@bitkyc08/opencodex 2.63.0 → 2.64.0-preview.20260923

package/src/providers/reasoning-metadata.ts

@@ -12,8 +12,8 @@
  * fallback ladder, so the Codex catalog AND the wire clamp agree with the model instead of a
  * hand-written guess.
  *
- * Failure policy: the network is never on the critical path. A missing, stale or corrupt
- * snapshot yields undefined, which leaves every hand-written contract untouched. The second
+ * Failure policy: a missing or corrupt snapshot yields undefined, and an expired snapshot still
+ * serves its last ladder while a best-effort refresh runs in the background. The second
  * cache records rungs the upstream actually rejected (400/403 naming reasoning_effort), so an
  * entitlement gap (muse-spark max needs an active Muse Code subscription) costs one rejected
  * request instead of failing every turn that selects that rung.
@@ -137,6 +137,11 @@ function metadataProviderKey(provider: OcxProviderConfig): string | undefined {
   return undefined;
 }
 
+/** Whether catalog sync should bootstrap metadata for this destination. */
+export function providerUsesReasoningMetadata(provider: OcxProviderConfig): boolean {
+  return metadataProviderKey(provider) !== undefined;
+}
+
 /**
  * Local mirror of `modelRecordValue()` from `src/reasoning-effort.ts`, which imports this
  * module and so cannot be imported back. Exact id, then the `family:` prefix, then a
@@ -487,19 +492,36 @@ export function planReasoningEffortDowngrade(args: {
  * kept so the gate can be checked against real data and widened without another format change.
  * Non-reasoning models carry no ladder and are dropped.
  */
-export async function refreshReasoningMetadata(options: { force?: boolean } = {}): Promise<{
+type RefreshOutcome = {
   ok: boolean;
   reason: string;
   providers?: number;
   models?: number;
-}> {
+};
+
+/** Bound a caller's wait without cancelling the shared refresh job. */
+function waitForRefresh(work: Promise<RefreshOutcome>, waitMs: number | undefined): Promise<RefreshOutcome> {
+  if (waitMs === undefined) return work;
+  if (!Number.isSafeInteger(waitMs) || waitMs < 0) {
+    return Promise.resolve({ ok: false, reason: "invalid wait budget" });
+  }
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  const deadline = new Promise<RefreshOutcome>(resolve => {
+    timer = setTimeout(() => resolve({ ok: false, reason: "wait budget exceeded" }), waitMs);
+    timer.unref?.();
+  });
+  return Promise.race([work, deadline]).finally(() => {
+    if (timer) clearTimeout(timer);
+  });
+}
+
+export async function refreshReasoningMetadata(options: { force?: boolean; waitMs?: number } = {}): Promise<RefreshOutcome> {
   const snapshot = loadSnapshot();
   if (!options.force && snapshot && Date.now() - snapshot.fetchedAt <= CACHE_TTL_MS) {
     return { ok: true, reason: "fresh" };
   }
   if (refreshInFlight) {
-    await refreshInFlight;
-    return { ok: true, reason: "coalesced" };
+    return waitForRefresh(refreshInFlight.then(() => ({ ok: true, reason: "coalesced" })), options.waitMs);
   }
   const job = (async () => {
     const response = await fetch(SOURCE_URL, {
@@ -548,21 +570,22 @@ export async function refreshReasoningMetadata(options: { force?: boolean } = {}
     return { ok: true, reason: "refreshed", providers: Object.keys(providers).length, models };
   })();
   refreshInFlight = job.catch(() => undefined).finally(() => { refreshInFlight = null; });
-  try {
-    return await job;
-  } catch (error) {
-    return { ok: false, reason: error instanceof Error ? error.message : String(error) };
-  }
+  const settled = job.catch((error): RefreshOutcome => ({
+    ok: false,
+    reason: error instanceof Error ? error.message : String(error),
+  }));
+  return waitForRefresh(settled, options.waitMs);
 }
 
 /**
- * Kick a background refresh when the snapshot is missing or stale. Called from the ladder read
- * path so both the long-lived proxy and short-lived ocx sync self-heal without a new CLI
- * surface. One refresh per process at a time; failures are ignored on purpose.
+ * Optional background refresh for callers that do not wait for a snapshot. Catalog sync uses
+ * refreshReasoningMetadata with a bounded wait; an expired ladder read can request this refresh.
+ * A classified toggle/budget model can have a fallback ladder without any snapshot, so this
+ * path must never bootstrap a missing snapshot. One refresh per process; failures are ignored.
  */
 export function ensureReasoningMetadataSnapshot(): void {
   const snapshot = loadSnapshot();
-  if (snapshot && Date.now() - snapshot.fetchedAt <= CACHE_TTL_MS) return;
+  if (!snapshot || Date.now() - snapshot.fetchedAt <= CACHE_TTL_MS) return;
   if (refreshInFlight) return;
   void refreshReasoningMetadata().catch(() => undefined);
 }

