@bitkyc08/opencodex 2.58.0 → 2.59.0

package/src/providers/key-failover.ts

@@ -29,6 +29,243 @@ interface KeyCooldown {
 const DEFAULT_COOLDOWN_MS = 60_000;
 const MAX_COOLDOWN_MS = 10 * 60_000; // cap at 10 min for api-key rotation
 
+/**
+ * Cap for a cooldown the upstream itself dated, as opposed to one we inferred.
+ *
+ * `MAX_COOLDOWN_MS` is deliberately short because an undated 429 is a guess: ten
+ * minutes bounds how long a transient limit can park a working key. A free-tier
+ * quota is not a guess — OpenRouter replies `Weekly/Monthly Limit Exhausted ...
+ * will reset at <date>`, and until that date the key cannot serve anything. Held
+ * for ten minutes instead, it comes back, takes a 429, and rotates again, every
+ * ten minutes for the rest of the week (#4024).
+ *
+ * 32 days rather than unbounded. The wording this parses is
+ * `Weekly/Monthly Limit Exhausted`, so the cap has to clear a monthly window —
+ * 31 days plus a day of slack for timezone and month length. An earlier 8-day
+ * cap looked generous against the weekly case in the issue and silently clamped
+ * every monthly reset to ~23 days early, which puts the key back into exactly
+ * the 429 loop this exists to stop. Caught by the cap's own test.
+ *
+ * Bounded at all because the date is upstream-controlled input: a malformed or
+ * hostile `reset at 2999-01-01` must not park a working key past any horizon an
+ * operator would think to look at.
+ */
+const MAX_QUOTA_COOLDOWN_MS = 32 * 24 * 60 * 60_000;
+const QUOTA_RESET_PEEK_TIMEOUT_MS = 250;
+
+interface QuotaResetReadOptions {
+  now?: number;
+  signal?: AbortSignal;
+  timeoutMs?: number;
+}
+
+function rebuiltResponse(response: Response, body: ReadableStream<Uint8Array>): Response {
+  return new Response(body, {
+    status: response.status,
+    statusText: response.statusText,
+    headers: response.headers,
+  });
+}
+
+function replayOnlyResponse(response: Response, chunks: readonly Uint8Array[]): Response {
+  return rebuiltResponse(response, new ReadableStream<Uint8Array>({
+    start(controller) {
+      for (const chunk of chunks) controller.enqueue(chunk);
+      controller.close();
+    },
+  }));
+}
+
+/**
+ * Read a bounded prefix of a 429 body and pull the upstream's declared reset instant.
+ *
+ * The returned response replays the bounded prefix and any boundary-chunk overflow
+ * before streaming the unread remainder. A rotation storm must not be gated on
+ * reading N full error payloads. Any failure — no body, already consumed, slow,
+ * malformed — returns undefined and leaves the `Retry-After` path in charge.
+ */
+export async function readQuotaResetAt(
+  response: Response,
+  nowOrOptions: number | QuotaResetReadOptions = {},
+): Promise<{ at: number | undefined; response: Response }> {
+  if (!response.body) return { at: undefined, response };
+  const options = typeof nowOrOptions === "number" ? { now: nowOrOptions } : nowOrOptions;
+  const now = options.now ?? Date.now();
+  let reader: ReadableStreamDefaultReader<Uint8Array>;
+  try {
+    reader = response.body.getReader();
+  } catch {
+    return { at: undefined, response };
+  }
+  const chunks: Uint8Array[] = [];
+  let transferred = false;
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  const deadline = new AbortController();
+  const timeoutReason = new DOMException("Quota reset body peek timed out", "TimeoutError");
+  try {
+    const decoder = new TextDecoder();
+    let seen = 0;
+    let text = "";
+    timer = setTimeout(() => deadline.abort(timeoutReason), options.timeoutMs ?? QUOTA_RESET_PEEK_TIMEOUT_MS);
+    const signal = options.signal
+      ? AbortSignal.any([options.signal, deadline.signal])
+      : deadline.signal;
+    while (seen < QUOTA_RESET_SCAN_BYTES) {
+      const read = reader.read();
+      let rejectAbort: ((reason: unknown) => void) | undefined;
+      const onAbort = () => rejectAbort?.(signal.reason);
+      const aborted = new Promise<never>((_resolve, reject) => {
+        rejectAbort = reject;
+        if (signal.aborted) reject(signal.reason);
+        else signal.addEventListener("abort", onAbort, { once: true });
+      });
+      // Derived from the reader rather than named directly: Bun's lib types
+      // `ReadableStreamDefaultReader.read()` as returning
+      // `ReadableStreamDefaultReadResult`, which is not assignable to the
+      // `ReadableStreamReadResult` alias.
+      let result: Awaited<ReturnType<typeof reader.read>>;
+      try {
+        result = await Promise.race([read, aborted]);
+      } finally {
+        signal.removeEventListener("abort", onAbort);
+      }
+      const { done, value } = result;
+      if (done) break;
+      const remaining = QUOTA_RESET_SCAN_BYTES - seen;
+      const prefix = value.byteLength > remaining ? value.subarray(0, remaining) : value;
+      const overflow = value.byteLength > remaining ? value.subarray(remaining) : undefined;
+      chunks.push(prefix);
+      if (overflow?.byteLength) chunks.push(overflow);
+      seen += prefix.byteLength;
+      text += decoder.decode(prefix, { stream: true });
+    }
+    // Hand back a Response carrying the bytes already pulled followed by whatever
+    // is left, so the caller can still read or cancel it. `response.clone()` is
+    // NOT usable here: it tees, and with the original branch undrained the tee
+    // stalls once its buffer fills — a 5MB error body hangs the rotation path,
+    // which is worse than the unbounded read this replaced.
+    const rest = new ReadableStream<Uint8Array>({
