@bitkyc08/opencodex 2.6.32 → 2.7.1-preview.20260710

package/src/providers/registry.ts

@@ -37,9 +37,12 @@ export interface ProviderRegistryEntry {
   noTemperatureModels?: string[];
   noTopPModels?: string[];
   noPenaltyModels?: string[];
+  /** Opt this provider into parallel tool calls (see OcxProviderConfig.parallelToolCalls). */
+  parallelToolCalls?: boolean;
   autoToolChoiceOnlyModels?: string[];
   preserveReasoningContentModels?: string[];
   thinkingToggleModels?: string[];
+  thinkingBudgetModels?: string[];
   escapeBuiltinToolNames?: boolean;
   oauthId?: string;
   jawcodeBundle?: string;
@@ -56,14 +59,16 @@ export type ProviderConfigSeed = Pick<
   | "liveModels" | "contextWindow" | "modelContextWindows" | "modelInputModalities"
   | "reasoningEfforts" | "modelReasoningEfforts" | "reasoningEffortMap" | "modelReasoningEffortMap"
   | "noVisionModels" | "noReasoningModels" | "noTemperatureModels" | "noTopPModels" | "noPenaltyModels"
-  | "autoToolChoiceOnlyModels" | "preserveReasoningContentModels" | "thinkingToggleModels" | "escapeBuiltinToolNames"
+  | "autoToolChoiceOnlyModels" | "preserveReasoningContentModels" | "thinkingToggleModels" | "thinkingBudgetModels" | "escapeBuiltinToolNames"
   | "googleMode" | "project" | "location"
 >;
 
 // Shared between the OAuth (Claude account) and API-key Anthropic entries so both expose the
 // same static model seed.
-const ANTHROPIC_MODELS = ["claude-sonnet-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6", "claude-sonnet-4-6", "claude-haiku-4-5"];
-const ANTHROPIC_MODEL_CONTEXT_WINDOWS: Record<string, number> = { "claude-sonnet-5": 1_000_000 };
+// 260709 refresh: claude-fable-5 added (official models overview); evidence in
+// devlog/model_update/260709_model_refresh/002_cursor_registry_drift.md.
+const ANTHROPIC_MODELS = ["claude-fable-5", "claude-sonnet-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6", "claude-sonnet-4-6", "claude-haiku-4-5"];
+const ANTHROPIC_MODEL_CONTEXT_WINDOWS: Record<string, number> = { "claude-sonnet-5": 1_000_000, "claude-fable-5": 1_000_000 };
 
 const ZAI_GLM_52_MODELS = ["glm-5.2", "glm-5.2[1m]"];
 const ZAI_GLM_52_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
@@ -83,11 +88,11 @@ const OPENROUTER_GPT56_CONTEXT_WINDOWS = {
 