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

@@ -14,6 +14,7 @@ import { cursorFastCapableBases } from "../../adapters/cursor/catalog";
 import { COMMAND_CODE_MODEL_REASONING_EFFORTS } from "../command-code-efforts";
 import { isCanonicalOpenRouterTarget } from "../openrouter-routing";
 import type { ProviderRegistryEntry } from "./types";
+import { ANTHROPIC_FAST_MODE_BETA } from "../anthropic-fast";
 import {
   ANTHROPIC_MODELS,
   ANTHROPIC_MODEL_CONTEXT_WINDOWS,
@@ -52,6 +53,7 @@ import {
   DEEPSEEK_NATIVE_THINKING_MODELS,
   DEEPSEEK_GATEWAY_THINKING_MODELS,
   DEEPSEEK_VISION_PREVIEW_MODEL,
+  COMMAND_CODE_MIMO_CONTEXT_WINDOWS,
   COMMAND_CODE_MODEL_INPUT_MODALITIES,
   deepseekThinkingEffortsFor,
   deepseekReasoningMapFor,
@@ -85,6 +87,27 @@ import {
   CLINE_PASS_MODEL_INPUT_MODALITIES,
 } from "./model-seeds";
 
+/**
+ * Claude fast mode (`speed: "fast"` + beta), shared by the OAuth and API-key Anthropic entries.
+ * Only the models Anthropic documents for the lane are classified; Opus 4.6 silently runs
+ * standard and Opus 4.7, Sonnet, Haiku and Fable reject `speed`, so they and future ids stay
+ * unclassified. Source: https://platform.claude.com/docs/en/build-with-claude/fast-mode
+ * (2026-09-23) and the live probe in devlog/_plan/260923_anthropic_fast_speed.
+ */
+const ANTHROPIC_FAST_WIRE = Object.freeze({
+  kind: "anthropic-speed" as const,
+  canonicalToWire: Object.freeze({ priority: "fast" }),
+  foreignCallerTiers: "drop" as const,
+  betas: Object.freeze([ANTHROPIC_FAST_MODE_BETA]),
+});
+const ANTHROPIC_FAST_MODELS: Readonly<Record<string, boolean>> = Object.freeze({
+  "claude-opus-5-5": true,
+  "claude-opus-5": true,
+  "claude-opus-4-8": true,
+});
+const ANTHROPIC_FAST_TIER_DESCRIPTION =
+  "Claude fast mode: faster output at 2x price; needs usage credits (subscription) or fast-mode access (API)";
+
 export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
   {
     id: "openai",
@@ -167,7 +190,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // (it is the current catalog, so its default ordering wins), then the ids
     // only the old devin entry carried. Degraded-mode seed only either way —
     // `liveModels` discovers the account's real roster.
-    models: ["swe-2", "swe-1-7", "gpt-5-6-sol", "gpt-6-astra", "gpt-6-sol", "gpt-6-luna", "claude-opus-5-5", "claude-opus-5", "claude-fable-5-1", "claude-sonnet-5", "glm-5-3", "kimi-k3", "gemini-3-8-flash", "grok-4-6", "swe-1-7-lightning", "gpt-5-6-luna", "gpt-5-6-terra", "claude-opus-4-8", "glm-5-2", "kimi-k2-7", "grok-4-5"],
+    models: ["swe-2", "swe-1-7", "gpt-5-6-sol", "gpt-6-astra", "gpt-6-sol", "gpt-6-luna", "claude-opus-5-5", "claude-opus-5", "claude-fable-5-1", "claude-sonnet-5", "glm-5-3", "kimi-k3", "gemini-3-8-flash", "grok-4-6", "grok-4-7", "swe-1-7-lightning", "gpt-5-6-luna", "gpt-5-6-terra", "claude-opus-4-8", "glm-5-2", "kimi-k2-7", "grok-4-5"],
     liveModels: true,
     defaultModel: "swe-2",
     modelContextWindows: DEVIN_MODEL_CONTEXT_WINDOWS,
@@ -198,7 +221,10 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // the OAuth lane. grok-4.20-multi-agent-0309 is deliberately absent: the gateway accepts
     // the field but answers service_tier "default" — a live downgrade, not a fast tier.
     // Unlisted and future-discovered ids stay unclassified.
+    // grok-4.7 applied and confirmed priority on OAuth Responses in the 2026-09-23
+    // live probe: devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
     modelSupportsServiceTier: {
+      "grok-4.7": true,
       "grok-4.6": true,
       "grok-4.5": true,
       "grok-4.3": true,
@@ -253,14 +279,19 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // than the seeded ones do.
     supportsVerbosity: false,
     defaultModel: "grok-4.5",
-    // Grok 4.6/4.5 subscription Responses callers use the native wire with the existing
+    // Grok 4.7/4.6/4.5 subscription Responses callers use the native wire with the existing
     // namespace/web-search/replay normalization. Chat remains an explicit modelAdapters
     // opt-in. Multi-agent has no Chat wire and uses Responses under both auth modes.
-    // grok-4.6/4.5 are classified OAuth fast-tier models (modelSupportsServiceTier above),
+    // grok-4.7/4.6/4.5 are classified OAuth fast-tier models (modelSupportsServiceTier above),
     // so a caller-sent service_tier:"priority" forwards on this lane — the Codex fast-toggle
     // path. Multi-agent keeps its pin: probed 2026-09-13, the gateway downgrades its tier to
     // "default", so forwarding a caller tier would advertise a tier it does not get.
