@bitkyc08/opencodex 2.6.31-preview.20260707 → 2.7.0

package/src/oauth/token-guardian.ts

@@ -18,9 +18,12 @@ import { getValidAccessTokenForAccount, listOAuthProviders, OAuthLoginRequiredEr
 import {
   getValidCodexToken,
   listCodexAccountIds,
+  markCodexAccountValidated,
+  markCodexAccountValidationFailed,
   readCodexAccountRecord,
   TokenRefreshError,
 } from "../codex/account-store";
+import { codexWarmupFailureReason, warmCodexAccount } from "../codex/warmup";
 
 export interface TokenGuardianHandle {
   stop(): void;
@@ -29,6 +32,7 @@ export interface TokenGuardianHandle {
 export interface GuardianSweepResult {
   enabled: boolean;
   refreshed: string[];
+  warmed: string[];
   failed: string[];
   skippedBackoff: string[];
 }
@@ -40,6 +44,8 @@ const DEFAULTS = {
   leadSeconds: 900,
   failureBackoffBaseSeconds: 300,
   failureBackoffMaxSeconds: 3600,
+  codexWarmupMaxAgeSeconds: 691_200, // 8d — matches Codex managed-auth last_refresh cadence.
+  codexWarmupModel: "gpt-5.4-mini",
 };
 
 interface BackoffEntry {
@@ -67,6 +73,9 @@ function resolved(g: OcxTokenGuardianConfig | undefined) {
     leadSeconds: num(g?.leadSeconds, DEFAULTS.leadSeconds, 0),
     backoffBaseSeconds: num(g?.failureBackoffBaseSeconds, DEFAULTS.failureBackoffBaseSeconds, 0),
     backoffMaxSeconds: num(g?.failureBackoffMaxSeconds, DEFAULTS.failureBackoffMaxSeconds, 0),
+    codexWarmupEnabled: g?.codexWarmupEnabled === true,
+    codexWarmupMaxAgeSeconds: num(g?.codexWarmupMaxAgeSeconds, DEFAULTS.codexWarmupMaxAgeSeconds, 60),
+    codexWarmupModel: g?.codexWarmupModel?.trim() || DEFAULTS.codexWarmupModel,
   };
 }
 
@@ -106,7 +115,7 @@ async function runWithConcurrency(tasks: Array<() => Promise<void>>, limit: numb
 export async function guardianSweep(nowMs: number = Date.now()): Promise<GuardianSweepResult> {
   const config: OcxConfig = loadConfig();
   const g = config.tokenGuardian;
-  const result: GuardianSweepResult = { enabled: !!g?.enabled, refreshed: [], failed: [], skippedBackoff: [] };
+  const result: GuardianSweepResult = { enabled: !!g?.enabled, refreshed: [], warmed: [], failed: [], skippedBackoff: [] };
   if (!g?.enabled) return result;
 
   const opts = resolved(g);
@@ -142,18 +151,34 @@ export async function guardianSweep(nowMs: number = Date.now()): Promise<Guardia
   if (resolveRefreshPolicy("chatgpt", config) === "proactive") {
     for (const id of listCodexAccountIds()) {
       const record = readCodexAccountRecord(id);
-      const cred = record?.deletedAt == null ? record?.credential : undefined;
+      if (!record || record.deletedAt != null) continue;
+      const cred = record.credential;
       if (!cred) continue;
-      if (cred.expiresAt > nowMs + horizonMs) continue;
+      const needsRefresh = cred.expiresAt <= nowMs + horizonMs;
+      const needsWarmup = opts.codexWarmupEnabled
+        && (record.lastCodexValidatedAt === undefined || nowMs - record.lastCodexValidatedAt > opts.codexWarmupMaxAgeSeconds * 1000);
+      if (!needsRefresh && !needsWarmup) continue;
       const key = `codex:${id}`;
       if (inBackoff(key, nowMs)) { result.skippedBackoff.push(key); continue; }
       tasks.push(async () => {
         try {
-          await getValidCodexToken(id);
+          const token = await getValidCodexToken(id);
+          if (needsRefresh) result.refreshed.push(key);
+          if (needsWarmup) {
+            await warmCodexAccount({
+              accessToken: token.accessToken,
+              chatgptAccountId: token.chatgptAccountId,
+              model: opts.codexWarmupModel,
+            });
+            markCodexAccountValidated(id, Date.now());
+            result.warmed.push(key);
+          }
           backoff.delete(key);
-          result.refreshed.push(key);
         } catch (err) {
           const permanent = err instanceof TokenRefreshError && (err.reason === "revoked" || err.reason === "expired");
+          if (needsWarmup && !(err instanceof TokenRefreshError)) {
+            markCodexAccountValidationFailed(id, codexWarmupFailureReason(err));
+          }
           recordFailure(key, nowMs, opts.backoffBaseSeconds, opts.backoffMaxSeconds, permanent);
           result.failed.push(key);
         }
@@ -186,8 +211,8 @@ export function startTokenGuardian(): TokenGuardianHandle {
   const runSweep = () => {
     void guardianSweep()
       .then(r => {
-        if (r.enabled && (r.refreshed.length || r.failed.length)) {
-          console.log(`🛡️  token-guardian: refreshed ${r.refreshed.length}, failed ${r.failed.length}`);
+        if (r.enabled && (r.refreshed.length || r.warmed.length || r.failed.length)) {
+          console.log(`🛡️  token-guardian: refreshed ${r.refreshed.length}, warmed ${r.warmed.length}, failed ${r.failed.length}`);
         }
       })
       .catch(err => console.log(`token-guardian sweep error: ${err instanceof Error ? err.message : String(err)}`))

