omnius 1.0.873 → 1.0.874

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 37773730 bytes
-sha256: f762c00e6a2c7407ca37ff6c15bacffd0028cd2c2927dd91fd7db512549090c8
+size: 37925419 bytes
+sha256: 60600931f277c7850afc402e56d9985628d48802a8c905961f9396a865666432

package/dist/library.js

@@ -340844,14 +340844,14 @@ var registrations = [
       label: "Built-in free",
       protocol: "openai-chat",
       local: false,
-      apiPrefix: "/inference/v1",
-      paths: { chat: "/inference/v1/chat/completions", models: "/inference/v1/models" },
+      apiPrefix: "/inference/openai/v1",
+      paths: { chat: "/inference/openai/v1/chat/completions", models: "/inference/openai/v1/models" },
       auth: { strategy: "bearer", required: false, staticHeaders: { Authorization: "Bearer public" } },
       capabilities: CHAT_CAPABILITIES,
-      defaultUrl: "https://opencode.ai/inference/v1",
+      defaultUrl: "https://opencode.ai/inference/openai/v1",
       description: "Built-in models with public access where permitted and optional provider-bound API credentials."
     }),
-    match: (url) => url.protocol === "https:" && url.hostname.toLowerCase() === "opencode.ai" && !url.port && /^\/inference\/v1(?:\/(?:models|chat\/completions|responses))?\/?$/.test(url.pathname)
+    match: (url) => url.protocol === "https:" && url.hostname.toLowerCase() === "opencode.ai" && !url.port && /^\/inference\/(?:openai\/)?v1(?:\/(?:models|chat\/completions|responses))?\/?$/.test(url.pathname)
   },
   {
     descriptor: descriptor({
@@ -341436,7 +341436,7 @@ function resolveProviderTransport(options) {
   }
   if (provider.id === "builtin-free") {
     const url = parseProviderUrl(rawUrl);
-    if (url.protocol !== "https:" || url.hostname.toLowerCase() !== "opencode.ai" || url.port || url.username || url.password || url.search || url.hash || !/^\/inference\/v1(?:\/(?:models|chat\/completions|responses))?\/?$/.test(url.pathname)) {
+    if (url.protocol !== "https:" || url.hostname.toLowerCase() !== "opencode.ai" || url.port || url.username || url.password || url.search || url.hash || !/^\/inference\/(?:openai\/)?v1(?:\/(?:models|chat\/completions|responses))?\/?$/.test(url.pathname)) {
       throw new UnknownProviderProtocolError("Built-in inference requires its exact registered HTTPS endpoint.");
     }
   }
@@ -341631,7 +341631,7 @@ var PROVIDER_PRESETS = [
 ];
 
 // packages/backend-vllm/dist/builtin-free-credentials.js
-var BUILTIN_INFERENCE_ENDPOINT = "https://opencode.ai/inference/v1";
+var BUILTIN_INFERENCE_ENDPOINTS = /* @__PURE__ */ new Set(["https://opencode.ai/inference/v1", "https://opencode.ai/inference/openai/v1"]);
 function hostEnvironment() {
   return globalThis.process?.env ?? {};
 }
@@ -341649,7 +341649,7 @@ function resolveBuiltInFreeApiKey(baseUrl, env = hostEnvironment()) {
   } catch {
     return void 0;
   }
-  if (url.username || url.password || url.search || url.hash || url.href.replace(/\/+$/, "") !== BUILTIN_INFERENCE_ENDPOINT)
+  if (url.username || url.password || url.search || url.hash || !BUILTIN_INFERENCE_ENDPOINTS.has(url.href.replace(/\/+$/, "")))
     return void 0;
   const key = env["OPENCODE_API_KEY"];
   if (key === void 0 || key === "" || key === "public")
@@ -343363,7 +343363,11 @@ init_dist5();
 // packages/orchestrator/dist/opencode-free-models.js
 import { createHash as createHash44, randomBytes as randomBytes6 } from "node:crypto";
 var OPENCODE_MODEL_CATALOG_URL = "https://models.opencode.ai/api.json";
-var OPENCODE_INFERENCE_BASE_URL = "https://opencode.ai/inference/v1";
+var OPENCODE_LEGACY_INFERENCE_BASE_URL = "https://opencode.ai/inference/v1";
+var OPENCODE_INFERENCE_BASE_URL = "https://opencode.ai/inference/openai/v1";
+function isOpenCodeInferenceBaseUrl(value) {
+  return value === OPENCODE_INFERENCE_BASE_URL || value === OPENCODE_LEGACY_INFERENCE_BASE_URL;
+}
 var OPENCODE_AVAILABLE_MODELS_URL = `${OPENCODE_INFERENCE_BASE_URL}/models`;
 var OPENCODE_COMPATIBILITY_CLIENT_VERSION = "2.0.27";
 var observedModelCompatibility = Object.freeze({
@@ -343382,6 +343386,10 @@ var observedModelCompatibility = Object.freeze({
     // Production chat + review accepted the unchanged host JSON contract only
     // when native response_format was omitted, observed 2026-10-13.
