billion-context 0.1.75-pr.484.303 → 0.1.75-pr.486.305

package/dist/index.js

@@ -48008,12 +48008,22 @@ import path3 from "path";
 // src/fetch-util.ts
 var MAX_REQUEST_BYTES = 100 * 1024 * 1024;
 var UPSTREAM_TIMEOUT_MS = 10 * 60 * 1e3;
+var liveUpstreamTimers = /* @__PURE__ */ new Set();
 async function fetchWithTimeout(url, opts, timeoutMs = UPSTREAM_TIMEOUT_MS, externalSignal) {
   const controller = new AbortController();
-  let timer2 = setTimeout(() => controller.abort(), timeoutMs);
+  const armTimer = () => {
+    const t = setTimeout(() => {
+      liveUpstreamTimers.delete(t);
+      controller.abort();
+    }, timeoutMs);
+    liveUpstreamTimers.add(t);
+    return t;
+  };
+  let timer2 = armTimer();
   const rearm = () => {
     clearTimeout(timer2);
-    timer2 = setTimeout(() => controller.abort(), timeoutMs);
+    liveUpstreamTimers.delete(timer2);
+    timer2 = armTimer();
   };
   let onExternalAbort = null;
   if (externalSignal) {
@@ -48025,6 +48035,7 @@ async function fetchWithTimeout(url, opts, timeoutMs = UPSTREAM_TIMEOUT_MS, exte
   }
   const cleanup = () => {
     clearTimeout(timer2);
+    liveUpstreamTimers.delete(timer2);
     if (onExternalAbort && externalSignal) externalSignal.removeEventListener("abort", onExternalAbort);
   };
   try {
@@ -54563,6 +54574,12 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
       return new Promise((r) => res.once("drain", () => r()));
     }
   };
+  const settleUsage = () => {
+    if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
+      applyUsageSample(session, acc, protocol);
+      markDirty(session);
+    }
+  };
   const pushField = (field, index, text) => {
     const s3 = filterFor(field, index);
     const clean = s3.filter.push(text);
@@ -54668,10 +54685,14 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
     }
     const rest = flushTails();
     if (rest.length > 0 && !res.destroyed && !res.writableEnded) await write(rest);
-    if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
-      applyUsageSample(session, acc, protocol);
-      markDirty(session);
+    settleUsage();
+  } catch (e) {
+    settleUsage();
+    if (res.destroyed || res.writableEnded) {
+      log2?.("client aborted mid-stream");
+      return;
     }
+    throw e;
   } finally {
     reader.releaseLock();
     res.end();
@@ -54705,6 +54726,12 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
       return new Promise((r) => res.once("drain", () => r()));
     }
   };
+  const settleUsage = () => {
+    if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
+      applyUsageSample(session, acc, "responses");
+      markDirty(session);
+    }
+  };
   let lastDeltaMeta = null;
   const flushTail = (after) => {
     const tail = tagFilter.flush();
@@ -54777,10 +54804,14 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
       const rest = flushTail("");
       if (rest.length > 0) await write(rest);
     }
-    if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
-      applyUsageSample(session, acc, "responses");
-      markDirty(session);
+    settleUsage();
+  } catch (e) {
+    settleUsage();
+    if (res.destroyed || res.writableEnded) {
+      log2?.("client aborted mid-stream");
+      return;
     }
+    throw e;
   } finally {
     reader.releaseLock();
     res.end();
@@ -54806,10 +54837,16 @@ function rebuildEvent(rawEvent, ev) {
 async function pipePluginJson(stream2, res, session, protocol) {
   const reader = stream2.getReader();
   const chunks = [];
-  for (; ; ) {
-    const { done, value } = await reader.read();
-    if (done) break;
-    if (value && value.length > 0) chunks.push(Buffer.from(value));
+  try {
+    for (; ; ) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      if (value && value.length > 0) chunks.push(Buffer.from(value));
+    }
+  } catch (e) {
+    if (!res.writableEnded) res.end();
+    if (res.destroyed || res.writableEnded) return;
+    throw e;
   }
   reader.releaseLock();
   const text = Buffer.concat(chunks).toString("utf8");
@@ -56460,7 +56497,10 @@ async function startServer(opts) {
       await handle(req, res, opts, core, config, log2, instanceId, instanceStartedAt);
     } catch (err2) {
       const msg = String(err2);
-      log2("error", msg);
+      const e = err2;
+      const clientAbort = (e?.name === "AbortError" || /abort/i.test(String(e?.message ?? ""))) && (res.destroyed || res.writableEnded);
+      if (clientAbort) log2("info", `client aborted mid-stream: ${msg}`);
+      else log2("error", msg);
       if (!res.headersSent) {
