@yansigit/opencodex 2.32.0 → 2.33.0

package/src/bridge.ts

@@ -19,6 +19,7 @@ import {
   awaitThoughtSignatureDurability,
 } from "./responses/thought-signature-replay";
 import { resolveStallTimeoutSec } from "./stall-timeout";
+import { normalizeDeclaredToolName } from "./types";
 import { usageDisplayTotalTokens } from "./usage/totals";
 import { appendSafeWebSearchSource, safeWebSearchSources } from "./web-search/sources";
 import {
@@ -28,6 +29,7 @@ import {
   type TranslatorBudget,
   type TranslatorBufferKind,
 } from "./lib/translator-budget";
+import { debugFingerprint, debugStreamDiagnostic, type DebugStreamDiagnosticContext } from "./lib/debug";
 
 function uuid(): string {
   return crypto.randomUUID().replace(/-/g, "");
@@ -118,7 +120,9 @@ function adapterFailureFromEvent(event: Extract<AdapterEvent, { type: "error" }>
   }
   const fallback = adapterFailureFromMessage(event.message);
   let httpStatus = event.status ?? fallback.httpStatus;
-  const error = classifyError(httpStatus, event.errorType ?? fallback.error.type, event.message);
+  const error = httpStatus === 400 && event.errorType === "upstream_error"
+    ? { message: event.message, type: "upstream_error", code: event.code ?? null }
+    : classifyError(httpStatus, event.errorType ?? fallback.error.type, event.message);
   if (event.errorType !== undefined) error.type = event.errorType;
   if (event.code !== undefined) error.code = event.code;
   // Codex maps cyber_policy on HTTP 400 (body) or mid-stream code; never leave it as 502.
@@ -166,6 +170,88 @@ interface OutputItem {
 
 export type ResponsesTerminalStatus = "completed" | "failed" | "incomplete";
 