 /**
  * Vendor thinking-toggle models (MiMo v2.x, GLM 5/5.1 on Zen Go): the wire knob is
- * `thinking: {type: enabled|disabled}` — a binary. Advertise a two-step Codex ladder
- * (low = thinking off, high = thinking on) and map efforts onto the toggle. Zen Go
+ * `thinking: {type: enabled|disabled}` — a binary. Advertise the full Codex picker ladder
+ * and map efforts onto the toggle. Zen Go
  * pass-through probed live 2026-07-07 (glm-5.2 toggle verified; mimo/minimax accept shape).
  */
-const THINKING_TOGGLE_EFFORTS = ["low", "high"];
+const THINKING_TOGGLE_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
 const THINKING_TOGGLE_MAP: Record<string, string> = {
   none: "disabled",
   minimal: "disabled",
@@ -100,8 +105,16 @@ const THINKING_TOGGLE_MAP: Record<string, string> = {
 const OPENCODE_GO_THINKING_TOGGLE_MODELS = [
   "mimo-v2.5", "mimo-v2.5-pro", "mimo-v2-omni", "mimo-v2-pro", "glm-5", "glm-5.1",
 ];
+const THINKING_BUDGET_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
+const THINKING_BUDGET_MODELS = [
+  "qwen3.5-397b", "qwen3.6-35b",
+  "qwen3.5-plus", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus",
+];
+const OPENCODE_GO_THINKING_BUDGET_MODELS = ["qwen3.5-plus", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus"];
 const DEEPSEEK_THINKING_MODELS = ["deepseek-v4-pro", "deepseek-v4-flash"];
-const DEEPSEEK_THINKING_EFFORTS = ["high", "xhigh"];
+// "max" is advertised too: the wire map routes xhigh->max and max->max, so the picker
+// should surface the max tier instead of hiding it behind xhigh.
+const DEEPSEEK_THINKING_EFFORTS = ["high", "xhigh", "max"];
 const DEEPSEEK_THINKING_REASONING_MAP: Record<string, string> = {
   low: "high",
   medium: "high",
@@ -109,7 +122,7 @@ const DEEPSEEK_THINKING_REASONING_MAP: Record<string, string> = {
   xhigh: "max",
   max: "max",
 };
-const KIMI_THINKING_MODELS = ["kimi-k2.7-code", "kimi-k2.7-code-highspeed", "kimi-k2.6", "kimi-k2.5", "kimi-k2-0905-preview"];
+const KIMI_THINKING_MODELS = ["kimi-k2.7-code", "kimi-k2.7-code-highspeed", "kimi-k2.6", "kimi-k2.5"];
 const KIMI_LOCKED_PARAMETER_MODELS = ["kimi-k2.7-code", "kimi-k2.7-code-highspeed", "kimi-k2.6", "kimi-k2.5"];
 const NEURALWATT_REASONING_HISTORY_MODELS = [
   "glm-5.2",
@@ -119,7 +132,6 @@ const NEURALWATT_REASONING_HISTORY_MODELS = [
 const UMANS_MODELS = [
   "umans-coder",
   "umans-kimi-k2.7",
-  "umans-kimi-k2.6",
   "umans-flash",
   "umans-glm-5.2",
   "umans-glm-5.1",
@@ -131,7 +143,6 @@ const UMANS_TEXT_ONLY_MODELS = ["umans-glm-5.2", "umans-glm-5.1"];
 const UMANS_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
   "umans-coder": 262_144,
   "umans-kimi-k2.7": 262_144,
-  "umans-kimi-k2.6": 262_144,
   "umans-flash": 262_144,
   "umans-glm-5.2": 405_504,
   "umans-glm-5.1": 202_752,
@@ -159,7 +170,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     authKind: "oauth",
     featured: false,
     dashboardPreset: true,
-    note: "Experimental Cursor bridge. Live transport and live model discovery are enabled after a standalone PKCE browser login via 'ocx login cursor'; native read/write/delete/shell/fetch execution stays disabled unless provider.unsafeAllowNativeLocalExec is explicitly set for a trusted local experiment.",
