@bitkyc08/opencodex 2.10.1 → 2.10.2

package/src/providers/registry.ts

@@ -14,6 +14,7 @@ import {
   cursorModelInputModalities,
   cursorModelReasoningEfforts,
 } from "../adapters/cursor/discovery";
+import { COMMAND_CODE_MODEL_REASONING_EFFORTS } from "./command-code-efforts";
 
 export type ProviderAuthKind = "forward" | "oauth" | "key" | "local";
 export type MetadataModelIdNormalize = "case-insensitive";
@@ -348,16 +349,89 @@ const THINKING_BUDGET_MODELS = [
 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 OPENCODE_FREE_DEEPSEEK_MODELS = ["deepseek-v4-flash-free"];
-// "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> = {
+/*
+ * Zen free models that reject `image_url` upstream (#1043, and the reproducible
+ * half of #1024).
+ *
+ * Zen publishes NO modality metadata — its `/v1/models` returns only id, object,
+ * created, owned_by — so this list is measured, not derived. Each id was probed
+ * once against https://opencode.ai/zen/v1 on 2026-08-05 with a text control first
+ * and then a 1x1 PNG; the six below failed the image request, four of them with
+ * `[404] No endpoints found that support image input` and `big-pickle` with the
+ * exact deserialize error quoted in #1043.
+ *
+ * `mimo-v2.5-free` and `longcat-2.0-free` ACCEPT images and are deliberately
+ * absent. Adding them would silently replace a working image with a caption,
+ * which is worse than the loud 400 this list exists to prevent — see the negative
+ * assertion in tests/provider-registry-parity.test.ts.
+ *
+ * Zen's roster is discovered live while this list is static, so it is a dated
+ * exception list, not a capability model. Re-probe before extending it.
+ * Evidence: devlog/_plan/260805_bug_fix_stack/002_zen_modality_probe.md
+ */
+const OPENCODE_ZEN_TEXT_ONLY_MODELS = [
+  "big-pickle",
+  "nemotron-3-ultra-free",
+  "ling-3.0-flash-free",
+  "north-mini-code-free",
+  "laguna-s-2.1-free",
+  "deepseek-v4-flash-free",
+];
+/*
+ * DeepSeek's Codex ladder is low/high/max, and the two V4 models resolve it
+ * DIFFERENTLY. From the official thinking-mode table (api-docs.deepseek.com,
+ * EN and zh-cn agree, re-verified 2026-08-06):
+ *
+ *   requested  | v4-flash | v4-pro
+ *   low        | low      | high
+ *   high       | high     | high
+ *   xhigh      | high     | max
+ *   max        | max      | max
+ *
+ * Two consequences (#1057):
+ *
+ * - `xhigh` is a COMPATIBILITY ALIAS, not a native tier. It stays in the wire maps
+ *   so existing requests and saved configs keep working, but it is not advertised.
+ * - Pro does NOT honor `low` — the vendor silently upgrades it to `high`. So Pro
+ *   advertises only the two levels it actually distinguishes. Advertising `low`
+ *   there would put a tier in the picker that costs `high`, which is the same
+ *   defect this fixes wearing a different value.
+ *
+ * The vendor page footnotes that Pro's mapping updates in early August 2026; as of
+ * the re-verification above it had not changed. When it does, Pro gains `low` here.
+ *
+ * `medium` has no row in the vendor table — mapping it to `high` is OUR
+ * compatibility choice for clients that only speak the OpenAI ladder.
+ */
+const DEEPSEEK_FLASH_THINKING_EFFORTS = ["low", "high", "max"];
+const DEEPSEEK_PRO_THINKING_EFFORTS = ["high", "max"];
+const DEEPSEEK_PRO_REASONING_MAP: Record<string, string> = {
   low: "high",
   medium: "high",
   high: "high",
   xhigh: "max",
   max: "max",
 };
+const DEEPSEEK_FLASH_REASONING_MAP: Record<string, string> = {
+  low: "low",
+  medium: "high",
+  high: "high",
+  xhigh: "high",
+  max: "max",
+};
+/**
+ * Flash-versus-Pro classification for DeepSeek V4 model ids, including prefixed
+ * (`deepseek/deepseek-v4-pro`) and suffixed (`deepseek-v4-flash-free`) forms.
+ * `tests/provider-registry-parity.test.ts` enumerates every id the registry
+ * actually passes here, so a future id this substring test would misread cannot
+ * land silently.
+ */
+const isDeepseekFlashModel = (modelId: string): boolean =>
+  modelId.toLowerCase().includes("flash");
+const deepseekThinkingEffortsFor = (modelId: string): string[] =>
