@bitkyc08/opencodex 2.56.0 → 2.58.0

package/src/providers/model-discovery.ts

@@ -23,6 +23,8 @@ import {
 
 const MODEL_DISCOVERY_MAX_FILTER_VALUES = 256;
 const MODEL_DISCOVERY_MAX_FILTER_STRING_LENGTH = 1_024;
+const TRAILING_SLASHES = /\/+$/;
+const TRAILING_MODELS = /\/models$/;
 
 export interface ResolvedProviderModelDiscovery {
   spec?: ProviderModelDiscoverySpec;
@@ -50,6 +52,20 @@ export type ModelEnvelopeRowsResult =
   | { ok: true; rows: unknown[] }
   | { ok: false; reason: "invalid_shape" | "too_many_models" };
 
+/**
+ * Build the default OpenAI-compatible model-discovery URL from a configured baseUrl.
+ *
+ * `baseUrl` is required on both `OcxProviderConfig` and the persisted-config schema, so a row
+ * without one is not a state configuration loading can produce. It is deliberately not tolerated
+ * here: the old template-literal join silently produced `"undefined/models"`, which is not a usable
+ * fallback either — it only ever survived because a static row returns before the URL is parsed.
+ */
+export function providerModelsUrl(baseUrl: string): string {
+  const trimmed = baseUrl.trim().replace(TRAILING_SLASHES, "");
+  const withoutEndpoint = trimmed.replace(TRAILING_MODELS, "");
+  return `${withoutEndpoint}/models`;
+}
+
 function positiveIntegerAtMost(value: number | undefined, hardLimit: number): number {
   if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) return hardLimit;
   return Math.min(Math.floor(value), hardLimit);
@@ -113,6 +129,14 @@ export function providerModelDiscoverySpecError(spec: ProviderModelDiscoverySpec
   if (queryEntries.some(([key, value]) => !key.trim() || key.length > 128 || typeof value !== "string" || value.length > 512)) {
     return "discovery query keys/values exceed their bounds";
   }
+  for (const [field, value] of [
+    ["envelopeKey", spec.envelopeKey],
+    ["idField", spec.idField],
+  ] as const) {
+    if (value !== undefined && (
+      typeof value !== "string" || !value || value !== value.trim() || value.length > 128
+    )) return `${field} must be a nonblank field name up to 128 characters`;
+  }
   for (const [field, value, hardLimit] of [
     ["maxResponseBytes", spec.maxResponseBytes, MODEL_DISCOVERY_MAX_RESPONSE_BYTES],
     ["maxModels", spec.maxModels, MODEL_DISCOVERY_MAX_MODELS],
@@ -406,7 +430,7 @@ export function extractModelEnvelopeRows(
   return { ok: true, rows };
 }
 
-/** Validate, bound, deduplicate, and declaratively filter OpenAI `{data:[...]}` or top-level arrays (Together `#617`). */
+/** Validate, bound, deduplicate, and filter the declared envelope or a top-level array (Together `#617`). */
 /**
  * Metadata a sibling `models[]` array may contribute to an ALREADY-ADMITTED
  * `data[]` row (#1797).
@@ -485,24 +509,26 @@ export function extractProviderModelItems(
   let data: unknown[];
   let siblings: SiblingIndex | null = null;
   if (Array.isArray(value)) {
-    // Together-style top-level /models arrays. Catalog discovery must not treat a stray
-    // `models` key on openai-chat responses as valid — only `data` envelopes or top-level arrays.
+    // Together-style top-level /models arrays. The default contract must not treat a stray
+    // `models` key on openai-chat responses as valid; only a provider spec may opt into it.
     if (value.length > limit) return { ok: false, reason: "too_many_models" };
     data = value;
   } else {
-    const envelope = extractModelEnvelopeRows(value, discovery.maxModels, ["data"]);
+    const envelopeKey = discovery.spec?.envelopeKey ?? "data";
+    const envelope = extractModelEnvelopeRows(value, discovery.maxModels, [envelopeKey]);
     if (!envelope.ok) return envelope;
     data = envelope.rows;
-    siblings = buildSiblingIndex(value, limit);
+    siblings = envelopeKey === "data" ? buildSiblingIndex(value, limit) : null;
   }
 