+    note: "Experimental Cursor bridge. Live transport and live model discovery are enabled after a standalone PKCE browser login via 'ocx login cursor'; native read/write/delete/shell/fetch execution stays disabled unless you set \"unsafeAllowNativeLocalExec\": true on providers.cursor in ~/.opencodex/config.json (dashboard: Providers → Cursor → Edit JSON) for a trusted local experiment.",
     models: cursorModelIds(CURSOR_STATIC_MODELS),
     liveModels: true,
     defaultModel: "auto",
@@ -182,9 +193,29 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     oauthId: "xai",
     jawcodeBundle: "xai",
     note: "Log in with your Grok account",
-    models: ["grok-4.3", "grok-4.20-0309-reasoning", "grok-4.20-0309-non-reasoning", "grok-build-0.1", "grok-composer-2.5-fast"],
-    defaultModel: "grok-4.3",
-    noReasoningModels: ["grok-build-0.1", "grok-composer-2.5-fast"],
+    // Parallel tool calls: officially supported and default-on per docs.x.ai function-calling
+    // (verified 260709, devlog/_plan/260709_parallel_tool_calls). Streamed calls arrive whole
+    // per chunk, so the buffered parser assembles them losslessly.
+    parallelToolCalls: true,
+    // Live /v1/models discovery is the authoritative lineup (verified 260709: returns grok-4.5);
+    // the static list below is the logged-out fallback seed.
+    liveModels: true,
+    // 260709 refresh: lineup + metadata from official docs.x.ai (grok-4.5 announced 07-08);
+    // grok-composer-2.5-fast kept as account-verified (absent from public docs). Evidence:
+    // devlog/model_update/260709_model_refresh/001_xai_lineup.md.
+    models: ["grok-4.5", "grok-4.3", "grok-4.20-multi-agent-0309", "grok-4.20-0309-reasoning", "grok-4.20-0309-non-reasoning", "grok-build-0.1", "grok-composer-2.5-fast"],
+    defaultModel: "grok-4.5",
+    noReasoningModels: ["grok-4.20-0309-non-reasoning", "grok-build-0.1", "grok-composer-2.5-fast"],
+    // grok-4.5 reasoning is always-on with low/medium/high control (no off tier upstream).
+    modelReasoningEfforts: { "grok-4.5": ["low", "medium", "high"] },
+    modelContextWindows: {
+      "grok-4.5": 500_000,
+      "grok-4.3": 1_000_000,
+      "grok-4.20-multi-agent-0309": 1_000_000,
+      "grok-4.20-0309-reasoning": 1_000_000,
+      "grok-4.20-0309-non-reasoning": 1_000_000,
+      "grok-build-0.1": 256_000,
+    },
     noVisionModels: ["grok-build-0.1", "grok-composer-2.5-fast"],
   },
   {
@@ -199,7 +230,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     note: "Log in with your Claude account",
     models: [...ANTHROPIC_MODELS],
     modelContextWindows: { ...ANTHROPIC_MODEL_CONTEXT_WINDOWS },
-    defaultModel: "claude-sonnet-4-6",
+    defaultModel: "claude-sonnet-5",
   },
   {
     id: "anthropic-apikey",
@@ -215,7 +246,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     models: [...ANTHROPIC_MODELS],
     liveModels: true,
     modelContextWindows: { ...ANTHROPIC_MODEL_CONTEXT_WINDOWS },
-    defaultModel: "claude-sonnet-4-6",
+    defaultModel: "claude-sonnet-5",
   },
   {
     id: "kimi",
@@ -252,7 +283,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelContextWindows: KIRO_MODEL_CONTEXT_WINDOWS,
     modelReasoningEfforts: KIRO_MODEL_REASONING_EFFORTS,
   },
-  { id: "openai-apikey", label: "OpenAI (API key)", adapter: "openai-responses", baseUrl: "https://api.openai.com/v1", authKind: "key", featured: true, dashboardUrl: "https://platform.openai.com/api-keys", defaultModel: "gpt-5.5", models: ["gpt-5.5", ...OPENAI_GPT56_MODELS], modelContextWindows: OPENAI_GPT56_CONTEXT_WINDOWS },