+      start(controller) {
+        for (const c of chunks) controller.enqueue(c);
+      },
+      async pull(controller) {
+        try {
+          const { done, value } = await reader.read();
+          if (done) {
+            controller.close();
+            reader.releaseLock();
+            return;
+          }
+          controller.enqueue(value);
+        } catch (error) {
+          controller.error(error);
+          reader.releaseLock();
+        }
+      },
+      async cancel(reason) {
+        try {
+          await reader.cancel(reason);
+        } finally {
+          reader.releaseLock();
+        }
+      },
+    });
+    transferred = true;
+    return { at: parseQuotaResetAt(text, now), response: rebuiltResponse(response, rest) };
+  } catch (error) {
+    const clientAborted = options.signal?.aborted === true;
+    void reader.cancel(error).catch(() => {}).finally(() => {
+      try { reader.releaseLock(); } catch { /* already released */ }
+    });
+    if (clientAborted) throw options.signal!.reason ?? error;
+    return { at: undefined, response: replayOnlyResponse(response, chunks) };
+  } finally {
+    if (timer !== undefined) clearTimeout(timer);
+    if (!transferred) {
+      try { reader.releaseLock(); } catch { /* already released */ }
+    }
+  }
+}
+
+/**
+ * How much of a 429 body is read and scanned for the reset instant.
+ *
+ * Bounds the READ, not just the parse: this runs on the rotation path, once per
+ * rotated key under a rate-limit storm, and the body is upstream-controlled.
+ * OpenRouter's rate_limit_error JSON is a few hundred bytes.
+ */
+const QUOTA_RESET_SCAN_BYTES = 4_096;
+
+/**
+ * Reset instant an upstream declared in a 429 *body*, in epoch ms.
+ *
+ * Only the body carries this: OpenRouter sends no `Retry-After` for a quota
+ * exhaustion, so the header path (`parseRetryAfterMs`) sees nothing and falls
+ * back to `DEFAULT_COOLDOWN_MS`. Returns undefined for anything it cannot read
+ * as a date, so an unparsable body keeps today's behaviour exactly.
+ */
+/**
+ * Whether `YYYY-MM-DD…` names a day that exists.
+ *
+ * `Date.parse` does NOT reject an out-of-range day: measured on Bun,
+ * `2026-02-30T00:00:00Z` yields March 2 and `2026-04-31T00:00:00Z` yields
+ * May 1, so a malformed upstream body would park a key past the instant it
+ * actually named. Only the month is rejected outright (`2026-13-01` is NaN).
+ *
+ * Checked on the date text alone rather than by round-tripping the parsed
+ * instant, because a value carrying an explicit offset (`…T23:00+05:30`)
+ * legitimately lands on a different UTC day than the one written.
+ */
+function isRealCalendarDate(value: string): boolean {
+  const [year, month, day] = value.slice(0, 10).split("-").map(Number);
+  if (month < 1 || month > 12 || day < 1) return false;
+  const leap = (year % 4 === 0 && year % 100 !== 0) || year % 400 === 0;
+  const lengths = [31, leap ? 29 : 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31];
+  return day <= lengths[month - 1]!;
+}
+
+export function parseQuotaResetAt(body: string | null | undefined, now = Date.now()): number | undefined {
+  const text = body?.slice(0, QUOTA_RESET_SCAN_BYTES);
+  if (!text) return undefined;
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(text);
+  } catch {
+    return undefined;
+  }
+  if (!parsed || typeof parsed !== "object") return undefined;
+  const error = (parsed as { error?: unknown }).error;
+  if (!error || typeof error !== "object") return undefined;
+  const code = (error as { code?: unknown }).code;
+  const message = (error as { message?: unknown }).message;
+  if ((code !== "rate_limit_error" && code !== 429) || typeof message !== "string") return undefined;
+  if (!/^(?:Weekly|Monthly) Limit Exhausted\b/i.test(message.trim())) return undefined;
+  // `will reset at 2026-09-09 03:30:06` / `... at 2026-09-09T03:30:06Z` / `resets at <date>`
+  const match = /reset[s]?\s+at\s+([0-9]{4}-[0-9]{2}-[0-9]{2}(?:[T ][0-9]{2}:[0-9]{2}(?::[0-9]{2})?(?:\.[0-9]+)?(?:Z|[+-][0-9]{2}:?[0-9]{2})?)?)/i.exec(message);
+  if (!match) return undefined;
+  // Pin a bare `YYYY-MM-DD hh:mm:ss` to UTC explicitly.
+  //
+  // ECMA-262 says a date-TIME form with no offset is LOCAL time, and Node follows
+  // that: `Date.parse("2026-09-09 03:30:06")` differs from the UTC reading by the
+  // host offset (7h on a PDT box, measured). Bun currently returns the UTC value
+  // for the same string, so on this runtime the normalisation is a no-op today —
+  // which is exactly why it is written out rather than relied upon. If Bun ever
+  // conforms, an un-normalised parse would silently shift every park-until by the
+  // operator's offset, and the early direction resumes the 429 loop.
+  //
+  // A consequence worth knowing: no Bun test can observe this branch being
+  // removed. The explicit-zone case below is the part the suite can pin.
+  const raw = match[1].includes("T") || /(?:Z|[+-][0-9]{2}:?[0-9]{2})$/.test(match[1])
+    ? match[1]
+    : `${match[1].replace(" ", "T")}Z`;
+  if (!isRealCalendarDate(match[1])) return undefined;
+  const at = Date.parse(raw);
+  if (!Number.isFinite(at)) return undefined;
+  // Already past, or beyond the cap: not usable as a park-until instant.
+  if (at <= now) return undefined;
+  return Math.min(at, now + MAX_QUOTA_COOLDOWN_MS);
+}
+
 /**
  * Default same-target 429 retry policy used when a provider opts in via a bare
  * `retryOn429: {}` (presence = opt-in with these defaults).
@@ -303,11 +540,11 @@ export function rateLimitRetryPolicyFor(
 