+  isDeepseekFlashModel(modelId) ? DEEPSEEK_FLASH_THINKING_EFFORTS : DEEPSEEK_PRO_THINKING_EFFORTS;
+const deepseekReasoningMapFor = (modelId: string): Record<string, string> =>
+  isDeepseekFlashModel(modelId) ? DEEPSEEK_FLASH_REASONING_MAP : DEEPSEEK_PRO_REASONING_MAP;
 // 260719 Alibaba Token Plan Personal Edition (China/Beijing). Keep it distinct from
 // Coding Plan: the products use different exact allowlists and different base URLs.
 // Evidence: https://help.aliyun.com/en/model-studio/token-plan-personal-overview
@@ -412,7 +486,7 @@ const TENCENT_CODING_PLAN_MODELS = ["tc-code-latest", "glm-5", "kimi-k2.5", "min
 // is documented as grounds for suspension — see the `note` on both Plan entries.
 // Report a break by opening an issue tagging the owner; the three things that rot first are the
 // static catalogs (liveModels:false cannot self-heal), the base URLs, and those Plan terms.
-// Full evidence ledger: devlog/_plan/260801_pr611_volcengine_evidence/000_evidence_ledger.md
+// Full evidence ledger: devlog/_fin/260801_pr611_volcengine_evidence/000_evidence_ledger.md
 const VOLCENGINE_ARK_MODELS = [
   "doubao-seed-2-1-pro-260628",
   "doubao-seed-2-1-turbo-260628",
@@ -537,7 +611,7 @@ const NVIDIA_NIM_KIMI_MODELS = [
  *
  * Two lists, both verified per-model against NVIDIA documentation on 2026-08-04
  * (build.nvidia.com model pages and docs.api.nvidia.com/nim/reference/*). Evidence and
- * the per-id audit: devlog/_plan/260804_stack7_service_vision/011_nim_id_audit.md.
+ * the per-id audit: devlog/_fin/260804_stack7_service_vision/011_nim_id_audit.md.
  *
  * Read `noVisionModels` carefully — it lists models that CANNOT see images, which is
  * what routes them through the proxy's vision sidecar (src/vision/index.ts) and makes the
@@ -799,6 +873,32 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     },
     noVisionModels: ["grok-build-0.1", "grok-composer-2.5-fast"],
   },
+  {
+    id: "command-code",
+    label: "Command Code - Auth",
+    adapter: "command-code",
+    baseUrl: "https://api.commandcode.ai",
+    authKind: "oauth",
+    oauthId: "command-code",
+    featured: true,
+    note: "Log in with your Command Code account",
+    // OAuth needs one initial selection, but the exposed catalog is always discovered from the
+    // signed-in account. Do not add a static model list here.
+    defaultModel: "deepseek/deepseek-v4-flash",
+    liveModels: true,
+    modelDiscovery: {
+      url: "https://api.commandcode.ai/provider/v1/models",
+      maxResponseBytes: 262_144,
+      maxModels: 256,
+    },
+    // These are capability facts from official Command Code model profiles, not seeded models.
+    // Unknown/new live models deliberately do not advertise a reasoning picker.
+    reasoningEfforts: [],
+    modelReasoningEfforts: COMMAND_CODE_MODEL_REASONING_EFFORTS,
+    defaultMaxOutputTokens: 64_000,
+    // The proprietary generate wire has no verified per-request serialization flag.
+    parallelToolCalls: false,
+  },
   {
     id: "anthropic",
     label: "Anthropic Claude",
@@ -938,7 +1038,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       "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])),
-      ...Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, DEEPSEEK_THINKING_EFFORTS])),
+      ...Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)])),
     },
     modelDefaultReasoningEfforts: { "kimi-k3": "max" },
     // glm-5.2 uses identity labels now that `max` is a native Codex level (no alias map);
@@ -946,7 +1046,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelReasoningEffortMap: {
       "kimi-k3": KIMI_CODING_K3_REASONING_EFFORT_MAP,
       ...Object.fromEntries(OPENCODE_GO_THINKING_TOGGLE_MODELS.map(id => [id, THINKING_TOGGLE_MAP])),
-      ...Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, DEEPSEEK_THINKING_REASONING_MAP])),
+      ...Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, deepseekReasoningMapFor(id)])),
     },
     thinkingToggleModels: OPENCODE_GO_THINKING_TOGGLE_MODELS,
     thinkingBudgetModels: THINKING_BUDGET_MODELS,