+  { id: "openai-apikey", label: "OpenAI (API key)", adapter: "openai-responses", baseUrl: "https://api.openai.com/v1", authKind: "key", featured: true, dashboardUrl: "https://platform.openai.com/api-keys", defaultModel: "gpt-5.5", models: ["gpt-5.5", ...OPENAI_GPT56_MODELS], liveModels: true, modelContextWindows: OPENAI_GPT56_CONTEXT_WINDOWS },
   {
     id: "umans",
     label: "Umans AI Coding Plan",
@@ -269,7 +300,6 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelReasoningEfforts: {
       "umans-coder": UMANS_REASONING_EFFORTS,
       "umans-kimi-k2.7": UMANS_REASONING_EFFORTS,
-      "umans-kimi-k2.6": UMANS_REASONING_EFFORTS,
       "umans-flash": UMANS_REASONING_EFFORTS,
       "umans-glm-5.2": UMANS_GLM_REASONING_EFFORTS,
       "umans-glm-5.1": UMANS_GLM_REASONING_EFFORTS,
@@ -287,6 +317,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       "kimi-k2.7-code": [],
       "kimi-k2.7-code-highspeed": [],
       ...Object.fromEntries(OPENCODE_GO_THINKING_TOGGLE_MODELS.map(id => [id, THINKING_TOGGLE_EFFORTS])),
+      ...Object.fromEntries(OPENCODE_GO_THINKING_BUDGET_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
     },
     // 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.
@@ -294,6 +325,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       ...Object.fromEntries(OPENCODE_GO_THINKING_TOGGLE_MODELS.map(id => [id, THINKING_TOGGLE_MAP])),
     },
     thinkingToggleModels: OPENCODE_GO_THINKING_TOGGLE_MODELS,
+    thinkingBudgetModels: THINKING_BUDGET_MODELS,
     noReasoningModels: ["kimi-k2.7-code", "kimi-k2.7-code-highspeed"],
     // Text-only Zen Go models (jawcode metadata) — the vision sidecar describes images for
     // every model listed here (and the catalog advertises image input on their behalf).
@@ -334,11 +366,14 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       "kimi-k2.6": [],
       "kimi-k2.6-fast": [],
       "kimi-k2.7-code": [],
-      "qwen3.5-397b": ["low", "medium", "high", "xhigh", "max"],
+      // Qwen3.x uses thinking_budget, NOT graded reasoning_effort; the adapter maps the five
+      // Codex picker levels onto budget fractions.
+      "qwen3.5-397b": THINKING_BUDGET_EFFORTS,
       "qwen3.5-397b-fast": [],
-      "qwen3.6-35b": ["low", "medium", "high", "xhigh", "max"],
+      "qwen3.6-35b": THINKING_BUDGET_EFFORTS,
       "qwen3.6-35b-fast": [],
     },
+    thinkingBudgetModels: THINKING_BUDGET_MODELS,
     noReasoningModels: ["glm-5.2-fast", "kimi-k2.5-fast", "kimi-k2.6-fast", "qwen3.5-397b-fast", "qwen3.6-35b-fast"],
     noVisionModels: ["glm-5.2", "glm-5.2-fast", "qwen3.5-397b", "qwen3.5-397b-fast"],
     noTemperatureModels: ["kimi-k2.7-code"],
@@ -381,7 +416,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
   {
     id: "moonshot", label: "Moonshot (Kimi API)", baseUrl: "https://api.moonshot.ai/v1", adapter: "openai-chat", authKind: "key",
     dashboardUrl: "https://platform.moonshot.ai/console/api-keys", defaultModel: "kimi-k2.7-code", jawcodeBundle: "moonshot",
-    models: ["kimi-k2.7-code", "kimi-k2.7-code-highspeed", "kimi-k2.6", "kimi-k2.5", "kimi-k2-0905-preview"],
+    models: ["kimi-k2.7-code", "kimi-k2.7-code-highspeed", "kimi-k2.6", "kimi-k2.5"],
     noReasoningModels: KIMI_THINKING_MODELS,
     modelReasoningEfforts: Object.fromEntries(KIMI_THINKING_MODELS.map(id => [id, []])),
     noTemperatureModels: KIMI_LOCKED_PARAMETER_MODELS,