     modelWireDefaults: {
+      "grok-4.7": {
+        wire: "openai-responses",
+        inbound: ["responses"],
+        authModes: ["oauth"],
+      },
       "grok-4.6": {
         wire: "openai-responses",
         inbound: ["responses"],
@@ -303,6 +334,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // the app blocks attachments client-side. grok-build-0.1 / grok-composer-2.5-fast stay out
     // (they are already listed in noVisionModels below).
     modelInputModalities: {
+      "grok-4.7": ["text", "image"],
       "grok-4.6": ["text", "image"],
       "grok-4.5": ["text", "image"],
       "grok-4.3": ["text", "image"],
@@ -315,18 +347,24 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // reasoning_content as the top cause of prompt-cache misses on multi-turn conversations
     // (docs.x.ai prompt-caching/multi-turn, verified 2026-07-13 — devlog/_plan/260713_grok_caching).
     // Models that never emit reasoning simply have no thinking parts to replay (no-op).
-    preserveReasoningContentModels: ["grok-4.6", "grok-4.5", "grok-4.3", "grok-4.20-0309-reasoning"],
+    preserveReasoningContentModels: ["grok-4.7", "grok-4.6", "grok-4.5", "grok-4.3", "grok-4.20-0309-reasoning"],
     // grok-4.5 reasoning is always-on with low/medium/high (no off tier, no xhigh).
     // grok-4.6 adds xhigh per docs.x.ai/developers/model-capabilities/text/reasoning;
     // multi-agent accepts the same four wire values to select 4 or 16 collaborators. xAI
     // documents high as the 4.6 default but no multi-agent default, so do not invent one.
     modelReasoningEfforts: {
+      // 2026-09-23 live probe accepted low..xhigh and rejected max on both wires;
+      // devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+      "grok-4.7": ["low", "medium", "high", "xhigh"],
       "grok-4.6": ["low", "medium", "high", "xhigh"],
       "grok-4.5": ["low", "medium", "high"],
       "grok-4.20-multi-agent-0309": ["low", "medium", "high", "xhigh"],
     },
-    modelDefaultReasoningEfforts: { "grok-4.6": "high" },
+    modelDefaultReasoningEfforts: { "grok-4.7": "high", "grok-4.6": "high" },
     modelContextWindows: {
+      // 500k confirmed by context_length_exceeded:
+      // devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+      "grok-4.7": 500_000,
       "grok-4.6": 500_000,
       "grok-4.5": 500_000,
       "grok-4.3": 1_000_000,
@@ -363,6 +401,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // merge into deepseek-v4-flash later.
     modelContextWindows: {
       [`deepseek/${DEEPSEEK_VISION_PREVIEW_MODEL}`]: 1_048_576,
+      ...COMMAND_CODE_MIMO_CONTEXT_WINDOWS,
     },
     modelInputModalities: COMMAND_CODE_MODEL_INPUT_MODALITIES,
     defaultMaxOutputTokens: 64_000,
@@ -404,6 +443,12 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // falls back to 8192, which truncates long answers with stop_reason=max_tokens.
     defaultMaxOutputTokens: ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS,
     defaultModel: "claude-sonnet-5",
+    // Claude fast mode on the subscription lane (Claude Code `/fast`): the OAuth route accepts
+    // `speed` and gates it on account entitlement (usage credits / org enablement), probed live
+    // 2026-09-23 (devlog/_plan/260923_anthropic_fast_speed/020_probe-evidence.md).
+    fastWire: ANTHROPIC_FAST_WIRE,
+    modelSupportsServiceTier: { ...ANTHROPIC_FAST_MODELS },
+    fastTierDescription: ANTHROPIC_FAST_TIER_DESCRIPTION,
   },
   {
     id: "anthropic-apikey",
@@ -423,6 +468,9 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     modelReasoningEfforts: { ...ANTHROPIC_MODEL_REASONING_EFFORTS },
     defaultMaxOutputTokens: ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS,
     defaultModel: "claude-sonnet-5",
+    fastWire: ANTHROPIC_FAST_WIRE,
+    modelSupportsServiceTier: { ...ANTHROPIC_FAST_MODELS },
+    fastTierDescription: ANTHROPIC_FAST_TIER_DESCRIPTION,
   },
   {
     id: "kimi",
@@ -466,6 +514,41 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     autoToolChoiceOnlyModels: KIMI_AUTO_TOOL_CHOICE_ONLY_MODELS,
     preserveReasoningContentModels: KIMI_THINKING_MODELS,
   },
+  {
+    id: "kimi-responses",
+    label: "Kimi (Responses)",
+    adapter: "openai-responses",
+    baseUrl: "https://api.kimi.com/coding/v1",
+    authKind: "oauth",
+    modelSuffixBracketStrip: true,
+    // Same wire-capability defaults as the Chat preset. promptCacheKey is copied here
+    // deliberately even though only the Chat adapter reads it today: the field is a
+    // stable session/task key Kimi documents for cache affinity, and the Responses
