@bitkyc08/opencodex 2.57.0 → 2.59.0

package/src/providers/registry/entries-core.ts

@@ -282,6 +282,17 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
         forwardCallerServiceTier: false,
       },
     },
+    // Grok 4.6/4.5 OAuth Responses replays Codex tool history. After a mid-stream 502/reset,
+    // the client can resend a function_call without a matching output, or with hook-injected
+    // developer context between the pair. Google already synthesizes a missing tool_result
+    // (#2199). xAI's Responses parser does not, so the next turns 400 and the thread snowballs.
+    // Reuse the existing adjacency capability (Kimi #4726, DeepSeek #1292). Do not set
+    // statelessResponses: xAI stores responses for 30 days and documents previous_response_id.
+    // https://docs.x.ai/developers/model-capabilities/text/comparison
+    requiresAdjacentResponsesToolResults: true,
+    // The dangling half of the same failure: a call whose output never arrived. Kimi accepts that
+    // shape, so this is a second capability rather than a widening of the one above.
+    requiresPairedResponsesToolResults: true,
     // Vision lineup per docs.x.ai model-capabilities/images/understanding: the grok-4.x chat
     // models accept image input (JPEG/PNG, URL or base64). Without this the catalog leaves
     // inputModalities undefined, and deriveComboCatalogModel defaults an undefined member to

package/src/providers/registry/entries-extended.ts

@@ -56,6 +56,11 @@ import {
   ALIBABA_TOKEN_PLAN_MODELS,
   ALIBABA_TOKEN_PLAN_QWEN_MODELS,
   ALIBABA_TOKEN_PLAN_INPUT_MODALITIES,
+  ALIBABA_TOKEN_PLAN_CONTEXT_WINDOWS,
+  ALIBABA_TOKEN_PLAN_MAX_OUTPUT_TOKENS,
+  ALIBABA_TOKEN_PLAN_NO_VISION,
+  ALIBABA_TOKEN_PLAN_PRESERVE_REASONING,
+  QWEN38_FAMILY,
   ALIBABA_INTL_TOKEN_PLAN_MODELS,
   ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS,
   TENCENT_CODING_PLAN_MODELS,
@@ -93,6 +98,10 @@ import {
   DIGITALOCEAN_CHAT_COMPLETION_MODELS,
   SCALEWAY_SERVERLESS_CHAT_MODELS,
   SCALEWAY_MODEL_INPUT_MODALITIES,
+  OPPER_MODELS,
+  OPPER_MODEL_CONTEXT_WINDOWS,
+  OPPER_MODEL_MAX_OUTPUT_TOKENS,
+  OPPER_MODEL_INPUT_MODALITIES,
 } from "./model-seeds";
 