@@ -1087,9 +1187,9 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     //   returned 200), so no noTemperatureModels entry is warranted here.
     modelReasoningEfforts: {
       "openai/gpt-5.5": ["low", "medium", "high", "xhigh"],
-      "deepseek/deepseek-v4-pro": DEEPSEEK_THINKING_EFFORTS,
+      "deepseek/deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek/deepseek-v4-pro"),
     },
-    modelReasoningEffortMap: { "deepseek/deepseek-v4-pro": DEEPSEEK_THINKING_REASONING_MAP },
+    modelReasoningEffortMap: { "deepseek/deepseek-v4-pro": deepseekReasoningMapFor("deepseek/deepseek-v4-pro") },
     preserveReasoningContentModels: ["deepseek/deepseek-v4-pro"],
     note: "OpenAI-compatible adaptive router. Default is a tool-capable model; orcarouter/auto (adaptive routing) is also selectable. Full catalog: https://www.orcarouter.ai/models",
   },
@@ -1182,8 +1282,8 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     - 대안 분석: Globally preserve reasoning_content for all OpenAI-compatible models; preserve it for legacy deepseek-reasoner too; mark only V4 thinking models in registry metadata.
     - 선택 근거: DeepSeek V4 thinking mode requires history replay, while older DeepSeek reasoner has different compatibility rules. A model-scoped registry flag fixes built-in and stale saved configs without broad provider regressions.
     */
-    modelReasoningEfforts: Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, DEEPSEEK_THINKING_EFFORTS])),
-    modelReasoningEffortMap: Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, DEEPSEEK_THINKING_REASONING_MAP])),
+    modelReasoningEfforts: Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)])),
+    modelReasoningEffortMap: Object.fromEntries(DEEPSEEK_THINKING_MODELS.map(id => [id, deepseekReasoningMapFor(id)])),
     preserveReasoningContentModels: DEEPSEEK_THINKING_MODELS,
     // Issue #88: every DeepSeek API model is text-only input (no image support upstream) — the
     // vision sidecar describes attached images for them, and the catalog advertises image input
@@ -1256,7 +1356,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
   },
   {
     id: "commandcode",
-    label: "Command Code",
+    label: "Command Code - API",
     adapter: "openai-chat",
     baseUrl: "https://api.commandcode.ai/provider/v1",
     authKind: "key",
@@ -1278,7 +1378,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     },
     // Verified 2026-08-03: public /provider/v1/models returns 51 rows; /chat/completions returns
     // 401 UNAUTHORIZED without a Bearer key. Primary source: https://commandcode.ai/docs/provider.
-    note: "Command Code Provider API (OpenAI-compatible); API access requires the Provider plan. CLI auth bridging for Go/Pro subscriptions is not yet available. Docs: https://commandcode.ai/docs/provider.",
+    note: "Command Code Provider API (OpenAI-compatible); API access requires the Provider plan. Use `ocx login command-code` for OAuth account login (imports an existing local Command Code CLI credential when present). Docs: https://commandcode.ai/docs/provider.",
   },
   // FREEZE 2026-07-10: exact serverless ids remain auth-gated/unverified. Evidence: devlog/_plan/260710_provider_hardening/003_research_aggregators.md.
   { id: "together", label: "Together", baseUrl: "https://api.together.xyz/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://api.together.xyz/settings/api-keys" },
@@ -1470,10 +1570,10 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelInputModalities: VOLCENGINE_PLAN_INPUT_MODALITIES,
     noVisionModels: VOLCENGINE_PLAN_TEXT_ONLY_MODELS,
     modelReasoningEfforts: Object.fromEntries(
-      DEEPSEEK_THINKING_MODELS.map(id => [id, DEEPSEEK_THINKING_EFFORTS]),
+      DEEPSEEK_THINKING_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)]),
     ),
     modelReasoningEffortMap: Object.fromEntries(
-      DEEPSEEK_THINKING_MODELS.map(id => [id, DEEPSEEK_THINKING_REASONING_MAP]),
+      DEEPSEEK_THINKING_MODELS.map(id => [id, deepseekReasoningMapFor(id)]),
     ),
     preserveReasoningContentModels: DEEPSEEK_THINKING_MODELS,