+    // endpoint already accepts it (live probe 260921: prompt_cache_key round-trips 200).
+    promptCacheKey: true,
+    // Kimi's Responses endpoint rejects hook-provided context between a tool call and
+    // its matching result (#4726); the flag is live on this wire.
+    requiresAdjacentResponsesToolResults: true,
+    featured: false,
+    // Shares the kimi OAuth account: the login flow and credential store are keyed by
+    // oauthId, so adding this preset after logging into kimi needs no second login.
+    oauthId: "kimi",
+    jawcodeBundle: "moonshot",
+    note: "Same Kimi account login, routed over the OpenAI Responses wire. Thinking content stays encrypted server-side; tool calls and results stay visible. Chat wire remains the default preset for transparency.",
+    models: KIMI_CODING_LIVE_MODELS,
+    defaultModel: "kimi-for-coding",
+    modelContextWindows: KIMI_CODING_MODEL_CONTEXT_WINDOWS,
+    modelInputModalities: KIMI_CODING_MODEL_INPUT_MODALITIES,
+    noReasoningModels: KIMI_CODING_NO_REASONING_MODELS,
+    modelReasoningEfforts: KIMI_CODING_REASONING_EFFORTS,
+    modelDefaultReasoningEfforts: KIMI_CODING_DEFAULT_REASONING_EFFORTS,
+    modelReasoningEffortMap: KIMI_CODING_REASONING_EFFORT_MAPS,
+    noTemperatureModels: KIMI_LOCKED_PARAMETER_MODELS,
+    noTopPModels: KIMI_LOCKED_PARAMETER_MODELS,
+    noPenaltyModels: KIMI_LOCKED_PARAMETER_MODELS,
+    autoToolChoiceOnlyModels: KIMI_AUTO_TOOL_CHOICE_ONLY_MODELS,
+    preserveReasoningContentModels: KIMI_THINKING_MODELS,
+  },
   {
     id: "kiro",
     label: "Kiro (AWS CodeWhisperer)",
@@ -680,7 +763,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     // Use explicit replay history and the existing stateless Responses policy.
     statelessResponses: true,
     /* [Decision Log]
-    - 목적과 의도: Route the exact models OpenCode Go documents on the Responses endpoint — GPT 5.6 Luna, Grok 4.6, and Muse Spark Contributor (#2617).
+    - 목적과 의도: Route the exact models OpenCode Go documents on the Responses endpoint — GPT 5.6 Luna, Grok 4.6/4.7, and Muse Spark Contributor (#2617; opencode.ai/docs/go).
     - 기존 구현 및 제약 조건: The provider is mixed-wire but its provider-wide `openai-chat` adapter sent Luna to `/chat/completions`; explicit user `modelAdapters` entries must remain authoritative.
     - 검토한 주요 대안: Change the whole provider to Responses; infer the wire from model-family names; add one registry-only exact-model default.
     - 선택한 방식: Declare only the named models as `openai-responses` through the existing registry default mechanism; the map stays an exact-model allowlist rather than a family or provider-wide rule.
@@ -690,6 +773,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     modelWireDefaults: {
       "gpt-5.6-luna": "openai-responses",
       "grok-4.6": "openai-responses",
+      "grok-4.7": "openai-responses",
       "muse-spark-1.3-contributor": "openai-responses",
       "muse-spark-1.2-contributor": "openai-responses",
     },
@@ -750,6 +834,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
     modelReasoningEfforts: {
       "gpt-5.6-luna": OPENAI_API_GPT56_REASONING_EFFORTS,
       "grok-4.6": ["low", "medium", "high", "xhigh"],
+      "grok-4.7": ["low", "medium", "high", "xhigh"],
       "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
       "glm-5.3-flash": ZAI_GLM_53_REASONING_EFFORTS,
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
@@ -761,7 +846,7 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
       ...Object.fromEntries(OPENCODE_GO_THINKING_BUDGET_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
       ...Object.fromEntries(DEEPSEEK_GATEWAY_THINKING_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)])),
     },
-    modelDefaultReasoningEfforts: { "grok-4.6": "high", "kimi-k3": "max" },
+    modelDefaultReasoningEfforts: { "grok-4.6": "high", "grok-4.7": "high", "kimi-k3": "max" },
     // glm-5.2 uses identity labels now that `max` is a native Codex level (no alias map);
     // the thinking-toggle map is a REAL wire alias (effort -> enabled/disabled) and stays.
     modelReasoningEffortMap: {
@@ -797,6 +882,9 @@ export const PROVIDER_REGISTRY_CORE: readonly ProviderRegistryEntry[] = [
       // deepseek-v4-flash stays listed — that route rejects image_url upstream.
       "deepseek-v4-flash",
       "mimo-v2-pro", "mimo-v2.5-pro",
+      // V2.6 is multimodal first-party, but image forwarding on this gateway is unprobed:
+      // the sidecar describes images until a route probe proves native input.
+      "mimo-v2.6-pro", "mimo-v2.6-flash",
       "minimax-m2.5", "minimax-m2.7",
       "qwen3.7-max",
     ],