   const items: ProviderModelsApiItem[] = [];
   const seen = new Set<string>();
+  const idField = discovery.spec?.idField ?? "id";
   for (const raw of data) {
     if (raw === null || typeof raw !== "object" || Array.isArray(raw)) {
       return { ok: false, reason: "invalid_shape" };
     }
-    const id = (raw as { id?: unknown }).id;
+    const id = (raw as Record<string, unknown>)[idField];
     if (!isValidModelDiscoveryModelId(id)) return { ok: false, reason: "invalid_shape" };
     const prefix = discovery.spec?.stripIdPrefix;
     let finalId = id;
@@ -510,7 +536,9 @@ export function extractProviderModelItems(
       finalId = finalId.slice(prefix.length);
       if (!isValidModelDiscoveryModelId(finalId)) continue;
     }
-    const item = finalId === id ? raw as ProviderModelsApiItem : { ...(raw as ProviderModelsApiItem), id: finalId };
+    const item = finalId === id && idField === "id"
+      ? raw as ProviderModelsApiItem
+      : { ...(raw as Record<string, unknown>), id: finalId };
     // Admission is decided on the ORIGINAL `data[]` row, before any sibling
     // enrichment. Merging first let a `models[]` entry supply the very field a
     // provider filter requires — reproduced against the real Chutes policy,

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

@@ -17,6 +17,7 @@ import type { ProviderRegistryEntry } from "./types";
 import {
   ANTHROPIC_MODELS,
   ANTHROPIC_MODEL_CONTEXT_WINDOWS,
+  ANTHROPIC_MODEL_INPUT_MODALITIES,
   ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS,
   ANTHROPIC_MODEL_REASONING_EFFORTS,
   ZAI_GLM_52_REASONING_EFFORTS,
@@ -281,6 +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
@@ -383,6 +395,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     note: "Log in with your Claude account",
     models: [...ANTHROPIC_MODELS],
     modelContextWindows: { ...ANTHROPIC_MODEL_CONTEXT_WINDOWS },
+    modelInputModalities: { ...ANTHROPIC_MODEL_INPUT_MODALITIES },
     modelReasoningEfforts: { ...ANTHROPIC_MODEL_REASONING_EFFORTS },
     // Codex omits max_output_tokens; without a provider budget the Anthropic adapter
     // falls back to 8192, which truncates long answers with stop_reason=max_tokens.
@@ -403,6 +416,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     models: [...ANTHROPIC_MODELS],
     liveModels: true,
     modelContextWindows: { ...ANTHROPIC_MODEL_CONTEXT_WINDOWS },
+    modelInputModalities: { ...ANTHROPIC_MODEL_INPUT_MODALITIES },
     modelReasoningEfforts: { ...ANTHROPIC_MODEL_REASONING_EFFORTS },
     defaultMaxOutputTokens: ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS,
     defaultModel: "claude-sonnet-5",
@@ -420,6 +434,10 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // or the one the Claude /v1/messages inbound derives); the adapter itself never invents one.
     // Evidence: https://platform.kimi.com/docs/api/chat
     promptCacheKey: true,
+    // Kimi's Responses endpoint rejects hook-provided context between a tool call and
+    // its matching result (#4726), the same strict shape DeepSeek exposed in #1292.
+    // The flag is inert while this preset uses the Chat wire.
+    requiresAdjacentResponsesToolResults: true,
     featured: true,
     oauthId: "kimi",
     jawcodeBundle: "moonshot",