     note: "Coding tools only. Volcengine restricts Coding Plan quota to supported AI coding tools and warns that using this key for general API calls may suspend the subscription or ban the account. Use the plan key issued by the Ark console.",
@@ -1519,9 +1619,9 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     modelReasoningEfforts: {
       ...Object.fromEntries(ALIBABA_TOKEN_PLAN_QWEN_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
-      "deepseek-v4-pro": DEEPSEEK_THINKING_EFFORTS,
+      "deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek-v4-pro"),
     },
-    modelReasoningEffortMap: { "deepseek-v4-pro": DEEPSEEK_THINKING_REASONING_MAP },
+    modelReasoningEffortMap: { "deepseek-v4-pro": deepseekReasoningMapFor("deepseek-v4-pro") },
     thinkingBudgetModels: ALIBABA_TOKEN_PLAN_QWEN_MODELS,
     preserveReasoningContentModels: ["glm-5.2", "deepseek-v4-pro", "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash"],
     noVisionModels: ["glm-5.2", "deepseek-v4-pro"],
@@ -1553,12 +1653,12 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       ...Object.fromEntries(ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
       "qwen3.8-max": ["low", "high", "xhigh"],
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
-      "deepseek-v4-pro": DEEPSEEK_THINKING_EFFORTS,
-      "deepseek-v4-flash": DEEPSEEK_THINKING_EFFORTS,
+      "deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek-v4-pro"),
+      "deepseek-v4-flash": deepseekThinkingEffortsFor("deepseek-v4-flash"),
     },
     modelReasoningEffortMap: {
-      "deepseek-v4-pro": DEEPSEEK_THINKING_REASONING_MAP,
-      "deepseek-v4-flash": DEEPSEEK_THINKING_REASONING_MAP,
+      "deepseek-v4-pro": deepseekReasoningMapFor("deepseek-v4-pro"),
+      "deepseek-v4-flash": deepseekReasoningMapFor("deepseek-v4-flash"),
     },
     thinkingBudgetModels: ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS,
     preserveReasoningContentModels: ["glm-5.2", "deepseek-v4-pro", "deepseek-v4-flash", "qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash"],
@@ -1649,7 +1749,17 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     autoToolChoiceOnlyModels: KIMI_AUTO_TOOL_CHOICE_ONLY_MODELS,
     preserveReasoningContentModels: KIMI_THINKING_MODELS,
   },
-  { id: "opencode-zen", label: "opencode zen", baseUrl: "https://opencode.ai/zen/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://opencode.ai/auth" },
+  {
+    id: "opencode-zen",
+    label: "opencode zen",
+    baseUrl: "https://opencode.ai/zen/v1",
+    adapter: "openai-chat",
+    authKind: "key",
+    dashboardUrl: "https://opencode.ai/auth",
+    // #1043: without this the proxy forwards image parts to text-only Zen models and
+    // the upstream rejects the whole request with a 400.
+    noVisionModels: OPENCODE_ZEN_TEXT_ONLY_MODELS,
+  },
   { id: "vercel-ai-gateway", label: "Vercel AI Gateway", baseUrl: "https://ai-gateway.vercel.sh/v1", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://vercel.com/dashboard" },
   {
     id: "opencode-free",
@@ -1665,10 +1775,12 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     staticHeaders: {
       "x-opencode-client": "desktop",
     },
-    modelReasoningEfforts: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, DEEPSEEK_THINKING_EFFORTS])),
-    modelReasoningEffortMap: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, DEEPSEEK_THINKING_REASONING_MAP])),
+    modelReasoningEfforts: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, deepseekThinkingEffortsFor(id)])),
+    modelReasoningEffortMap: Object.fromEntries(OPENCODE_FREE_DEEPSEEK_MODELS.map(id => [id, deepseekReasoningMapFor(id)])),
     preserveReasoningContentModels: OPENCODE_FREE_DEEPSEEK_MODELS,
-    noVisionModels: OPENCODE_FREE_DEEPSEEK_MODELS,
+    // Same Zen roster behind the same base URL, so it carries the same measured
+    // text-only list rather than only its DeepSeek member (#1043).
+    noVisionModels: OPENCODE_ZEN_TEXT_ONLY_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" },
   { id: "kilo", label: "Kilo", baseUrl: "https://api.kilo.ai/api/gateway", adapter: "openai-chat", authKind: "key", dashboardUrl: "https://kilo.ai" },