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

@@ -47,6 +47,7 @@ import {
   DEEPSEEK_V4_LEGACY_MODELS,
   DEEPSEEK_GATEWAY_THINKING_MODELS,
   DEEPSEEK_VISION_PREVIEW_MODEL,
+  COMMAND_CODE_MIMO_CONTEXT_WINDOWS,
   COMMAND_CODE_MODEL_INPUT_MODALITIES,
   OPENCODE_FREE_DEEPSEEK_MODELS,
   OPENCODE_ZEN_TEXT_ONLY_MODELS,
@@ -170,6 +171,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     // (merges into v4-flash later).
     modelContextWindows: {
       [`deepseek/${DEEPSEEK_VISION_PREVIEW_MODEL}`]: 1_048_576,
+      ...COMMAND_CODE_MIMO_CONTEXT_WINDOWS,
     },
     modelInputModalities: COMMAND_CODE_MODEL_INPUT_MODALITIES,
     modelDiscovery: {
@@ -1139,7 +1141,11 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     // narrower table that silently falls behind whenever the keyed one is updated.
     noJsonSchemaModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_MODELS],
   },
-  { id: "xiaomi", label: "Xiaomi MiMo", baseUrl: "https://api.xiaomimimo.com/anthropic", adapter: "anthropic", authKind: "key", dashboardUrl: "https://xiaomimimo.com", defaultModel: "mimo-v2.5-pro" },
+  // Xiaomi retires mimo-v2.5 and mimo-v2.5-pro on 2026-10-21 with no redirect
+  // (https://mimo.mi.com/docs/en-US/updates/deprecate), so the first-party presets default to V2.6.
+  // Saved defaults are not rewritten; V2.5 stays listed until it stops answering.
+  // Both first-party presets read the xiaomi metadata bundle for window, output, modalities and price.
+  { id: "xiaomi", label: "Xiaomi MiMo", baseUrl: "https://api.xiaomimimo.com/anthropic", adapter: "anthropic", authKind: "key", dashboardUrl: "https://xiaomimimo.com", defaultModel: "mimo-v2.6-pro", models: ["mimo-v2.6-pro", "mimo-v2.6-flash", "mimo-v2.6-pro-ultraspeed", "mimo-v2.5-pro", "mimo-v2.5"], jawcodeBundle: "xiaomi" },
   // Xiaomi's public OpenAI-compatible endpoint is a distinct transport from both the Anthropic
   // preset above and the paid token-plan host below. Keep a separate fixed-destination contract
   // so existing custom providers are never retargeted while the official route receives the
@@ -1151,8 +1157,9 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     adapter: "openai-chat",
     authKind: "key",
     dashboardUrl: "https://platform.xiaomimimo.com/console/balance",
-    defaultModel: "mimo-v2.5",
-    models: ["mimo-v2.5"],
+    defaultModel: "mimo-v2.6-flash",
+    models: ["mimo-v2.6-flash", "mimo-v2.6-pro", "mimo-v2.6-pro-ultraspeed", "mimo-v2.5"],
+    jawcodeBundle: "xiaomi",
     reasoningEfforts: ["low", "medium", "high"],
     reasoningEffortMap: { xhigh: "high", max: "high", ultra: "high" },
     preserveCustomDestination: true,
@@ -1192,8 +1199,11 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     adapter: "openai-chat",
     authKind: "key",
     dashboardUrl: "https://xiaomimimo.com",
-    defaultModel: "mimo-v2.5-pro",
-    models: ["mimo-v2.5-pro", "mimo-v2.5"],
+    // Token-plan roster per Xiaomi's token-plan model list (V2.6 Pro and Flash). No jawcodeBundle,
+    // so no plan-specific facts are claimed; usage estimates still come from the model-level vendor
+    // price fallback (the pay-as-you-go equivalent), exactly as they did for V2.5.
+    defaultModel: "mimo-v2.6-pro",
+    models: ["mimo-v2.6-pro", "mimo-v2.6-flash", "mimo-v2.5-pro", "mimo-v2.5"],
     // The gateway validates the ladder strictly and rejects anything above `high`.
     reasoningEfforts: ["low", "medium", "high"],
     reasoningEffortMap: { xhigh: "high", max: "high", ultra: "high" },
@@ -1324,9 +1334,10 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     // private console endpoint — the approach closed in #687 and left in draft in #2244.
     // baseUrl is the canonical region identity: the adapter fails closed if it is overridden, so a
     // global key is never sent to the CN environment (that is the separate `codebuddy-cn` entry).
-    // v1 runs tools-disabled so Codex keeps tool ownership; this provider is text/reasoning only
-    // until the control-protocol tool bridge lands (see docs). Free/trial/promotional/subscription
-    // credits draw from the same official API-key pool. Requires the CLI: `npm i -g @tencent-ai/codebuddy-code`.
+    // The CLI always runs tools-disabled; a capture-only MCP bridge advertises the request's
+    // Codex tool catalog, so approval, sandboxing, and execution stay with the client.
+    // Free/trial/promotional/subscription credits draw from the same official API-key pool.
+    // Requires the CLI: `npm i -g @tencent-ai/codebuddy-code`.
     // GOVERNANCE: whether routing this vendor automation surface behind a proxy for a third-party
     // agent satisfies CodeBuddy's AUP is an open question flagged for maintainer security review.
     id: "codebuddy",
@@ -1346,7 +1357,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     reasoningEfforts: CODEBUDDY_REASONING_EFFORTS,
     modelReasoningEfforts: CODEBUDDY_GLOBAL_MODEL_REASONING_EFFORTS,
     modelDefaultReasoningEfforts: CODEBUDDY_GLOBAL_MODEL_DEFAULT_REASONING_EFFORTS,
-    note: "Official CodeBuddy Code CLI (Tencent Cloud), global/public environment. Uses the documented CODEBUDDY_API_KEY + headless CLI surface; never reads desktop sessions or private console endpoints. Region-isolated from codebuddy-cn. v1 disables CLI tools (--tools \"\") so Codex retains tool ownership: text/reasoning only for now. Requires `npm i -g @tencent-ai/codebuddy-code`. AUP/routing authorization flagged for maintainer security review.",
+    note: "Official CodeBuddy Code CLI (Tencent Cloud), global/public environment. Uses the documented CODEBUDDY_API_KEY + headless CLI surface; never reads desktop sessions or private console endpoints. Region-isolated from codebuddy-cn. The CLI always runs tools-disabled (--tools \"\"); a capture-only MCP bridge surfaces the request's Codex tool catalog as capturable calls, with approval and execution kept by the client. Requires `npm i -g @tencent-ai/codebuddy-code`. AUP/routing authorization flagged for maintainer security review.",
   },
   {
     // Official CodeBuddy Code CLI provider, CHINA / `internal` environment. Identical adapter and
@@ -1371,7 +1382,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     modelReasoningEfforts: CODEBUDDY_CN_MODEL_REASONING_EFFORTS,
     modelDefaultReasoningEfforts: CODEBUDDY_CN_MODEL_DEFAULT_REASONING_EFFORTS,
     noVisionModels: CODEBUDDY_CN_NO_VISION_MODELS,
-    note: "Official CodeBuddy Code CLI (Tencent Cloud), China/internal environment. Uses the documented CODEBUDDY_API_KEY + headless CLI surface; never reads desktop sessions or private console endpoints. Region-isolated from codebuddy (Global); credentials are never exchanged across regions. v1 disables CLI tools (--tools \"\"): text/reasoning only for now. Requires `npm i -g @tencent-ai/codebuddy-code`. AUP/routing authorization flagged for maintainer security review.",
+    note: "Official CodeBuddy Code CLI (Tencent Cloud), China/internal environment. Uses the documented CODEBUDDY_API_KEY + headless CLI surface; never reads desktop sessions or private console endpoints. Region-isolated from codebuddy (Global); credentials are never exchanged across regions. The CLI always runs tools-disabled (--tools \"\"); a capture-only MCP bridge surfaces the request's Codex tool catalog as capturable calls, with approval and execution kept by the client. Requires `npm i -g @tencent-ai/codebuddy-code`. AUP/routing authorization flagged for maintainer security review.",
   },
   {
     id: "stepfun",