 export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
@@ -214,6 +223,72 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     },
     note: "Shared Token Factory text-output inference only; live discovery excludes embedding and image-generation rows.",
   },
+  {
+    // Primary sources checked 2026-09-11:
+    // - https://docs.crusoecloud.com/quickstart/getting-started-with-serverless-inference documents
+    //   the fixed OpenAI-compatible host https://api.inference.crusoecloud.com/v1, Bearer API keys
+    //   created in the Cloud console (Intelligence Foundry > Inference > Create API Key), and an
+    //   OpenAI SDK chat.completions example against meta-llama/Llama-3.3-70B-Instruct.
+    // - https://docs.crusoecloud.com/serverless-inference/available-models lists the served models
+    //   with slash-delimited ids; https://docs.crusoecloud.com/serverless-inference/rate-limits
+    //   documents per-project, per-model TPM/RPM limits (429 when exceeded, 503 under shared load).
+    // - GET /v1/models rejects unauthenticated requests with 401 {"errors":["Authentication failed"]},
+    //   so a successful authenticated list response is evidence that the supplied key is valid.
+    //   An authenticated capture on 2026-09-12 returned 18 rows shaped like OpenRouter's catalog
+    //   (`is_public`, `type`, `context_length`, `architecture.modality` of "text" or "multimodal",
+    //   `tags`, `pricing`, `supported_parameters`); 17 were public serverless models and one was an
+    //   account-private dedicated deployment with empty `type`/`modality`. `type` is blank on one
+    //   public model, so the filter keys on `is_public` plus `architecture.modality` instead.
+    // - https://legal.crusoe.ai/ hosts the Crusoe Cloud Platform Terms of Service v1.10 (effective
+    //   2026-08-10), which name Crusoe Technologies LLC as the contracting entity, and the Service
+    //   Specific Terms v5.0 (effective 2026-07-14), whose Crusoe Intelligence Foundry Terms cover the
+    //   Managed Inference Service reached through the Crusoe API.
+    // - https://models.dev/api.json (provider "crusoe") records openai/gpt-oss-120b as the one served
+    //   model with a low/medium/high reasoning_effort ladder; the other reasoning models expose an
+    //   on/off toggle only.
+    // Maintainer: @acheamponge, who works at Crusoe (affiliation disclosed) and also maintains the
+    // models.dev crusoe entry.
+    id: "crusoe",
+    label: "Crusoe",
+    baseUrl: "https://api.inference.crusoecloud.com/v1",
+    adapter: "openai-chat",
+    authKind: "key",
+    dashboardUrl: "https://console.crusoecloud.com",
+    liveModels: true,
+    preserveCustomDestination: true,
+    // The getting-started guide documents tools through the OpenAI SDK but no provider-wide
+    // parallel tool-call contract.
+    parallelToolCalls: false,
+    // Only gpt-oss-120b has a real effort ladder; toggle-style reasoning models must not be promoted
+    // to Codex's full fallback ladder.
+    reasoningEfforts: [],
+    modelReasoningEfforts: { "openai/gpt-oss-120b": ["low", "medium", "high"] },
+    directReasoningEffortModels: ["openai/gpt-oss-120b"],
+    // The catalog reports `architecture.modality: "multimodal"` without an input list. Four rows
+    // also carry the explicit "image text to text" tag; yutori/n2 instead reports multimodal
+    // type/modality plus browser/computer-use tags. Those five captured rows are classified here.
+    modelInputModalities: {
+      "google/gemma-4-31b-it": ["text", "image"],
+      "moonshotai/Kimi-K2.6": ["text", "image"],
+      "nvidia/Nemotron-3-Nano-Omni-Reasoning-30B-A3B": ["text", "image"],
+      "yutori/n2": ["text", "image"],
+      "zai-org/GLM-5.3-Flash": ["text", "image"],
+    },
+    modelDiscovery: {
+      path: "models",
+      maxResponseBytes: 256 * 1024,
+      maxModels: 256,
+      filter: {
+        // Keep public serverless rows whose architecture produces text; account-private
+        // deployments (blank modality) and any embedding or media rows fail closed.
+        allOf: [
+          { path: ["is_public"], equalsAny: [true] },
+          { path: ["architecture", "modality"], equalsAny: ["text", "multimodal"] },
+        ],
+      },
+    },
+    note: "Public Serverless Inference chat models on the shared OpenAI-compatible host; account-private and self-serve dedicated deployments are excluded from discovery and out of scope.",
+  },
   {
     id: "digitalocean",
     label: "DigitalOcean Serverless Inference",
