@autorag/librarian 2.13.2 → 2.13.3

package/dist/cli/index.js

@@ -2381,6 +2381,51 @@ var init_credential = __esm(() => {
   init_models();
 });
 
+// src/cloud/context-overflow.ts
+import {
+  createAssistantMessageEventStream,
+  isContextOverflow
+} from "@earendil-works/pi-ai";
+import { getApiProvider } from "@earendil-works/pi-ai/compat";
+function recordRequestEstimate(model, estimatedTokens) {
+  lastEstimates.set(modelKey(model), estimatedTokens);
+}
+function noteContextOverflow(model) {
+  const key = modelKey(model);
+  const failed = lastEstimates.get(key);
+  if (failed === undefined)
+    return;
+  ceilings.set(key, Math.min(Math.floor(failed / 2), ceilings.get(key) ?? Number.POSITIVE_INFINITY));
+}
+function contextOverflowCeiling(model) {
+  return ceilings.get(modelKey(model));
+}
+function streamDazziSimple(model, context, options) {
+  const api = getApiProvider(model.api);
+  if (api === undefined)
+    throw new Error(`No API provider registered for api: ${model.api}`);
+  const inner = api.streamSimple(model, context, options);
+  const stream = createAssistantMessageEventStream();
+  (async () => {
+    try {
+      for await (const event of inner) {
+        if (event.type === "error" && event.error.stopReason === "error" && isContextOverflow(event.error)) {
+          noteContextOverflow(model);
+        }
+        stream.push(event);
+      }
+    } finally {
+      stream.end();
+    }
+  })();
+  return stream;
+}
+var lastEstimates, ceilings, modelKey = (model) => `${model.provider}/${model.id}`;
+var init_context_overflow = __esm(() => {
+  lastEstimates = new Map;
+  ceilings = new Map;
+});
+
 // src/cloud/authorize.ts
 function buildAuthorizeUrl(params) {
   const url = new URL(`${normalizeBaseUrl(params.baseUrl)}${AUTHORIZE_PATH}`);
@@ -2736,6 +2781,7 @@ function createDazziProviderConfig(env = process.env) {
     name: PROVIDER_NAME,
     baseUrl: endpoints.apiBaseUrl,
     api: "openai-responses",
+    streamSimple: streamDazziSimple,
     oauth: {
       name: PROVIDER_NAME,
       login: (callbacks) => loginDazzi(callbacks, { baseUrl }),
@@ -2751,6 +2797,7 @@ async function registerDazziProvider(runtime, env = process.env) {
 }
 var init_provider = __esm(() => {
   init_config2();
+  init_context_overflow();
   init_models();
   init_oauth_login();
 });
@@ -28880,12 +28927,20 @@ function withBoundedOutput(model) {
   return { ...model, maxTokens: MAX_OUTPUT_TOKENS };
 }
 function guardAgentContextMessages(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
+  const guarded = guardWithinBudget(model, messages, useReportedUsage, fixedOverheadTokens);
+  if (model !== undefined && guarded.length > 0) {
+    recordRequestEstimate(model, estimateContextMessageTokens(guarded, false) + fixedOverheadTokens);
+  }
+  return guarded;
+}
+function guardWithinBudget(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
   if (model === undefined || !Number.isFinite(model.contextWindow) || model.contextWindow <= 0 || messages.length === 0) {
     return messages;
   }
   const usableContext = Math.floor(model.contextWindow / CONTEXT_HEADROOM);
   const outputReserve = Math.min(Math.floor(usableContext / 4), MAX_OUTPUT_TOKENS, Number.isFinite(model.maxTokens) ? Math.max(0, model.maxTokens) : 0);
-  const inputBudget = Math.max(0, usableContext - outputReserve);
+  const overflowCeiling = contextOverflowCeiling(model);
+  const inputBudget = Math.min(Math.max(0, usableContext - outputReserve), overflowCeiling ?? Number.POSITIVE_INFINITY);
   const budget = Math.max(0, inputBudget - fixedOverheadTokens);
   const reported = useReportedUsage && findLastAssistantUsage(messages) !== undefined;
   if (estimateContextMessageTokens(messages, useReportedUsage) <= (reported ? inputBudget : budget)) {
@@ -29198,6 +29253,7 @@ function splitCandidates(text, sectionStart, sectionEnd, pattern = /^\[e?\d+\] [
 }
 var CONTEXT_HEADROOM = 1.2, MAX_OUTPUT_TOKENS = 32000, MIN_QUERY_TOKENS = 2000, GUARD_MARKER = "[AutoRAG context guard: omitted context to stay within the model context budget.]", RETRIEVAL_TOOL_PREFIX = "search_datasource_", RETRIEVAL_TOOL_NAMES;
 var init_context_token_guard = __esm(() => {
+  init_context_overflow();
   RETRIEVAL_TOOL_NAMES = {
     jikji_find: true,
     search_all_documents: true,