         const status = msg.includes("exceeds") ? 413 : 502;
         res.writeHead(status, { "content-type": "application/json" });
@@ -57849,187 +57889,205 @@ ${hdrText}
     return;
   }
   if (prepared?.pluginMode) {
-    if (prepared.stream) {
-      if (prepared.protocol === "responses") {
-        await pipePluginResponsesWithStrip(upstream.body, res, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
-      } else {
-        await pipePluginChatWithStrip(upstream.body, res, prepared.protocol, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
-      }
-    } else {
-      await pipePluginJson(upstream.body, res, prepared.session, prepared.protocol);
-    }
-    clearUpstreamTimer();
-    return;
-  }
-  let responseBody = upstream.body;
-  if (prepared !== null && prepared.stream && maxFakeCompletionRetries() > 0) {
-    const resolvedBuf = await resolveFakeCompletion(upstream.body, {
-      protocol: prepared.protocol,
-      body,
-      upstreamUrl,
-      reqHeaders: buildForwardHeaders(headers),
-      proxyUrl,
-      signal: clientAbort.signal,
-      session: prepared.session,
-      log: log2
-    });
-    responseBody = bufferToStream(resolvedBuf);
-  }
-  const useRewriter = prepared !== null && prepared.compressInjected && prepared.processedMessages.length > 0;
-  if (!useRewriter || prepared === null) {
-    if (prepared && prepared.resetAfterSuccess) {
-      const [toClient, toObserve] = responseBody.tee();
-      const observed = observeResponsesTerminalState(toObserve, prepared.stream);
-      const tagLog = (msg) => log2("info", `[${prepared.session.id}] ${msg}`);
-      if ((upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
-        await pipePluginResponsesWithStrip(toClient, res, void 0, tagLog);
-      } else {
-        await pipeThrough(toClient, res);
-      }
-      clearUpstreamTimer();
-      const terminal = await observed;
-      if (terminal === "completed") {
-        markNativeCompactionBoundary(prepared.session);
-        log2("info", `[${prepared.session.id}] native Responses compact completed; rebase scheduled for next Responses turn`);
-      } else {
-        log2("warn", `[${prepared.session.id}] native compact response terminal=${terminal}; rebase NOT scheduled`);
+    try {
+      let pluginBody = upstream.body;
+      if (prepared.stream && maxFakeCompletionRetries() > 0) {
+        const resolvedBuf = await resolveFakeCompletion(pluginBody, {
+          protocol: prepared.protocol,
+          body,
+          upstreamUrl,
+          reqHeaders: buildForwardHeaders(headers),
+          proxyUrl,
+          signal: clientAbort.signal,
+          session: prepared.session,
+          log: log2
+        });
+        pluginBody = bufferToStream(resolvedBuf);
       }
-    } else if (prepared && prepared.stream && (upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
-      const p2 = prepared;
-      const tagLog = (msg) => log2("info", `[${p2.session.id}] ${msg}`);
-      if (p2.protocol === "responses") {
-        await pipePluginResponsesWithStrip(responseBody, res, void 0, tagLog);
+      if (prepared.stream) {
+        if (prepared.protocol === "responses") {
+          await pipePluginResponsesWithStrip(pluginBody, res, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
+        } else {
+          await pipePluginChatWithStrip(pluginBody, res, prepared.protocol, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
+        }
       } else {
-        await pipePluginChatWithStrip(responseBody, res, p2.protocol, void 0, tagLog);
+        await pipePluginJson(pluginBody, res, prepared.session, prepared.protocol);
       }
-      clearUpstreamTimer();
-    } else if (prepared && (upstream.headers.get("content-type") ?? "").includes("application/json")) {
-      await pipePluginJson(responseBody, res, void 0, prepared.protocol);
-      clearUpstreamTimer();
-    } else {
-      await pipeThrough(responseBody, res);
+    } finally {
       clearUpstreamTimer();
     }
     return;
   }
-  const ctx = {
-    core,
-    config,
-    messages: prepared.originalMessages,
-    session: prepared.session,
-    log: (msg) => log2("info", `[${prepared.session.id}] ${msg}`),
-    debug: opts.debug
-  };
-  if (prepared.stream) {
-    let streamToRead = responseBody;
-    let dumpRaw;
-    if (opts.dumpSse) {
-      const [a, b2] = responseBody.tee();
-      streamToRead = a;
-      dumpRaw = dumpStreamToFile(b2, opts.dumpSse, `${Date.now()}-${safeSessionId(prepared.session.id)}-raw.sse`);
-    }
-    try {