package/src/routing/capability.ts

@@ -111,6 +111,7 @@ const TOOL_CAPABLE_ADAPTERS = new Set([
   "azure",
   "kiro",
   "mimo-free",
+  "command-code",
 ]);
 
 function localRemoteEvidence(baseUrl: string | undefined): Pick<RouteCapabilityEvidence, "localOnly" | "remoteAllowed"> {

package/src/server/adapter-resolve.ts

@@ -5,6 +5,7 @@ import { createGoogleAdapter } from "../adapters/google";
 import { createKiroAdapter } from "../adapters/kiro";
 import { createMimoFreeAdapter } from "../adapters/mimo-free";
 import { createOpenAIChatAdapter } from "../adapters/openai-chat";
+import { createCommandCodeAdapter } from "../adapters/command-code";
 import { createResponsesPassthroughAdapter } from "../adapters/openai-responses";
 import type { OcxProviderConfig } from "../types";
 import { isWirePinnedModel, MODEL_ADAPTER_OVERRIDE_ALLOWED, pinnedWireAdapter } from "../types";
@@ -57,6 +58,8 @@ export function resolveWireProtocolOverride(
 /** Build the provider adapter for a resolved provider config. */
 export function resolveAdapter(providerConfig: OcxProviderConfig, cacheRetention?: "none" | "short" | "long") {
   switch (providerConfig.adapter) {
+    case "command-code":
+      return createCommandCodeAdapter(providerConfig);
     case "openai-chat":
       return createOpenAIChatAdapter(providerConfig);
     case "anthropic":

package/src/server/auth-cors.ts

@@ -15,10 +15,11 @@ import {
 } from "../config";
 import { providerDestinationConfigError } from "../lib/destination-policy";
 import { redactSecretString } from "../lib/redact";
-import { getProviderRegistryEntry, providerCodexAccountMode, providerMatchesRegistryTransport, registryEntryForProviderDestination } from "../providers/registry";
+import { effectiveGoogleMode, getProviderRegistryEntry, providerCodexAccountMode, providerMatchesRegistryTransport, registryEntryForProviderDestination } from "../providers/registry";
 import { providerConfigSeed } from "../providers/derive";
 import type { OcxConfig, OcxProviderConfig } from "../types";
 import { openRouterRoutingConfigError } from "../providers/openrouter-routing";
+import { googleVertexLocationConfigError } from "../providers/google-vertex-location";
 