package/dist/cloud/context-overflow.d.ts

@@ -0,0 +1,29 @@
+/**
+ * Context-overflow recovery for the hosted Dazzi Models API.
+ *
+ * The hosted server reports a request that exceeds the model's window as
+ * `400 context_length_exceeded` (or HTTP 413 for an oversized body), which Pi
+ * already recognises: it compacts, then retries once.
+ *
+ * That recovery is not enough on its own. Pi compacts by summarising older
+ * turns and keeping the latest ones whole, so an oversized tool result that is
+ * the newest turn survives compaction and the retry is rejected again. This
+ * module remembers which model just overflowed and how big the rejected request
+ * was, so the context guard can hold the retry to a fraction of that size.
+ */
+import { type Api, type AssistantMessageEventStream, type Model, type SimpleStreamOptions, type TranscriptContext } from "@earendil-works/pi-ai";
+/** The context guard reports the estimated size (its chars/4 unit, system prompt and tools included) of each request it releases. */
+export declare function recordRequestEstimate(model: Pick<Model<Api>, "provider" | "id">, estimatedTokens: number): void;
+/**
+ * Remember that `model` just overflowed, so later requests are held to half the
+ * estimated size of the one that failed. The ceiling only ever ratchets down,
+ * and is not set for a model whose requests the guard never sized.
+ */
+export declare function noteContextOverflow(model: Pick<Model<Api>, "provider" | "id">): void;
+/** The most estimated tokens a request to `model` may carry after an overflow, if it has overflowed. */
+export declare function contextOverflowCeiling(model: Pick<Model<Api>, "provider" | "id">): number | undefined;
+/**
+ * Pi's built-in `openai-responses` stream, unchanged except that a context
+ * overflow it reports is noted for {@link contextOverflowCeiling}.
+ */
+export declare function streamDazziSimple(model: Model<Api>, context: TranscriptContext, options?: SimpleStreamOptions): AssistantMessageEventStream;

package/dist/core.js

@@ -15098,6 +15098,51 @@ var init_permission_gate = __esm(() => {
   };
 });
 