 /**
  * Normalize a provider's `transientRetryOn5xx` policy, or return null when it is absent,
- * explicitly disabled, not key-auth, or not the `openai-chat` adapter.
+ * explicitly disabled, not key-auth, or not an adapter this policy governs.
  *
- * The adapter gate is part of the accepted scope, not incidental: this first version covers
- * key-auth `openai-chat` only, and without an explicit check any generic key-auth adapter
- * could opt in. Auth mode follows the same fail-closed rule as `rateLimitRetryPolicyFor` —
+ * The adapter gate is part of the accepted scope, not incidental: it names the adapters whose
+ * lanes actually read this policy, so no generic key-auth adapter can opt in by accident.
+ * Auth mode follows the same fail-closed rule as `rateLimitRetryPolicyFor` —
  * explicit `key` or the documented omitted default, never OAuth, forward, local, or an
  * unknown value.
  */
@@ -316,7 +553,14 @@ export function transientRetryPolicyFor(
 ): Required<TransientRetryPolicy> | null {
   const policy = provider.transientRetryOn5xx;
   if (!policy || policy.enabled === false) return null;
-  if (provider.adapter !== "openai-chat") return null;
+  // Both adapters this policy governs. The first version covered chat only, which left a
+  // key-auth Responses provider unable to tune its ladder in either direction, because the
+  // Responses passthrough lane hard-coded TRANSIENT_RETRY_MAX_ATTEMPTS (#4893). Widening this
+  // gate is necessary and not sufficient: the lane also has to call this function, which it
+  // now does. Still an explicit list, so no generic key-auth adapter opts in by accident, and
+  // the auth check below keeps the ChatGPT forward pool out -- those providers are
+  // `authMode: "forward"` and keep the default ladder they have always had.
+  if (provider.adapter !== "openai-chat" && provider.adapter !== "openai-responses") return null;
   if (provider.authMode !== undefined && provider.authMode !== "key") return null;
   return {
     enabled: policy.enabled ?? DEFAULT_TRANSIENT_RETRY.enabled,
@@ -363,6 +607,7 @@ function rotateKeyAfterFailure(
   now = Date.now(),
   attemptedKey?: string,
   attemptedSelection?: ProviderApiKeySelection,
+  quotaResetAt?: number,
 ): OcxProviderConfig | null {
   const provider = config.providers[providerName];
   if (!provider) return null;
@@ -421,7 +666,12 @@ function rotateKeyAfterFailure(
     // full cap instead of the 429 default so a dead key is not re-tried once a minute.
     const cooldownMs = failureStatus === 401
       ? MAX_COOLDOWN_MS
-      : parseRetryAfterMs(retryAfterHeader, now) ?? DEFAULT_COOLDOWN_MS;
+      // A reset instant the upstream dated outranks both the header and the
+      // default: it is the only one of the three that knows when the quota
+      // actually returns (#4024).
+      : quotaResetAt !== undefined
+        ? Math.max(quotaResetAt - now, 1)
+        : parseRetryAfterMs(retryAfterHeader, now) ?? DEFAULT_COOLDOWN_MS;
     keyCooldowns.set(cooldownKey(providerName, outcome.value.failedId), { cooldownUntil: now + cooldownMs });
     sweepExpiredOnWrite(now);
   }
@@ -448,8 +698,9 @@ export function rotateKeyOn429(
   now = Date.now(),
   attemptedKey?: string,
   attemptedSelection?: ProviderApiKeySelection,
+  quotaResetAt?: number,
 ): OcxProviderConfig | null {
-  return rotateKeyAfterFailure(config, providerName, 429, retryAfterHeader, now, attemptedKey, attemptedSelection);
+  return rotateKeyAfterFailure(config, providerName, 429, retryAfterHeader, now, attemptedKey, attemptedSelection, quotaResetAt);
 }
 