package/src/reasoning-effort.ts

@@ -1,17 +1,24 @@
 import type { OcxProviderConfig } from "./types";
 import { modelInList } from "./types";
 
+// Descriptions mirror the upstream bundled models.json canonical wording (openai/codex PR #31684).
 export const CODEX_REASONING_LEVELS: { effort: string; description: string }[] = [
   { effort: "low", description: "Fast responses with lighter reasoning" },
-  { effort: "medium", description: "Balances speed and reasoning depth" },
+  { effort: "medium", description: "Balances speed and reasoning depth for everyday tasks" },
   { effort: "high", description: "Greater reasoning depth for complex problems" },
-  { effort: "xhigh", description: "Extended reasoning for the hardest problems" },
-  { effort: "max", description: "Maximum reasoning for the hardest problems" },
+  { effort: "xhigh", description: "Extra high reasoning depth for complex problems" },
+  { effort: "max", description: "Maximum reasoning depth for the hardest problems" },
+  { effort: "ultra", description: "Maximum reasoning with automatic task delegation" },
 ];
 
 const CODEX_REASONING_ORDER = CODEX_REASONING_LEVELS.map(l => l.effort);
 const CODEX_REASONING_SET = new Set(CODEX_REASONING_ORDER);
 
+/** True when `effort` is a member of the Codex reasoning ladder (low..ultra). */
+export function isCodexReasoningEffort(effort: string): boolean {
+  return CODEX_REASONING_SET.has(effort);
+}
+
 export function modelRecordValue<T>(record: Record<string, T> | undefined, modelId: string): T | undefined {
   if (!record) return undefined;
   if (Object.prototype.hasOwnProperty.call(record, modelId)) return record[modelId];
@@ -46,11 +53,25 @@ export function sanitizeCodexReasoningEfforts(efforts: readonly string[] | undef
 export function configuredReasoningEfforts(provider: OcxProviderConfig, modelId: string): string[] | undefined {
   if (modelInList(provider.noReasoningModels, modelId)) return [];
   const modelEfforts = modelRecordValue(provider.modelReasoningEfforts, modelId);
-  if (modelEfforts !== undefined) return sanitizeCodexReasoningEfforts(modelEfforts) ?? [];
-  if (provider.reasoningEfforts !== undefined) return sanitizeCodexReasoningEfforts(provider.reasoningEfforts) ?? [];
+  if (modelEfforts !== undefined) return healMaxTier(provider, modelId, sanitizeCodexReasoningEfforts(modelEfforts) ?? []);
+  if (provider.reasoningEfforts !== undefined) return healMaxTier(provider, modelId, sanitizeCodexReasoningEfforts(provider.reasoningEfforts) ?? []);
   return undefined;
 }
 
+/**
+ * Stale-ladder self-heal: saved configs seeded before `max` became a native Codex level can
+ * advertise a ladder that stops at `xhigh` while the wire map already routes xhigh -> max
+ * (e.g. opencode-go glm-5.2, deepseek thinking models). When the map proves the provider
+ * accepts wire `max`, append `max` so the picker actually shows the top tier. Thinking-toggle
+ * maps (xhigh -> "enabled") never match, so binary-toggle models stay two-step.
+ */
+function healMaxTier(provider: OcxProviderConfig, modelId: string, efforts: string[]): string[] {
+  if (efforts.includes("max") || !efforts.includes("xhigh")) return efforts;
+  const wireMap = reasoningEffortMapFor(provider, modelId);
+  if (wireMap?.xhigh !== "max" && wireMap?.max !== "max") return efforts;
+  return sanitizeCodexReasoningEfforts([...efforts, "max"]) ?? efforts;
+}
+
 function requestToCodexEffort(requested: string): string | undefined {
   if (requested === "none") return undefined;
   if (requested === "minimal") return "low";
@@ -89,13 +110,20 @@ export function mapReasoningEffort(provider: OcxProviderConfig, modelId: string,
   if (!requested) return undefined;
   if (modelInList(provider.noReasoningModels, modelId)) return undefined;
 
+  // Upstream codex-rs converts ultra -> max before ANY provider request (core/src/client.rs
+  // `reasoning_effort_for_request`), so "ultra" must never influence the provider wire — not even
+  // through a raw alias. Apply the boundary before alias/clamp resolution.
+  const boundary = requested === "ultra" ? "max" : requested;
+
   const wireMap = reasoningEffortMapFor(provider, modelId);
-  if (wireMap && Object.prototype.hasOwnProperty.call(wireMap, requested)) return wireMap[requested];
+  if (wireMap && Object.prototype.hasOwnProperty.call(wireMap, boundary)) return wireMap[boundary];
 
   const supported = configuredReasoningEfforts(provider, modelId);
-  const codexEffort = supported !== undefined ? clampToSupportedCodexEffort(requested, supported) : requestToCodexEffort(requested);
+  const codexEffort = supported !== undefined ? clampToSupportedCodexEffort(boundary, supported) : requestToCodexEffort(boundary);
   if (!codexEffort) return undefined;
 
-  if (wireMap && Object.prototype.hasOwnProperty.call(wireMap, codexEffort)) return wireMap[codexEffort];
-  return codexEffort;
+  // Belt for the odd config where the supported ladder is ultra-only and the clamp lands on it.
+  const wire = codexEffort === "ultra" ? "max" : codexEffort;
+  if (wireMap && Object.prototype.hasOwnProperty.call(wireMap, wire)) return wireMap[wire];
+  return wire;
 }

