TypeBox, built-in catalogs, or provider SDKs.
- 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))
### Fixed
- Fixed context overflow detection for Z.AI CN endpoint `Prompt exceeds max length` errors ([#10208](https://github.com/earendil-works/pi/issues/10208))
- Fixed Anthropic requests failing when a `strict: "prefer"` tool schema uses keywords Anthropic strict tool use rejects, such as `minimum`/`maximum`; such tools are now sent non-strict ([#9953](https://github.com/earendil-works/pi/issues/9953))
- Fixed provider retries firing immediately when a `Retry-After` header contains an unparseable date; they now use exponential backoff ([#9571](https://github.com/earendil-works/pi/issues/9571))
## [0.99.1] - 2026-09-29
### Added
- Added GPT-6.1 Sol (`gpt-6.1-sol`) to the OpenAI, Azure OpenAI Responses, and OpenAI Codex providers.
## [0.99.0] - 2026-09-29
### Breaking Changes
- Unified image models into the regular `Provider`/`Models` surface. The separate `ImagesModels` collection is removed: `createImagesModels()`, `createImagesProvider()`, `ImagesProvider`, `openrouterImagesProvider()`, `builtinImagesProviders()`, and `builtinImagesModels()` are gone. Use `builtinModels()`, `models.getModelOfType("image", ...)`, `models.generateImages()`, and `createProvider({ models, images })` instead. Existing unqualified reads remain chat-only.
- Image models are now `ImageModel` with a required `type: "image"` and share `BaseModel` with chat models. The old plural image type names (`ImagesModel`, `ImagesApi`, `KnownImagesApi`, `KnownImagesProvider`, and `ImagesProviderId`) are removed. `generateImages()` accepts only image models. Output modalities (`output`) remain on image models only.
- Generated model data schema is now version 6: every entry carries `type`, operation-specific catalogs include chat, image, and classifier models, and one upstream ID may have separate entries per type. OpenRouter image models live in the `openrouter-images` api group of `openrouter.json`; `image-models.generated.ts` and `scripts/generate-image-models.ts` are removed. Run `npm run hydrate:model-data`.
### Added
- Added `Models.generateImages()` with provider-resolved auth, `Provider.generateImages?`, and `createProvider({ images })` keyed by `model.api`. `createProvider()` `models` and `fetchModels` accept models of every type, and `api` is optional when `images` or `classifiers` is given.
- Added an optional model `type` (`"chat"`, `"image"`, or `"classifier"`). Chat models may omit it, so existing chat models, providers, and stores keep working unchanged. Narrow mixed lists with the new `isModelType()` guard or read the effective type with `getModelType()`.
- Added `getModelsOfType()`, `getModelOfType()`, `getAvailableOfType()`, `getAllModels()`, and `getAllAvailable()` on `Models`; optional `Provider.getAllModels()` and `Provider.filterAllModels()`; corresponding generated-catalog accessors; and the `AnyModel` and `ModelTypeMap` types. `hasApi()`, `calculateCost()`, and `modelsAreEqual()` accept `AnyModel`.
- Added support for models of every type in `ModelsStoreEntry.models`. Stored and fetched models of unknown types are dropped instead of failing a refresh.
- Added classifier models and `Models.classify()` with a provider-neutral JEV-style `choice`/`score`/`bool` contract. The built-in TypeSafe provider exposes models.dev's `jev-latest` through the System One API and translates public `bool` questions to TypeSafe's `noul` wire format.
- 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.
- Added Jev classifier models on Vercel AI Gateway (`typesafe-ai/jev`, generated from its evaluation model catalog) and OpenCode Zen (`jev-1.13`, `jev-1.13-free`) through their TypeSafe-compatible System One endpoints.
- Added `usage` to `ClassifierResult`: System One classifications report token counts, priced from the model catalog like chat usage.
- Added a runtime chat-model check to the `Models` stream entry points so non-chat models fail with a clear `ModelsError` instead of a missing-api stream error.