@@ -427,6 +502,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     // model_access_denied, which is why the Chat path cannot simply hang off the new base.
     responsesPath: "/api/v1/responses",
     chatCompletionsPath: "/api/coding/paas/v4/chat/completions",
+    modelDiscovery: { path: "/api/v1/models", envelopeKey: "models", idField: "slug" },
     // The address this row occupied before the move. A saved custom provider still pointing
     // at the Chat endpoint keeps receiving this row's metadata (#1100).
     destinationAliases: [{ baseUrl: "https://api.z.ai/api/coding/paas/v4", adapter: "openai-chat" }],
@@ -724,22 +800,36 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     liveModels: false,
     note: "Token Plan Personal Edition · China (Beijing)",
     modelInputModalities: ALIBABA_TOKEN_PLAN_INPUT_MODALITIES,
-    modelContextWindows: {
-      "qwen3.8-max": 983_616, "qwen3.7-max": 1_000_000, "qwen3.7-plus": 1_000_000,
-      "qwen3.6-flash": 1_000_000, "glm-5.3": 1_000_000, "glm-5.3-flash": 1_000_000, "glm-5.2": 1_000_000, 
-    },
+    modelContextWindows: ALIBABA_TOKEN_PLAN_CONTEXT_WINDOWS,
+    modelMaxOutputTokens: ALIBABA_TOKEN_PLAN_MAX_OUTPUT_TOKENS,
     modelReasoningEfforts: {
       ...Object.fromEntries(ALIBABA_TOKEN_PLAN_QWEN_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
-      "qwen3.8-max": QWEN38_REASONING_EFFORTS,
-      "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
-      "glm-5.3-flash": ZAI_GLM_53_REASONING_EFFORTS,
+      ...Object.fromEntries(QWEN38_FAMILY.map(id => [id, QWEN38_REASONING_EFFORTS])),
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
+      "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
+      "deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek-v4-pro"),
+      "deepseek-v4-pro-0813": deepseekThinkingEffortsFor("deepseek-v4-pro-0813"),
+      "deepseek-v4-flash-0731": deepseekThinkingEffortsFor("deepseek-v4-flash-0731"),
+      "deepseek-v4.1-flash": deepseekThinkingEffortsFor("deepseek-v4.1-flash"),
+    },
+    modelReasoningEffortMap: {
+      "deepseek-v4-pro": deepseekReasoningMapFor("deepseek-v4-pro"),
+      "deepseek-v4-pro-0813": deepseekReasoningMapFor("deepseek-v4-pro-0813"),
+      "deepseek-v4-flash-0731": deepseekReasoningMapFor("deepseek-v4-flash-0731"),
+      "deepseek-v4.1-flash": deepseekReasoningMapFor("deepseek-v4.1-flash"),
     },
-    modelDefaultReasoningEfforts: { "qwen3.8-max": "xhigh" },
-    directReasoningEffortModels: ["qwen3.8-max"],
-    thinkingBudgetModels: ALIBABA_TOKEN_PLAN_QWEN_MODELS.filter(id => id !== "qwen3.8-max"),
-    preserveReasoningContentModels: ["glm-5.3", "glm-5.3-flash", "glm-5.2", "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash"],
-    noVisionModels: ["glm-5.3", "glm-5.2"],
+    // Probed 260915 on the plan gateway: json_object returns valid JSON, strict
+    // json_schema is rejected 400 ("This response_format type is unavailable now")
+    // in both thinking modes, so requests downgrade to json_object rather than
+    // sending a schema the gateway refuses.
+    noJsonSchemaModels: ["deepseek-v4.1-flash"],
+    modelDefaultReasoningEfforts: Object.fromEntries(QWEN38_FAMILY.map(id => [id, "xhigh"])),
+    directReasoningEffortModels: QWEN38_FAMILY,
+    thinkingBudgetModels: ALIBABA_TOKEN_PLAN_QWEN_MODELS.filter(id => !QWEN38_FAMILY.includes(id)),
+    preserveReasoningContentModels: ALIBABA_TOKEN_PLAN_PRESERVE_REASONING,
+    noVisionModels: ALIBABA_TOKEN_PLAN_NO_VISION,
+    // The gateway accepts prompt_cache_key on every Token Plan chat model (probed 260902).
+    promptCacheKey: true,
   },
   {
     id: "alibaba-token-plan-intl",
@@ -756,31 +846,35 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
    note: "Token Plan Team Edition · Singapore (ap-southeast-1)",
     metadataModelIdNormalize: "case-insensitive",
    modelInputModalities: ALIBABA_INTL_TOKEN_PLAN_INPUT_MODALITIES,
-    modelContextWindows: {
-      "qwen3.8-max": 983_616,
-      "qwen3.7-max": 1_000_000, "qwen3.7-plus": 1_000_000, "qwen3.6-plus": 1_000_000, "qwen3.6-flash": 1_000_000,
-      "deepseek-v4-flash": 1_000_000, "deepseek-v3.2": 131_072,
-      "kimi-k2.7-code": 262_144, "kimi-k2.6": 262_144, "kimi-k2.5": 262_144,
-      "glm-5.3": 1_000_000, "glm-5.3-flash": 1_000_000, "glm-5.2": 1_000_000, "glm-5.1": 1_000_000, "glm-5": 1_000_000,
-      "MiniMax-M2.5": 204_800,
-    },
+    modelContextWindows: ALIBABA_TOKEN_PLAN_CONTEXT_WINDOWS,
+    modelMaxOutputTokens: ALIBABA_TOKEN_PLAN_MAX_OUTPUT_TOKENS,
     modelReasoningEfforts: {
       ...Object.fromEntries(ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
-      "qwen3.8-max": QWEN38_REASONING_EFFORTS,