+export interface BridgeDiagnosticSequence { value: number }
+
+export interface BridgeDiagnosticContext extends DebugStreamDiagnosticContext {
+  sequence?: BridgeDiagnosticSequence;
+}
+
+export function adapterEventDiagnosticDetails(event: AdapterEvent): Record<string, unknown> {
+  switch (event.type) {
+    case "text_delta":
+      return { byteLength: Buffer.byteLength(event.text), fingerprint: debugFingerprint(event.text) };
+    case "thinking_delta":
+      return { byteLength: Buffer.byteLength(event.thinking), fingerprint: debugFingerprint(event.thinking) };
+    case "reasoning_raw_delta":
+      return { byteLength: Buffer.byteLength(event.text), fingerprint: debugFingerprint(event.text) };
+    case "thinking_signature":
+    case "redacted_thinking":
+    case "kiro_redacted_reasoning": {
+      const content = event.type === "thinking_signature" ? event.signature : event.data;
+      return { byteLength: Buffer.byteLength(content), fingerprint: debugFingerprint(content) };
+    }
+    case "tool_call_delta":
+      return { byteLength: Buffer.byteLength(event.arguments), fingerprint: debugFingerprint(event.arguments) };
+    case "tool_call_start":
+      return {
+        idByteLength: Buffer.byteLength(event.id),
+        idFingerprint: debugFingerprint(event.id),
+        nameByteLength: Buffer.byteLength(event.name),
+        nameFingerprint: debugFingerprint(event.name),
+      };
+    case "web_search_call_begin":
+      return { idByteLength: Buffer.byteLength(event.id), idFingerprint: debugFingerprint(event.id) };
+    case "web_search_call_end": {
+      const queries = JSON.stringify(event.queries);
+      return {
+        idByteLength: Buffer.byteLength(event.id),
+        idFingerprint: debugFingerprint(event.id),
+        byteLength: Buffer.byteLength(queries),
+        fingerprint: debugFingerprint(queries),
+        status: event.status,
+      };
+    }
+    case "error":
+      return {
+        byteLength: Buffer.byteLength(event.message),
+        fingerprint: debugFingerprint(event.message),
+        ...(event.status !== undefined ? { status: event.status } : {}),
+        ...(event.code !== undefined
+          ? { codeByteLength: Buffer.byteLength(event.code), codeFingerprint: debugFingerprint(event.code) }
+          : {}),
+        ...(event.retryable !== undefined ? { retryable: event.retryable } : {}),
+      };
+    case "incomplete":
+      return {
+        ...(event.message !== undefined ? { byteLength: Buffer.byteLength(event.message), fingerprint: debugFingerprint(event.message) } : {}),
+        reasonByteLength: Buffer.byteLength(event.reason),
+        reasonFingerprint: debugFingerprint(event.reason),
+        ...(event.retryable !== undefined ? { retryable: event.retryable } : {}),
+      };
+    case "done":
+      return {
+        ...(event.stopReason !== undefined
+          ? { stopReasonByteLength: Buffer.byteLength(event.stopReason), stopReasonFingerprint: debugFingerprint(event.stopReason) }
+          : {}),
+        ...(event.endTurn !== undefined ? { endTurn: event.endTurn } : {}),
+      };
+    default:
+      return {};
+  }
+}
+
+/** Emit one adapter-stage diagnostic while preserving one sequence across sidecar iterations. */
+export function diagnoseAdapterEvent(context: BridgeDiagnosticContext, event: AdapterEvent): void {
+  const sequence = context.sequence ??= { value: 0 };
+  debugStreamDiagnostic(
+    context,
+    "adapter",
+    ++sequence.value,
+    event.type,
+    adapterEventDiagnosticDetails(event),
+  );
+}
+
 export function bridgeToResponsesSSE(
   events: AsyncIterable<AdapterEvent>,
   modelId: string,
@@ -226,6 +312,8 @@ export function bridgeToResponsesSSE(
       setInterval: (handler: () => void, ms: number) => unknown;
       clearInterval: (id: unknown) => void;
     };
+    /** Internal, opt-in structural stream diagnostics. */
+    diagnostic?: BridgeDiagnosticContext;
   },
 ): ReadableStream<Uint8Array> {
   const replayCacheScope = options?.replayCacheScope;
@@ -326,6 +414,7 @@ export function bridgeToResponsesSSE(
   };
   const responseId = options?.responseId ?? `resp_${uuid()}`;
   let seq = 0;
+  let diagnosticSequence = 0;
   // Set once the client is gone (cancel) or an enqueue throws on a torn-down controller, so we
   // never enqueue again and never throw a second time inside start() — the RC2 double-throw that
   // otherwise surfaced as proxy-side stream noise on every client disconnect.
@@ -858,6 +947,17 @@ export function bridgeToResponsesSSE(
           }
           if (next.done) { upstreamDone = true; break; }
           const event = next.value;
+          if (options?.diagnostic) {
+            debugStreamDiagnostic(
+              options.diagnostic,
+              "bridge",
+              options.diagnostic.sequence
+                ? ++options.diagnostic.sequence.value
+                : ++diagnosticSequence,
+              event.type,
+              adapterEventDiagnosticDetails(event),
+            );
+          }
           let terminalEvent = false;
           // Invisible adapter heartbeats (and buffered web-search progress) count as upstream
           // liveness only — they must not suppress wire keepalives that re-arm Codex idle timers.
@@ -1041,13 +1141,14 @@ export function bridgeToResponsesSSE(
                 rememberReasoningForCall(event.id, rawReasoningForNextToolCall, replayCacheScope);
               }
               if (currentToolCall) closeCurrentToolCall();
-              const mapped = toolNsMap?.get(event.name);
-              const realName = mapped?.name ?? event.name;
-              if (options?.declaredToolNames && !options.declaredToolNames.has(event.name)) {
+              const effectiveName = normalizeDeclaredToolName(event.name, options?.declaredToolNames);
+              const mapped = toolNsMap?.get(effectiveName);
+              const realName = mapped?.name ?? effectiveName;
+              if (options?.declaredToolNames && !options.declaredToolNames.has(effectiveName)) {
                 const failure = responseError(
                   502,
                   "upstream_error",
-                  `routed provider emitted undeclared client tool "${event.name}"; only request-declared tools may be called`,
+                  `routed provider emitted undeclared client tool "${effectiveName}"; only request-declared tools may be called`,
                 );
                 emit("response.failed", {
                   response: {
@@ -1783,7 +1884,7 @@ function buildResponseJSONWithBudget(
           ));
         }
         break;
-      case "tool_call_start":
+      case "tool_call_start": {
         if (currentText) flushText("commentary");
         if (currentSummaryReasoning) flushSummaryReasoning();
         if (currentRawReasoning) flushRawReasoning();
@@ -1791,10 +1892,11 @@ function buildResponseJSONWithBudget(
           rememberReasoningForCall(e.id, rawReasoningForNextToolCall, replayCacheScope);
         }
         flushToolCall();
-        if (options?.declaredToolNames && !options.declaredToolNames.has(e.name)) {
+        const effectiveName = normalizeDeclaredToolName(e.name, options?.declaredToolNames);
+        if (options?.declaredToolNames && !options.declaredToolNames.has(effectiveName)) {
           errorEvent = {
             type: "error",
-            message: `routed provider emitted undeclared client tool "${e.name}"; only request-declared tools may be called`,
+            message: `routed provider emitted undeclared client tool "${effectiveName}"; only request-declared tools may be called`,
             status: 502,
             errorType: "upstream_error",
           };
@@ -1802,11 +1904,12 @@ function buildResponseJSONWithBudget(
         }
         currentToolCallId = e.id;
         budget?.openCall(e.id);
-        currentToolCallName = e.name;
+        currentToolCallName = effectiveName;
         currentToolCallArgs = "";
         currentToolCallArgsBytes = 0;
         currentToolCallProviderMetadata = e.providerMetadata;
         break;
+      }
       case "tool_call_delta":
         {
           ({ value: currentToolCallArgs, bytes: currentToolCallArgsBytes } = appendBatchString(