package/src/providers/derive.ts

@@ -23,6 +23,8 @@ export interface DerivedKeyLoginProvider {
   noPenaltyModels?: string[];
   autoToolChoiceOnlyModels?: string[];
   preserveReasoningContentModels?: string[];
+  thinkingToggleModels?: string[];
+  thinkingBudgetModels?: string[];
   escapeBuiltinToolNames?: boolean;
 }
 
@@ -80,9 +82,11 @@ export function providerConfigSeed(entry: ProviderRegistryEntry): OcxProviderCon
     ...(entry.noTemperatureModels ? { noTemperatureModels: [...entry.noTemperatureModels] } : {}),
     ...(entry.noTopPModels ? { noTopPModels: [...entry.noTopPModels] } : {}),
     ...(entry.noPenaltyModels ? { noPenaltyModels: [...entry.noPenaltyModels] } : {}),
+    ...(entry.parallelToolCalls !== undefined ? { parallelToolCalls: entry.parallelToolCalls } : {}),
     ...(entry.autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels: [...entry.autoToolChoiceOnlyModels] } : {}),
     ...(entry.preserveReasoningContentModels ? { preserveReasoningContentModels: [...entry.preserveReasoningContentModels] } : {}),
     ...(entry.thinkingToggleModels ? { thinkingToggleModels: [...entry.thinkingToggleModels] } : {}),
+    ...(entry.thinkingBudgetModels ? { thinkingBudgetModels: [...entry.thinkingBudgetModels] } : {}),
     ...(entry.escapeBuiltinToolNames !== undefined ? { escapeBuiltinToolNames: entry.escapeBuiltinToolNames } : {}),
     ...(entry.googleMode ? { googleMode: entry.googleMode } : {}),
     ...(entry.project ? { project: entry.project } : {}),
@@ -117,6 +121,8 @@ export function deriveKeyLoginMap(): Record<string, DerivedKeyLoginProvider> {
       ...(entry.noPenaltyModels ? { noPenaltyModels: [...entry.noPenaltyModels] } : {}),
       ...(entry.autoToolChoiceOnlyModels ? { autoToolChoiceOnlyModels: [...entry.autoToolChoiceOnlyModels] } : {}),
       ...(entry.preserveReasoningContentModels ? { preserveReasoningContentModels: [...entry.preserveReasoningContentModels] } : {}),
+      ...(entry.thinkingToggleModels ? { thinkingToggleModels: [...entry.thinkingToggleModels] } : {}),
+      ...(entry.thinkingBudgetModels ? { thinkingBudgetModels: [...entry.thinkingBudgetModels] } : {}),
       ...(entry.escapeBuiltinToolNames !== undefined ? { escapeBuiltinToolNames: entry.escapeBuiltinToolNames } : {}),
       ...(entry.googleMode ? { googleMode: entry.googleMode } : {}),
       ...(entry.project ? { project: entry.project } : {}),
@@ -177,9 +183,11 @@ export function enrichProviderFromRegistry(name: string, prov: OcxProviderConfig
   if (!prov.noTemperatureModels && seed.noTemperatureModels) prov.noTemperatureModels = [...seed.noTemperatureModels];
   if (!prov.noTopPModels && seed.noTopPModels) prov.noTopPModels = [...seed.noTopPModels];
   if (!prov.noPenaltyModels && seed.noPenaltyModels) prov.noPenaltyModels = [...seed.noPenaltyModels];
+  if (prov.parallelToolCalls === undefined && seed.parallelToolCalls !== undefined) prov.parallelToolCalls = seed.parallelToolCalls;
   if (!prov.autoToolChoiceOnlyModels && seed.autoToolChoiceOnlyModels) prov.autoToolChoiceOnlyModels = [...seed.autoToolChoiceOnlyModels];
   if (!prov.preserveReasoningContentModels && seed.preserveReasoningContentModels) prov.preserveReasoningContentModels = [...seed.preserveReasoningContentModels];
   if (!prov.thinkingToggleModels && seed.thinkingToggleModels) prov.thinkingToggleModels = [...seed.thinkingToggleModels];
+  if (!prov.thinkingBudgetModels && seed.thinkingBudgetModels) prov.thinkingBudgetModels = [...seed.thinkingBudgetModels];
   if (prov.escapeBuiltinToolNames === undefined && seed.escapeBuiltinToolNames !== undefined) prov.escapeBuiltinToolNames = seed.escapeBuiltinToolNames;
 }
 