     "big-pickle": Object.freeze({ nativeStructuredOutput: false }),
+    // Canonical chat rejects the real child completion schema unless every
+    // object forbids additional properties (2026-10-17). Preserve its original
+    // host contract through prompt-only framing instead of tightening it.
+    "beam-free": Object.freeze({ nativeStructuredOutput: false }),
     // Step accepts simple native JSON but rejects complex Telegram control
     // schemas with HTTP400. The identical generation + review contract passed
     // when only native framing was omitted (2026-10-13); prefer prompt-only
@@ -343500,7 +343508,7 @@ function tryParseOpenCodeFreeModelCatalog(json5) {
     if (!object2(catalog) || !object2(catalog.opencode))
       return void 0;
     const provider = catalog.opencode;
-    if (provider.id !== "opencode" || provider.api !== OPENCODE_INFERENCE_BASE_URL || !["@ai-sdk/openai-compatible", "@ai-sdk/openai"].includes(String(provider.npm)) || !object2(provider.models))
+    if (provider.id !== "opencode" || !isOpenCodeInferenceBaseUrl(provider.api) || !["@ai-sdk/openai-compatible", "@ai-sdk/openai"].includes(String(provider.npm)) || !object2(provider.models))
       return void 0;
     const entries = Object.entries(provider.models);
     if (entries.length > MAX_OPENCODE_CATALOG_MODELS)
@@ -343511,7 +343519,7 @@ function tryParseOpenCodeFreeModelCatalog(json5) {
         continue;
       if (model.limit.input !== void 0 && (!positiveInteger2(model.limit.input) || model.limit.input > model.limit.context))
         continue;
-      if (model.provider !== void 0 && (!object2(model.provider) || model.provider.api !== void 0 && model.provider.api !== OPENCODE_INFERENCE_BASE_URL))
+      if (model.provider !== void 0 && (!object2(model.provider) || model.provider.api !== void 0 && !isOpenCodeInferenceBaseUrl(model.provider.api)))
         continue;
       const transport = object2(model.provider) ? model.provider : {};
       const npm = transport.npm ?? provider.npm;
@@ -343752,14 +343760,15 @@ function parseKeylessInferenceRegistry(html) {
 
 // packages/orchestrator/dist/builtin-free-snapshot.js
 var rows = [
-  ["muse-spark-1.3-contributor-free", "Muse Spark 1.3 Free", 1048576, 131072, void 0, true, true],
-  ["step-5-preview-free", "Step 5 Preview Free", 1e6, 65536, 1e6, true],
-  ["nemotron-3-ultra-free", "Nemotron 3 Ultra Free", 1e6, 128e3, void 0, void 0],
-  ["mimo-v2.6-flash-free", "MiMo-V2.6-Flash Free", 2e5, 32e3, void 0, void 0],
-  ["nemotron-3.5-lightning-free", "Nemotron 3.5 Lightning Free", 262144, 262144, void 0, true],
+  ["beam-free", "Beam Free", 1048576, 131072, void 0, true],
   ["big-pickle", "Big Pickle", 2e5, 32e3, 16e4, true],
-  ["cat-jam-free", "Cat Jam Free", 262144, 131072, void 0, true],
+  ["deepseek-v4.1-flash-free", "DeepSeek V4.1 Flash Free", 2e5, 32e3, void 0, true],
+  ["glm-5.3-flash-free", "GLM-5.3-Flash Free", 2e5, 32e3, void 0, true],
   ["longcat-2.5-preview-free", "LongCat 2.5 Preview Free", 1e6, 131072, void 0, void 0],
+  ["mimo-v2.6-flash-free", "MiMo-V2.6-Flash Free", 2e5, 32e3, void 0, void 0],
+  ["nemotron-3-ultra-free", "Nemotron 3 Ultra Free", 1e6, 128e3, void 0, void 0],
+  ["nemotron-3.5-lightning-free", "Nemotron 3.5 Lightning Free", 262144, 262144, void 0, true],
+  ["nemotron-3.5-super-vl-free", "Nemotron 3.5 Super VL Free", 1048576, 131072, 1048576, void 0],
   ["space-bunny-free", "Space Bunny Free", 1048576, 524288, 524288, void 0]
 ];
 var metadata = Object.fromEntries(rows.map(([id5, name, context, output, input, structured_output, responses]) => [id5, {
@@ -343779,6 +343788,15 @@ var BUILTIN_FREE_MODEL_SNAPSHOT = Object.freeze(parseOpenCodeFreeModelCatalog(JS
   models: metadata
 } })).map((model) => Object.freeze(model)));
 var BUILTIN_FREE_MODEL_IDS = new Set(BUILTIN_FREE_MODEL_SNAPSHOT.map((model) => model.model));
+var BUILTIN_FREE_LEGACY_MODEL_IDS = /* @__PURE__ */ new Set([
+  "cat-jam-free",
+  "muse-spark-1.3-contributor-free",
+  "step-5-preview-free"
+]);
+var BUILTIN_FREE_RECOGNIZED_MODEL_IDS = /* @__PURE__ */ new Set([
+  ...BUILTIN_FREE_MODEL_IDS,
+  ...BUILTIN_FREE_LEGACY_MODEL_IDS
+]);
 
 // packages/orchestrator/dist/inference-provider-error.js
 import { createHash as createHash45 } from "node:crypto";