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,
@@ -110,6 +115,13 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     // Baseten says models outside its reasoning table do not support reasoning. Keep
     // unknown/new live slugs conservative until an official-docs registry refresh proves it.
     reasoningEfforts: [],
+    // `text.verbosity` is an OpenAI Responses parameter. Baseten documents its Model
+    // APIs as Chat Completions compatible, so there is nothing on that wire for it to
+    // become, and a routed row must not inherit the Codex template's verbosity picker
+    // (#4630: Codex sent `text: { verbosity: "low" }` and the turn 400'd before any
+    // model output). Provider-wide rather than per-model because this catalog is live-
+    // discovered: a slug that arrives tomorrow supports it no more than the seeded ones.
+    supportsVerbosity: false,
     modelReasoningEfforts: BASETEN_MODEL_REASONING_EFFORTS,
     modelReasoningEffortMap: BASETEN_MODEL_REASONING_EFFORT_MAP,
     modelDefaultReasoningEfforts: BASETEN_MODEL_DEFAULT_REASONING_EFFORTS,
@@ -420,6 +432,7 @@ 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" }],
@@ -717,22 +730,35 @@ 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,
+      "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",
@@ -749,31 +775,34 @@ 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,
+      "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.
@@ -883,6 +912,8 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     modelSuffixBracketStrip: true,
     // API-key form of the same Kimi Code Plan transport; keep cache affinity identical to OAuth.
     promptCacheKey: true,
+    // Keep Responses tool-result adjacency aligned with the OAuth preset (#4726).
+    requiresAdjacentResponsesToolResults: true,
     models: KIMI_CODING_MODELS,
     modelContextWindows: KIMI_CODING_MODEL_CONTEXT_WINDOWS,
     modelInputModalities: KIMI_CODING_MODEL_INPUT_MODALITIES,

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

@@ -8,6 +8,10 @@ import type { ProviderModelDiscoverySpec } from "./types";
 // always on, per the official models overview and pricing page (platform.claude.com).
 export const ANTHROPIC_MODELS = ["claude-fable-5-1", "claude-fable-5", "claude-sonnet-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6", "claude-sonnet-4-6", "claude-haiku-4-5"];
 export const ANTHROPIC_MODEL_CONTEXT_WINDOWS: Record<string, number> = { "claude-fable-5-1": 1_000_000, "claude-sonnet-5": 1_000_000, "claude-fable-5": 1_000_000, "claude-opus-5": 1_000_000, "claude-opus-4-8": 1_000_000, "claude-opus-4-7": 1_000_000, "claude-opus-4-6": 1_000_000, "claude-sonnet-4-6": 1_000_000, "claude-haiku-4-5": 200_000 };
+// All seeded Claude models support vision: https://platform.claude.com/docs/en/models/overview
+export const ANTHROPIC_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
+  ANTHROPIC_MODELS.map(id => [id, ["text", "image"]]),
+);
 // Every current Claude family accepts at least 64k output tokens (Haiku 4.5 / Sonnet 4.x
 // through Opus 5 and Fable 5). Anthropic caps max_tokens per model server-side, so a
 // larger request never over-allocates; it only stops the 8192 truncation.
@@ -433,20 +437,42 @@ 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).
+// glm-5.3 / glm-5.3-flash removed from both Token Plan catalogs: they exist on Z.AI
+// endpoints but the Token Plan gateway has never served either id (the 260826 seed
+// propagated them across every GLM-carrying catalog; a selected row 404s).
+// 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",
 ];
 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"],
 };
 
@@ -454,15 +480,18 @@ export const ALIBABA_TOKEN_PLAN_INPUT_MODALITIES: Record<string, string[]> = {
 // 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-5.1", "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
@@ -539,24 +568,49 @@ 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-5.1": 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-5.1": 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-5.1", "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",
+];
+
 // 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.

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/reasoning-envelope.ts

@@ -28,9 +28,12 @@ export interface ReasoningEnvelope {
    */
   txt?: string;
   /**
-   * Kiro `reasoningContentEvent.redactedContent`: a KMS-encrypted reasoning blob that is opaque to
-   * the proxy. Kiro's own CLI replays it on the matching `assistantResponseMessage` to preserve
-   * model reasoning across turns, so it round-trips here the same way a signature does.
+   * Kiro's reasoning blob from `reasoningContentEvent`: a KMS-encrypted value that is opaque to the
+   * proxy (the GPT-5.6 family sends it as `signature`, other models as the base64
+   * `redactedContent`, and the value carries a tag naming which one — see
+   * src/adapters/kiro/reasoning.ts). Kiro's own CLI replays it on the matching
+   * `assistantResponseMessage` to preserve model reasoning across turns, so it round-trips here the
+   * same way a signature does.
    */
   krc?: string;
 }

package/src/responses/spill-store.ts

@@ -159,6 +159,23 @@ function spillNow(): number {
   return spillNowOverride?.() ?? Date.now();
 }
 