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

@@ -118,6 +118,7 @@ export const REGISTRY_FIELD_MODEL_ID_ROLES = {
   requiresReasoningPlaceholderModels: NONE,
   showThinkingSummary: NONE,
   reasoningSplitModels: NONE,
+  inlineThinkTagModels: NONE,
   reasoningDetailsModels: NONE,
   thinkingToggleModels: NONE,
   thinkingBudgetModels: NONE,

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

@@ -231,6 +231,7 @@ export const OPENAI_DAYBREAK_REASONING_EFFORTS: Record<string, string[]> = Objec
 );
 export const OPENROUTER_GPT56_MODELS = OPENAI_GPT56_MODELS.map(id => `openai/${id}`);
 export const XAI_MODELS = [
+  "grok-4.7",
   "grok-4.6",
   "grok-4.5",
   "grok-4.3",
@@ -270,6 +271,8 @@ export const THINKING_TOGGLE_MAP: Record<string, string> = {
 };
 export const OPENCODE_GO_THINKING_TOGGLE_MODELS = [
   "mimo-v2.5", "mimo-v2.5-pro", "glm-5", "glm-5.1",
+  // V2.6 keeps the vendor's thinking toggle; listed ahead of a Go probe (preemptive, 2026-09-23).
+  "mimo-v2.6-pro", "mimo-v2.6-flash",
 ];
 /**
  * Zhipu's domestic BigModel platform. Text families first, then the vision member: modalities are
@@ -329,7 +332,7 @@ export const DEEPSEEK_VISION_PREVIEW_MODEL = "deepseek-v4-flash-vision-exp";
  * CommandCode routes verified to accept image input end-to-end (#2406).
  *
  * Verified-negative and therefore deliberately ABSENT: deepseek/deepseek-v4-flash,
- * zai-org/GLM-5.2, zai-org/GLM-5.3, xai/grok-4.6. Those
+ * zai-org/GLM-5.2, zai-org/GLM-5.3. Those
  * routes accept the request and drop the image, which is worse than declining it — the
  * model answers about an image it never saw. Do not add an id here on family resemblance;
  * capability intersection trusts this map.
@@ -351,10 +354,16 @@ export const COMMAND_CODE_IMAGE_MODELS = [
   "meta/muse-spark-1.3-contributor",
   "meta/muse-spark-1.2",
   "meta/muse-spark-1.2-contributor",
+  // Live 2026-09-23 3x3 random-color grid (180x180) via ocx 2.62.0:
+  // 4.7 read 9/9 in user messages and tool results; 4.6 read 9/9 and 8/9.
+  // Neither route requested a vision sidecar. Evidence:
+  // devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
+  "xai/grok-4.6",
+  "xai/grok-4.7",
   // Native Z.AI VLM (docs.z.ai/guides/vlm/glm-5.3-flash). This exact id is already
   // classified as natively vision-capable in NVIDIA_NIM_VISION_MODELS in this file;
   // it is not one of the verified-negative ids the header names (those are
-  // deepseek/deepseek-v4-flash, zai-org/GLM-5.2, zai-org/GLM-5.3, xai/grok-4.6 —
+  // deepseek/deepseek-v4-flash, zai-org/GLM-5.2, zai-org/GLM-5.3 —
   // different ids). Adding it on the shared GLM-5.3 prefix would be the family-
   // resemblance mistake the header forbids; the VLM docs are the evidence (#4505).
   "z-ai/glm-5.3-flash",
@@ -374,6 +383,19 @@ export const COMMAND_CODE_IMAGE_MODELS = [
  * the user-message and tool-result paths (see the note at that entry). The
  * mechanism stays for the next route that measures text-only.
  */
+/**
+ * Command Code MiMo context windows from the live /provider/v1/models catalog (2026-09-23 fixture,
+ * tests/fixtures/commandcode-models.json). Model-keyed registry facts double as the router's native
+ * decode ids, so a cold start or failed discovery still turns `command-code/xiaomi-mimo-v2.6-pro`
+ * into `xiaomi/mimo-v2.6-pro` instead of sending the flattened slug upstream. They are not a roster.
+ */
+export const COMMAND_CODE_MIMO_CONTEXT_WINDOWS: Record<string, number> = {