@@ -343988,13 +344006,13 @@ function packet(value, completeResponse = false) {
       if (delta[field] !== void 0 && delta[field] !== null && typeof delta[field] !== "string")
         throw new OpenAIChatResponseError();
     }
-    if (delta.tool_calls !== void 0 && !Array.isArray(delta.tool_calls))
+    if (delta.tool_calls !== void 0 && delta.tool_calls !== null && !Array.isArray(delta.tool_calls))
       throw new OpenAIChatResponseError();
     const calls = Array.isArray(delta.tool_calls) ? delta.tool_calls.map((call2, index) => {
-      if (!object5(call2) || call2.function !== void 0 && !object5(call2.function) || call2.id !== void 0 && typeof call2.id !== "string" || call2.index !== void 0 && (!Number.isSafeInteger(call2.index) || Number(call2.index) < 0))
+      if (!object5(call2) || call2.function !== void 0 && !object5(call2.function) || call2.id !== void 0 && !(call2.id === null && !complete) && typeof call2.id !== "string" || call2.index !== void 0 && (!Number.isSafeInteger(call2.index) || Number(call2.index) < 0))
         throw new OpenAIChatResponseError();
       const fn = object5(call2.function) ? call2.function : {};
-      if (fn.name !== void 0 && typeof fn.name !== "string")
+      if (fn.name !== void 0 && !(fn.name === null && !complete) && typeof fn.name !== "string")
         throw new OpenAIChatResponseError();
       const args = fn.arguments;
       if (args !== void 0 && typeof args !== "string" && !object5(args))
@@ -344009,11 +344027,13 @@ function packet(value, completeResponse = false) {
         if (typeof call2.id !== "string" || !call2.id.trim() || typeof fn.name !== "string" || !fn.name.trim() || !object5(parsed))
           throw new OpenAIChatResponseError();
       }
-      return {
-        ...call2,
-        index: call2.index ?? index,
-        function: { ...fn, ...object5(args) ? { arguments: JSON.stringify(args) } : {} }
-      };
+      const normalizedCall = { ...call2 };
+      const normalizedFunction = { ...fn, ...object5(args) ? { arguments: JSON.stringify(args) } : {} };
+      if (!complete && normalizedCall.id === null)
+        delete normalizedCall.id;
+      if (!complete && normalizedFunction.name === null)
+        delete normalizedFunction.name;
+      return { ...normalizedCall, index: call2.index ?? index, function: normalizedFunction };
     }) : void 0;
     return {
       ...choice,
@@ -345833,9 +345853,24 @@ var InvalidInferenceResponse = class extends Error {
 };
 var InvalidDiscoveryMetadata = class extends Error {
 };
+var CatalogModelUnavailable = class extends Error {
+};
 var OutputReservationUnavailable = class extends Error {
 };
+var InferenceRoutePolicyError = class extends Error {
+  failure;
+  code = "inference_route_failed";
+  fatal = true;
+  retryable = false;
+  constructor(failure) {
+    super(`Inference request failed (${failure.category}${failure.status === void 0 ? "" : `; HTTP ${failure.status}`}${failure.validationReason ? `; ${failure.validationReason}` : failure.reason && failure.reason !== "terminal" ? `; ${failure.reason}` : ""}).`);
+    this.failure = failure;
+    this.name = "InferenceRoutePolicyError";
+  }
+};
 var cooldowns = /* @__PURE__ */ new Map();
+var successfulRoutes = /* @__PURE__ */ new WeakMap();
+var SAME_MODEL_ATTEMPTS = 3;
 var requestTransports = /* @__PURE__ */ new WeakMap();
 var responseValidators = /* @__PURE__ */ new WeakMap();
 var nextResponseValidator = 0;
@@ -345908,12 +345943,25 @@ function errorChain(error) {
     chain.push(current2);
   return chain;
 }
+function explicitModelUnavailable(error) {
+  return errorChain(error).some((value) => {
+    if (value instanceof InferenceProviderError && value.status === 429 && value.reason === "upstream")
+      return true;
+    if (!object12(value) || value.name !== "InferenceHttpError" || ![400, 404, 422, 503].includes(Number(value.status)) || !object12(value.responseJson) || !object12(value.responseJson.error))
+      return false;
+    return ["model_not_found", "model_not_available", "model_unavailable", "unsupported_model"].includes(String(value.responseJson.error.code));
+  });
+}
+function sameModelTransient(error, request) {
+  const failure = safeFailure(error, request);