-      "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
-      "glm-5.3-flash": ZAI_GLM_53_REASONING_EFFORTS,
+      ...Object.fromEntries(QWEN38_FAMILY.map(id => [id, QWEN38_REASONING_EFFORTS])),
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
+      "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
+      "deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek-v4-pro"),
+      "deepseek-v4-pro-0813": deepseekThinkingEffortsFor("deepseek-v4-pro-0813"),
       "deepseek-v4-flash": deepseekThinkingEffortsFor("deepseek-v4-flash"),
+      "deepseek-v4-flash-0731": deepseekThinkingEffortsFor("deepseek-v4-flash-0731"),
+      "deepseek-v4.1-flash": deepseekThinkingEffortsFor("deepseek-v4.1-flash"),
     },
     modelReasoningEffortMap: {
+      "deepseek-v4-pro": deepseekReasoningMapFor("deepseek-v4-pro"),
+      "deepseek-v4-pro-0813": deepseekReasoningMapFor("deepseek-v4-pro-0813"),
       "deepseek-v4-flash": deepseekReasoningMapFor("deepseek-v4-flash"),
+      "deepseek-v4-flash-0731": deepseekReasoningMapFor("deepseek-v4-flash-0731"),
+      "deepseek-v4.1-flash": deepseekReasoningMapFor("deepseek-v4.1-flash"),
     },
-    directReasoningEffortModels: ["qwen3.8-max"],
-    thinkingBudgetModels: ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS.filter(id => id !== "qwen3.8-max"),
-    preserveReasoningContentModels: ["glm-5.3", "glm-5.3-flash", "glm-5.2", "deepseek-v4-flash", "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash"],
-    noVisionModels: ["deepseek-v4-flash", "deepseek-v3.2", "glm-5.3", "glm-5.2", "glm-5.1", "glm-5", "MiniMax-M2.5"],
+    // Same 260915 json_schema rejection probe as the Beijing entry.
+    noJsonSchemaModels: ["deepseek-v4.1-flash"],
+    directReasoningEffortModels: QWEN38_FAMILY,
+    thinkingBudgetModels: ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS.filter(id => !QWEN38_FAMILY.includes(id)),
+    preserveReasoningContentModels: ALIBABA_TOKEN_PLAN_PRESERVE_REASONING,
+    noVisionModels: ALIBABA_TOKEN_PLAN_NO_VISION,
     noReasoningModels: ["kimi-k2.7-code", "kimi-k2.6", "kimi-k2.5", "deepseek-v3.2", "glm-5.1", "glm-5", "MiniMax-M2.5"],
-    modelDefaultReasoningEfforts: { "qwen3.8-max": "xhigh" },
+    modelDefaultReasoningEfforts: Object.fromEntries(QWEN38_FAMILY.map(id => [id, "xhigh"])),
+    promptCacheKey: true,
   },
   // NEEDS_HUMAN 2026-07-10: kept for config compatibility, but this is a dashboard URL,
   // no /models endpoint is documented, and tools are silently ignored upstream per docs.parallel.ai.
@@ -935,6 +1029,30 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     noJsonSchemaModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS],
   },
   { id: "vercel-ai-gateway", label: "Vercel AI Gateway", baseUrl: "https://ai-gateway.vercel.sh/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://vercel.com/dashboard" },
+  {
+    // Opper: EU-hosted AI gateway (Opper AI AB, Stockholm). One OpenAI-compatible endpoint and one
+    // key in front of 30+ upstream providers. Seeded ids are Opper *pools* (bare names such as
+    // `claude-sonnet-4-6`): the gateway chooses the provider/region per request, and a
+    // `vendor/model` id (`anthropic/claude-sonnet-4-6`, `aws/claude-sonnet-4-6-eu`) pins one route.
+    // The original provider author reported on 2026-09-08 that GET /v3/compat/models answers 401
+    // without a key, so the default discovery URL doubles as key validation. Windows, output caps
+    // and modalities live in model-seeds.ts (smallest value / shared modality across each pool's
+    // members); live discovery owns which models exist.
+    id: "opper",
+    label: "Opper",
+    adapter: "openai-chat",
+    baseUrl: "https://api.opper.ai/v3/compat",
+    authKind: "key",
+    dashboardUrl: "https://platform.opper.ai",
+    liveModels: true,
+    preserveCustomDestination: true,
+    defaultModel: "claude-sonnet-4-6",
+    models: OPPER_MODELS,
+    modelContextWindows: OPPER_MODEL_CONTEXT_WINDOWS,
+    modelMaxOutputTokens: OPPER_MODEL_MAX_OUTPUT_TOKENS,
+    modelInputModalities: OPPER_MODEL_INPUT_MODALITIES,
+    note: "EU-hosted AI gateway: one OpenAI-compatible endpoint and one key in front of 30+ providers. Bare model ids are pools (claude-sonnet-4-6, gpt-5.5) and Opper picks the route per request; vendor/model ids (anthropic/claude-sonnet-4-6) pin one provider. The catalogue is discovered live from /v3/compat/models with your key; the public list is at opper.ai/models. Token rates are the model providers' rates with no markup; Opper charges a 3% fee when you buy credits.",
+  },
   {
     id: "opencode-free",
     label: "OpenCode Free",