- Added array-based `models.all.json` and `providers/{id}.all.json` variants to the generated and published JSON catalog, allowing the same upstream ID once per model type; the existing keyed `models.json` and `providers/{id}.json` stay chat-only for released clients.
- 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), [#9901](https://github.com/earendil-works/pi/pull/9901) by [@davidbrai](https://github.com/davidbrai)).
- Added Claude Sonnet 5.5 to the built-in Anthropic model catalog with adaptive thinking, mid-conversation effort, 1M context, and official pricing metadata.
- Added Sign in with ChatGPT to the `openai` provider: an OAuth login that uses a ChatGPT subscription with the OpenAI API. `Models.login()` accepts `LoginOptions` with `getDeviceId()`, which supplies a stable installation ID to login flows that need one. Subscription usage-limit errors are not retried and link to the ChatGPT usage page; temporary usage errors are retried.
- Added the `llama-cpp-classify` classifier API, which answers classifier questions from llama-server's next-token probabilities for single-token answer labels.
- Added optional `AssistantMessage.thinkingLevel`, which records the thinking level the agent loop requested for a response.
### Changed
- Renamed the OpenAI Codex provider to "OpenAI Codex (legacy)"; Sign in with ChatGPT on the `openai` provider supersedes it.
- Unified the Anthropic, OpenAI Codex, OpenRouter, and Radius browser sign-in callback servers into one shared implementation with the same browser pages. The OAuth page helpers are now available as `@earendil-works/pi-ai/utils/oauth-page`.
- Changed Radius browser sign-in to exchange the authorization code before showing the browser page, so token exchange failures are shown in the browser.
### Fixed
- 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)).
- Fixed model-level `samplingParams` being dropped by direct `stream()`/`complete()` calls on OpenAI-compatible APIs ([#9506](https://github.com/earendil-works/pi/issues/9506)).
- 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)).
- Fixed OpenAI Fast mode requests being priced at the standard rate when the response reports `service_tier: "fast"`, as GPT-6 models do ([#10034](https://github.com/earendil-works/pi/issues/10034)).
- 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)).
- 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)).
- 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)).
- Fixed Anthropic and OpenAI Codex browser sign-in waiting indefinitely after the provider redirected with an authorization error; sign-in now fails with the provider's error description.
- Fixed Anthropic browser sign-in failing when its callback port is in use; it now falls back to pasting the redirect URL.
- Fixed GitHub Copilot Claude Opus 5.5 offering unsupported thinking levels when upstream model metadata is incomplete; it now offers low through max.
## [0.87.1] - 2026-09-22
### Added
- Added Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna to the GitHub Copilot catalog.
- 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.
- Added Claude Opus 5.5 to the built-in Anthropic model catalog with adaptive thinking, 1M context, and official pricing metadata.
- Added Grok 4.7 to the built-in xAI model catalog with long-context pricing metadata.
### Fixed
- 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))
- Fixed Anthropic OAuth requests reporting an outdated Claude Code version.
## [0.87.0] - 2026-09-21
### Added
- Added model image-input limit and cache-safe resize metadata to the generated catalog ([#9631](https://github.com/earendil-works/pi/issues/9631)).
### Fixed
- Fixed unknown OpenAI-compatible Chat Completions endpoints receiving strict tool schemas unless they explicitly advertise support, while preserving strict tools for capable built-in models ([#9816](https://github.com/earendil-works/pi/issues/9816)).
## [0.86.1] - 2026-09-20
### Added
- Added Meta provider (Model API key and Muse subscription OAuth) with Muse Spark models ([#9096](https://github.com/earendil-works/pi/pull/9096) by [@xl0](https://github.com/xl0)).
### Fixed
- Fixed z.ai `Prompt too long` errors not being recognized as context overflow ([#9805](https://github.com/earendil-works/pi/issues/9805)).
- Fixed Cerebras models advertising unsupported strict tool schemas, which caused HTTP 400 errors when strict and non-strict tools were mixed ([#9804](https://github.com/earendil-works/pi/pull/9804) by [@EdenGottlieb](https://github.com/EdenGottlieb)).