+  return !errorChain(error).some((value) => object12(value) && (value.fatal === true || value.retryable === false) || value instanceof AttemptTimeout && (value.deadlineKind === "remote_aggregate" || value.deadlineKind === "caller_request") || value instanceof SyntaxError || value instanceof InvalidInferenceResponse || value instanceof ExpectedInferenceResponseError || object12(value) && value.name === "OpenAIChatResponseError") && (failure.category === "rate_limit" || failure.category === "timeout" || failure.category === "upstream_unavailable" || failure.category === "empty_response" || failure.category === "transport" && transportFailure(error));
+}
 function safeFailure(error, request, route, streamFrames) {
   const status2 = errorStatus(error);
   const chain = errorChain(error);
   const upstream429 = status2 === 429 && chain.some((value) => value instanceof InferenceProviderError && value.reason === "upstream");
   const access = chain.find((value) => value instanceof InferenceProviderAccessError);
-  const category = access ? "access_denied" : upstream429 ? "upstream_unavailable" : status2 === 429 ? "rate_limit" : status2 === 401 || status2 === 403 ? "authentication" : status2 !== void 0 && status2 >= 400 && status2 < 500 ? "request_rejected" : status2 !== void 0 && status2 >= 500 ? "upstream_unavailable" : chain.some((value) => value instanceof AttemptTimeout || object12(value) && (value.name === "TimeoutError" || value.name === "AbortError" || value.code === "ETIMEDOUT") || value instanceof Error && /^stream timeout: no response or chunk within [0-9]+s$/.test(value.message)) ? "timeout" : chain.some((value) => value instanceof EmptyInferenceResponse) ? "empty_response" : chain.some((value) => value instanceof SyntaxError || value instanceof InvalidInferenceResponse || value instanceof ExpectedInferenceResponseError || object12(value) && value.name === "OpenAIChatResponseError") ? "invalid_response" : "transport";
+  const category = chain.some((value) => value instanceof CatalogModelUnavailable) ? "model_unavailable" : access ? "access_denied" : upstream429 ? "upstream_unavailable" : status2 === 429 ? "rate_limit" : status2 === 401 || status2 === 403 ? "authentication" : status2 !== void 0 && status2 >= 400 && status2 < 500 ? "request_rejected" : status2 !== void 0 && status2 >= 500 ? "upstream_unavailable" : chain.some((value) => value instanceof AttemptTimeout || object12(value) && (value.name === "TimeoutError" || value.name === "AbortError" || value.code === "ETIMEDOUT") || value instanceof Error && /^stream timeout: no response or chunk within [0-9]+s$/.test(value.message)) ? "timeout" : chain.some((value) => value instanceof EmptyInferenceResponse) ? "empty_response" : chain.some((value) => value instanceof SyntaxError || value instanceof InvalidInferenceResponse || value instanceof ExpectedInferenceResponseError || object12(value) && value.name === "OpenAIChatResponseError") ? "invalid_response" : "transport";
   const validation = chain.find((value) => value instanceof ExpectedInferenceResponseError);
   const providerSchemaError = chain.find((value) => value instanceof InferenceProviderError && value.reason === "unsupported_schema");
   const malformedEnvelope = chain.some((value) => object12(value) && value.name === "OpenAIChatResponseError");
@@ -345923,6 +345971,7 @@ function safeFailure(error, request, route, streamFrames) {
   return {
     category,
     ...status2 === void 0 ? {} : { status: status2 },
+    ...chain.some((value) => value instanceof CatalogModelUnavailable) ? { catalogReason: "not_in_active_free_catalog" } : {},
     ...access instanceof InferenceProviderAccessError ? { accessReason: access.reason, responseBodySha256: access.responseBodySha256 } : {},
     ...deadline instanceof AttemptTimeout ? { deadlineKind: deadline.deadlineKind } : providerIdle ? { deadlineKind: "provider_idle" } : {},
     ...streamFrames ? {
@@ -346006,6 +346055,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
   policy;
   builtInPrimaryRequest = false;
   builtInAlternativesApproved = false;
+  stickyRouting = false;
   now;
   env;
   lastRoute;
@@ -346038,6 +346088,55 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
   capacityPolicyScope() {
     return createHash47("sha256").update(canonicalContract(this.policy)).digest("hex");