 let _corsOrigin = "http://localhost:10100";
 export function setCorsOrigin(port: number): void { _corsOrigin = `http://localhost:${port}`; }
@@ -425,6 +426,10 @@ export function providerManagementConfigError(name: unknown, provider: unknown):
   const typed = provider as unknown as OcxProviderConfig;
   const baseUrlError = providerBaseUrlConfigError(typed.baseUrl);
   if (baseUrlError) return `provider ${name} ${baseUrlError}`;
+  if (effectiveGoogleMode(name, typed) === "vertex" && typed.location !== undefined) {
+    const locationError = googleVertexLocationConfigError(typed.location);
+    if (locationError) return `provider ${name} ${locationError}`;
+  }
   const destinationError = providerDestinationConfigError(name, typed);
   if (destinationError) return `provider ${name} ${destinationError}`;
   const headersError = providerHeadersConfigError(typed.headers);

package/src/server/index.ts

@@ -174,6 +174,12 @@ import {
   requireManagementAuth,
   type ManagementAuthState,
 } from "./management-auth";
+import {
+  LOCAL_ATTESTATION_CHALLENGE_HEADER,
+  LOCAL_ATTESTATION_PROOF_HEADER,
+  createLocalAttestationProof,
+  createLocalAttestationSecret,
+} from "../lib/local-management-attestation";
 
 const MAX_WS_FRAME_BYTES = 50 * 1024 * 1024;
 const WEBSOCKET_IDLE_TIMEOUT_SECONDS = 0;
@@ -357,9 +363,33 @@ export interface StartServerDeps {
   nativeMainStartup?: NativeMainStartupGateDeps;
   /** Test-only seam for an upstream that cannot complete its WebSocket close handshake. */
   liveSidebandWebSocketFactory?: LiveSidebandWebSocketFactory;
+  /** Test-only seam; production derives a fresh local-attestation secret per process. */
+  localAttestationSecret?: string;
+}
+
+/*
+ * #1046. `startServer` rewrites the Codex models cache during boot, and an
+ * app-server that started earlier keeps its own in-memory model list. The stale
+ * warning is not emitted here: `handleStart` runs a catalog sync moments later,
+ * so warning now would read an mtime that write is about to move, and both sites
+ * calling the helper independently would warn twice. This records the fact; the
+ * CLI start path owns the single decision.
+ *
+ * A caller that starts a server without `handleStart` (tests, embedded use)
+ * deliberately gets no warning — lifecycle diagnostics belong to whoever owns
+ * the lifecycle.
+ */
+let startupCacheInvalidationWrote = false;
+
+/** #1046: did this process's startup cache invalidation actually write? */
+export function consumeStartupCacheInvalidationWrite(): boolean {
+  const wrote = startupCacheInvalidationWrote;
+  startupCacheInvalidationWrote = false;
+  return wrote;
 }
 
 export function startServer(port?: number, deps: StartServerDeps = {}) {
+  const localAttestationSecret = deps.localAttestationSecret ?? createLocalAttestationSecret();
   const config = runAlibabaRegionStartupMigration(runOpenAiTierStartupMigration(loadConfig()));
   setLiveStateStoreConfig(config);
   applyProxyEnv(config);
@@ -407,8 +437,13 @@ export function startServer(port?: number, deps: StartServerDeps = {}) {
   // otherwise turn "no Codex installed" into "proxy will not start".
   try {
     const startupCodexHome = getCodexHome();
-    withCatalogWriteSerialization(startupCodexHome, permit =>
+    // #1046: record whether this actually rewrote the cache. `handleStart` ORs this
+    // with the later startup sync and warns ONCE about stale app-servers; warning
+    // here instead would read a catalog mtime the sync is about to move.
+    const outcome = withCatalogWriteSerialization(startupCodexHome, permit =>
       invalidateCodexModelsCacheWithPermit(permit, startupCodexHome));
+    // A refused permit is not a write; only a completed run that returned true is.
+    startupCacheInvalidationWrote = outcome.kind === "completed" && outcome.value === true;
   } catch { /* no readable Codex home: nothing to invalidate */ }
   // Arm the `claudeCode` hand-edit guard (devlog 260726_claude_auth_auto/040 H1) BEFORE
   // the server can serve a request, and AFTER the startup migrations above — those run
@@ -545,7 +580,14 @@ export function startServer(port?: number, deps: StartServerDeps = {}) {
 