package/src/providers/registry/model-seeds.ts

@@ -315,6 +315,13 @@ export const DEEPSEEK_VISION_PREVIEW_MODEL = "deepseek-v4-flash-vision-exp";
  */
 export const COMMAND_CODE_IMAGE_MODELS = [
   `deepseek/${DEEPSEEK_VISION_PREVIEW_MODEL}`,
+  // Probed 2026-09-18 through a running 2.58.0 proxy: a 3x3 random-color grid
+  // (180x180 PNG, six candidate colors) came back 9/9 correct both as a user
+  // message and as a tool_result, and the request logs show the route served
+  // the image natively — no vision-sidecar call in either window. #4505 asked
+  // for exactly this upstream probe before promoting the id. The sibling
+  // deepseek/deepseek-v4-flash route remains verified-negative above.
+  "deepseek/deepseek-v4.1-flash",
   "gpt-5.6-luna",
   "gpt-5.6-sol",
   "MiniMaxAI/MiniMax-M3",
@@ -334,20 +341,19 @@ export const COMMAND_CODE_IMAGE_MODELS = [
 /**
  * Native image stays sourced from COMMAND_CODE_IMAGE_MODELS. Text-only routes
  * sit beside that list so the catalog can still advertise sidecar coverage
- * without claiming the gateway itself accepts a picture.
- *
- * The gateway-prefixed DeepSeek V4.1 Flash route has no verified native image
- * support, so declaring it image-capable would hand it a picture it drops. A
- * positive text-only declaration makes it a vision-sidecar consumer
+ * without claiming the gateway itself accepts a picture. A positive text-only
+ * declaration makes the route a vision-sidecar consumer
  * (src/vision/eligibility.ts), so the catalog advertises image input on its
- * behalf and the four-target combo in #4505 intersects to ["text","image"]
- * instead of ["text"] — without claiming native vision. modelInputModalities
- * is per-key filled, so this reaches an existing install even when
- * noVisionModels was persisted before the id joined that list.
+ * behalf — without claiming native vision — and modelInputModalities is
+ * per-key filled, so that reaches an existing install even when noVisionModels
+ * was persisted before the id joined a list.
+ *
+ * Empty as of 2026-09-18. Its only entry, deepseek/deepseek-v4.1-flash, moved
+ * to COMMAND_CODE_IMAGE_MODELS once the #4505-requested probe passed on both
+ * the user-message and tool-result paths (see the note at that entry). The
+ * mechanism stays for the next route that measures text-only.
  */
-export const COMMAND_CODE_TEXT_ONLY_MODELS = [
-  "deepseek/deepseek-v4.1-flash",
-] as const;
+export const COMMAND_CODE_TEXT_ONLY_MODELS = [] as const;
 export const COMMAND_CODE_MODEL_INPUT_MODALITIES: Record<string, ["text"] | ["text", "image"]> = {
   ...Object.fromEntries(COMMAND_CODE_IMAGE_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
   ...Object.fromEntries(COMMAND_CODE_TEXT_ONLY_MODELS.map(id => [id, ["text"] as ["text"]])),
@@ -437,36 +443,68 @@ export const deepseekReasoningMapFor = (modelId: string): Record<string, string>
 // Coding Plan: the products use different exact allowlists and different base URLs.
 // Evidence: https://help.aliyun.com/en/model-studio/token-plan-personal-overview
 //           https://help.aliyun.com/en/model-studio/token-plan-quickstart
+// 260909 refresh, re-probed against the live gateway (both regions, both tiers):
+// https://github.com/oliver-mee/alibaba-token-plan-wiki (machine-readable catalog).
+// 260918: glm-5.3 returns. The 260909 removal was correct at the time (the id
+// 404'd on every plan key), but the gateway started serving glm-5.3 on 260917:
+// it now appears on /models for global Team, global Personal, and CN Team, and
+// answers a completion on a Personal key (probed 260918). Contract on the plan
+// gateway: effort low/high/max (default max), thinking always-on (the gateway
+// rejects enable_thinking:false with 400), 1M context, 131,072 max output,
+// text-only input, strict json_schema accepted. glm-5.3-flash REMAINS OUT:
+// still never served by the Token Plan gateway (docs.z.ai VLM id, not plan
+// entitlement).
+// The Beijing preset keeps the Personal Edition subset; non-chat ids (audio/image/
+// video families) stay out: they answer only on async endpoints openai-chat cannot
+// reach. deepseek-v4-pro-0813 is callable but NOT listed by /models, which is the
+// reason liveModels must stay false for this provider. deepseek-v4.1-flash is the
+// 260910 DeepSeek rename row: listed on /models on both tiers and regions from 260915,
+// hybrid thinking, vision via user message and tool result, json_object but not
+// json_schema (see noJsonSchemaModels on the entries).
+// Beijing serves the Personal Edition, so this is the Personal-tier roster probed
+// 260909 (a strict subset of Team). deepseek-v4-pro-0813 stays out of the Beijing
+// entry: its callability is only proven on Team keys, and no Personal key has been
+// shown to reach it. The Beijing entry also shares the intl maps, so it carries a
+// few orphan keys (kimi/glm-5/MiniMax rows); harmless, and one map beats two
+// drifting ones.
 export const ALIBABA_TOKEN_PLAN_MODELS = [
-  "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash",
-  "glm-5.3", "glm-5.3-flash", "glm-5.2", 
+  "qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash",
+  "deepseek-v4-pro", "deepseek-v4-flash-0731", "deepseek-v4.1-flash", "glm-5.2", "glm-5.3",
 ];
 export const ALIBABA_TOKEN_PLAN_QWEN_MODELS = [
-  "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash",
+  "qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash",
 ];
 export const ALIBABA_TOKEN_PLAN_INPUT_MODALITIES: Record<string, string[]> = {
   "qwen3.8-max": ["text", "image"],
-  "qwen3.7-max": ["text", "image"],
+  "qwen3.8-flash": ["text", "image"],
+  "qwen3.7-max": ["text"],
   "qwen3.7-plus": ["text", "image"],
   "qwen3.6-flash": ["text", "image"],
-  "glm-5.3": ["text"],
-  "glm-5.3-flash": ["text", "image"],
+  "deepseek-v4-pro": ["text"],
+  "deepseek-v4-pro-0813": ["text"],
+  "deepseek-v4-flash-0731": ["text"],
+  // Vision probed on the plan gateway 260915 (user message and tool result, both 200).
+  "deepseek-v4.1-flash": ["text", "image"],
   "glm-5.2": ["text"],
+  "glm-5.3": ["text"],
 };
 