   }
+  stickyKey() {
+    return createHash47("sha256").update(canonicalContract({
+      primary: this.capacityPrimaryScope(),
+      policy: this.capacityPolicyScope(),
+      credential: resolveBuiltInFreeApiKey(OPENCODE_INFERENCE_BASE_URL, this.env) ?? "public"
+    })).digest("hex");
+  }
+  rememberRoute(route) {
+    if (!this.stickyRouting && !this.builtInPrimaryRequest)
+      return;
+    const fetcher = this.options.fetch ?? fetch;
+    let entries = successfulRoutes.get(fetcher);
+    if (!entries)
+      successfulRoutes.set(fetcher, entries = /* @__PURE__ */ new Map());
+    if (entries.size >= 64)
+      entries.delete(entries.keys().next().value);
+    const { policyAttempt: _attempt, transientFailures: _failures, retryPendingMs: _retry, schemaRecoveryPending: _schema, modelAccessRestricted: _access, lastAttemptRateLimited: _rate, ...retained } = route;
+    entries.set(this.stickyKey(), retained);
+  }
+  /** One bounded decision for both unary and streaming attempts. No transcript writes. */
+  handleStickyFailure(request, candidate, deadline, error, frames) {
+    let disposition = "surface_error";
+    const category = safeFailure(error, request, candidate, frames).category;
+    let reason = category === "authentication" || category === "access_denied" ? "authentication" : category === "request_rejected" ? "request_rejected" : category === "invalid_response" ? "response_contract" : category === "transport" ? "adapter_or_configuration" : "terminal";
+    let retryDelayMs;
+    if (this.learnNativeSchemaRestriction(request, candidate, error)) {
+      disposition = "repair_same_model";
+      reason = "unsupported_capability";
+    } else if (explicitModelUnavailable(error) || errorChain(error).some((value) => value instanceof InferenceProviderError && value.reason === "unsupported_schema")) {
+      disposition = "rotate_model";
+      reason = explicitModelUnavailable(error) ? "model_unavailable" : "unsupported_capability";
+      const entries = successfulRoutes.get(this.options.fetch ?? fetch), key = this.stickyKey();
+      if (entries?.get(key) && routeKey(entries.get(key)) === routeKey(candidate))
+        entries.delete(key);
+    } else if (sameModelTransient(error, request)) {
+      reason = "transient";
+      candidate.transientFailures = (candidate.transientFailures ?? 0) + 1;
+      const hint = errorChain(error).find((value) => object12(value) && typeof value.retryAfterMs === "number" && Number.isFinite(value.retryAfterMs));
+      retryDelayMs = Math.max(1e3 * 2 ** (candidate.transientFailures - 1), object12(hint) && typeof hint.retryAfterMs === "number" ? Math.max(0, hint.retryAfterMs) : 0);
+      if (candidate.transientFailures < SAME_MODEL_ATTEMPTS && retryDelayMs < deadline - this.now()) {
+        disposition = "retry_same_model";
+        candidate.retryPendingMs = retryDelayMs;
+      }
+    }
+    const detail = { disposition, reason, attempt: candidate.policyAttempt ?? 1, ...retryDelayMs === void 0 ? {} : { retryDelayMs } };
+    this.lifecycle(request, "failed", candidate, error, frames, void 0, detail);
+    if (disposition === "surface_error")
+      throw new InferenceRoutePolicyError({ ...safeFailure(error, request, candidate, frames), ...detail });
+  }
   /** Read-only route discovery; never loads or generates with any local/remote model. */
   async prepareContextCapacity(request, options = {}) {
     if (request.configuredEndpointOnly === true || request.maxTokens <= 0 || !Number.isFinite(request.maxTokens))
@@ -346160,7 +346259,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     return localInferenceBrokerReason(reason);
   }
   builtInScope() {
-    if (this.primary.providerId !== "builtin-free" || this.primary.baseUrl !== OPENCODE_INFERENCE_BASE_URL)
+    if (this.primary.providerId !== "builtin-free" || !isOpenCodeInferenceBaseUrl(this.primary.baseUrl))
       throw new Error("Built-in inference requires its exact registered public endpoint.");
     this.reloadPolicy();
     const scope = new _RemoteInferenceFallback(this.primary, this.makeBackend, { ...this.options, policy: this.policy });
@@ -346256,7 +346355,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     request = { ...request };
     requestTransports.set(request, "unary");