package/src/claude/context-windows.ts

@@ -118,17 +118,24 @@ export function buildClaudeContextWindows(
     put(desktop3pAlias("native", slug), window);
     put(aliasForNative(slug), window);
   }
+  // Anthropic passthrough guard (audit 021 #3): canonical claude ids ride the
+  // subscription passthrough — marking a sub-1M one would strap [1m]/1M-beta onto
+  // a model that cannot host it. Register anthropic rows only at >=1M.
+  const registrable = routedModels.filter(
+    m =>
+      typeof m.contextWindow === "number" &&
+      m.contextWindow > 0 &&
+      !(m.provider === "anthropic" && m.contextWindow < ONE_MILLION),
+  );
   // Bare routed ids are registered only when unambiguous across providers (audit
-  // 021 #5) — natives are registered first, so a native slug always wins the bare key.
+  // 021 #5) — natives are registered first, so a native slug always wins the bare
+  // key. Counted over the rows that can actually claim the key: a row this loop
+  // skips contributes no window, so letting it veto the bare key withholds an
+  // answer that was never in doubt.
   const bareCounts = new Map<string, number>();
-  for (const m of routedModels) bareCounts.set(m.id, (bareCounts.get(m.id) ?? 0) + 1);
-  for (const m of routedModels) {
-    const window = m.contextWindow;
-    if (typeof window !== "number" || window <= 0) continue;
-    // Anthropic passthrough guard (audit 021 #3): canonical claude ids ride the
-    // subscription passthrough — marking a sub-1M one would strap [1m]/1M-beta onto
-    // a model that cannot host it. Register anthropic rows only at >=1M.
-    if (m.provider === "anthropic" && window < ONE_MILLION) continue;
+  for (const m of registrable) bareCounts.set(m.id, (bareCounts.get(m.id) ?? 0) + 1);
+  for (const m of registrable) {
+    const window = m.contextWindow as number;
     put(`${m.provider}/${m.id}`, window);
     put(desktop3pAlias(m.provider, m.id), window);
     put(aliasForRoute(m.provider, m.id), window);

package/src/cli/doctor.ts

@@ -42,7 +42,7 @@ import {
   resolveEffectiveUserIdentity,
 } from "../codex/user-identity";
 import { collectProjectCodexConfigWarnings, formatProjectCodexConfigWarningsForDoctor } from "../codex/project-config-warnings";
-import { collectStartupHealth, startupHealthSummary } from "../codex/autostart-health";
+import { collectStartupHealth, formatStartupRoutingDetail, startupHealthSummary } from "../codex/autostart-health";
 import {
   displayCodexRuntimePath,
   loadLastEffortClamp,
@@ -983,7 +983,7 @@ export async function runDoctor(args: string[] = []): Promise<void> {
   const startup = collectStartupHealth(doctorConfig);
   console.log("\nCodex restart safety");
   console.log(`  ${startup.rebootSafe ? "ok " : "!! "} ${startupHealthSummary(startup)}`);
-  console.log(`       routing=${startup.routingKind}, service=${startup.serviceViable ? "viable" : startup.serviceInstalled ? "installed-but-unhealthy" : "absent"}, shim=${startup.shimHealthy ? "healthy" : startup.shimInstalled ? "stale" : "absent"}`);
+  console.log(`       ${formatStartupRoutingDetail(startup)}`);
 