 // 260721 Alibaba Token Plan International (ap-southeast-1 / Singapore, hardened 260721).
 // Multi-vendor lineup distinct from Beijing — includes DeepSeek V4 flash, Kimi K2.7, MiniMax.
 // Evidence: https://www.alibabacloud.com/help/en/model-studio/token-plan-overview
 //           https://qwencloud.com/pricing/token-plan (qwen3.8 metadata)
+// The Team Edition roster (Singapore), verified identical to the CN Team set on 260909.
+// deepseek-v4-pro is restored: it remains callable on the plan gateway (probed 260909,
+// listed on /models on both regions) after being dropped as "retired" upstream.
 export const ALIBABA_INTL_TOKEN_PLAN_MODELS = [
-  "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
-  "deepseek-v4-flash", "deepseek-v3.2",
+  "qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
+  "deepseek-v4-pro", "deepseek-v4-pro-0813", "deepseek-v4-flash", "deepseek-v4-flash-0731", "deepseek-v4.1-flash", "deepseek-v3.2",
   "kimi-k2.7-code", "kimi-k2.6", "kimi-k2.5",
-  "glm-5.3", "glm-5.3-flash", "glm-5.2", "glm-5.1", "glm-5",
+  "glm-5.2", "glm-5.3", "glm-5.1", "glm-5",
   "MiniMax-M2.5",
 ];
 export const ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS = [
-  "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
+  "qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
 ];
 
 // 260722 Tencent Cloud Coding Plan. The plan's model set is explicitly dynamic; these are the
@@ -543,24 +581,49 @@ export const VOLCENGINE_PLAN_TEXT_ONLY_MODELS = [
   "doubao-seed-2.0-pro",
 ];
 export const ALIBABA_INTL_TOKEN_PLAN_INPUT_MODALITIES: Record<string, string[]> = {
-  "qwen3.8-max": ["text", "image"],
-  "qwen3.7-max": ["text", "image"],
-  "qwen3.7-plus": ["text", "image"],
+  ...ALIBABA_TOKEN_PLAN_INPUT_MODALITIES,
   "qwen3.6-plus": ["text", "image"],
-  "qwen3.6-flash": ["text", "image"],
   "deepseek-v4-flash": ["text"],
   "deepseek-v3.2": ["text"],
   "kimi-k2.7-code": ["text", "image"],
   "kimi-k2.6": ["text", "image"],
   "kimi-k2.5": ["text", "image"],
-  "glm-5.3": ["text"],
-  "glm-5.3-flash": ["text", "image"],
-  "glm-5.2": ["text"],
   "glm-5.1": ["text"],
   "glm-5": ["text"],
   "MiniMax-M2.5": ["text"],
 };
 