+// src/cloud/context-overflow.ts
+import {
+  createAssistantMessageEventStream,
+  isContextOverflow
+} from "@earendil-works/pi-ai";
+import { getApiProvider } from "@earendil-works/pi-ai/compat";
+function recordRequestEstimate(model, estimatedTokens) {
+  lastEstimates.set(modelKey(model), estimatedTokens);
+}
+function noteContextOverflow(model) {
+  const key = modelKey(model);
+  const failed = lastEstimates.get(key);
+  if (failed === undefined)
+    return;
+  ceilings.set(key, Math.min(Math.floor(failed / 2), ceilings.get(key) ?? Number.POSITIVE_INFINITY));
+}
+function contextOverflowCeiling(model) {
+  return ceilings.get(modelKey(model));
+}
+function streamDazziSimple(model, context, options) {
+  const api = getApiProvider(model.api);
+  if (api === undefined)
+    throw new Error(`No API provider registered for api: ${model.api}`);
+  const inner = api.streamSimple(model, context, options);
+  const stream = createAssistantMessageEventStream();
+  (async () => {
+    try {
+      for await (const event of inner) {
+        if (event.type === "error" && event.error.stopReason === "error" && isContextOverflow(event.error)) {
+          noteContextOverflow(model);
+        }
+        stream.push(event);
+      }
+    } finally {
+      stream.end();
+    }
+  })();
+  return stream;
+}
+var lastEstimates, ceilings, modelKey = (model) => `${model.provider}/${model.id}`;
+var init_context_overflow = __esm(() => {
+  lastEstimates = new Map;
+  ceilings = new Map;
+});
+
 // src/cloud/authorize.ts
 function buildAuthorizeUrl(params) {
   const url = new URL(`${normalizeBaseUrl(params.baseUrl)}${AUTHORIZE_PATH}`);
@@ -15453,6 +15498,7 @@ function createDazziProviderConfig(env = process.env) {
     name: PROVIDER_NAME,
     baseUrl: endpoints.apiBaseUrl,
     api: "openai-responses",
+    streamSimple: streamDazziSimple,
     oauth: {
       name: PROVIDER_NAME,
       login: (callbacks) => loginDazzi(callbacks, { baseUrl }),
@@ -15468,6 +15514,7 @@ async function registerDazziProvider(runtime, env = process.env) {
 }
 var init_provider2 = __esm(() => {
   init_config();
+  init_context_overflow();
   init_models();
   init_oauth_login();
 });
@@ -26850,12 +26897,20 @@ function withBoundedOutput(model) {
   return { ...model, maxTokens: MAX_OUTPUT_TOKENS };
 }
 function guardAgentContextMessages(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
+  const guarded = guardWithinBudget(model, messages, useReportedUsage, fixedOverheadTokens);
+  if (model !== undefined && guarded.length > 0) {
+    recordRequestEstimate(model, estimateContextMessageTokens(guarded, false) + fixedOverheadTokens);
+  }
+  return guarded;
+}
+function guardWithinBudget(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
   if (model === undefined || !Number.isFinite(model.contextWindow) || model.contextWindow <= 0 || messages.length === 0) {
     return messages;
   }
   const usableContext = Math.floor(model.contextWindow / CONTEXT_HEADROOM);
   const outputReserve = Math.min(Math.floor(usableContext / 4), MAX_OUTPUT_TOKENS, Number.isFinite(model.maxTokens) ? Math.max(0, model.maxTokens) : 0);
-  const inputBudget = Math.max(0, usableContext - outputReserve);
+  const overflowCeiling = contextOverflowCeiling(model);
+  const inputBudget = Math.min(Math.max(0, usableContext - outputReserve), overflowCeiling ?? Number.POSITIVE_INFINITY);
   const budget = Math.max(0, inputBudget - fixedOverheadTokens);
   const reported = useReportedUsage && findLastAssistantUsage(messages) !== undefined;
   if (estimateContextMessageTokens(messages, useReportedUsage) <= (reported ? inputBudget : budget)) {
@@ -27168,6 +27223,7 @@ function splitCandidates(text, sectionStart, sectionEnd, pattern = /^\[e?\d+\] [
 }
 var CONTEXT_HEADROOM = 1.2, MAX_OUTPUT_TOKENS = 32000, MIN_QUERY_TOKENS = 2000, GUARD_MARKER = "[AutoRAG context guard: omitted context to stay within the model context budget.]", RETRIEVAL_TOOL_PREFIX = "search_datasource_", RETRIEVAL_TOOL_NAMES;
 var init_context_token_guard = __esm(() => {
+  init_context_overflow();
   RETRIEVAL_TOOL_NAMES = {
     jikji_find: true,
     search_all_documents: true,

package/dist/index.js

@@ -15098,6 +15098,51 @@ var init_permission_gate = __esm(() => {
   };
 });
 