   console.log("\nCodex runtime selection");
   {

package/src/cli/index.ts

@@ -25,7 +25,7 @@ import {
   writePid,
   writeRuntimePort,
 } from "../config/process-state";
-import { collectStatus } from "./status";
+import { collectStatus, unusedProxyWarningLines } from "./status";
 
 import {
   discoverStableProxyForRestart,
@@ -46,7 +46,7 @@ import { runCli } from "./root";
 import { ProxyOwnershipRefusedError, stopProxy } from "../lib/process-control";
 import { loadServiceTokenFromFile } from "../lib/service-secrets";
 import { diagnoseService, isServiceOwnershipError, serviceCommand, serviceEnvironmentOwnedHere, serviceStartableFromTray, serviceStatusSummary, stopServiceIfInstalled, uninstallServiceIfInstalled } from "../service";
-import { startupHealthSummary } from "../codex/autostart-health";
+import { formatStartupRoutingDetail, startupHealthSummary } from "../codex/autostart-health";
 import { drainAndShutdown, isRecyclingForExit, startServer } from "../server";
 import { injectSystemEnv, reconcileShellHook, revertSystemEnv, uninstallShellHook } from "../server/system-env";
 import { buildDesktop3pRegistry } from "../claude/desktop-3p";
@@ -849,6 +849,12 @@ async function handleStatus() {
     console.log(`❌ Proxy: ${status.proxyLabel}`);
   }
   console.log(`   Health: ${status.healthLabel}`);
+  for (const line of unusedProxyWarningLines({
+    proxyUp: Boolean(status.json.proxy.pid || status.json.proxy.health.ok),
+    routingKind: status.json.startup.routingKind,
+  })) {
+    console.log(`   ${line}`);
+  }
   if (!(status.json.proxy.pid || status.json.proxy.health.ok)) {
     console.log("   ↳ Not running — Codex/Claude requests will fail with connection errors.");
     // The service summary a few lines below already tells a registered-but-not-serving
@@ -868,6 +874,7 @@ async function handleStatus() {
   console.log(`   Default provider: ${status.json.defaultProvider}`);
   console.log(`   Codex autostart: ${status.json.codexAutostart ? "enabled" : "disabled"}`);
   console.log(`   Restart safety: ${startupHealthSummary(status.json.startup)}`);
+  console.log(`   ${formatStartupRoutingDetail(status.json.startup)}`);
   console.log(`   Service: ${status.json.service.summary}`);
   console.log(`   ${status.json.codexShim.summary}`);
   console.log(`   Codex runtime: ${status.json.codexRuntime.path}`);

package/src/cli/provider.ts

@@ -360,6 +360,12 @@ function handleShow(args: string[]): void {
     ...(prov.modelCosts !== undefined ? { modelCosts: sanitizeModelCostsForDisplay(prov.modelCosts) } : {}),
     ...(prov.apiKey ? { apiKey: maskSecret(prov.apiKey) } : {}),
     ...(prov.apiKeyPool ? { apiKeyPool: prov.apiKeyPool.map(e => ({ ...e, key: maskSecret(e.key) })) } : {}),
+    ...(prov.azureCredential ? {
+      azureCredential: {
+        type: prov.azureCredential.type,
+        hasManagedIdentityClientId: Boolean(prov.azureCredential.managedIdentityClientId?.trim()),
+      },
+    } : {}),
   };
 
   if (wantsJson) {

package/src/cli/status.ts

@@ -118,6 +118,29 @@ export function proxyHealthFailureReason(error: unknown, signal: AbortSignal): "
     : "unreachable";
 }
 
+/**
+ * `ocx status` greens on process liveness alone, so a proxy that answers
+ * /healthz reads healthy even when Codex is not pointed at it and every routed
+ * request goes to OpenAI instead (#2411). The proxy line is not wrong — the
+ * listener really is up — so it keeps its check, and this supplies the signal
+ * that was missing rather than corrupting the one that was already honest.
+ *
+ * Only `native` warns. `custom-local` and `unknown` are also "this proxy is
+ * unused", but startupHealthSummary already renders both as AT RISK with a
+ * remedy command, and `custom-remote` is a deliberate operator choice. Warning
+ * on all four would teach operators to skip the line that matters.
+ */
+export function unusedProxyWarningLines(input: {
+  proxyUp: boolean;
+  routingKind: StartupHealth["routingKind"];
+}): string[] {
+  if (!input.proxyUp || input.routingKind !== "native") return [];
+  return [
+    "⚠️  Codex routing is native — the running proxy is unused.",
+    "   Codex requests go to OpenAI, not this proxy. Re-point with: ocx start",
+  ];
+}
+
 async function checkProxyHealth(target: ListenTarget): Promise<HealthCheck> {
   const url = target.healthUrl;
   const controller = new AbortController();

