omnius 1.0.905 → 1.0.906

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 39009730 bytes
-sha256: d59577959512a0d5ed8a9cb0f7d9c73caee41656b6d0d2b9dc4d29fc27c2de6a
+size: 39000951 bytes
+sha256: 3c496332771e553c64eb7f05f2a3e919c35e70795e46a92b38b1715c09ec1a96

package/dist/library.js

@@ -345149,21 +345149,7 @@ 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, providerLabel) {
-    const provider = failure.category === "authentication" && providerLabel && /^[A-Za-z0-9][A-Za-z0-9 ._()-]{0,63}$/.test(providerLabel) ? providerLabel : "Inference";
-    super(`${provider} request failed (${failure.category}${failure.status === void 0 ? "" : `; HTTP ${failure.status}`}${failure.authReason === "expired_key" ? "; API key expired" : failure.accessReason ? `; ${failure.accessReason}` : 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;
@@ -345233,19 +345219,6 @@ 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 (!object8(value) || value.name !== "InferenceHttpError" || ![400, 404, 422, 503].includes(Number(value.status)) || !object8(value.responseJson) || !object8(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) => object8(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 || object8(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);
@@ -345345,7 +345318,6 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
   makeBackend;
   options;
   policy;
-  stickyRouting = false;
   now;
   env;
   lastRoute;
@@ -345378,51 +345350,6 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
   capacityPolicyScope() {
     return createHash45("sha256").update(canonicalContract(this.policy)).digest("hex");
   }
-  stickyKey() {
-    return createHash45("sha256").update(canonicalContract({ primary: this.capacityPrimaryScope(), policy: this.capacityPolicyScope() })).digest("hex");
-  }
-  rememberRoute(route) {
-    if (!this.stickyRouting)
-      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" ? "authentication" : category === "access_denied" ? "access_restricted" : 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) => object8(value) && typeof value.retryAfterMs === "number" && Number.isFinite(value.retryAfterMs));
-      retryDelayMs = Math.max(1e3 * 2 ** (candidate.transientFailures - 1), object8(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 }, errorChain(error).map(inferenceAuthenticationDiagnostic).find((value) => value !== void 0)?.providerLabel ?? candidate.providerId);
-  }
   /** Read-only route discovery; never loads or generates with any local/remote model. */
   async prepareContextCapacity(request, options = {}) {
     if (request.configuredEndpointOnly === true || request.brokerResumeOnly === true || request.maxTokens <= 0 || !Number.isFinite(request.maxTokens))
@@ -345657,7 +345584,6 @@ 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) {
@@ -345666,10 +345592,6 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
           this.discovery(request, candidate, { stage: "eligibility", outcome: "no_eligible_models", modelCount: 0 });
           continue;
         }
-        if (this.stickyRouting) {
-          this.handleStickyFailure(request, candidate, deadline, error);
-          continue;
-        }
         this.lifecycle(request, "failed", candidate, error);
         if (errorChain(error).some((value) => value instanceof AttemptTimeout && value.deadlineKind === "remote_aggregate"))
           break;
@@ -345809,7 +345731,6 @@ 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;
       } catch (error) {
@@ -345822,10 +345743,6 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
           this.lifecycle(request, "failed", candidate, error, frames);
           throw new StreamInterruptedError(safeFailure(error, request, candidate, frames));
         }
-        if (this.stickyRouting) {
-          this.handleStickyFailure(request, candidate, deadline, error, frames);
-          continue;
-        }
         this.lifecycle(request, "failed", candidate, error, frames);
         if (errorChain(error).some((value) => value instanceof AttemptTimeout && value.deadlineKind === "remote_aggregate"))
           break;
@@ -346429,12 +346346,6 @@ ${line}`).join("");
     };
     for await (const candidate of candidates()) {
       const key = routeKey(candidate);
-      if (this.stickyRouting) {
-        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()) {
@@ -346498,60 +346409,8 @@ ${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((resolve42) => {
-        timer = setTimeout(resolve42, 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;
+    yield* this.catalogCandidates(request, deadline, includeCooling);
   }
   async *catalogCandidates(request, deadline, includeCooling = false) {
     const seen = /* @__PURE__ */ new Set();
@@ -346673,10 +346532,7 @@ ${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 || object8(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 remembered = successfulRoutes.get(fetcher)?.get(this.stickyKey());
-        const rememberedHere = remembered && remembered.baseUrl === route.baseUrl && this.compatible(request, remembered, true);
-        const preferredModel = rememberedHere ? remembered.model : route.model;
-        const preferred = unique4.find((model) => model.id === preferredModel);
+        const preferred = unique4.find((model) => model.id === route.model);
         const ordered = unique4.sort((a, b) => compareFallbackCatalogModels(a, b, route.model));
         const compatible = [];
         const modelLimit = MAX_MODELS_PER_PROVIDER;

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.905",
+  "version": "1.0.906",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.905",
+      "version": "1.0.906",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.905",
+  "version": "1.0.906",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",