+// src/cloud/context-overflow.ts
+import {
+  createAssistantMessageEventStream,
+  isContextOverflow
+} from "@earendil-works/pi-ai";
+import { getApiProvider } from "@earendil-works/pi-ai/compat";
+function recordRequestEstimate(model, estimatedTokens) {
+  lastEstimates.set(modelKey(model), estimatedTokens);
+}
+function noteContextOverflow(model) {
+  const key = modelKey(model);
+  const failed = lastEstimates.get(key);
+  if (failed === undefined)
+    return;
+  ceilings.set(key, Math.min(Math.floor(failed / 2), ceilings.get(key) ?? Number.POSITIVE_INFINITY));
+}
+function contextOverflowCeiling(model) {
+  return ceilings.get(modelKey(model));
+}
+function streamDazziSimple(model, context, options) {
+  const api = getApiProvider(model.api);
+  if (api === undefined)
+    throw new Error(`No API provider registered for api: ${model.api}`);
+  const inner = api.streamSimple(model, context, options);
+  const stream = createAssistantMessageEventStream();
+  (async () => {
+    try {
+      for await (const event of inner) {
+        if (event.type === "error" && event.error.stopReason === "error" && isContextOverflow(event.error)) {
+          noteContextOverflow(model);
+        }
+        stream.push(event);
+      }
+    } finally {
+      stream.end();
+    }
+  })();
+  return stream;
+}
+var lastEstimates, ceilings, modelKey = (model) => `${model.provider}/${model.id}`;
+var init_context_overflow = __esm(() => {
+  lastEstimates = new Map;
+  ceilings = new Map;
+});
+
 // src/cloud/authorize.ts
 function buildAuthorizeUrl(params) {
   const url = new URL(`${normalizeBaseUrl(params.baseUrl)}${AUTHORIZE_PATH}`);
@@ -15453,6 +15498,7 @@ function createDazziProviderConfig(env = process.env) {
     name: PROVIDER_NAME,
     baseUrl: endpoints.apiBaseUrl,
     api: "openai-responses",
+    streamSimple: streamDazziSimple,
     oauth: {
       name: PROVIDER_NAME,
       login: (callbacks) => loginDazzi(callbacks, { baseUrl }),
@@ -15468,6 +15514,7 @@ async function registerDazziProvider(runtime, env = process.env) {
 }
 var init_provider2 = __esm(() => {
   init_config();
+  init_context_overflow();
   init_models();
   init_oauth_login();
 });
@@ -26850,12 +26897,20 @@ function withBoundedOutput(model) {
   return { ...model, maxTokens: MAX_OUTPUT_TOKENS };
 }
 function guardAgentContextMessages(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
+  const guarded = guardWithinBudget(model, messages, useReportedUsage, fixedOverheadTokens);
+  if (model !== undefined && guarded.length > 0) {
+    recordRequestEstimate(model, estimateContextMessageTokens(guarded, false) + fixedOverheadTokens);
+  }
+  return guarded;
+}
+function guardWithinBudget(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
   if (model === undefined || !Number.isFinite(model.contextWindow) || model.contextWindow <= 0 || messages.length === 0) {
     return messages;
   }
   const usableContext = Math.floor(model.contextWindow / CONTEXT_HEADROOM);
   const outputReserve = Math.min(Math.floor(usableContext / 4), MAX_OUTPUT_TOKENS, Number.isFinite(model.maxTokens) ? Math.max(0, model.maxTokens) : 0);
-  const inputBudget = Math.max(0, usableContext - outputReserve);
+  const overflowCeiling = contextOverflowCeiling(model);
+  const inputBudget = Math.min(Math.max(0, usableContext - outputReserve), overflowCeiling ?? Number.POSITIVE_INFINITY);
   const budget = Math.max(0, inputBudget - fixedOverheadTokens);
   const reported = useReportedUsage && findLastAssistantUsage(messages) !== undefined;
   if (estimateContextMessageTokens(messages, useReportedUsage) <= (reported ? inputBudget : budget)) {
@@ -27168,6 +27223,7 @@ function splitCandidates(text, sectionStart, sectionEnd, pattern = /^\[e?\d+\] [
 }
 var CONTEXT_HEADROOM = 1.2, MAX_OUTPUT_TOKENS = 32000, MIN_QUERY_TOKENS = 2000, GUARD_MARKER = "[AutoRAG context guard: omitted context to stay within the model context budget.]", RETRIEVAL_TOOL_PREFIX = "search_datasource_", RETRIEVAL_TOOL_NAMES;
 var init_context_token_guard = __esm(() => {
+  init_context_overflow();
   RETRIEVAL_TOOL_NAMES = {
     jikji_find: true,
     search_all_documents: true,

package/dist/mcp/index.js

@@ -15099,6 +15099,51 @@ var init_permission_gate = __esm(() => {
   };
 });
 