package/src/codex/auth-api.ts

@@ -1806,7 +1806,7 @@ export async function handleCodexAuthAPI(
   }
 
   if (url.pathname === "/api/codex-auth/login" && req.method === "POST") {
-    const body = (await req.json().catch(() => ({}))) as { id?: string; reauth?: boolean };
+    const body = (await req.json().catch(() => ({}))) as { id?: string; reauth?: boolean; openBrowser?: unknown };
     const requestedAccountId = body.id?.trim();
     const reauth = body.reauth === true;
     if (requestedAccountId && !isValidCodexAccountId(requestedAccountId)) {
@@ -1840,7 +1840,9 @@ export async function handleCodexAuthAPI(
 
       // Open the browser server-side (same pattern as /api/oauth/login in management-api.ts).
       // The GUI's window.open is popup-blocked because it runs after an await, not a direct click.
-      if (result.url) {
+      // Both login routes share one resolver so this surface cannot drift from the other.
+      const { shouldOpenBrowserForLogin } = await import("../oauth/open-browser-choice");
+      if (result.url && shouldOpenBrowserForLogin(body.openBrowser, runtimeConfig)) {
         const { openUrl } = await import("../lib/open-url");
         openUrl(result.url);
       }

package/src/codex/autostart-health.ts

@@ -154,3 +154,19 @@ export function startupHealthSummary(health: StartupHealth): string {
   if (health.serviceInstalled && !health.serviceViable) return `AT RISK after restart (installed service is disabled, stopped, or unhealthy; run '${command}')`;
   return `AT RISK after restart (no viable background service; run '${command}')`;
 }
+
+/**
+ * The routing/service/shim token `ocx doctor` prints under restart safety.
+ * Extracted so `ocx status` can show the same string rather than growing a
+ * second copy that drifts (#2411). Two management routes computing the same
+ * thing separately is exactly how #2457 happened.
+ */
+export function formatStartupRoutingDetail(health: StartupHealth): string {
+  const service = health.serviceViable
+    ? "viable"
+    : health.serviceInstalled ? "installed-but-unhealthy" : "absent";
+  const shim = health.shimHealthy
+    ? "healthy"
+    : health.shimInstalled ? "stale" : "absent";
+  return `routing=${health.routingKind}, service=${service}, shim=${shim}`;
+}

