@bastani/pi-ai 0.9.72-alpha.1 → 0.9.73-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 CHANGED
@@ -4,6 +4,62 @@ This package is a Bastani fork of `@earendil-works/pi-ai`. Upstream history at t
4
4
 
5
5
  ## [Unreleased]
6
6
 
7
+ ## [0.9.72] - 2026-11-08
8
+
9
+ GPT Image 2.6, typed Copilot failures, and thinking `off` that every provider honors.
10
+
11
+ - **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)).
12
+ - **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)).
13
+ - **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)).
14
+ - **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)).
15
+
16
+ **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.
17
+
18
+ ### Breaking Changes
19
+
20
+ - `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))
21
+
22
+ ### Added
23
+
24
+ - 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)).
25
+ - `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)).
26
+ - 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)).
27
+ - `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`.
28
+ - 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)).
29
+ - 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)).
30
+ - `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))
31
+ - `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))
32
+ - 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))
33
+ - 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))
34
+ - 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))
35
+ - 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))
36
+
37
+ ### Changed
38
+
39
+ - `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)).
40
+ - 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))
41
+ - 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))
42
+ - 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))
43
+ - The `openai-codex` provider is named "OpenAI Codex" again instead of "OpenAI Codex (legacy)" (upstream [`36142c1c`](https://github.com/earendil-works/pi/commit/36142c1c7eab1d756e977b5f39149a634cc472ac))
44
+
45
+ ### Fixed
46
+
47
+ - 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)).
48
+ - `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))
49
+ - 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))
50
+ - 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))
51
+ - 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))
52
+ - 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))
53
+ - 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))
54
+ - 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))
55
+ - 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))
56
+ - 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))
57
+ - 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))
58
+ - 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))
59
+ - 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))
60
+ - 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))
61
+ - 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))
62
+
7
63
  ## [0.9.72-alpha.1] - 2026-11-08
8
64
 
9
65
  GPT Image 2.6, typed Copilot failures, and thinking `off` that every provider honors.
@@ -1 +1 @@
1