package/src/providers/kiro-models.ts

@@ -38,10 +38,10 @@ export const KIRO_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
   "qwen3-coder-next": 256_000,
 };
 
-const KIRO_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
+const KIRO_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
 
-// Codex does not accept raw "max" in catalog metadata; Kiro xhigh already maps to the maximum
-// fake-thinking budget in src/adapters/kiro.ts.
+// Kiro has no upstream reasoning_effort enum; these labels map to fake-thinking budgets in
+// src/adapters/kiro.ts.
 export const KIRO_MODEL_REASONING_EFFORTS: Record<string, string[]> = Object.fromEntries(
   KIRO_MODELS.map(id => [id, KIRO_REASONING_EFFORTS]),
 );

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;
@@ -56,37 +59,40 @@ 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"
 >;
 
-
-const OLLAMA_REASONING_MAP: Record<string, string> = { xhigh: "max" };
-
 // 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"];
-const ZAI_GLM_52_REASONING_MAP: Record<string, string> = {
-  none: "none",
-  minimal: "none",
-  low: "high",
-  medium: "high",
-  high: "high",
-  xhigh: "max",
-  max: "max",
+const ZAI_GLM_52_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
+const OPENAI_GPT56_MODELS = ["gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"];
+const OPENAI_GPT56_CONTEXT_WINDOW = 372_000;
+const OPENAI_GPT56_CONTEXT_WINDOWS = {
+  "gpt-5.6-sol": OPENAI_GPT56_CONTEXT_WINDOW,
+  "gpt-5.6-terra": OPENAI_GPT56_CONTEXT_WINDOW,
+  "gpt-5.6-luna": OPENAI_GPT56_CONTEXT_WINDOW,
+};
+const OPENROUTER_GPT56_MODELS = OPENAI_GPT56_MODELS.map(id => `openai/${id}`);
+const OPENROUTER_GPT56_CONTEXT_WINDOWS = {
+  "openai/gpt-5.6-sol": OPENAI_GPT56_CONTEXT_WINDOW,
+  "openai/gpt-5.6-terra": OPENAI_GPT56_CONTEXT_WINDOW,
+  "openai/gpt-5.6-luna": OPENAI_GPT56_CONTEXT_WINDOW,
 };
 