+/**
+ * The spill deadline clock, shared with the shutdown drain in `state/spill-queue.ts`.
+ *
+ * Every deadline the shutdown path enforces has to read the same clock the work it
+ * budgets reads. When the drain measured its reserve on `Date.now()` while the ACL
+ * harden it was budgeting ran on this injected clock, a test could freeze the clock,
+ * believe it had removed wall time from the case, and still lose an 80 ms reserve to
+ * real elapsed time on a loaded runner — which is what turned
+ * `shutdown fallback prices the job-owned superseded generation before publishing`
+ * red on macOS 2/2 in run 35137850114 while the assertion it was written for never ran.
+ *
+ * Production is unchanged: with no override installed this is `Date.now()`.
+ */
+export function responseSpillNow(): number {
+  return spillNow();
+}
+
 function record(event: "write" | "fsync" | "close" | "harden" | "publish" | "dir-fsync" | "stub-swap"): void {
   spillIoForTest?.record?.(event);
 }

package/src/responses/state/body-policy.ts

@@ -0,0 +1,25 @@
+/**
+ * Request bodies that must never enter the continuation cache.
+ *
+ * The cache is persisted to `responses-state.json`, so anything recorded here reaches disk.
+ * Encrypted-agent-task recovery decrypts task text into the request body and promises
+ * in-memory, TTL-bounded retention; recording that body would put the plaintext on disk with
+ * no TTL and break the promise.
+ *
+ * A WeakSet rather than a body field on purpose: `_rawBody` is serialized verbatim by the
+ * native passthrough, so any marker written into the body itself would be sent upstream.
+ * Marking is enforced once here rather than at each call site, because every recording path
+ * (streaming, non-streaming, passthrough, forced) funnels through `rememberResponseState` —
+ * a new call site cannot reintroduce the leak by forgetting a guard.
+ */
+const nonPersistableBodies = new WeakSet<object>();
+
+/** Bar this exact request body from the continuation cache, and therefore from disk. */
+export function markBodyNonPersistable(body: unknown): void {
+  if (body && typeof body === "object") nonPersistableBodies.add(body as object);
+}
+
+/** Test the body's in-memory persistence restriction without adding a wire marker. */
+export function isBodyNonPersistable(body: unknown): boolean {
+  return !!body && typeof body === "object" && nonPersistableBodies.has(body);
+}

package/src/responses/state/spill-queue.ts

@@ -7,6 +7,7 @@ import {
   MAX_RESPONSE_SPILL_PAYLOAD_BYTES,
   prospectiveResponseSpillBytes,
   responseSpillPayloadCap,
+  responseSpillNow,
   type ResponseSpillPublicationControl,
   type ResponseSpillRef,
   writeResponseSpillDurably,
@@ -377,7 +378,7 @@ function responseSpillShutdownBudget(): { totalMs: number; fallbackReserveMs: nu
 }
 
 function awaitResponseSpillTailUntil(observed: Promise<void>, deadline: number): Promise<boolean> {
-  const remaining = deadline - Date.now();
+  const remaining = deadline - responseSpillNow();
   if (remaining <= 0) return Promise.resolve(false);
   return new Promise(resolve => {
     let finished = false;
@@ -554,12 +555,13 @@ function terminalizeExhaustedShutdownFallback(
 }
 