 /**
@@ -482,6 +733,8 @@ export function sweepExpiredApiKeyCooldowns(now = Date.now()): number {
 
 interface RotateProviderTransportOptions {
   retryAfter?: string | null;
+  /** Epoch ms from `parseQuotaResetAt`, when the upstream dated the reset in its body. */
+  quotaResetAt?: number;
   now?: number;
   attemptedKey?: string;
   attemptedSelection?: ProviderApiKeySelection;
@@ -507,6 +760,7 @@ export function rotateProviderTransportOn429(
     options.now,
     options.attemptedKey,
     options.attemptedSelection ?? routedProvider._apiKeyAttempt,
+    options.quotaResetAt,
   );
   if (!rotated) return null;
   return applyRotatedTransport(providerName, routedProvider, rotated, options.promptCacheKey);

package/src/providers/model-rename-migration.ts

@@ -86,6 +86,7 @@ const MODEL_KEYED_RECORDS = [
   "modelMaxOutputTokens",
   "modelInputModalities",
   "modelReasoningEfforts",
+  "modelSuppressSyntheticMax",
   "modelDefaultReasoningEfforts",
   "modelReasoningEffortMap",
 ] as const;

package/src/providers/openai-sidecar.ts

@@ -32,6 +32,8 @@ export interface ResolvedOpenAiForwardSidecar extends OpenAiForwardSidecarCandid
   authContext: CodexAuthContext;
   headers: Headers;
   recordOutcome?: (outcome: CodexUpstreamOutcome) => void;
+  /** Hand back an acquired recovery probe when no sidecar request reached upstream. */
+  releaseProbeLease?: () => void;
 }
 