     const completedBinding = completedCapacityBinding(request);
-    const deadline = Math.min(this.now() + this.policy.totalTimeoutMs, this.preparedCapacity(request)?.deadline ?? Infinity);
+    const deadline = Math.min(this.now() + this.policy.totalTimeoutMs, this.preparedCapacity(request)?.deadline ?? Infinity, request.timeoutMs > 0 ? this.now() + request.timeoutMs : Infinity);
     for await (const candidate of this.retryableCandidates(request, deadline)) {
       if (this.now() >= deadline)
         break;
@@ -346276,6 +346375,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
           this.fenceRemoteDeadline(bounded3, deadline);
           return remote.chatCompletion(remoteRequest);
         }, remaining <= this.policy.requestTimeoutMs ? "remote_aggregate" : "remote_attempt");
+        this.fenceRemoteDeadline(request, deadline);
         if (!usableResult(result))
           throw new EmptyInferenceResponse();
         const validation = validateRemoteMessage(result.choices[0]?.message, request);
@@ -346286,6 +346386,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
         cooldowns.delete(routeKey(candidate));
         cooldowns.delete(rateLimitKey(candidate));
         cooldowns.delete(contractCooldownKey(candidate, request));
+        this.rememberRoute(candidate);
         this.lifecycle(request, "succeeded", candidate, void 0, void 0, completedBinding);
         return this.scrubber(candidate)(result);
       } catch (error) {
@@ -346294,6 +346395,10 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
           this.discovery(request, candidate, { stage: "eligibility", outcome: "no_eligible_models", modelCount: 0 });
           continue;
         }
+        if (this.stickyRouting || this.builtInPrimaryRequest) {
+          this.handleStickyFailure(request, candidate, deadline, error);
+          continue;
+        }
         this.lifecycle(request, "failed", candidate, error);
         if (this.restrictedModel(candidate, error))
           continue;
@@ -346373,7 +346478,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     requestTransports.set(request, "stream");
     const completedBinding = completedCapacityBinding(request);
     let emitted = false;
-    const deadline = Math.min(this.now() + this.policy.totalTimeoutMs, this.preparedCapacity(request)?.deadline ?? Infinity);
+    const deadline = Math.min(this.now() + this.policy.totalTimeoutMs, this.preparedCapacity(request)?.deadline ?? Infinity, request.timeoutMs > 0 ? this.now() + request.timeoutMs : Infinity);
     for await (const candidate of this.retryableCandidates(request, deadline)) {
       if (this.now() >= deadline)
         break;
@@ -346411,9 +346516,11 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
         for await (const chunk of this.scrubStream(track(), candidate)) {
           yield chunk;
         }
+        this.fenceRemoteDeadline(request, deadline);
         cooldowns.delete(routeKey(candidate));
         cooldowns.delete(rateLimitKey(candidate));
         cooldowns.delete(contractCooldownKey(candidate, request));
+        this.rememberRoute(candidate);
         this.lifecycle(request, "succeeded", candidate, void 0, void 0, completedBinding);
         return;
       } catch (error) {
@@ -346422,9 +346529,15 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
           this.discovery(request, candidate, { stage: "eligibility", outcome: "no_eligible_models", modelCount: 0 });
           continue;
         }
-        this.lifecycle(request, "failed", candidate, error, frames);
-        if (emitted)
+        if (emitted) {
+          this.lifecycle(request, "failed", candidate, error, frames);
           throw new StreamInterruptedError(safeFailure(error, request, candidate, frames));
+        }
+        if (this.stickyRouting || this.builtInPrimaryRequest) {
+          this.handleStickyFailure(request, candidate, deadline, error, frames);
+          continue;
+        }
+        this.lifecycle(request, "failed", candidate, error, frames);
         if (this.restrictedModel(candidate, error))
           continue;
         if (errorChain(error).some((value) => value instanceof AttemptTimeout && value.deadlineKind === "remote_aggregate"))
@@ -346687,7 +346800,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     }