package/src/codex/catalog/aggregation.ts

@@ -13,6 +13,7 @@ import { getModelMetadata, getModelMetadataCaseInsensitive, listModelMetadata, r
 import { enrichProviderFromRegistry, shouldCaseFoldMetadataModelId } from "../../providers/derive";
 import { getProviderRegistryEntry } from "../../providers/registry";
 import { applyProviderContextCap, providerContextCap } from "../../providers/context-cap";
+import { clampAutoCompactTokenLimit } from "../../providers/auto-compact-budget";
 import { routedSlug, slugEquals, slugEquivalenceKey, slugsEquivalent } from "../../providers/slug-codec";
 import { CODEX_GPT5_IDENTITY_LINE } from "../../adapters/identity";
 import { filterCursorConfiguredModelsByLiveDiscovery } from "../../adapters/cursor/discovery";
@@ -152,19 +153,17 @@ export function deriveComboCatalogModel(
   // A combo is cap-limited only when every member defining its effective minimum was
   // itself reduced by a provider cap. An uncapped member at the same minimum means the
   // combo would have the same window even without the cap.
-  const contextCapped = limitingMembers.every(member => (
-    member.contextCapped === true
-    || (
-      typeof member.detectedContextWindow === "number"
-      && typeof member.contextWindow === "number"
-      && member.detectedContextWindow > member.contextWindow
-    )
-  ));
+  const contextCapped = limitingMembers.every(member => member.contextCapped === true);
   const maxInputTokens = Math.min(
+    contextWindow,
     ...members.map(member => member.maxInputTokens ?? member.contextWindow!),
   );
-  const detectedContextWindow = Math.min(
-    ...limitingMembers.map(member => member.detectedContextWindow ?? member.contextWindow!),
+  const autoCompactTokenLimit = Math.min(
+    ...members.map(member => clampAutoCompactTokenLimit(
+      member.contextWindow!,
+      member.maxInputTokens,
+      member.autoCompactTokenLimit,
+    )),
   );
   const defaultReasoningEffort = effectiveComboDefault(
     combo.defaultEffort,
@@ -177,8 +176,7 @@ export function deriveComboCatalogModel(
     owned_by: COMBO_NAMESPACE,
     contextWindow,
     maxInputTokens,
-    metadataSource: "derived",
-    detectedContextWindow,
+    autoCompactTokenLimit,
     ...(hasLimitingContextCapMetadata ? { contextCapped } : {}),
     inputModalities,
     reasoningEfforts,
@@ -222,6 +220,7 @@ export function comboCatalogWarningSignature(
       key,
       contextWindow: member?.contextWindow ?? null,
       maxInputTokens: member?.maxInputTokens ?? null,
+      autoCompactTokenLimit: member?.autoCompactTokenLimit ?? null,
       inputModalities: [...new Set(member?.inputModalities ?? [])].sort(),
       reasoningEfforts: [...new Set(member?.reasoningEfforts ?? [])].sort(),
       parallelToolCalls: member?.parallelToolCalls === true,
@@ -311,6 +310,7 @@ export function normalizedOpenAiApiSignature(model: CatalogModel): string {
     id: model.id,
     contextWindow: model.contextWindow ?? null,
     maxInputTokens: model.maxInputTokens ?? null,
+    autoCompactTokenLimit: model.autoCompactTokenLimit ?? null,
     inputModalities: [...new Set(model.inputModalities ?? [])].sort(),
     reasoningEfforts: [...new Set(model.reasoningEfforts ?? [])].sort(),
     ownedBy: model.owned_by ?? null,

package/src/codex/catalog/effort.ts

@@ -13,6 +13,7 @@ import { getModelMetadata, getModelMetadataCaseInsensitive, listModelMetadata, r
 import { enrichProviderFromRegistry, shouldCaseFoldMetadataModelId } from "../../providers/derive";
 import { getProviderRegistryEntry } from "../../providers/registry";
 import { applyProviderContextCap, providerContextCap, resolveUnknownRoutedContextWindow } from "../../providers/context-cap";
+import { clampAutoCompactTokenLimit } from "../../providers/auto-compact-budget";
 import { routedSlug, slugEquals, slugsEquivalent } from "../../providers/slug-codec";
 import { CODEX_GPT5_IDENTITY_LINE } from "../../adapters/identity";
 import { filterCursorConfiguredModelsByLiveDiscovery } from "../../adapters/cursor/discovery";
@@ -129,9 +130,23 @@ export function applyCatalogModelMetadata(entry: RawEntry, model?: CatalogModel)
   if (typeof resolvedContext === "number" && resolvedContext > 0) {
     entry.context_window = resolvedContext;
     entry.max_context_window = resolvedContext;
-    entry.auto_compact_token_limit = Math.min(
-      Math.floor(resolvedContext * 0.9),
-      model.maxInputTokens ?? Number.POSITIVE_INFINITY,
+    entry.auto_compact_token_limit = clampAutoCompactTokenLimit(
+      resolvedContext,
+      model.maxInputTokens,
+      model.autoCompactTokenLimit,
+    );
+  } else if (
+    typeof entry.context_window === "number"
+    && entry.context_window > 0
+    && typeof model.maxInputTokens === "number"
+    && model.maxInputTokens > 0
+  ) {
+    // A conservative routed fallback is not evidence for applying the optional soft policy,
+    // but a measured/configured input ceiling is still a hard bound. Compact before that
+    // ceiling even when the provider supplied no authoritative context window.
+    entry.auto_compact_token_limit = clampAutoCompactTokenLimit(
+      entry.context_window,
+      model.maxInputTokens,
     );
   }
   if (Array.isArray(model.inputModalities) && model.inputModalities.length > 0) {

