@code-yeongyu/senpi-ai 2026.9.3 → 2026.9.4-3

package/dist/api/context-room.d.ts

@@ -0,0 +1,20 @@
+import type { Api, Context, Model } from "../types.ts";
+export declare const CONTEXT_SAFETY_TOKENS = 4096;
+/** Tokens always left for the answer when a thinking budget shares the response ceiling. */
+export declare const MIN_ANSWER_TOKENS = 1024;
+/** Windows too small to hold the safety margin plus one answer keep the legacy one-token floor. */
+export declare const CONTEXT_GUARD_MIN_WINDOW: number;
+export declare class ContextWindowExhaustedError extends Error {
+    readonly estimatedTokens: number;
+    readonly contextWindow: number;
+    constructor(estimatedTokens: number, contextWindow: number);
+}
+/**
+ * Fit the requested output budget into the room the context leaves in the model window.
+ * For windows of at least {@link CONTEXT_GUARD_MIN_WINDOW} tokens this throws
+ * {@link ContextWindowExhaustedError} instead of shrinking the budget below
+ * {@link MIN_ANSWER_TOKENS}: such a request can only return a truncated tool call or an
+ * empty "length" stop while still billing the whole prompt.
+ */
+export declare function clampMaxTokensToContext(model: Model<Api>, context: Context, maxTokens: number): number;
+//# sourceMappingURL=context-room.d.ts.map

package/dist/api/context-room.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"context-room.d.ts","sourceRoot":"","sources":["../../src/api/context-room.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,GAAG,EAAE,OAAO,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAGvD,eAAO,MAAM,qBAAqB,OAAO,CAAC;AAC1C,4FAA4F;AAC5F,eAAO,MAAM,iBAAiB,OAAO,CAAC;AAEtC,mGAAmG;AACnG,eAAO,MAAM,wBAAwB,QAA4C,CAAC;AAElF,qBAAa,2BAA4B,SAAQ,KAAK;IACrD,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;gBAEnB,eAAe,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM;CAU1D;AAED;;;;;;GAMG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,OAAO,EAAE,OAAO,EAAE,SAAS,EAAE,MAAM,GAAG,MAAM,CActG"}

package/dist/api/context-room.js

@@ -0,0 +1,40 @@
+import { estimateContextTokens } from "../utils/estimate.js";
+export const CONTEXT_SAFETY_TOKENS = 4096;
+/** Tokens always left for the answer when a thinking budget shares the response ceiling. */
+export const MIN_ANSWER_TOKENS = 1024;
+const MIN_MAX_TOKENS = 1;
+/** Windows too small to hold the safety margin plus one answer keep the legacy one-token floor. */
+export const CONTEXT_GUARD_MIN_WINDOW = CONTEXT_SAFETY_TOKENS + MIN_ANSWER_TOKENS;
+export class ContextWindowExhaustedError extends Error {
+    constructor(estimatedTokens, contextWindow) {
+        super(`Context window exhausted: the conversation is estimated at ${estimatedTokens} of ${contextWindow} tokens, ` +
+            `leaving fewer than ${MIN_ANSWER_TOKENS} tokens for a response. ` +
+            "Compact the conversation, enable auto-compaction, or start a new session before retrying.");
+        this.name = "ContextWindowExhaustedError";
+        this.estimatedTokens = estimatedTokens;
+        this.contextWindow = contextWindow;
+    }
+}
+/**
+ * Fit the requested output budget into the room the context leaves in the model window.
+ * For windows of at least {@link CONTEXT_GUARD_MIN_WINDOW} tokens this throws
+ * {@link ContextWindowExhaustedError} instead of shrinking the budget below
+ * {@link MIN_ANSWER_TOKENS}: such a request can only return a truncated tool call or an
+ * empty "length" stop while still billing the whole prompt.
+ */
+export function clampMaxTokensToContext(model, context, maxTokens) {
+    if (model.contextWindow <= 0) {
+        return Number.isFinite(maxTokens) && maxTokens > 0
+            ? Math.max(MIN_MAX_TOKENS, Math.floor(maxTokens))
+            : MIN_MAX_TOKENS;
+    }
+    const estimatedTokens = estimateContextTokens(context).tokens;
+    const available = model.contextWindow - estimatedTokens - CONTEXT_SAFETY_TOKENS;
+    if (model.contextWindow >= CONTEXT_GUARD_MIN_WINDOW && available < MIN_ANSWER_TOKENS) {
+        throw new ContextWindowExhaustedError(estimatedTokens, model.contextWindow);
+    }
+    const safeAvailable = Math.max(MIN_MAX_TOKENS, available);
+    const requested = Number.isFinite(maxTokens) && maxTokens > 0 ? Math.floor(maxTokens) : safeAvailable;
+    return Math.min(requested, safeAvailable);
+}
+//# sourceMappingURL=context-room.js.map