     this.lifecycle(request, "selected", route);
   }
-  lifecycle(request, phase, route, error, frames, completedBinding) {
+  lifecycle(request, phase, route, error, frames, completedBinding, decision2) {
     this.fence(request);
     if (!request.onInferenceRouteLifecycle)
       return;
@@ -346707,6 +346820,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
       } } : {},
       ...phase === "failed" ? { failure: {
         ...safeFailure(error, request, route, frames),
+        ...decision2,
         ...!route && this.localBrokerReason(error) ? { brokerReason: this.localBrokerReason(error) } : {}
       } } : {}
     };
@@ -347074,6 +347188,14 @@ ${line}`).join("");
     };
     for await (const candidate of candidates()) {
       const key = routeKey(candidate);
+      if (candidate.baseUrl === OPENCODE_INFERENCE_BASE_URL && candidate.keyless?.clientIdentity === "opencode")
+        this.stickyRouting = true;
+      if (this.stickyRouting || this.builtInPrimaryRequest) {
+        if (this.now() >= deadline)
+          return;
+        yield* this.stickyAttempts(request, candidate, deadline);
+        continue;
+      }
       if (!priority.has(key))
         priority.set(key, priority.size);
       if ((routeCooldown(candidate, request)?.until ?? 0) > this.now()) {
@@ -347137,7 +347259,62 @@ ${line}`).join("");
         defer(candidate);
     }
   }
+  async *stickyAttempts(request, candidate, deadline) {
+    let schemaRetries = 0;
+    while (this.now() < deadline) {
+      this.fence(request);
+      candidate.policyAttempt = (candidate.policyAttempt ?? 0) + 1;
+      yield candidate;
+      if (candidate.schemaRecoveryPending && schemaRetries < SUPPORTED_PROVIDER_SCHEMA_ADAPTATIONS.length) {
+        candidate.schemaRecoveryPending = false;
+        schemaRetries++;
+        continue;
+      }
+      if (candidate.retryPendingMs === void 0)
+        return;
+      const delay2 = candidate.retryPendingMs;
+      candidate.retryPendingMs = void 0;
+      await this.waitForRetry(request, delay2);
+      this.fence(request);
+      if (this.now() >= deadline)
+        throw new InferenceRoutePolicyError({ category: "timeout", deadlineKind: "remote_aggregate", disposition: "surface_error", reason: "transient", attempt: candidate.policyAttempt });
+    }
+  }
+  async waitForRetry(request, delay2) {
+    let timer;
+    try {
+      const pause = new Promise((resolve40) => {
+        timer = setTimeout(resolve40, delay2);
+      });
+      await (request.signal ? this.abortable(pause, request.signal) : pause);
+    } finally {
+      clearTimeout(timer);
+    }
+  }
   async *candidates(request, deadline, includeCooling = false) {
+    const sticky = successfulRoutes.get(this.options.fetch ?? fetch)?.get(this.stickyKey());
+    const eligible = sticky && this.compatible(request, sticky, true) ? sticky : void 0;
+    const pending = [];
+    let matched = !eligible;
+    for await (const candidate of this.catalogCandidates(request, deadline, includeCooling)) {
+      if (!matched && routeKey(candidate) !== routeKey(eligible)) {
+        pending.push(candidate);
+        continue;
+      }
+      if (!matched) {
+        this.stickyRouting = true;
+        matched = true;
+        yield candidate;
+        for (const prior of pending)
+          yield prior;
+        pending.length = 0;
+      } else
+        yield candidate;
+    }
+    for (const candidate of pending)
+      yield candidate;
+  }
+  async *catalogCandidates(request, deadline, includeCooling = false) {
     const seen = /* @__PURE__ */ new Set();
     const configured = [];
     for (const profile3 of this.policy.providers.slice(0, MAX_CONFIGURED_PROVIDERS)) {
@@ -347312,9 +347489,23 @@ ${line}`).join("");
         const zeroPrice = (value) => (typeof value === "number" || typeof value === "string" && value.trim().length > 0) && Number.isFinite(Number(value)) && Number(value) === 0;
         const models = raw.data.filter((model) => chatCatalogModel(model, anonymousTier) && (!kiloCatalog || object12(model.pricing) && zeroPrice(model.pricing.prompt) && zeroPrice(model.pricing.completion) && Object.values(model.pricing).every(zeroPrice)));
         const unique4 = [...new Map(models.map((model) => [String(model.id), model])).values()];