- {"schemaVersion":6,"generatedAt":"2026-11-08T06:40:47.767Z","structureHash":"b00c568d9dca2b7e226e47e18440354ccd5941ca464488dfbd20b76d18318f24","files":{"amazon-bedrock.json":"91e42fa911ae81bc79e84633cfb1c9d3270e3e21e75967b22b213ccaae566741","ant-ling.json":"628feab3ef6c7d80ce96f75999804e1d9f05fec3851e5302e30fff3b69db90f6","anthropic.json":"0d76d762a56e45f5ba605c0bb0bd4f4bc13e73cba8531cf4832b1665d72903c6","azure.json":"4b667155b271dc3d58dedaf62382be5172ce2f47d6cab12dc149f6b08a1385f5","baseten.json":"47d10c8050601eb88757def4fc957cdba769dacec4aeb3b0c3f39c41e442330e","cerebras.json":"2eb969b2956844f9a641d3de52c19f1791293c58f607282fb3b0ebfe85ff66f8","cloudflare-ai-gateway.json":"c4671ad861ab6e7e45ec844ea513a2088a390c533dba8e8bd9a54c1a47641861","cloudflare-workers-ai.json":"0dc5eda87ff6de3ee75ee6d547a60162319bb2d51ed09b1e7d47d8ae9ef8548e","deepseek.json":"10a296fb3e898f7715c80890c8af0af4f5fd58f22dbcd6612fdc5d762157ccdc","fireworks.json":"8a5b946e0fbfd1889570591f50a8e38bd01b19fe83135b23dca66138c3ad8f8c","github-copilot.json":"581141741866f4b334bad472229b4ded73999877d58d7748ab95c5dbe7c29118","google-vertex.json":"d8f5698a02ca594f02d9af5a8be1a45881201d097821873305ae79dc4211f147","google.json":"fa30d97f32e63794206759806d5aac0374072144a65d288611d02ccf7d56617f","groq.json":"087df37206e9eed8b10f1e558e4c496289bef60074b7e8afba4f20fb102b4ec2","huggingface.json":"209cdcf205961ef2ff203486da614232a86a069ebe0e98266189a14a235fac8f","kimi-coding.json":"e6d5bce9045cf533ca9e0d563a7d30a6a1ed48c89a0f539ed4c9d3ac53d4ac8b","meta.json":"563f5ec07973f52ab8c3eb26ef91a36ad8a79e950b37174935666cdcb624bf35","minimax-cn.json":"fa15d1f67a07046d38e191af9bd97ed922ba45a99a11af4ad281151cc6b762aa","minimax.json":"1e0459dda3858ea24151182106fc4179b95f43a0a8dffc1c3d2e7db4c2d9206e","mistral.json":"7e67f69ac2b609fca9f7b721243fb70fa6f4feade3cb20eedcd5786f9eac64f1","moonshotai-cn.json":"1a15468b1edf9205956e4c5ea41fc4cacc14eabf00523b9143e23a38264f57f7","moonshotai.json":"f31cdf2e6c81a403c4cb0e6b7e759ceff2820f966b08dbec80cfafa3f80dfc63","nvidia.json":"078fae404fc622ca6461f16a74791fedafe44976a6c3323b6ba779729823d4ff","openai-codex.json":"79856e655a4cdf5ad32f178dd001cbfda5d13e68006ad8aade0bc3636f5faabb","openai-decisions.json":"0c7d54fcbac5e7388872a85c0db7f336c5a57c82d403266e9ae04f87561daf4f","openai-images.json":"95e906730507ae8adcd49afa81fc21f7cee13c12842a71ba04ab62c59637bdb6","openai.json":"c51bffed14b4bdfb15fbffef068a9c07e973d36dd7490397d5bb7d00483752b3","opencode-go.json":"57ac9e903ba673130c11d3f50928d0019473849a261c6319e90cedda6a8321d5","opencode.json":"d20226ba8646d85a1101ce773ed6731711d32f94228652537261aaa2584bfdd1","openrouter.json":"8ed917db7c511e7a5c88a3850d3bf601a7cd7c03235f2f2500de5a2bc292852c","qwen-token-plan-cn.json":"64427a6adeb3e7dd1857d68bef9bb12097632b646ed151b740e209143e31341e","qwen-token-plan-individual-cn.json":"4d5af1f18c8c2306b2158bda5052a9b347cce4dfa7e97245700f1c84798d1cfb","qwen-token-plan-individual.json":"3c3825ba2d19642b691b4c51c79bc93fcf501d39e9f0566d05a0f135aab8a6f6","qwen-token-plan.json":"37b9a76c54090228d1026a8f0fcb4d304b00d64b1a3a94cb2ae6e677d2c52b13","radius.json":"b757791113c3ac1c7b43501d7b4de2ae51b7dad443e34c78650632de395bf925","together.json":"cd63c051ea61bdb210c6e826e5fc892e4c1f9c9c0ab6eb91933a0620701d17e1","typesafe.json":"938c89dfb6f636dce1483ec6acc739af4b881f61003d007807a74414eb2bf357","vercel-ai-gateway.json":"aefbd8fef0f37f3fe69fd84385d1c9008c8c50374032087af723f007a9a15df0","xai.json":"59f8a7bdf14e79ba1b0d0a37e56251bbf100420bdac7d6af5942508fdb990f11","xiaomi-token-plan-ams.json":"bfd33741da693bdb8e8ebd4e7b90c3e90820aa18432ad081a5a98db13c6bd296","xiaomi-token-plan-cn.json":"4b61d78feae3a926fda918c3fd3a9afaf2ca60c1bba0af8270a5398fb50a983e","xiaomi-token-plan-sgp.json":"b05a197ce24fd69fc9440f4ba929984ac7a0fefbe8083268734b422e721987ec","xiaomi.json":"495b5fddf3c3ec8904fecc15c4c19536f4961dce7b51512b6909d2df2e739824","zai-coding-cn.json":"cc9e0768e726ef55d2b7af853ac5f3585dc72575932cb38840a3877d18736b78","zai.json":"58e6642fdfb736f0b9a4a277c760c536a3cea66a030abbfb2615f7b2851d5c6f"}}
1