+// Shared Token Plan metadata (260909 gateway probes; output ceilings are max_tokens
+// boundary probes: accept at N, reject at N+1).
+export const QWEN38_FAMILY = ["qwen3.8-max", "qwen3.8-flash"];
+export const ALIBABA_TOKEN_PLAN_CONTEXT_WINDOWS: Record<string, number> = {
+  "qwen3.8-max": 1_000_000, "qwen3.8-flash": 1_000_000, "qwen3.7-max": 1_000_000, "qwen3.7-plus": 1_000_000,
+  "qwen3.6-plus": 1_000_000, "qwen3.6-flash": 1_000_000,
+  "deepseek-v4-pro": 1_000_000, "deepseek-v4-pro-0813": 1_000_000, "deepseek-v4-flash": 1_000_000,
+  "deepseek-v4-flash-0731": 1_000_000, "deepseek-v4.1-flash": 1_000_000, "deepseek-v3.2": 131_072,
+  "kimi-k2.7-code": 262_144, "kimi-k2.6": 262_144, "kimi-k2.5": 262_144,
+  "glm-5.2": 1_000_000, "glm-5.3": 1_000_000, "glm-5.1": 202_752, "glm-5": 202_752,
+  "MiniMax-M2.5": 196_608,
+};
+export const ALIBABA_TOKEN_PLAN_MAX_OUTPUT_TOKENS: Record<string, number> = {
+  "qwen3.8-max": 131_072, "qwen3.8-flash": 131_072, "qwen3.7-max": 131_072, "qwen3.7-plus": 131_072,
+  "qwen3.6-plus": 65_536, "qwen3.6-flash": 65_536,
+  "deepseek-v4-pro": 393_216, "deepseek-v4-pro-0813": 393_216, "deepseek-v4-flash": 393_216,
+  "deepseek-v4-flash-0731": 393_216, "deepseek-v4.1-flash": 393_216, "deepseek-v3.2": 65_536,
+  "kimi-k2.7-code": 262_144, "kimi-k2.6": 262_144, "kimi-k2.5": 98_304,
+  "glm-5.2": 131_072, "glm-5.3": 131_072, "glm-5.1": 128_000, "glm-5": 16_384,
+  "MiniMax-M2.5": 32_768,
+};
+export const ALIBABA_TOKEN_PLAN_NO_VISION = [
+  "qwen3.7-max", "deepseek-v4-pro", "deepseek-v4-pro-0813", "deepseek-v4-flash",
+  "deepseek-v4-flash-0731", "deepseek-v3.2", "glm-5.2", "glm-5.3", "glm-5.1", "glm-5", "MiniMax-M2.5",
+];
+export const ALIBABA_TOKEN_PLAN_PRESERVE_REASONING = [
+  "qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
+  "deepseek-v4-pro", "deepseek-v4-pro-0813", "deepseek-v4-flash", "deepseek-v4-flash-0731",
+  "deepseek-v4.1-flash", "glm-5.2", "glm-5.3",
+];
+
 // 260717 Kimi K3: the subscription endpoint uses one upstream id (`k3`) for both
 // entitlement tiers. Bare `k3` advertises the Moderato 256K ceiling; the local `[1m]`
 // alias advertises Allegretto's 1M ceiling and is stripped before the upstream request.
@@ -910,3 +973,47 @@ export const CLINE_PASS_TEXT_ONLY_MODELS = CLINE_PASS_MODALITY_KNOWN_MODELS.filt
 export const CLINE_PASS_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
   CLINE_PASS_MODALITY_KNOWN_MODELS.map(id => [id, CLINE_PASS_IMAGE_MODELS.has(id) ? ["text", "image"] : ["text"]]),
 );
+
+// Opper seed: bare *pool* names. A pool is every provider Opper serves that model through; Opper
+// picks the route per request. Each name is the `.model` of a `pooled: true` entry in the public
+// catalogue snapshot supplied by the original provider author
+// (https://api.opper.ai/v3/models?limit=2000, captured 2026-09-14); `vendor/model` ids
+// (anthropic/claude-sonnet-4-6) pin one route and stay valid, they are just not seeded.
+export const OPPER_MODELS = [
+  "claude-sonnet-4-6",
+  "claude-opus-5",
+  "gpt-5.5",
+  "gpt-5.4-mini",
+  "gemini-3.8-flash",
+  "deepseek-v4-pro",
+  "kimi-k3",
+  "mistral-large-2512",
+];
+// Smallest value across each pool's members in that snapshot, capped at the lab model's own limit
+// (kimi-k3 output); live discovery owns which models exist.
+export const OPPER_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
+  "claude-sonnet-4-6": 1_000_000,
+  "claude-opus-5": 1_000_000,
+  "gpt-5.5": 1_050_000,
+  "gpt-5.4-mini": 400_000,
+  "gemini-3.8-flash": 1_048_576,
+  "deepseek-v4-pro": 1_000_000,
+  "kimi-k3": 1_048_576,
+  "mistral-large-2512": 256_000,
+};
+export const OPPER_MODEL_MAX_OUTPUT_TOKENS: Record<string, number> = {
+  "claude-sonnet-4-6": 64_000,
+  "claude-opus-5": 128_000,
+  "gpt-5.5": 128_000,
+  "gpt-5.4-mini": 128_000,
+  "gemini-3.8-flash": 65_536,
+  "deepseek-v4-pro": 65_536,
+  "kimi-k3": 131_072,
+  "mistral-large-2512": 8_192,
+};
+// Pools whose members do not all accept image input (deepseek-v4-pro: no member does; kimi-k3: the
+// sference route is text-only), so the shared modality set is text.
+export const OPPER_TEXT_ONLY_MODELS = ["deepseek-v4-pro", "kimi-k3"];
+export const OPPER_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
+  OPPER_MODELS.map(id => [id, OPPER_TEXT_ONLY_MODELS.includes(id) ? ["text"] : ["text", "image"]]),
+);

package/src/providers/registry/types.ts

@@ -64,6 +64,10 @@ export interface ProviderModelDiscoveryFilter {
 interface ProviderModelDiscoverySharedSpec {
   /** Query parameters applied to the resolved discovery URL. */
   query?: Readonly<Record<string, string>>;
+  /** Top-level response key containing model rows; defaults to `data`. */
+  envelopeKey?: string;
+  /** Model-row field containing the provider-native identifier; defaults to `id`. */
+  idField?: string;
   /** Declarative eligibility rules evaluated against each untrusted model row. */
   filter?: ProviderModelDiscoveryFilter;
   /** Optional lower byte ceiling; the process-wide hard ceiling still wins. */
@@ -217,6 +221,11 @@ export interface ProviderRegistryEntry {
    * to stay contiguous. This is seeded/backfilled like other fixed wire capabilities.
    */
   requiresAdjacentResponsesToolResults?: boolean;
+  /**
+   * Responses upstream that also rejects a tool call with no matching output anywhere in the
+   * replayed input. Seeded/backfilled like other fixed wire capabilities.
+   */
+  requiresPairedResponsesToolResults?: boolean;
   /**
    * When enabled, tool results that are present but empty are annotated on the wire.
    * Seeded/backfilled like other fixed wire capabilities.