-        const preferredModel = this.builtInPrimaryRequest && openCodeCatalog ? this.primary.model : route.model;
+        const remembered = successfulRoutes.get(fetcher)?.get(this.stickyKey());
+        const rememberedHere = remembered && remembered.baseUrl === route.baseUrl && this.compatible(request, remembered, true);
+        const preferredModel = rememberedHere ? remembered.model : this.builtInPrimaryRequest && openCodeCatalog ? this.primary.model : route.model;
         const preferred = unique4.find((model) => model.id === preferredModel);
-        const ordered = unique4.sort((a, b) => this.builtInPrimaryRequest && openCodeCatalog && a.id === preferredModel !== (b.id === preferredModel) ? a.id === preferredModel ? -1 : 1 : compareFallbackCatalogModels(a, b, route.model));
+        if (openCodeCatalog && !preferred && (rememberedHere || this.builtInPrimaryRequest) && cache3.get(modelsUrl)?.source !== "bundled") {
+          const removed = { ...route, model: preferredModel };
+          this.discovery(request, removed, {
+            stage: "eligibility",
+            outcome: "unavailable",
+            modelCount: 0,
+            reason: "selected_model_not_in_active_free_catalog"
+          });
+          this.lifecycle(request, "failed", removed, new CatalogModelUnavailable(), void 0, void 0, { disposition: "rotate_model", reason: "model_unavailable", attempt: 0 });
+          if (rememberedHere)
+            successfulRoutes.get(fetcher)?.delete(this.stickyKey());
+        }
+        const ordered = unique4.sort((a, b) => (rememberedHere || this.builtInPrimaryRequest) && openCodeCatalog && a.id === preferredModel !== (b.id === preferredModel) ? a.id === preferredModel ? -1 : 1 : compareFallbackCatalogModels(a, b, route.model));
         const compatible = [];
         const modelLimit = openCodeCatalog ? MAX_OPENCODE_CATALOG_MODELS : MAX_MODELS_PER_PROVIDER;
         for (const selected of ordered) {
@@ -352995,11 +353186,6 @@ var StructuredMemoryExtractionQueue = class {
   }
 };
 
-// packages/orchestrator/dist/inference-provider-presentation.js
-function inferenceProviderDisplayName(providerId) {
-  return providerId === "opencode" || providerId === "builtin-free" ? "built-in free" : providerId ?? "remote";
-}
-
 // packages/orchestrator/dist/native-ollama-tool-calls.js
 function object13(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
@@ -393265,12 +393451,6 @@ ${read.content}`;
         onInferenceRoute: (route) => {
           assertCurrent();
           outbound.onInferenceRoute?.(route);
-          this.emit({
-            type: "status",
-            content: `Local inference unavailable; using ${inferenceProviderDisplayName(route.providerId)} (${route.model}).`,
-            turn: operation.turn,
-            timestamp: (/* @__PURE__ */ new Date()).toISOString()
-          });
         },
         onInferenceRouteLifecycle: (receipt) => {
           assertCurrent();
@@ -427998,7 +428178,7 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
     this._inferenceFallbackApprovalHandler = handler;
   }
   isBuiltInPrimary() {
-    return this._provider.providerId === "builtin-free" && this._provider.urls.chat === `${OPENCODE_INFERENCE_BASE_URL}/chat/completions`;
+    return this._provider.providerId === "builtin-free" && isOpenCodeInferenceBaseUrl(this._provider.urls.chat?.replace(/\/chat\/completions$/, ""));
   }
   remoteFallback() {
     return this._remoteFallback ??= new RemoteInferenceFallback({
@@ -428378,7 +428558,7 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
     };
   }
   usesOpenCodeCompatibility(request) {
-    return request.remoteClientIdentity === "opencode" && request.configuredEndpointOnly === true && this._provider.protocol === "openai-chat" && this._provider.urls.chat === `${OPENCODE_INFERENCE_BASE_URL}/chat/completions`;
+    return request.remoteClientIdentity === "opencode" && request.configuredEndpointOnly === true && this._provider.protocol === "openai-chat" && isOpenCodeInferenceBaseUrl(this._provider.urls.chat?.replace(/\/chat\/completions$/, ""));
   }
   usesOpenCodeResponses(request) {
     return this.usesOpenCodeCompatibility(request) && request.remoteOpenCodeApiShape === "responses";