-      const parsedReq = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
-      const reqHeaders = buildForwardHeaders(headers);
-      const textProtocol = prepared.protocol === "responses" && !!prepared.responsesTextProtocol;
-      const systemPrompt = textProtocol ? buildCompressHybridSystemPrompt(prepared.prompts ?? defaultPrompts) : buildCompressSystemPrompt(prepared.prompts ?? defaultPrompts);
-      const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText);
-      const refreshFolded = (current) => {
-        const turn = core.processTurn({
-          messages: prepared.originalMessages,
-          state: prepared.session.state,
-          config,
-          tokenCount: prepared.session.stats.lastInputTokens,
-          renderTags: prepared.renderTags ?? "text-only"
-        });
-        prepared.session.state = turn.state;
-        const records = current.filter((m2) => typeof m2.id === "string" && m2.id.startsWith("acp_loop_"));
-        return stripKernelSummaries([...turn.messages, ...records], turn.state);
-      };
-      const loop = runCompressLoop(
-        streamToRead,
-        { core, config, messages: prepared.processedMessages.length > 0 ? prepared.processedMessages : prepared.originalMessages, compressMessages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, protocol: prepared.protocol, textProtocol, debug: opts.debug, nudge: prepared.nudge, refreshFolded },
-        parsedReq,
-        { url: upstreamUrl, headers: reqHeaders },
-        adapter,
-        systemPrompt,
-        clientAbort.signal
-      );
-      let protocolFragmentWarned = false;
-      for await (const chunk of loop) {
-        if (res.destroyed || res.writableEnded) break;
-        {
-          const s3 = chunk.toString("utf8");
-          if (s3.includes("<acp ") || s3.includes("</acp")) {
-            log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains <acp tag`);
-          } else if (!protocolFragmentWarned && containsToolCallXmlFragment(s3)) {
-            protocolFragmentWarned = true;
-            log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains tool-call XML fragment (possible tag echo; not stripped)`);
-          }
-        }
-        res.write(chunk);
-        if (res.writableNeedDrain) {
-          await Promise.race([
-            new Promise((r) => res.once("drain", () => r())),
-            new Promise((r) => res.once("close", () => r()))
-          ]);
+  let responseBody = upstream.body;
+  try {
+    if (prepared !== null && prepared.stream && maxFakeCompletionRetries() > 0) {
+      const resolvedBuf = await resolveFakeCompletion(upstream.body, {
+        protocol: prepared.protocol,
+        body,
+        upstreamUrl,
+        reqHeaders: buildForwardHeaders(headers),
+        proxyUrl,
+        signal: clientAbort.signal,
+        session: prepared.session,
+        log: log2
+      });
+      responseBody = bufferToStream(resolvedBuf);
+    }
+    const useRewriter = prepared !== null && prepared.compressInjected && prepared.processedMessages.length > 0;
+    if (!useRewriter || prepared === null) {
+      if (prepared && prepared.resetAfterSuccess) {
+        const [toClient, toObserve] = responseBody.tee();
+        const observed = observeResponsesTerminalState(toObserve, prepared.stream);
+        const tagLog = (msg) => log2("info", `[${prepared.session.id}] ${msg}`);
+        if ((upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
+          await pipePluginResponsesWithStrip(toClient, res, void 0, tagLog);
+        } else {
+          await pipeThrough(toClient, res);
+        }
+        const terminal = await observed;
+        if (terminal === "completed") {
+          markNativeCompactionBoundary(prepared.session);
+          log2("info", `[${prepared.session.id}] native Responses compact completed; rebase scheduled for next Responses turn`);
+        } else {
+          log2("warn", `[${prepared.session.id}] native compact response terminal=${terminal}; rebase NOT scheduled`);
+        }
+      } else if (prepared && prepared.stream && (upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
+        const p2 = prepared;
+        const tagLog = (msg) => log2("info", `[${p2.session.id}] ${msg}`);
+        if (p2.protocol === "responses") {
+          await pipePluginResponsesWithStrip(responseBody, res, void 0, tagLog);
+        } else {
+          await pipePluginChatWithStrip(responseBody, res, p2.protocol, void 0, tagLog);
         }
-        if (res.destroyed || res.writableEnded) break;