       if (url.pathname === "/healthz" && req.method === "GET") {
         // service/pid/port let CLI liveness reject foreign 200s and verify pid identity.
-        return jsonResponse({ status: "ok", service: "opencodex", version: VERSION, uptime: process.uptime(), pid: process.pid, port: listenPort }, 200, req, config);
+        const healthPort = server.port ?? listenPort;
+        const response = jsonResponse({ status: "ok", service: "opencodex", version: VERSION, uptime: process.uptime(), pid: process.pid, port: healthPort }, 200, req, config);
+        const challenge = req.headers.get(LOCAL_ATTESTATION_CHALLENGE_HEADER);
+        if (challenge) {
+          const proof = createLocalAttestationProof(localAttestationSecret, challenge, process.pid, healthPort);
+          if (proof) response.headers.set(LOCAL_ATTESTATION_PROOF_HEADER, proof);
+        }
+        return response;
       }
 
       if (url.pathname.startsWith("/api/")) {

package/src/server/management/integration-routes.ts

@@ -6,7 +6,7 @@
  * belongs to src/integrations/writer.ts. Duplicating any of it here is how the
  * API and the writer would start disagreeing about what happened to a file.
  *
- * Design of record: devlog/_plan/260802_client_toggle_api/040_wp4_management_api.md.
+ * Design of record: devlog/_fin/260802_client_toggle_api/040_wp4_management_api.md.
  */
 import { readFileSync } from "node:fs";
 import type { IntegrationIO } from "../../integrations/config-io";

package/src/server/management/native-integration-routes.ts

@@ -9,12 +9,12 @@
  * route — turning them back on is the undo.
  *
  * That conclusion cost eleven audit rounds; the reasoning is in
- * devlog/_plan/260803_integrations_toggle_all/, and 007 records why Codex and
+ * devlog/_fin/260803_integrations_toggle_all/, and 007 records why Codex and
  * Claude Desktop are NOT here: their state spans several artifacts and a live
  * database, so they need a durable operation record this module deliberately
  * does not have.
  *
- * Design of record: devlog/_plan/260803_integrations_toggle_all/030 (routes),
+ * Design of record: devlog/_fin/260803_integrations_toggle_all/030 (routes),
  * 011 (Claude Code), 012 (Grok).
  */
 import { readRuntimePort, saveConfigPreservingClaudeCode } from "../../config";

package/src/server/responses/core.ts

@@ -747,10 +747,18 @@ const UNREADABLE_ENCRYPTED_AGENT_TASK_MESSAGE =
 // branch of the passthrough return path). 32 MiB matches the continuation snapshot read
 // bound and is far above any legitimate non-streaming completion, including base64 image
 // payloads. The stall deadlines only govern the body transfer — generation time before
-// the response headers is untouched.
+// the response headers is untouched. Generation after early/chunked headers but before
+// the first body byte previously used the 30-second inactivity deadline; this call site
+// gives it the full body deadline instead.
 const MAX_UPSTREAM_JSON_BODY_BYTES = 32 * 1024 * 1024;
 const UPSTREAM_JSON_BODY_TOTAL_TIMEOUT_MS = 180_000;
 const UPSTREAM_JSON_BODY_INACTIVITY_TIMEOUT_MS = 30_000;
+export const UPSTREAM_JSON_BODY_READ_OPTIONS = {
+  maxBytes: MAX_UPSTREAM_JSON_BODY_BYTES,
+  totalTimeoutMs: UPSTREAM_JSON_BODY_TOTAL_TIMEOUT_MS,
+  inactivityTimeoutMs: UPSTREAM_JSON_BODY_INACTIVITY_TIMEOUT_MS,
+  firstByteTimeoutMs: UPSTREAM_JSON_BODY_TOTAL_TIMEOUT_MS,
+};
 
 function unreadableEncryptedAgentTaskResponse(): Response {
   return new Response(
@@ -1375,8 +1383,8 @@ async function handleResponsesInner(
   logCtx.configuredServiceTier = readConfiguredCodexServiceTier();
   logCtx.configuredSpeedLabel = requestLogSpeedLabel(logCtx.configuredServiceTier);
 