package/src/codex/catalog/metadata.ts

@@ -14,6 +14,7 @@ import { getModelMetadata, getModelMetadataCaseInsensitive, listModelMetadata, r
 import { enrichProviderFromRegistry, shouldCaseFoldMetadataModelId } from "../../providers/derive";
 import { getProviderRegistryEntry, providerCodexAccountMode } from "../../providers/registry";
 import { applyProviderContextCap, providerContextCap } from "../../providers/context-cap";
+import { clampAutoCompactTokenLimit } from "../../providers/auto-compact-budget";
 import { routedSlug, slugEquals, slugsEquivalent } from "../../providers/slug-codec";
 import { identifyRoutedModel } from "../../adapters/identity";
 import { filterCursorConfiguredModelsByLiveDiscovery } from "../../adapters/cursor/discovery";
@@ -204,6 +205,8 @@ export interface NativeContextLimits {
   readonly providerWindow?: number;
   /** `providers.openai.modelContextWindows` — per-model, wins over `providerWindow`. */
   readonly modelWindows?: Readonly<Record<string, number>>;
+  /** `providers.openai.modelAutoCompactTokenLimits` — soft, lowering-only budgets. */
+  readonly modelAutoCompactTokenLimits?: Readonly<Record<string, number>>;
 }
 
 export type NativeContextLimitsInput = NativeContextLimits | number | undefined;
@@ -227,12 +230,18 @@ export function nativeContextLimits(
     const window = positiveInt(value);
     if (window !== undefined) modelWindows[slug] = window;
   }
+  const modelAutoCompactTokenLimits: Record<string, number> = {};
+  for (const [slug, value] of Object.entries(provider?.modelAutoCompactTokenLimits ?? {})) {
+    const budget = positiveInt(value);
+    if (budget !== undefined) modelAutoCompactTokenLimits[slug] = budget;
+  }
   return {
     ...(positiveInt(providerContextCap(config, OPENAI_CODEX_PROVIDER_ID)) !== undefined
       ? { cap: providerContextCap(config, OPENAI_CODEX_PROVIDER_ID) }
       : {}),
     ...(positiveInt(provider?.contextWindow) !== undefined ? { providerWindow: provider!.contextWindow } : {}),
     ...(Object.keys(modelWindows).length > 0 ? { modelWindows } : {}),
+    ...(Object.keys(modelAutoCompactTokenLimits).length > 0 ? { modelAutoCompactTokenLimits } : {}),
   };
 }
 
@@ -277,6 +286,21 @@ export function nativeOpenAiMaxInputTokens(slug: string, limits?: NativeContextL
   return window === undefined ? narrowed : Math.min(narrowed, window);
 }
 
+/** Effective native soft budget after every hard window/input limit is resolved. */
+export function nativeOpenAiAutoCompactTokenLimit(
+  slug: string,
+  limits?: NativeContextLimitsInput,
+): number | undefined {
+  const contextWindow = nativeOpenAiContextWindow(slug, limits);
+  if (contextWindow === undefined) return undefined;
+  const configured = positiveInt(asLimits(limits).modelAutoCompactTokenLimits?.[slug]);
+  return clampAutoCompactTokenLimit(
+    contextWindow,
+    nativeOpenAiMaxInputTokens(slug, limits),
+    configured,
+  );
+}
+
 export function nativeInputModalities(slug: string): string[] {
   const upstream = PINNED_NATIVE_CAPABILITY_ENTRIES.get(slug);
   if (Array.isArray(upstream?.input_modalities) && upstream!.input_modalities!.length > 0) {
@@ -387,7 +411,7 @@ export function desktopVisibleNativeSlugs(
   ]);
 }
 
-export function nativeModelRows(config: Pick<OcxConfig, "disabledModels" | "combos" | "providerContextCaps" | "providers">): Array<{ slug: string; disabled: boolean; contextWindow?: number; maxInputTokens?: number }> {
+export function nativeModelRows(config: Pick<OcxConfig, "disabledModels" | "combos" | "providerContextCaps" | "providers">): Array<{ slug: string; disabled: boolean; contextWindow?: number; maxInputTokens?: number; autoCompactTokenLimit?: number }> {
   const disabled = disabledNativeSlugs(config);
   const shadowed = configuredNativeAliasSlugs(config);
   // Both user levers, not just the cap: a per-model window set from the dashboard has to show
@@ -403,11 +427,13 @@ export function nativeModelRows(config: Pick<OcxConfig, "disabledModels" | "comb
     .filter(slug => !shadowed.has(slug)).map(slug => {
     const contextWindow = nativeOpenAiContextWindow(slug, limits);
     const maxInputTokens = nativeOpenAiMaxInputTokens(slug, limits);
+    const autoCompactTokenLimit = nativeOpenAiAutoCompactTokenLimit(slug, limits);
     return {
       slug,
       disabled: disabled.has(slug),
       ...(contextWindow !== undefined ? { contextWindow } : {}),
       ...(maxInputTokens !== undefined ? { maxInputTokens } : {}),
+      ...(autoCompactTokenLimit !== undefined ? { autoCompactTokenLimit } : {}),
     };
   });
 }