 /**
  * 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",
@@ -99,8 +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",
@@ -108,7 +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",
@@ -118,28 +132,17 @@ 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",
   "umans-qwen3.6-35b-a3b",
 ];
-const UMANS_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
-const UMANS_GLM_REASONING_EFFORTS = ["high", "xhigh"];
-const UMANS_GLM_REASONING_MAP: Record<string, string> = {
-  none: "high",
-  minimal: "high",
-  low: "high",
-  medium: "high",
-  high: "high",
-  xhigh: "max",
-  max: "max",
-};
+const UMANS_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
+const UMANS_GLM_REASONING_EFFORTS = ["high", "xhigh", "max"];
 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,
@@ -167,7 +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",
@@ -190,9 +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"],
   },
   {
@@ -207,7 +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",
@@ -223,7 +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",
@@ -260,7 +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" },
+  { 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",
@@ -277,15 +300,10 @@ 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": ["low", "medium", "high"],
+      "umans-flash": UMANS_REASONING_EFFORTS,
       "umans-glm-5.2": UMANS_GLM_REASONING_EFFORTS,
       "umans-glm-5.1": UMANS_GLM_REASONING_EFFORTS,
-      "umans-qwen3.6-35b-a3b": ["low", "medium", "high"],
-    },
-    modelReasoningEffortMap: {
-      "umans-glm-5.2": UMANS_GLM_REASONING_MAP,
-      "umans-glm-5.1": UMANS_GLM_REASONING_MAP,
+      "umans-qwen3.6-35b-a3b": UMANS_REASONING_EFFORTS,
     },
     noVisionModels: UMANS_TEXT_ONLY_MODELS,
     escapeBuiltinToolNames: true,
@@ -299,12 +317,15 @@ 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.
     modelReasoningEffortMap: {
-      "glm-5.2": ZAI_GLM_52_REASONING_MAP,
       ...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).
@@ -345,12 +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"],
+      // 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"],
+      "qwen3.6-35b": THINKING_BUDGET_EFFORTS,
       "qwen3.6-35b-fast": [],
     },
-    modelReasoningEffortMap: { "glm-5.2": ZAI_GLM_52_REASONING_MAP },
+    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"],
@@ -359,13 +382,13 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     autoToolChoiceOnlyModels: ["kimi-k2.7-code"],
     preserveReasoningContentModels: NEURALWATT_REASONING_HISTORY_MODELS,
   },
-  { id: "openrouter", label: "OpenRouter", adapter: "openai-chat", baseUrl: "https://openrouter.ai/api/v1", authKind: "key", featured: true, dashboardUrl: "https://openrouter.ai/keys", jawcodeBundle: "openrouter", models: ["anthropic/claude-sonnet-5"], modelContextWindows: { "anthropic/claude-sonnet-5": 1_000_000 } },
+  { id: "openrouter", label: "OpenRouter", adapter: "openai-chat", baseUrl: "https://openrouter.ai/api/v1", authKind: "key", featured: true, dashboardUrl: "https://openrouter.ai/keys", jawcodeBundle: "openrouter", models: ["anthropic/claude-sonnet-5", ...OPENROUTER_GPT56_MODELS], modelContextWindows: { "anthropic/claude-sonnet-5": 1_000_000, ...OPENROUTER_GPT56_CONTEXT_WINDOWS } },
   { id: "groq", label: "Groq", adapter: "openai-chat", baseUrl: "https://api.groq.com/openai/v1", authKind: "key", featured: true, dashboardUrl: "https://console.groq.com/keys" },
   { id: "google", label: "Google Gemini", adapter: "google", baseUrl: "https://generativelanguage.googleapis.com", authKind: "key", featured: true, dashboardUrl: "https://aistudio.google.com/apikey", defaultModel: "gemini-3-pro", jawcodeBundle: "google", extraMetadataAliases: ["gemini"] },
   { id: "google-vertex", label: "Google Vertex AI", adapter: "google", baseUrl: "https://aiplatform.googleapis.com", authKind: "key", dashboardUrl: "https://console.cloud.google.com/vertex-ai", defaultModel: "gemini-3-pro", googleMode: "vertex", jawcodeBundle: "google", extraMetadataAliases: ["gemini-vertex"] },
   { id: "google-antigravity", label: "Google Antigravity", adapter: "google", baseUrl: "https://daily-cloudcode-pa.googleapis.com", authKind: "oauth", dashboardUrl: "https://antigravity.google", models: ANTIGRAVITY_MODELS, defaultModel: "gemini-3.5-flash-low", modelContextWindows: ANTIGRAVITY_MODEL_CONTEXT_WINDOWS, googleMode: "cloud-code-assist", jawcodeBundle: "google", extraMetadataAliases: ["antigravity", "gemini-antigravity"] },
   { id: "azure-openai", label: "Azure OpenAI", adapter: "azure-openai", baseUrl: "https://{resource}.openai.azure.com/openai", authKind: "key", featured: true, dashboardUrl: "https://portal.azure.com" },
-  { id: "ollama", label: "Ollama (local)", adapter: "openai-chat", baseUrl: "http://localhost:11434/v1", authKind: "local", featured: true, note: "Local — key usually blank", reasoningEffortMap: OLLAMA_REASONING_MAP },
+  { id: "ollama", label: "Ollama (local)", adapter: "openai-chat", baseUrl: "http://localhost:11434/v1", authKind: "local", featured: true, note: "Local — key usually blank" },
   { id: "vllm", label: "vLLM (local)", adapter: "openai-chat", baseUrl: "http://localhost:8000/v1", authKind: "local", featured: true, note: "Local — key usually blank" },
   { id: "lm-studio", label: "LM Studio (local)", adapter: "openai-chat", baseUrl: "http://localhost:1234/v1", authKind: "local", featured: true, note: "Local — no key needed" },
   {
@@ -393,7 +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,
@@ -412,7 +435,6 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     models: ["glm-5.2", "glm-5.2[1m]", "glm-5.1", "glm-5", "glm-4.6"],
     noVisionModels: ZAI_GLM_52_MODELS,
     modelReasoningEfforts: Object.fromEntries(ZAI_GLM_52_MODELS.map(id => [id, ZAI_GLM_52_REASONING_EFFORTS])),
-    modelReasoningEffortMap: Object.fromEntries(ZAI_GLM_52_MODELS.map(id => [id, ZAI_GLM_52_REASONING_MAP])),
     preserveReasoningContentModels: ZAI_GLM_52_MODELS,
   },
   { id: "nanogpt", label: "NanoGPT", baseUrl: "https://nano-gpt.com/api/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://nano-gpt.com/api" },
@@ -430,7 +452,6 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     adapter: "openai-chat",
     authKind: "key",
     dashboardUrl: "https://ollama.com/settings/keys",
-    reasoningEffortMap: OLLAMA_REASONING_MAP,
     models: ["glm-5.2", "deepseek-v4-pro", "qwen3-coder", "gpt-oss:120b", "kimi-k2.6", "minimax-m3", "qwen3.5", "gemma4"],
     defaultModel: "glm-5.2",
     noVisionModels: [