-  // Shadow call intercept: rewrite Codex's hard-coded helper calls
-  // (gpt-5.4-mini on older clients, gpt-5.6-luna on 0.145.0+)
+  // Shadow call intercept: rewrite Codex 0.145.0+ helper calls (gpt-5.6-luna).
+  // Ancient clients using gpt-5.4-mini remain configurable via sourceModels.
   const _sci = config.shadowCallIntercept;
   if (_sci?.enabled && _sci.model && shouldInterceptShadowCall(
     parsed.modelId,
@@ -2039,7 +2047,7 @@ async function handleResponsesInner(
     // `streamMode: "eager-relay"` opt-in. Darwin `auto` always stays tee. The
     // eager shape skips tee and uses one bounded reader with inline inspection
     // (src/server/relay-eager.ts; policy:
-    // devlog/_plan/260731_macos_rss_retention/100_darwin_eager_optin.md).
+    // devlog/_fin/260731_macos_rss_retention/100_darwin_eager_optin.md).
     // The bundled known-bad runtime remains on tee by default on both platforms.
     if (isEventStream && upstreamResponse.body) {
       const repairConfig = route.provider.responsesItemIdRepair;
@@ -2226,11 +2234,7 @@ async function handleResponsesInner(
       // without limit. This path is no longer rare — WebSocket turns for models whose
       // streaming terminal event is unreliable are deliberately answered with bounded JSON.
       // Oversize and stall deadlines both fail closed; a partial body is never parsed.
-      const bounded = await readBoundedResponseBody(upstreamResponse, {
-        maxBytes: MAX_UPSTREAM_JSON_BODY_BYTES,
-        totalTimeoutMs: UPSTREAM_JSON_BODY_TOTAL_TIMEOUT_MS,
-        inactivityTimeoutMs: UPSTREAM_JSON_BODY_INACTIVITY_TIMEOUT_MS,
-      });
+      const bounded = await readBoundedResponseBody(upstreamResponse, UPSTREAM_JSON_BODY_READ_OPTIONS);
       if (bounded.oversized) {
         return formatErrorResponse(502, "upstream_error", "upstream JSON response exceeded the safe body limit");
       }

package/src/storage/scanner.ts

@@ -16,7 +16,7 @@ const IMMUTABLE_READONLY_FLAGS = constants.SQLITE_OPEN_READONLY | constants.SQLI
 
 /**
  * Read-only CODEX_HOME storage scanner — Phase 1 of the Storage page epic
- * (devlog/_plan/500_storage-page-session-cleanup). Pure measurement: sizes via
+ * (devlog/_fin/500_storage-page-session-cleanup). Pure measurement: sizes via
  * fs.stat walks, DB row counts via immutable readonly opens that degrade to
  * null on lock/corruption. Performs zero writes under CODEX_HOME.
  */

package/src/types.ts

@@ -1066,6 +1066,12 @@ export interface OcxProviderConfig {
    * the legacy `/v1/responses` construction.
    */
   responsesPath?: string;
+  /**
+   * Command Code protocol version sent as `x-command-code-version` on /alpha/generate requests.
+   * The internal endpoint's schema drifts with the CLI version; operators can pin a known-good
+   * version here instead of waiting for a code change. Absent uses the adapter's current default.
+   */
+  commandCodeVersion?: string;
   /**
    * Responses upstream that stores nothing server-side (DeepSeek documents "the API
    * is stateless"). Stateful request parameters are dropped, `store` is pinned false,

package/src/usage/log.ts

@@ -421,7 +421,7 @@ export type UsageLogRevision = {
 let usageReadCacheStats = { fullReads: 0, tailReads: 0, parsedLines: 0 };
 const MANAGEMENT_USAGE_MAX_READ_BYTES = 64 * 1024 * 1024;
 const MANAGEMENT_USAGE_READ_CHUNK_BYTES = 1024 * 1024;
-const MANAGEMENT_USAGE_MAX_ENTRIES = 200_000;
+const MANAGEMENT_USAGE_MAX_ENTRIES = 500_000;
 const MANAGEMENT_USAGE_FLIGHT_STALE_MS = 30_000;
 export interface ManagementUsageSnapshot {
   entries: PersistedUsageEntry[];