+ {"schemaVersion":6,"generatedAt":"2026-11-08T17:27:33.158Z","structureHash":"b00c568d9dca2b7e226e47e18440354ccd5941ca464488dfbd20b76d18318f24","files":{"amazon-bedrock.json":"91e42fa911ae81bc79e84633cfb1c9d3270e3e21e75967b22b213ccaae566741","ant-ling.json":"628feab3ef6c7d80ce96f75999804e1d9f05fec3851e5302e30fff3b69db90f6","anthropic.json":"0d76d762a56e45f5ba605c0bb0bd4f4bc13e73cba8531cf4832b1665d72903c6","azure.json":"4b667155b271dc3d58dedaf62382be5172ce2f47d6cab12dc149f6b08a1385f5","baseten.json":"47d10c8050601eb88757def4fc957cdba769dacec4aeb3b0c3f39c41e442330e","cerebras.json":"2eb969b2956844f9a641d3de52c19f1791293c58f607282fb3b0ebfe85ff66f8","cloudflare-ai-gateway.json":"c4671ad861ab6e7e45ec844ea513a2088a390c533dba8e8bd9a54c1a47641861","cloudflare-workers-ai.json":"0dc5eda87ff6de3ee75ee6d547a60162319bb2d51ed09b1e7d47d8ae9ef8548e","deepseek.json":"10a296fb3e898f7715c80890c8af0af4f5fd58f22dbcd6612fdc5d762157ccdc","fireworks.json":"8a5b946e0fbfd1889570591f50a8e38bd01b19fe83135b23dca66138c3ad8f8c","github-copilot.json":"581141741866f4b334bad472229b4ded73999877d58d7748ab95c5dbe7c29118","google-vertex.json":"d8f5698a02ca594f02d9af5a8be1a45881201d097821873305ae79dc4211f147","google.json":"fa30d97f32e63794206759806d5aac0374072144a65d288611d02ccf7d56617f","groq.json":"087df37206e9eed8b10f1e558e4c496289bef60074b7e8afba4f20fb102b4ec2","huggingface.json":"209cdcf205961ef2ff203486da614232a86a069ebe0e98266189a14a235fac8f","kimi-coding.json":"e6d5bce9045cf533ca9e0d563a7d30a6a1ed48c89a0f539ed4c9d3ac53d4ac8b","meta.json":"563f5ec07973f52ab8c3eb26ef91a36ad8a79e950b37174935666cdcb624bf35","minimax-cn.json":"fa15d1f67a07046d38e191af9bd97ed922ba45a99a11af4ad281151cc6b762aa","minimax.json":"1e0459dda3858ea24151182106fc4179b95f43a0a8dffc1c3d2e7db4c2d9206e","mistral.json":"7e67f69ac2b609fca9f7b721243fb70fa6f4feade3cb20eedcd5786f9eac64f1","moonshotai-cn.json":"1a15468b1edf9205956e4c5ea41fc4cacc14eabf00523b9143e23a38264f57f7","moonshotai.json":"f31cdf2e6c81a403c4cb0e6b7e759ceff2820f966b08dbec80cfafa3f80dfc63","nvidia.json":"078fae404fc622ca6461f16a74791fedafe44976a6c3323b6ba779729823d4ff","openai-codex.json":"79856e655a4cdf5ad32f178dd001cbfda5d13e68006ad8aade0bc3636f5faabb","openai-decisions.json":"0c7d54fcbac5e7388872a85c0db7f336c5a57c82d403266e9ae04f87561daf4f","openai-images.json":"95e906730507ae8adcd49afa81fc21f7cee13c12842a71ba04ab62c59637bdb6","openai.json":"c51bffed14b4bdfb15fbffef068a9c07e973d36dd7490397d5bb7d00483752b3","opencode-go.json":"57ac9e903ba673130c11d3f50928d0019473849a261c6319e90cedda6a8321d5","opencode.json":"d20226ba8646d85a1101ce773ed6731711d32f94228652537261aaa2584bfdd1","openrouter.json":"6d9cfcf3e5af15520ec3dddd5e4ca0883a6caebdc4b1e7c4debedc6f9854a62d","qwen-token-plan-cn.json":"64427a6adeb3e7dd1857d68bef9bb12097632b646ed151b740e209143e31341e","qwen-token-plan-individual-cn.json":"4d5af1f18c8c2306b2158bda5052a9b347cce4dfa7e97245700f1c84798d1cfb","qwen-token-plan-individual.json":"3c3825ba2d19642b691b4c51c79bc93fcf501d39e9f0566d05a0f135aab8a6f6","qwen-token-plan.json":"37b9a76c54090228d1026a8f0fcb4d304b00d64b1a3a94cb2ae6e677d2c52b13","radius.json":"b757791113c3ac1c7b43501d7b4de2ae51b7dad443e34c78650632de395bf925","together.json":"cd63c051ea61bdb210c6e826e5fc892e4c1f9c9c0ab6eb91933a0620701d17e1","typesafe.json":"938c89dfb6f636dce1483ec6acc739af4b881f61003d007807a74414eb2bf357","vercel-ai-gateway.json":"aefbd8fef0f37f3fe69fd84385d1c9008c8c50374032087af723f007a9a15df0","xai.json":"59f8a7bdf14e79ba1b0d0a37e56251bbf100420bdac7d6af5942508fdb990f11","xiaomi-token-plan-ams.json":"bfd33741da693bdb8e8ebd4e7b90c3e90820aa18432ad081a5a98db13c6bd296","xiaomi-token-plan-cn.json":"4b61d78feae3a926fda918c3fd3a9afaf2ca60c1bba0af8270a5398fb50a983e","xiaomi-token-plan-sgp.json":"b05a197ce24fd69fc9440f4ba929984ac7a0fefbe8083268734b422e721987ec","xiaomi.json":"495b5fddf3c3ec8904fecc15c4c19536f4961dce7b51512b6909d2df2e739824","zai-coding-cn.json":"cc9e0768e726ef55d2b7af853ac5f3585dc72575932cb38840a3877d18736b78","zai.json":"58e6642fdfb736f0b9a4a277c760c536a3cea66a030abbfb2615f7b2851d5c6f"}}