package/src/reasoning-effort.ts

@@ -1,11 +1,14 @@
 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: "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);
@@ -45,15 +48,28 @@ 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";
-  if (requested === "max") return "xhigh";
   return CODEX_REASONING_SET.has(requested) ? requested : undefined;
 }
 
@@ -82,21 +98,27 @@ export function reasoningEffortMapFor(provider: OcxProviderConfig, modelId: stri
 }
 
 /**
- * Translate Codex's reasoning label into the provider's real wire value. The Codex catalog must only
- * advertise labels Codex itself accepts (`low`/`medium`/`high`/`xhigh`), but some upstreams use
- * different values (`max`) or a smaller subset (`low`/`medium`/`high`).
+ * Translate Codex's reasoning label into the provider's real wire value. Prefer identity labels
+ * (`xhigh` stays `xhigh`, `max` stays `max`); provider maps are only for real upstream aliases.
  */
 export function mapReasoningEffort(provider: OcxProviderConfig, modelId: string, requested: string | undefined): string | undefined {
   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

@@ -29,9 +29,8 @@ const MODEL_PROVIDER_PATTERNS: Array<{ providerNames: string[]; prefixes: string
   },
 ];
 
-// Merge registry-default effort maps under user values so persisted built-in provider configs
-// that predate reasoningEffortMap/modelReasoningEffortMap still get correct wire translations
-// (e.g. ollama-cloud xhigh -> max) without a disk migration. User overrides win per-key.
+// Merge registry-default effort maps under user values so built-in provider configs can
+// carry real upstream aliases without a disk migration. User overrides win per-key.
 function mergeRecord(
   seed: Record<string, string> | undefined,
   user: Record<string, string> | undefined,
@@ -97,6 +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,
@@ -113,6 +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 } : {}),
@@ -126,6 +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. */