+  "xiaomi/mimo-v2.6-pro": 1_048_576,
+  "xiaomi/mimo-v2.6-pro-ultraspeed": 1_048_576,
+  "xiaomi/mimo-v2.6-flash": 1_048_576,
+  "xiaomi/mimo-v2.5-pro": 1_000_000,
+  "xiaomi/mimo-v2.5": 1_000_000,
+};
 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"]])),
@@ -649,11 +671,13 @@ export const ALIBABA_TOKEN_PLAN_PRESERVE_REASONING = [
 // 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.
 // The separately billed Moonshot API uses `kimi-k3`.
+// 260921: `k3-256k` is the same K3 served under the explicit ceiling id (verified live
+// 260921: same 988-token scaffold and identity answer as bare `k3` on the same input).
 // Evidence: https://www.kimi.com/code/docs/en/kimi-code/models.html
 //           https://www.kimi.com/code/docs/en/kimi-code/error-reference.html
 export const KIMI_K3_STANDARD_CONTEXT_WINDOW = 262_144;
 export const KIMI_K3_1M_CONTEXT_WINDOW = 1_048_576;
-export const KIMI_CODING_K3_MODELS = ["k3", "k3[1m]"];
+export const KIMI_CODING_K3_MODELS = ["k3", "k3[1m]", "k3-256k"];
 // 260921 Kimi K2.8: `kimi-for-coding` is the stable subscription alias Moonshot re-points
 // at each coding release. Live GET /coding/v1/models lists only kimi-for-coding[-highspeed],
 // k3, k3-256k — the k2.x ids are retired from the subscription endpoint. Since K2.8 Preview
@@ -940,6 +964,8 @@ export const CLINE_PASS_MODELS = [
   "cline-pass/kimi-k2.7-code",
   "cline-pass/kimi-k2.6",
   "cline-pass/deepseek-v4-flash",
+  "cline-pass/mimo-v2.6-pro",
+  "cline-pass/mimo-v2.6-flash",
   "cline-pass/mimo-v2.5",
   "cline-pass/mimo-v2.5-pro",
   "cline-pass/minimax-m3",
@@ -988,6 +1014,8 @@ export const CLINE_PASS_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
   "cline-pass/kimi-k2.7-code": 262_144,
   "cline-pass/kimi-k2.6": 262_144,
   "cline-pass/deepseek-v4-flash": 1_048_576,
+  "cline-pass/mimo-v2.6-pro": 1_048_576,
+  "cline-pass/mimo-v2.6-flash": 1_048_576,
   "cline-pass/mimo-v2.5": 1_050_000,
   "cline-pass/mimo-v2.5-pro": 1_050_000,
   "cline-pass/minimax-m3": 1_048_576,

package/src/providers/registry/types.ts

@@ -335,6 +335,8 @@ export interface ProviderRegistryEntry {
    */
   showThinkingSummary?: boolean;
   reasoningSplitModels?: string[];
+  /** See OcxProviderConfig.inlineThinkTagModels. */
+  inlineThinkTagModels?: string[];
   reasoningDetailsModels?: string[];
   thinkingToggleModels?: string[];
   thinkingBudgetModels?: string[];
@@ -358,6 +360,6 @@ export type ProviderConfigSeed = Pick<
   | "modelMaxInputTokens" | "defaultMaxOutputTokens" | "modelMaxOutputTokens"
   | "reasoningEfforts" | "modelReasoningEfforts" | "modelDefaultReasoningEfforts" | "reasoningEffortMap" | "modelReasoningEffortMap" | "reasoningWireFormat"
   | "noVisionModels" | "noReasoningModels" | "noTemperatureModels" | "noTopPModels" | "noPenaltyModels"
-  | "autoToolChoiceOnlyModels" | "preserveReasoningContentModels" | "requiresReasoningPlaceholderModels" | "reasoningSplitModels" | "reasoningDetailsModels" | "thinkingToggleModels" | "thinkingBudgetModels" | "escapeBuiltinToolNames" | "openaiChatEofTolerance" | "showThinkingSummary"
+  | "autoToolChoiceOnlyModels" | "preserveReasoningContentModels" | "requiresReasoningPlaceholderModels" | "reasoningSplitModels" | "inlineThinkTagModels" | "reasoningDetailsModels" | "thinkingToggleModels" | "thinkingBudgetModels" | "escapeBuiltinToolNames" | "openaiChatEofTolerance" | "showThinkingSummary"
   | "googleMode" | "project" | "location" | "headers"
 >;