+// src/cloud/context-overflow.ts
+import {
+  createAssistantMessageEventStream,
+  isContextOverflow
+} from "@earendil-works/pi-ai";
+import { getApiProvider } from "@earendil-works/pi-ai/compat";
+function recordRequestEstimate(model, estimatedTokens) {
+  lastEstimates.set(modelKey(model), estimatedTokens);
+}
+function noteContextOverflow(model) {
+  const key = modelKey(model);
+  const failed = lastEstimates.get(key);
+  if (failed === undefined)
+    return;
+  ceilings.set(key, Math.min(Math.floor(failed / 2), ceilings.get(key) ?? Number.POSITIVE_INFINITY));
+}
+function contextOverflowCeiling(model) {
+  return ceilings.get(modelKey(model));
+}
+function streamDazziSimple(model, context, options) {
+  const api = getApiProvider(model.api);
+  if (api === undefined)
+    throw new Error(`No API provider registered for api: ${model.api}`);
+  const inner = api.streamSimple(model, context, options);
+  const stream = createAssistantMessageEventStream();
+  (async () => {
+    try {
+      for await (const event of inner) {
+        if (event.type === "error" && event.error.stopReason === "error" && isContextOverflow(event.error)) {
+          noteContextOverflow(model);
+        }
+        stream.push(event);
+      }
+    } finally {
+      stream.end();
+    }
+  })();
+  return stream;
+}
+var lastEstimates, ceilings, modelKey = (model) => `${model.provider}/${model.id}`;
+var init_context_overflow = __esm(() => {
+  lastEstimates = new Map;
+  ceilings = new Map;
+});
+
 // src/cloud/authorize.ts
 function buildAuthorizeUrl(params) {
   const url = new URL(`${normalizeBaseUrl(params.baseUrl)}${AUTHORIZE_PATH}`);
@@ -15454,6 +15499,7 @@ function createDazziProviderConfig(env = process.env) {
     name: PROVIDER_NAME,
     baseUrl: endpoints.apiBaseUrl,
     api: "openai-responses",
+    streamSimple: streamDazziSimple,
     oauth: {
       name: PROVIDER_NAME,
       login: (callbacks) => loginDazzi(callbacks, { baseUrl }),
@@ -15469,6 +15515,7 @@ async function registerDazziProvider(runtime, env = process.env) {
 }
 var init_provider2 = __esm(() => {
   init_config();
+  init_context_overflow();
   init_models();
   init_oauth_login();
 });
@@ -26851,12 +26898,20 @@ function withBoundedOutput(model) {
   return { ...model, maxTokens: MAX_OUTPUT_TOKENS };
 }
 function guardAgentContextMessages(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
+  const guarded = guardWithinBudget(model, messages, useReportedUsage, fixedOverheadTokens);
+  if (model !== undefined && guarded.length > 0) {
+    recordRequestEstimate(model, estimateContextMessageTokens(guarded, false) + fixedOverheadTokens);
+  }
+  return guarded;
+}
+function guardWithinBudget(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
   if (model === undefined || !Number.isFinite(model.contextWindow) || model.contextWindow <= 0 || messages.length === 0) {
     return messages;
   }
   const usableContext = Math.floor(model.contextWindow / CONTEXT_HEADROOM);
   const outputReserve = Math.min(Math.floor(usableContext / 4), MAX_OUTPUT_TOKENS, Number.isFinite(model.maxTokens) ? Math.max(0, model.maxTokens) : 0);
-  const inputBudget = Math.max(0, usableContext - outputReserve);
+  const overflowCeiling = contextOverflowCeiling(model);
+  const inputBudget = Math.min(Math.max(0, usableContext - outputReserve), overflowCeiling ?? Number.POSITIVE_INFINITY);
   const budget = Math.max(0, inputBudget - fixedOverheadTokens);
   const reported = useReportedUsage && findLastAssistantUsage(messages) !== undefined;
   if (estimateContextMessageTokens(messages, useReportedUsage) <= (reported ? inputBudget : budget)) {
@@ -27169,6 +27224,7 @@ function splitCandidates(text, sectionStart, sectionEnd, pattern = /^\[e?\d+\] [
 }
 var CONTEXT_HEADROOM = 1.2, MAX_OUTPUT_TOKENS = 32000, MIN_QUERY_TOKENS = 2000, GUARD_MARKER = "[AutoRAG context guard: omitted context to stay within the model context budget.]", RETRIEVAL_TOOL_PREFIX = "search_datasource_", RETRIEVAL_TOOL_NAMES;
 var init_context_token_guard = __esm(() => {
+  init_context_overflow();
   RETRIEVAL_TOOL_NAMES = {
     jikji_find: true,
     search_all_documents: true,

package/dist/profiles.js

@@ -968,6 +968,51 @@ var init_credential = __esm(() => {
   init_models();
 });
 