package/dist/api/context-room.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"context-room.js","sourceRoot":"","sources":["../../src/api/context-room.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,qBAAqB,EAAE,MAAM,sBAAsB,CAAC;AAE7D,MAAM,CAAC,MAAM,qBAAqB,GAAG,IAAI,CAAC;AAC1C,4FAA4F;AAC5F,MAAM,CAAC,MAAM,iBAAiB,GAAG,IAAI,CAAC;AACtC,MAAM,cAAc,GAAG,CAAC,CAAC;AACzB,mGAAmG;AACnG,MAAM,CAAC,MAAM,wBAAwB,GAAG,qBAAqB,GAAG,iBAAiB,CAAC;AAElF,MAAM,OAAO,2BAA4B,SAAQ,KAAK;IAIrD,YAAY,eAAuB,EAAE,aAAqB;QACzD,KAAK,CACJ,8DAA8D,eAAe,OAAO,aAAa,WAAW;YAC3G,sBAAsB,iBAAiB,0BAA0B;YACjE,2FAA2F,CAC5F,CAAC;QACF,IAAI,CAAC,IAAI,GAAG,6BAA6B,CAAC;QAC1C,IAAI,CAAC,eAAe,GAAG,eAAe,CAAC;QACvC,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;IACpC,CAAC;CACD;AAED;;;;;;GAMG;AACH,MAAM,UAAU,uBAAuB,CAAC,KAAiB,EAAE,OAAgB,EAAE,SAAiB;IAC7F,IAAI,KAAK,CAAC,aAAa,IAAI,CAAC,EAAE,CAAC;QAC9B,OAAO,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,SAAS,GAAG,CAAC;YACjD,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,cAAc,EAAE,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;YACjD,CAAC,CAAC,cAAc,CAAC;IACnB,CAAC;IACD,MAAM,eAAe,GAAG,qBAAqB,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC;IAC9D,MAAM,SAAS,GAAG,KAAK,CAAC,aAAa,GAAG,eAAe,GAAG,qBAAqB,CAAC;IAChF,IAAI,KAAK,CAAC,aAAa,IAAI,wBAAwB,IAAI,SAAS,GAAG,iBAAiB,EAAE,CAAC;QACtF,MAAM,IAAI,2BAA2B,CAAC,eAAe,EAAE,KAAK,CAAC,aAAa,CAAC,CAAC;IAC7E,CAAC;IACD,MAAM,aAAa,GAAG,IAAI,CAAC,GAAG,CAAC,cAAc,EAAE,SAAS,CAAC,CAAC;IAC1D,MAAM,SAAS,GAAG,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC;IACtG,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,aAAa,CAAC,CAAC;AAC3C,CAAC","sourcesContent":["import type { Api, Context, Model } from \"../types.ts\";\nimport { estimateContextTokens } from \"../utils/estimate.ts\";\n\nexport const CONTEXT_SAFETY_TOKENS = 4096;\n/** Tokens always left for the answer when a thinking budget shares the response ceiling. */\nexport const MIN_ANSWER_TOKENS = 1024;\nconst MIN_MAX_TOKENS = 1;\n/** Windows too small to hold the safety margin plus one answer keep the legacy one-token floor. */\nexport const CONTEXT_GUARD_MIN_WINDOW = CONTEXT_SAFETY_TOKENS + MIN_ANSWER_TOKENS;\n\nexport class ContextWindowExhaustedError extends Error {\n\treadonly estimatedTokens: number;\n\treadonly contextWindow: number;\n\n\tconstructor(estimatedTokens: number, contextWindow: number) {\n\t\tsuper(\n\t\t\t`Context window exhausted: the conversation is estimated at ${estimatedTokens} of ${contextWindow} tokens, ` +\n\t\t\t\t`leaving fewer than ${MIN_ANSWER_TOKENS} tokens for a response. ` +\n\t\t\t\t\"Compact the conversation, enable auto-compaction, or start a new session before retrying.\",\n\t\t);\n\t\tthis.name = \"ContextWindowExhaustedError\";\n\t\tthis.estimatedTokens = estimatedTokens;\n\t\tthis.contextWindow = contextWindow;\n\t}\n}\n\n/**\n * Fit the requested output budget into the room the context leaves in the model window.\n * For windows of at least {@link CONTEXT_GUARD_MIN_WINDOW} tokens this throws\n * {@link ContextWindowExhaustedError} instead of shrinking the budget below\n * {@link MIN_ANSWER_TOKENS}: such a request can only return a truncated tool call or an\n * empty \"length\" stop while still billing the whole prompt.\n */\nexport function clampMaxTokensToContext(model: Model<Api>, context: Context, maxTokens: number): number {\n\tif (model.contextWindow <= 0) {\n\t\treturn Number.isFinite(maxTokens) && maxTokens > 0\n\t\t\t? Math.max(MIN_MAX_TOKENS, Math.floor(maxTokens))\n\t\t\t: MIN_MAX_TOKENS;\n\t}\n\tconst estimatedTokens = estimateContextTokens(context).tokens;\n\tconst available = model.contextWindow - estimatedTokens - CONTEXT_SAFETY_TOKENS;\n\tif (model.contextWindow >= CONTEXT_GUARD_MIN_WINDOW && available < MIN_ANSWER_TOKENS) {\n\t\tthrow new ContextWindowExhaustedError(estimatedTokens, model.contextWindow);\n\t}\n\tconst safeAvailable = Math.max(MIN_MAX_TOKENS, available);\n\tconst requested = Number.isFinite(maxTokens) && maxTokens > 0 ? Math.floor(maxTokens) : safeAvailable;\n\treturn Math.min(requested, safeAvailable);\n}\n"]}

package/dist/api/cursor-agent.js

@@ -532,10 +532,6 @@ export const stream = (model, context, options) => {
             if (h2Settled)
                 return;
             h2Settled = true;
-            if (streamHealthTimer) {
-                clearTimeout(streamHealthTimer);
-                streamHealthTimer = null;
-            }
             if (error !== undefined) {
                 rejectH2(error);
                 return;
@@ -958,6 +954,10 @@ export const stream = (model, context, options) => {
                     clearInterval(heartbeatTimer);
                     heartbeatTimer = null;
                 }
+                if (streamHealthTimer) {
+                    clearTimeout(streamHealthTimer);
+                    streamHealthTimer = null;
+                }
                 h2Request?.close();
                 h2Client?.close();
                 h2Request = null;