package/src/responses/apply-patch-envelope.ts

@@ -2,9 +2,9 @@
 //
 // Some routed models decorate the first and last lines as
 // `*** Begin Patch ***` / `*** End Patch ***`. Codex rejects those otherwise
-// valid custom-tool payloads. Repair is deliberately limited to a complete,
-// structurally recognizable top-level patch: arbitrary `exec` JavaScript is
-// caller-authored executable input and must remain byte-identical.
+// valid custom-tool payloads. Repair of executable bodies is deliberately limited to
+// unambiguous wrapper mistakes: one recognized alternate field or one complete outer
+// Markdown fence. Ordinary `exec` JavaScript remains byte-identical.
 //
 // This is the same intent boundary as `src/lib/tool-argument-integers.ts`:
 // repair the one faithful reading, leave genuine patch content alone.
@@ -22,20 +22,52 @@ const PATCH_BEGIN = "*** Begin Patch";
 const PATCH_END = "*** End Patch";
 const TOP_LEVEL_PATCH_ENVELOPE = /^(\*\*\* Begin Patch(?: \*\*\*)?)(\r?\n)([\s\S]*)(\r?\n)(\*\*\* End Patch(?: \*\*\*)?)(\r?\n)?$/;
 const PATCH_OPERATION_LINE = /^\*\*\* (?:Add|Update|Delete) File: .+$/m;
+const OUTER_MARKDOWN_CODE_FENCE = /^```[^\r\n]*\r?\n([\s\S]*?)\r?\n```$/;
+const FREEFORM_FALLBACK_KEYS: Readonly<Record<string, readonly string[]>> = {
+  exec: ["code", "script", "js", "javascript", "command", "cmd", "content"],
+  apply_patch: ["patch", "content"],
+};
+
+function stripMarkdownCodeFence(text: string, toolName: string): string {
+  if (toolName !== "exec" && toolName !== "apply_patch") return text;
+  const match = OUTER_MARKDOWN_CODE_FENCE.exec(text.trim());
+  return match ? match[1] : text;
+}
+
+/**
+ * The single-field wrappers `unwrapFreeformToolInput` accepts for one tool name, besides the
+ * canonical `input`.
+ *
+ * Exported so the streaming side can hold a buffer that is still turning into one of these.
+ * A second list of key names beside this one is how the streamed bytes and the completed item
+ * come to disagree, which is the defect it exists to prevent (#5047).
+ */
+export function freeformFallbackKeys(toolName: string): readonly string[] {
+  return FREEFORM_FALLBACK_KEYS[toolName] ?? [];
+}
 
 /** Unwrap the `{input:string}` function-call wrapper used for freeform tools. */
-export function unwrapFreeformToolInput(argumentsText: unknown): string {
+export function unwrapFreeformToolInput(argumentsText: unknown, toolName = ""): string {
   if (typeof argumentsText !== "string") return "";
   try {
     const parsed: unknown = JSON.parse(argumentsText);
     if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-      const input = (parsed as { input?: unknown }).input;
-      if (typeof input === "string") return input;
+      const record = parsed as Record<string, unknown>;
+      if (Object.prototype.hasOwnProperty.call(record, "input")) {
+        return typeof record.input === "string"
+          ? stripMarkdownCodeFence(record.input, toolName)
+          : argumentsText;
+      }
+      const fallbackKeys = FREEFORM_FALLBACK_KEYS[toolName] ?? [];
+      const candidates = fallbackKeys.filter(key => typeof record[key] === "string");
+      if (candidates.length === 1) {
+        return stripMarkdownCodeFence(record[candidates[0]] as string, toolName);
+      }
     }
   } catch {
     // The string is the freeform body, not nested JSON.
   }
-  return argumentsText;
+  return stripMarkdownCodeFence(argumentsText, toolName);
 }
 
 /**
@@ -92,17 +124,18 @@ export function mayBecomePatchEnvelope(text: string): boolean {
 /**
  * Repair freeform input before Codex sees it.
  *
- * Only a bare or reserved-`functions` `apply_patch` payload may receive delimiter
- * repair. Remote namespaces own their grammar; those bodies and every other
- * freeform input are unwrapped and left byte-exact.
+ * Only a bare or reserved-`functions` tool may receive fallback-field or outer-fence
+ * repair, and only `apply_patch` may receive delimiter repair. Remote namespaces own
+ * their grammar; those bodies and every other freeform input are unwrapped and left
+ * byte-exact.
  */
 export function repairFreeformToolInput(
   argumentsText: unknown,
   toolName = "",
   namespace?: string,
 ): string {
-  const unwrapped = unwrapFreeformToolInput(argumentsText);
   const ownsApplyPatchGrammar = namespace === undefined || namespace === "functions";
+  const unwrapped = unwrapFreeformToolInput(argumentsText, ownsApplyPatchGrammar ? toolName : "");
   return ownsApplyPatchGrammar && toolName === "apply_patch"
     ? normalizeApplyPatchDelimiters(unwrapped)
     : unwrapped;