+// src/cloud/context-overflow.ts
+import {
+  createAssistantMessageEventStream,
+  isContextOverflow
+} from "@earendil-works/pi-ai";
+import { getApiProvider } from "@earendil-works/pi-ai/compat";
+function recordRequestEstimate(model, estimatedTokens) {
+  lastEstimates.set(modelKey(model), estimatedTokens);
+}
+function noteContextOverflow(model) {
+  const key = modelKey(model);
+  const failed = lastEstimates.get(key);
+  if (failed === undefined)
+    return;
+  ceilings.set(key, Math.min(Math.floor(failed / 2), ceilings.get(key) ?? Number.POSITIVE_INFINITY));
+}
+function contextOverflowCeiling(model) {
+  return ceilings.get(modelKey(model));
+}
+function streamDazziSimple(model, context, options) {
+  const api = getApiProvider(model.api);
+  if (api === undefined)
+    throw new Error(`No API provider registered for api: ${model.api}`);
+  const inner = api.streamSimple(model, context, options);
+  const stream = createAssistantMessageEventStream();
+  (async () => {
+    try {
+      for await (const event of inner) {
+        if (event.type === "error" && event.error.stopReason === "error" && isContextOverflow(event.error)) {
+          noteContextOverflow(model);
+        }
+        stream.push(event);
+      }
+    } finally {
+      stream.end();
+    }
+  })();
+  return stream;
+}
+var lastEstimates, ceilings, modelKey = (model) => `${model.provider}/${model.id}`;
+var init_context_overflow = __esm(() => {
+  lastEstimates = new Map;
+  ceilings = new Map;
+});
+
 // src/cloud/authorize.ts
 function buildAuthorizeUrl(params) {
   const url = new URL(`${normalizeBaseUrl(params.baseUrl)}${AUTHORIZE_PATH}`);
@@ -1323,6 +1368,7 @@ function createDazziProviderConfig(env = process.env) {
     name: PROVIDER_NAME,
     baseUrl: endpoints.apiBaseUrl,
     api: "openai-responses",
+    streamSimple: streamDazziSimple,
     oauth: {
       name: PROVIDER_NAME,
       login: (callbacks) => loginDazzi(callbacks, { baseUrl }),
@@ -1338,6 +1384,7 @@ async function registerDazziProvider(runtime, env = process.env) {
 }
 var init_provider = __esm(() => {
   init_config();
+  init_context_overflow();
   init_models();
   init_oauth_login();
 });
@@ -29245,12 +29292,20 @@ function withBoundedOutput(model) {
   return { ...model, maxTokens: MAX_OUTPUT_TOKENS };
 }
 function guardAgentContextMessages(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
+  const guarded = guardWithinBudget(model, messages, useReportedUsage, fixedOverheadTokens);
+  if (model !== undefined && guarded.length > 0) {
+    recordRequestEstimate(model, estimateContextMessageTokens(guarded, false) + fixedOverheadTokens);
+  }
+  return guarded;
+}
+function guardWithinBudget(model, messages, useReportedUsage = true, fixedOverheadTokens = 0) {
   if (model === undefined || !Number.isFinite(model.contextWindow) || model.contextWindow <= 0 || messages.length === 0) {
     return messages;
   }
   const usableContext = Math.floor(model.contextWindow / CONTEXT_HEADROOM);
   const outputReserve = Math.min(Math.floor(usableContext / 4), MAX_OUTPUT_TOKENS, Number.isFinite(model.maxTokens) ? Math.max(0, model.maxTokens) : 0);
-  const inputBudget = Math.max(0, usableContext - outputReserve);
+  const overflowCeiling = contextOverflowCeiling(model);
+  const inputBudget = Math.min(Math.max(0, usableContext - outputReserve), overflowCeiling ?? Number.POSITIVE_INFINITY);
   const budget = Math.max(0, inputBudget - fixedOverheadTokens);
   const reported = useReportedUsage && findLastAssistantUsage(messages) !== undefined;
   if (estimateContextMessageTokens(messages, useReportedUsage) <= (reported ? inputBudget : budget)) {
@@ -29563,6 +29618,7 @@ function splitCandidates(text, sectionStart, sectionEnd, pattern = /^\[e?\d+\] [
 }
 var CONTEXT_HEADROOM = 1.2, MAX_OUTPUT_TOKENS = 32000, MIN_QUERY_TOKENS = 2000, GUARD_MARKER = "[AutoRAG context guard: omitted context to stay within the model context budget.]", RETRIEVAL_TOOL_PREFIX = "search_datasource_", RETRIEVAL_TOOL_NAMES;
 var init_context_token_guard = __esm(() => {
+  init_context_overflow();
   RETRIEVAL_TOOL_NAMES = {
     jikji_find: true,
     search_all_documents: true,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@autorag/librarian",
-  "version": "2.13.2",
+  "version": "2.13.3",
   "description": "Personal librarian agent for office work on the Pi framework, with pluggable retrieval methods",
   "license": "Apache-2.0",
   "repository": {