## [0.86.0] - 2026-09-19
### Breaking Changes
- Changed provider-facing `ProviderStreams` and `StreamFunction` inputs from `Context` to normalized `TranscriptContext` values. System prompts and tool declarations now live in transcript system messages; custom providers must read them with `getCurrentSystemPrompt()` and `getCurrentTools()`. See [System Messages](README.md#system-messages).
- Restricted `ToolCall.arguments` and `ToolResultMessage.details` to JSON-compatible values, changed `ToolResultMessage` into a conditional type, and made `JsonValue` arrays readonly.
### Added
- Added transcript-backed mid-conversation system prompt and tool changes, with native replay on supported models and automatic collapse for other providers. See [System Messages](README.md#system-messages) ([#9548](https://github.com/earendil-works/pi/pull/9548)).
- Added a generated public Radius model catalog for synchronous API lookup, with cached and live gateway catalogs overlaid at runtime.
- 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)).
- Added `RetryPolicy.maxAgentDelayMs` support to cap shared assistant retry backoff for summarization calls ([#8826](https://github.com/earendil-works/pi/issues/8826)).
- Added `Model.promptCache` lifetime metadata for short and long retention tiers to support prompt-cache warming decisions ([#9668](https://github.com/earendil-works/pi/pull/9668)).
### Fixed
- Fixed GitHub Copilot GPT models, including GPT-6 Astra, using the Chat Completions adapter instead of the required Responses adapter ([#9253](https://github.com/earendil-works/pi/pull/9253) by [@petrroll](https://github.com/petrroll)).
- Fixed DeepSeek V4.1 thinking levels on OpenRouter and OpenCode Go preserving provider effort metadata ([#9485](https://github.com/earendil-works/pi/issues/9485)).
- Fixed bodyless HTTP 400/413 errors from non-Cerebras providers being misclassified as context overflow ([#9482](https://github.com/earendil-works/pi/issues/9482)).
- Fixed Vercel AI Gateway replaying unsigned thinking as assistant text ([#9676](https://github.com/earendil-works/pi/issues/9676)).
- Fixed Google Generative AI and Vertex AI using unsupported thinking levels when reasoning is omitted or when model capabilities differ within a Gemini family ([#9455](https://github.com/earendil-works/pi/issues/9455)).
- Fixed Anthropic-compatible relays breaking signed thinking replay when they report a different response model, while preserving fallback pricing ([#9188](https://github.com/earendil-works/pi/issues/9188)).
- Fixed Amazon Bedrock one-hour cache writes being priced at the five-minute rate ([#9457](https://github.com/earendil-works/pi/issues/9457)).
- Fixed quadratic CPU usage when draining buffered `EventStream` events ([#9055](https://github.com/earendil-works/pi/issues/9055)).
- Fixed Mistral Medium reasoning requests to use `reasoning_effort` for all reasoning-capable `mistral-medium-*` model IDs instead of the unsupported `prompt_mode` ([#8700](https://github.com/earendil-works/pi/issues/8700)).
- Fixed OpenCode and OpenCode Go requests to send `x-opencode-session` from `sessionId` across all supported API adapters ([#9326](https://github.com/earendil-works/pi/issues/9326)).
- 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)).
- 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)).
- 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)).
- 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)).
- 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)).
- 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)).
- Fixed Baseten requests to send session-affinity headers from `sessionId` for automatic prompt-cache routing ([#9629](https://github.com/earendil-works/pi/issues/9629)).
- Fixed retry classification for Cloudflare 520 responses ([#9627](https://github.com/earendil-works/pi/issues/9627)).
- Fixed retry classification for transient Azure peak-load capacity errors ([#9669](https://github.com/earendil-works/pi/issues/9669)).
### Removed
- 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)).
## [0.85.1] - 2026-09-05
### Added
@@ -24,6 +300,7 @@
- Added an optional timestamp argument to `uuidv7()` for follower IDs.
- Added narrow `api`, `providers`, and `utils` subpath exports for direct imports without loading the package barrel.
- 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.
- Added Meta provider (Model API key and Muse subscription OAuth) with Muse Spark models.
### Fixed