package/src/responses/parser.ts

@@ -464,8 +464,13 @@ export function parseRequest(body: unknown): OcxParsedRequest {
   const tc = mapToolChoice(data.tool_choice);
   if (tc !== undefined) options.toolChoice = tc;
   if (data.parallel_tool_calls !== undefined) options.parallelToolCalls = data.parallel_tool_calls;
-  if (data.reasoning?.effort && REASONING_EFFORTS.has(data.reasoning.effort)) {
-    options.reasoning = data.reasoning.effort;
+  // Upstream codex-rs converts "ultra" to "max" at the inference boundary (core/src/client.rs
+  // `reasoning_effort_for_request`), so current clients never send it — but a catalog that
+  // advertises ultra plus an older/direct caller can. Degrade it to max like upstream instead of
+  // silently dropping reasoning altogether.
+  const requestedEffort = data.reasoning?.effort === "ultra" ? "max" : data.reasoning?.effort;
+  if (requestedEffort && REASONING_EFFORTS.has(requestedEffort)) {
+    options.reasoning = requestedEffort;
   }
   const summaryMode = data.reasoning?.summary;
   if (!summaryMode || summaryMode === "none") options.hideThinkingSummary = true;

package/src/router.ts

@@ -96,6 +96,7 @@ function routedProviderConfig(providerName: string, provider: OcxProviderConfig)
   const autoToolChoiceOnlyModels = mergeStringArray(registryEntry.autoToolChoiceOnlyModels, provider.autoToolChoiceOnlyModels);
   const preserveReasoningContentModels = mergeStringArray(registryEntry.preserveReasoningContentModels, provider.preserveReasoningContentModels);
   const thinkingToggleModels = mergeStringArray(registryEntry.thinkingToggleModels, provider.thinkingToggleModels);
+  const thinkingBudgetModels = mergeStringArray(registryEntry.thinkingBudgetModels, provider.thinkingBudgetModels);
 
   return {
     ...provider,
@@ -112,6 +113,9 @@ function routedProviderConfig(providerName: string, provider: OcxProviderConfig)
     ...(provider.contextWindow === undefined && registryEntry.contextWindow !== undefined ? { contextWindow: registryEntry.contextWindow } : {}),
     ...(provider.reasoningEfforts === undefined && registryEntry.reasoningEfforts !== undefined ? { reasoningEfforts: registryEntry.reasoningEfforts } : {}),
     ...(provider.escapeBuiltinToolNames === undefined && registryEntry.escapeBuiltinToolNames !== undefined ? { escapeBuiltinToolNames: registryEntry.escapeBuiltinToolNames } : {}),
+    // Scalar backfill: a persisted config created before the flag shipped inherits the registry
+    // opt-in, while an explicit user `false` keeps overriding registry `true`.
+    ...(provider.parallelToolCalls === undefined && registryEntry.parallelToolCalls !== undefined ? { parallelToolCalls: registryEntry.parallelToolCalls } : {}),
     ...(modelContextWindows ? { modelContextWindows } : {}),
     ...(modelInputModalities ? { modelInputModalities } : {}),
     ...(modelReasoningEfforts ? { modelReasoningEfforts } : {}),
@@ -125,6 +129,7 @@ function routedProviderConfig(providerName: string, provider: OcxProviderConfig)
     ...(autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels } : {}),
     ...(preserveReasoningContentModels ? { preserveReasoningContentModels } : {}),
     ...(thinkingToggleModels ? { thinkingToggleModels } : {}),
+    ...(thinkingBudgetModels ? { thinkingBudgetModels } : {}),
   };
 }
 

package/src/server/adapter-resolve.ts

@@ -10,7 +10,7 @@ import type { OcxProviderConfig } from "../types";
 /** Providers whose listed model ids must be driven over the Anthropic wire even if the provider's
  *  configured adapter is something else (the upstream only speaks Anthropic for these models). */
 const ANTHROPIC_WIRE_MODELS: Record<string, Set<string>> = {
-  "opencode-go": new Set(["minimax-m2.5", "minimax-m2.7", "minimax-m3", "qwen3.5-plus", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus"]),
+  "opencode-go": new Set(["minimax-m2.5", "minimax-m2.7", "minimax-m3"]),
 };
 