package/src/providers/resolved-model-policy-merge.ts

@@ -32,7 +32,15 @@ export function mapFill<T>(
   operator: Readonly<Record<string, T>> | undefined,
 ): [Record<string, T> | undefined, StaticPolicySource] {
   if (!registry && !operator) return [undefined, "unknown"];
-  return [detachedClone({ ...(registry ?? {}), ...(operator ?? {}) }), operator ? "operator" : "registry"];
+  // Per-model lookups fold case (legacyModelValue), so a case-varied operator key must claim the
+  // registry row here; leaving both keys lets the earlier registry entry shadow the override.
+  const claimed = new Set(Object.keys(operator ?? {}).map(key => key.toLowerCase()));
+  const merged: Record<string, T> = {};
+  for (const [key, value] of Object.entries(registry ?? {})) {
+    if (!claimed.has(key.toLowerCase())) merged[key] = value;
+  }
+  for (const [key, value] of Object.entries(operator ?? {})) merged[key] = value;
+  return [detachedClone(merged), operator ? "operator" : "registry"];
 }
 
 export function nestedMapFill(
@@ -40,10 +48,19 @@ export function nestedMapFill(
   operator: Readonly<Record<string, Record<string, string>>> | undefined,
 ): [Record<string, Record<string, string>> | undefined, StaticPolicySource] {
   if (!registry && !operator) return [undefined, "unknown"];
+  // The outer model key folds case like mapFill: a case-varied operator key claims the registry
+  // row instead of shadowing behind it, while the claimed row's inner entries still fill
+  // underneath the operator's inner map.
+  const claimed = new Set(Object.keys(operator ?? {}).map(key => key.toLowerCase()));
   const merged: Record<string, Record<string, string>> = {};
-  for (const [key, value] of Object.entries(registry ?? {})) merged[key] = { ...value };
+  const claimedInner: Record<string, Record<string, string>> = {};
+  for (const [key, value] of Object.entries(registry ?? {})) {
+    const folded = key.toLowerCase();
+    if (claimed.has(folded)) claimedInner[folded] = { ...(claimedInner[folded] ?? {}), ...value };
+    else merged[key] = { ...value };
+  }
   for (const [key, value] of Object.entries(operator ?? {})) {
-    merged[key] = { ...(merged[key] ?? {}), ...value };
+    merged[key] = { ...(claimedInner[key.toLowerCase()] ?? merged[key] ?? {}), ...value };
   }
   return [merged, operator ? "operator" : "registry"];
 }
@@ -52,12 +69,19 @@ export function positiveCapMap(
   registry: Readonly<Record<string, number>> | undefined,
   operator: Readonly<Record<string, number>> | undefined,
 ): [Record<string, number> | undefined, StaticPolicySource] {
-  if (!registry && !operator) return [undefined, "unknown"];
-  const merged = { ...(registry ?? {}) };
+  const [merged, source] = mapFill(registry, operator);
+  if (!merged) return [undefined, source];
+  // The operator owns a case-equal row, as in mapFill, but cannot widen its registry cap.
+  const registryCaps = new Map<string, number>();
+  for (const [key, value] of Object.entries(registry ?? {})) {
+    const folded = key.toLowerCase();
+    registryCaps.set(folded, Math.min(registryCaps.get(folded) ?? value, value));
+  }
   for (const [key, value] of Object.entries(operator ?? {})) {
-    merged[key] = typeof merged[key] === "number" ? Math.min(merged[key]!, value) : value;
+    const cap = registryCaps.get(key.toLowerCase());
+    merged[key] = cap === undefined ? value : Math.min(cap, value);
   }
-  return [merged, operator ? "operator" : "registry"];
+  return [merged, source];
 }
 
 export function stableUnion(
@@ -118,7 +142,13 @@ export function legacyModelSource<T>(
   registry: Readonly<Record<string, T>> | undefined,
   modelId: string,
 ): StaticPolicySource {
-  const merged = { ...(registry ?? {}), ...(operator ?? {}) };
+  // Match mapFill's case-folded ownership so provenance follows the value lookup.
+  const claimed = new Set(Object.keys(operator ?? {}).map(key => key.toLowerCase()));
+  const merged: Record<string, T> = {};
+  for (const [key, value] of Object.entries(registry ?? {})) {
+    if (!claimed.has(key.toLowerCase())) merged[key] = value;
+  }
+  for (const [key, value] of Object.entries(operator ?? {})) merged[key] = value;
   let winningKey: string | undefined;
   if (Object.hasOwn(merged, modelId)) winningKey = modelId;
   const colon = modelId.indexOf(":");