 function fallbackPendingResponseSpills(reserveMs: number): Error[] {
-  const deadline = Date.now() + reserveMs;
+  // Same clock as the harden work this reserve is budgeting — see `responseSpillNow`.
+  const deadline = responseSpillNow() + reserveMs;
   const failures: Error[] = [];
   for (;;) {
     const pending = pendingShutdownFallbackCandidates();
     if (pending.length === 0) return failures;
-    if (Date.now() >= deadline) {
+    if (responseSpillNow() >= deadline) {
       terminalizeExhaustedShutdownFallback(pending, failures);
       return failures;
     }
@@ -569,7 +571,7 @@ function fallbackPendingResponseSpills(reserveMs: number): Error[] {
     for (let index = 0; index < pending.length; index += 1) {
       const { job, candidate } = pending[index]!;
       if (requireStore().currentEntry(job.id) !== candidate) continue;
-      const remaining = deadline - Date.now();
+      const remaining = deadline - responseSpillNow();
       if (remaining <= 0) {
         reserveExhausted = true;
         for (const exhausted of pending.slice(index)) {
@@ -587,7 +589,7 @@ function fallbackPendingResponseSpills(reserveMs: number): Error[] {
     requireStore().recomputeOldestResident();
     requireStore().pruneResponses();
     enforceAppOwnedMemoryBudget();
-    if (reserveExhausted || Date.now() >= deadline) {
+    if (reserveExhausted || responseSpillNow() >= deadline) {
       terminalizeExhaustedShutdownFallback(pendingShutdownFallbackCandidates(), failures);
       return failures;
     }
@@ -597,7 +599,7 @@ function fallbackPendingResponseSpills(reserveMs: number): Error[] {
 export async function drainResponseSpillPublications(): Promise<void> {
   const budget = responseSpillShutdownBudget();
   const fallbackReserveMs = Math.min(budget.totalMs, Math.max(1, budget.fallbackReserveMs));
-  const drainDeadline = Date.now() + Math.max(0, budget.totalMs - fallbackReserveMs);
+  const drainDeadline = responseSpillNow() + Math.max(0, budget.totalMs - fallbackReserveMs);
 
   for (;;) {
     if (pendingResponseSpills.size === 0) return;

package/src/responses/state.ts

@@ -23,6 +23,8 @@ import type { ResponseSpillWriteFailureCode, ResponseSpillWriteStatus, ResponseS
 export { responseAdmissionCountersForTests } from "./state/spill-failure";
 import { admissionCounters, noteSpillWriteFailure, noteSpillWriteSuccess, spillCounters, spillWriteHealth } from "./state/spill-failure";
 import { loadSnapshotEntry } from "./state/snapshot-codec";
+import { isBodyNonPersistable } from "./state/body-policy";
+export { isBodyNonPersistable, markBodyNonPersistable } from "./state/body-policy";
 export { flushPendingResponseSpillsForTests, awaitResponseSpillPublicationTailForTests, pendingResponseSpillMetricsForTests, setResponseSpillShutdownBudgetForTests, setResponseSpillAsyncAclAttemptBudgetForTests, setResponseSpillShutdownTerminalizationPassLimitForTests } from "./state/spill-queue";
 import {
   bindSpillQueueStore,
@@ -1235,27 +1237,6 @@ export function responseStateMetrics(): ResponseStateMetrics {
  * Cache completed output and max_output_tokens partial output for previous_response_id replay.
  * Content-filtered incomplete and failed output are not authoritative replay history.
  */
-/**
- * Request bodies that must never enter the continuation cache.
- *
- * The cache is persisted to `responses-state.json`, so anything recorded here reaches disk.
- * Encrypted-agent-task recovery decrypts task text into the request body and promises
- * in-memory, TTL-bounded retention; recording that body would put the plaintext on disk with
- * no TTL and break the promise.
- *
- * A WeakSet rather than a body field on purpose: `_rawBody` is serialized verbatim by the
- * native passthrough, so any marker written into the body itself would be sent upstream.
- * Marking is enforced once here rather than at each call site, because every recording path
- * (streaming, non-streaming, passthrough, forced) funnels through `rememberResponseState` —
- * a new call site cannot reintroduce the leak by forgetting a guard.
- */
-const nonPersistableBodies = new WeakSet<object>();
-
-/** Bar this exact request body from the continuation cache, and therefore from disk. */
-export function markBodyNonPersistable(body: unknown): void {
-  if (body && typeof body === "object") nonPersistableBodies.add(body as object);
-}
-
 export function rememberResponseState(
   requestBody: unknown,
   response: { id?: unknown; output?: unknown; status?: unknown; incomplete_details?: unknown },
@@ -1264,7 +1245,7 @@ export function rememberResponseState(
 ): void {
   if (!requestBody || typeof requestBody !== "object" || Array.isArray(requestBody)) return;
   const request = requestBody as Record<string, unknown>;
-  if (nonPersistableBodies.has(request)) return;
+  if (isBodyNonPersistable(request)) return;
   // `force` bypasses only the store:false skip: Codex sends `store:false` on every non-Azure
   // HTTP request (and WS inherits it), yet its WS turns still chain with previous_response_id.
   // The passthrough branch records with force so those chains can be expanded locally; the