 /** Return a provider config whose adapter is forced to "anthropic" when the model id is wire-pinned. */

package/src/server/index.ts

@@ -69,6 +69,7 @@ export {
   addFinalRequestLog,
   filterRequestLogs,
   httpStatusForTerminalStatus,
+  httpStatusFromTerminalError,
   nextRequestLogId,
   requestLogErrorCode,
   requestLogSpeedLabel,
@@ -157,6 +158,24 @@ export function startServer(port?: number) {
     config.subagentModels = [...DEFAULT_SUBAGENT_MODELS];
     saveConfig(config);
   }
+  // Sidecar model migration (KST 2026-07-10 06:00 = UTC 2026-07-09 21:00): auto-migrate the old
+  // gpt-5.4-mini default to gpt-5.6-luna for both search and vision sidecars. Only touches configs
+  // still on the old default — explicit user choices are preserved.
+  {
+    const SIDECAR_MIGRATION_CUTOFF = Date.UTC(2026, 6, 9, 21, 0); // July 9 21:00 UTC = KST July 10 06:00
+    if (Date.now() >= SIDECAR_MIGRATION_CUTOFF) {
+      let migrated = false;
+      if (config.webSearchSidecar?.model === "gpt-5.4-mini") {
+        config.webSearchSidecar = { ...config.webSearchSidecar, model: "gpt-5.6-luna" };
+        migrated = true;
+      }
+      if (config.visionSidecar?.model === "gpt-5.4-mini") {
+        config.visionSidecar = { ...config.visionSidecar, model: "gpt-5.6-luna" };
+        migrated = true;
+      }
+      if (migrated) saveConfig(config);
+    }
+  }
   invalidateCodexModelsCache();
 
   const listenPort = port ?? config.port ?? 10100;
@@ -247,7 +266,7 @@ export function startServer(port?: number) {
           return withCors(formatErrorResponse(403, "origin_rejected", "cross-origin data-plane request blocked"), req, config);
         }
         const goModels = await fetchAllModels(config);
-        const { buildCatalogEntries, loadCatalogTemplate, nativeOpenAiSlugs, orderForSubagents, filterCatalogVisibleModels } = await import("../codex/catalog");
+        const { applyNativeVisibility, buildCatalogEntries, disabledNativeSlugs, loadCatalogTemplate, nativeOpenAiSlugs, orderForSubagents, filterCatalogVisibleModels, visibleNativeSlugs } = await import("../codex/catalog");
         const nativeSlugs = nativeOpenAiSlugs();
         const goEnabled = filterCatalogVisibleModels(goModels, config);
         const goOrdered = orderForSubagents(goEnabled, config.subagentModels);
@@ -255,11 +274,16 @@ export function startServer(port?: number) {
           // Codex client → Codex catalog shape: native gpt + namespaced routed models,
           // cloned from a native template so required fields (base_instructions, etc.) are present.
           // Pass the subagent picks so featured models lead by priority (matches the on-disk file).
-          return jsonResponse({ models: buildCatalogEntries(loadCatalogTemplate(), nativeSlugs, goOrdered, config.subagentModels, websocketsEnabled(config)) }, 200, req, config);
+          // Disabled natives stay in the catalog shape with visibility "hide" (mirrors the
+          // on-disk sync; codex-rs keeps them out of the picker itself).
+          const maMode = config.multiAgentMode === "v1" || config.multiAgentMode === "v2" ? config.multiAgentMode : "default";
+          const entries = buildCatalogEntries(loadCatalogTemplate(), nativeSlugs, goOrdered, config.subagentModels, websocketsEnabled(config), maMode as "v1" | "default" | "v2");
+          return jsonResponse({ models: applyNativeVisibility(entries, disabledNativeSlugs(config)) }, 200, req, config);
         }
         // OpenAI list shape: native gpt bare + routed models namespaced "<provider>/<id>"
+        // (pure availability list — disabled natives are omitted entirely).
         const data = [
-          ...nativeSlugs.map(id => ({ id, object: "model", created: 0, owned_by: "openai" })),
+          ...visibleNativeSlugs(config).map(id => ({ id, object: "model", created: 0, owned_by: "openai" })),
           ...goOrdered.map(m => ({ id: `${m.provider}/${m.id}`, object: "model", created: 0, owned_by: m.owned_by ?? m.provider })),
         ];
         return jsonResponse({ object: "list", data }, 200, req, config);