package/src/codex/catalog/model-metadata.ts

@@ -5,6 +5,8 @@ import { isModelCacheGenerationCurrent } from "../model-cache";
 import type { GenerationContext } from "../../lib/state-store-sweeper";
 import { captureConfigGeneration } from "../../lib/state-store-sweeper";
 import { assertNotRealHomeUnderTest } from "../../lib/test-home-guard";
+import { CURSOR_STATIC_MODELS } from "../../adapters/cursor/discovery";
+import { cursorModelEffortLadder } from "../../adapters/cursor/effort-map";
 import { generatedModelMetadata, type CatalogModel } from "./parsing";
 
 /** Codex strict-parser placeholder; must never be treated as discovered evidence. */
@@ -470,16 +472,30 @@ export function reconcileModelMetadataSnapshot(context: GenerationContext): numb
 
 export function registryLayerForModel(provider: string, modelId: string): ModelMetadataLayer | undefined {
   const meta = generatedModelMetadata(provider, modelId);
-  if (!meta) return undefined;
-  return {
-    ...(typeof meta.contextWindow === "number" && meta.contextWindow > 0
-      ? { contextWindow: meta.contextWindow }
-      : {}),
-    ...(typeof meta.maxTokens === "number" && meta.maxTokens > 0
-      ? { maxOutputTokens: meta.maxTokens }
-      : {}),
-    ...(Array.isArray(meta.input) && meta.input.length > 0 ? { inputModalities: [...meta.input] } : {}),
-  };
+  if (meta) {
+    return {
+      ...(typeof meta.contextWindow === "number" && meta.contextWindow > 0
+        ? { contextWindow: meta.contextWindow }
+        : {}),
+      ...(typeof meta.maxTokens === "number" && meta.maxTokens > 0
+        ? { maxOutputTokens: meta.maxTokens }
+        : {}),
+      ...(Array.isArray(meta.input) && meta.input.length > 0 ? { inputModalities: [...meta.input] } : {}),
+    };
+  }
+  if (provider === "cursor") {
+    const cursorModel = CURSOR_STATIC_MODELS.find(m => m.id === modelId || m.id.toLowerCase() === modelId.toLowerCase());
+    if (cursorModel) {
+      return {
+        contextWindow: cursorModel.contextWindow,
+        ...(cursorModel.inputModalities ? { inputModalities: [...cursorModel.inputModalities] } : {}),
+        ...(cursorModel.supportsReasoningEffort
+          ? { reasoningEfforts: cursorModelEffortLadder(cursorModel.id) ?? [] }
+          : {}),
+      };
+    }
+  }
+  return undefined;
 }
 
 function layerFromCatalogModel(model: CatalogModel): ModelMetadataLayer | undefined {
@@ -547,12 +563,23 @@ export function enrichCatalogModelMetadata(
     && model.detectedContextWindow > (model.contextWindow ?? 0)
     ? { ...fromModel, contextWindow: model.detectedContextWindow }
     : fromModel;
+  const registry = registryLayerForModel(model.provider, model.id);
+  // Cursor's GetUsableModels RPC returns ids only. Any capability lists already present on a
+  // Cursor row therefore come from explicit provider configuration; keep them ahead of the static
+  // registry backfill, including an intentional empty reasoning ladder.
+  const effectiveRegistry = registry && model.provider === "cursor"
+    ? {
+      ...registry,
+      ...(Array.isArray(model.inputModalities) ? { inputModalities: undefined } : {}),
+      ...(Array.isArray(model.reasoningEfforts) ? { reasoningEfforts: undefined } : {}),
+    }
+    : registry;
   const resolved = resolveModelMetadata({
     ...(liveFresh && liveLayer ? { live: liveLayer } : {}),
     liveFresh,
     ...(snapshot ? { snapshot } : {}),
-    ...(registryLayerForModel(model.provider, model.id)
-      ? { registry: registryLayerForModel(model.provider, model.id) }
+    ...(effectiveRegistry
+      ? { registry: effectiveRegistry }
       : {}),
     ...(options?.caps ? { caps: options.caps } : {}),
   });