 /**
@@ -189,6 +191,7 @@ export async function resolveFirstUsableOpenAiSidecar(
             ...(authContext.kind === "pool" ? { credentialGeneration: authContext.generation } : {}),
           },
         ),
+        releaseProbeLease: () => releaseCodexAuthContextProbeLease(authContext),
       };
     }
     if (candidate.accountMode === "direct") {
@@ -237,6 +240,7 @@ export async function resolveFirstUsableOpenAiSidecar(
               ...(authContext.kind === "pool" ? { credentialGeneration: authContext.generation } : {}),
             },
           ),
+          releaseProbeLease: () => releaseCodexAuthContextProbeLease(authContext),
         }
         : {}),
     };

package/src/providers/opencode-go-transport.ts

@@ -12,10 +12,12 @@ function hasHeaderCaseInsensitive(
   return Object.keys(headers ?? {}).some(key => key.toLowerCase() === target);
 }
 
-/** Derive a provider-scoped opaque value without exposing Codex task or subagent ids. */
-export function deriveOpenCodeGoSessionId(sessionLane: string): string {
+/** Derive a provider- and wire-scoped opaque value without exposing Codex task or subagent ids. */
+export function deriveOpenCodeGoSessionId(sessionLane: string, wireProtocol: string): string {
   const digest = createHash("sha256")
-    .update("opencodex/opencode-go/session/v1\0")
+    .update("opencodex/opencode-go/session/v2\0")
+    .update(wireProtocol)
+    .update("\0")
     .update(sessionLane)
     .digest("hex")
     .slice(0, 32);
@@ -30,12 +32,19 @@ export function deriveOpenCodeGoSessionId(sessionLane: string): string {
  * request reaching this helper from the proxy always carries a lane. The `!sessionLane` guard stays
  * for direct callers that have no request context; it is not a per-request identity of its own, and
  * minting one here would hand each retry a different value.
+ *
+ * `provider` is already settled onto the final wire, while `destinationProvider` is the original
+ * routed row. Keeping both explicit prevents an Anthropic hard pin from defeating the registry's
+ * adapter-sensitive destination recognition.
  */
 export function resolveOpenCodeGoTransport<T extends OcxProviderConfig>(
   provider: T,
   sessionLane: string | undefined,
+  destinationProvider:
+    Pick<OcxProviderConfig, "baseUrl" | "adapter">
+    & Partial<Pick<OcxProviderConfig, "authMode">>,
 ): T {
-  if (registryEntryForProviderDestination(provider)?.id !== "opencode-go") return provider;
+  if (registryEntryForProviderDestination(destinationProvider)?.id !== "opencode-go") return provider;
   if (!sessionLane) return provider;
   if (hasHeaderCaseInsensitive(provider.headers, OPENCODE_GO_SESSION_HEADER)) return provider;
 
@@ -43,7 +52,7 @@ export function resolveOpenCodeGoTransport<T extends OcxProviderConfig>(
     ...provider,
     headers: {
       ...(provider.headers ?? {}),
-      [OPENCODE_GO_SESSION_HEADER]: deriveOpenCodeGoSessionId(sessionLane),
+      [OPENCODE_GO_SESSION_HEADER]: deriveOpenCodeGoSessionId(sessionLane, provider.adapter),
     },
   };
 }

package/src/providers/quota/report-cache.ts

@@ -151,6 +151,9 @@ export function providerQuotaFromCodexQuota(
   const projected: CodexCapacityQuota = {
     ...(quota.shortPercent !== undefined ? { fiveHourPercent: quota.shortPercent } : {}),
     ...(quota.shortResetAt !== undefined ? { fiveHourResetAt: quota.shortResetAt } : {}),
+    // Freshness for the reset-less terminal rule. Without it the dashboard evaluates that rule
+    // with no evidence and returns null while routing refuses the same account (#5045).
+    ...(quota.shortObservedAt !== undefined ? { shortObservedAt: quota.shortObservedAt } : {}),
     ...(quota.weeklyPercent !== undefined ? { weeklyPercent: quota.weeklyPercent } : {}),
     ...(quota.weeklyResetAt !== undefined ? { weeklyResetAt: quota.weeklyResetAt } : {}),
     ...(quota.monthlyPercent !== undefined ? { monthlyPercent: quota.monthlyPercent } : {}),

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

@@ -98,6 +98,10 @@ import {
   DIGITALOCEAN_CHAT_COMPLETION_MODELS,
   SCALEWAY_SERVERLESS_CHAT_MODELS,
   SCALEWAY_MODEL_INPUT_MODALITIES,
+  OPPER_MODELS,
+  OPPER_MODEL_CONTEXT_WINDOWS,
+  OPPER_MODEL_MAX_OUTPUT_TOKENS,
+  OPPER_MODEL_INPUT_MODALITIES,
 } from "./model-seeds";
 
 export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
@@ -219,6 +223,72 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     },
     note: "Shared Token Factory text-output inference only; live discovery excludes embedding and image-generation rows.",
   },
+  {
+    // Primary sources checked 2026-09-11:
+    // - https://docs.crusoecloud.com/quickstart/getting-started-with-serverless-inference documents
+    //   the fixed OpenAI-compatible host https://api.inference.crusoecloud.com/v1, Bearer API keys
+    //   created in the Cloud console (Intelligence Foundry > Inference > Create API Key), and an
+    //   OpenAI SDK chat.completions example against meta-llama/Llama-3.3-70B-Instruct.
+    // - https://docs.crusoecloud.com/serverless-inference/available-models lists the served models
+    //   with slash-delimited ids; https://docs.crusoecloud.com/serverless-inference/rate-limits
+    //   documents per-project, per-model TPM/RPM limits (429 when exceeded, 503 under shared load).
+    // - GET /v1/models rejects unauthenticated requests with 401 {"errors":["Authentication failed"]},
+    //   so a successful authenticated list response is evidence that the supplied key is valid.
+    //   An authenticated capture on 2026-09-12 returned 18 rows shaped like OpenRouter's catalog
+    //   (`is_public`, `type`, `context_length`, `architecture.modality` of "text" or "multimodal",
+    //   `tags`, `pricing`, `supported_parameters`); 17 were public serverless models and one was an
+    //   account-private dedicated deployment with empty `type`/`modality`. `type` is blank on one
+    //   public model, so the filter keys on `is_public` plus `architecture.modality` instead.
+    // - https://legal.crusoe.ai/ hosts the Crusoe Cloud Platform Terms of Service v1.10 (effective
+    //   2026-08-10), which name Crusoe Technologies LLC as the contracting entity, and the Service
+    //   Specific Terms v5.0 (effective 2026-07-14), whose Crusoe Intelligence Foundry Terms cover the
+    //   Managed Inference Service reached through the Crusoe API.
+    // - https://models.dev/api.json (provider "crusoe") records openai/gpt-oss-120b as the one served
+    //   model with a low/medium/high reasoning_effort ladder; the other reasoning models expose an
+    //   on/off toggle only.
+    // Maintainer: @acheamponge, who works at Crusoe (affiliation disclosed) and also maintains the
+    // models.dev crusoe entry.
+    id: "crusoe",
+    label: "Crusoe",
+    baseUrl: "https://api.inference.crusoecloud.com/v1",
+    adapter: "openai-chat",
+    authKind: "key",
+    dashboardUrl: "https://console.crusoecloud.com",
+    liveModels: true,
+    preserveCustomDestination: true,
+    // The getting-started guide documents tools through the OpenAI SDK but no provider-wide
+    // parallel tool-call contract.
+    parallelToolCalls: false,
+    // Only gpt-oss-120b has a real effort ladder; toggle-style reasoning models must not be promoted
+    // to Codex's full fallback ladder.
+    reasoningEfforts: [],
+    modelReasoningEfforts: { "openai/gpt-oss-120b": ["low", "medium", "high"] },
+    directReasoningEffortModels: ["openai/gpt-oss-120b"],
+    // The catalog reports `architecture.modality: "multimodal"` without an input list. Four rows
+    // also carry the explicit "image text to text" tag; yutori/n2 instead reports multimodal
+    // type/modality plus browser/computer-use tags. Those five captured rows are classified here.
+    modelInputModalities: {
+      "google/gemma-4-31b-it": ["text", "image"],
+      "moonshotai/Kimi-K2.6": ["text", "image"],
+      "nvidia/Nemotron-3-Nano-Omni-Reasoning-30B-A3B": ["text", "image"],
+      "yutori/n2": ["text", "image"],
+      "zai-org/GLM-5.3-Flash": ["text", "image"],
+    },
+    modelDiscovery: {
+      path: "models",
+      maxResponseBytes: 256 * 1024,
+      maxModels: 256,
+      filter: {
+        // Keep public serverless rows whose architecture produces text; account-private
+        // deployments (blank modality) and any embedding or media rows fail closed.
+        allOf: [
+          { path: ["is_public"], equalsAny: [true] },
+          { path: ["architecture", "modality"], equalsAny: ["text", "multimodal"] },
+        ],
+      },
+    },
+    note: "Public Serverless Inference chat models on the shared OpenAI-compatible host; account-private and self-serve dedicated deployments are excluded from discovery and out of scope.",
+  },
   {
     id: "digitalocean",
     label: "DigitalOcean Serverless Inference",
@@ -736,6 +806,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
       ...Object.fromEntries(ALIBABA_TOKEN_PLAN_QWEN_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
       ...Object.fromEntries(QWEN38_FAMILY.map(id => [id, QWEN38_REASONING_EFFORTS])),
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
+      "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
       "deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek-v4-pro"),
       "deepseek-v4-pro-0813": deepseekThinkingEffortsFor("deepseek-v4-pro-0813"),
       "deepseek-v4-flash-0731": deepseekThinkingEffortsFor("deepseek-v4-flash-0731"),
@@ -781,6 +852,7 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
       ...Object.fromEntries(ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS.map(id => [id, THINKING_BUDGET_EFFORTS])),
       ...Object.fromEntries(QWEN38_FAMILY.map(id => [id, QWEN38_REASONING_EFFORTS])),
       "glm-5.2": ZAI_GLM_52_REASONING_EFFORTS,
+      "glm-5.3": ZAI_GLM_53_REASONING_EFFORTS,
       "deepseek-v4-pro": deepseekThinkingEffortsFor("deepseek-v4-pro"),
       "deepseek-v4-pro-0813": deepseekThinkingEffortsFor("deepseek-v4-pro-0813"),
       "deepseek-v4-flash": deepseekThinkingEffortsFor("deepseek-v4-flash"),
@@ -957,6 +1029,30 @@ export const PROVIDER_REGISTRY_EXTENDED: readonly ProviderRegistryEntry[] = [
     noJsonSchemaModels: [...DEEPSEEK_GATEWAY_THINKING_MODELS, ...OPENCODE_FREE_DEEPSEEK_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" },
+  {
+    // Opper: EU-hosted AI gateway (Opper AI AB, Stockholm). One OpenAI-compatible endpoint and one
+    // key in front of 30+ upstream providers. Seeded ids are Opper *pools* (bare names such as
+    // `claude-sonnet-4-6`): the gateway chooses the provider/region per request, and a
+    // `vendor/model` id (`anthropic/claude-sonnet-4-6`, `aws/claude-sonnet-4-6-eu`) pins one route.
+    // The original provider author reported on 2026-09-08 that GET /v3/compat/models answers 401
+    // without a key, so the default discovery URL doubles as key validation. Windows, output caps
+    // and modalities live in model-seeds.ts (smallest value / shared modality across each pool's
+    // members); live discovery owns which models exist.
+    id: "opper",
+    label: "Opper",
+    adapter: "openai-chat",
+    baseUrl: "https://api.opper.ai/v3/compat",
+    authKind: "key",
+    dashboardUrl: "https://platform.opper.ai",
+    liveModels: true,
+    preserveCustomDestination: true,
+    defaultModel: "claude-sonnet-4-6",
+    models: OPPER_MODELS,
+    modelContextWindows: OPPER_MODEL_CONTEXT_WINDOWS,
+    modelMaxOutputTokens: OPPER_MODEL_MAX_OUTPUT_TOKENS,
+    modelInputModalities: OPPER_MODEL_INPUT_MODALITIES,
+    note: "EU-hosted AI gateway: one OpenAI-compatible endpoint and one key in front of 30+ providers. Bare model ids are pools (claude-sonnet-4-6, gpt-5.5) and Opper picks the route per request; vendor/model ids (anthropic/claude-sonnet-4-6) pin one provider. The catalogue is discovered live from /v3/compat/models with your key; the public list is at opper.ai/models. Token rates are the model providers' rates with no markup; Opper charges a 3% fee when you buy credits.",
+  },
   {
     id: "opencode-free",
     label: "OpenCode Free",