+      } else if (prepared && (upstream.headers.get("content-type") ?? "").includes("application/json")) {
+        await pipePluginJson(responseBody, res, void 0, prepared.protocol);
+      } else {
+        await pipeThrough(responseBody, res);
       }
-      res.end();
-    } catch (e) {
-      emitStreamError(res, prepared.protocol, e?.message ?? String(e), (m2) => log2("error", `[${prepared.session.id}] ${m2}`));
-    } finally {
-      clearUpstreamTimer();
-      if (dumpRaw) await dumpRaw;
+      return;
     }
-  } else {
-    try {
-      const buf = await upstream.arrayBuffer();
-      const text = Buffer.from(buf).toString("utf8");
+    const ctx = {
+      core,
+      config,
+      messages: prepared.originalMessages,
+      session: prepared.session,
+      log: (msg) => log2("info", `[${prepared.session.id}] ${msg}`),
+      debug: opts.debug
+    };
+    if (prepared.stream) {
+      let streamToRead = responseBody;
+      let dumpRaw;
+      if (opts.dumpSse) {
+        const [a, b2] = responseBody.tee();
+        streamToRead = a;
+        dumpRaw = dumpStreamToFile(b2, opts.dumpSse, `${Date.now()}-${safeSessionId(prepared.session.id)}-raw.sse`);
+      }
       try {
-        let json = JSON.parse(text);
-        if (prepared.protocol === "responses" && prepared.responsesTextProtocol) {
-          const requestBody = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
-          const requestHeaders = buildForwardHeaders(headers);
-          json = await compressLoopResponsesJson(
-            json,
-            { core, config, messages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, textProtocol: true },
-            requestBody,
-            { url: upstreamUrl, headers: requestHeaders }
-          );
-        }
-        const u2 = json.usage ?? {};
-        const { total, cached } = usageTotals(prepared.protocol, u2);
-        if (typeof total === "number") {
-          prepared.session.stats.inputTokens += total;
-          prepared.session.stats.lastInputTokens = Math.max(
-            0,
-            total - (prepared.session.stats.compressCreditTokens ?? 0)
-          );
-          if (typeof cached === "number") {
-            prepared.session.stats.cachedTokens += cached;
-            prepared.session.stats.cacheSamples += 1;
+        const parsedReq = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
+        const reqHeaders = buildForwardHeaders(headers);
+        const textProtocol = prepared.protocol === "responses" && !!prepared.responsesTextProtocol;
+        const systemPrompt = textProtocol ? buildCompressHybridSystemPrompt(prepared.prompts ?? defaultPrompts) : buildCompressSystemPrompt(prepared.prompts ?? defaultPrompts);
+        const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText);
+        const refreshFolded = (current) => {
+          const turn = core.processTurn({
+            messages: prepared.originalMessages,
+            state: prepared.session.state,
+            config,
+            tokenCount: prepared.session.stats.lastInputTokens,
+            renderTags: prepared.renderTags ?? "text-only"
+          });
+          prepared.session.state = turn.state;
+          const records = current.filter((m2) => typeof m2.id === "string" && m2.id.startsWith("acp_loop_"));
+          return stripKernelSummaries([...turn.messages, ...records], turn.state);
+        };
+        const loop = runCompressLoop(
+          streamToRead,
+          { core, config, messages: prepared.processedMessages.length > 0 ? prepared.processedMessages : prepared.originalMessages, compressMessages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, protocol: prepared.protocol, textProtocol, debug: opts.debug, nudge: prepared.nudge, refreshFolded },
+          parsedReq,
+          { url: upstreamUrl, headers: reqHeaders },
+          adapter,
+          systemPrompt,
+          clientAbort.signal
+        );
+        let protocolFragmentWarned = false;
+        for await (const chunk of loop) {
+          if (res.destroyed || res.writableEnded) break;
+          {
+            const s3 = chunk.toString("utf8");
+            if (s3.includes("<acp ") || s3.includes("</acp")) {
+              log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains <acp tag`);
+            } else if (!protocolFragmentWarned && containsToolCallXmlFragment(s3)) {
+              protocolFragmentWarned = true;
+              log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains tool-call XML fragment (possible tag echo; not stripped)`);
+            }
+          }
+          res.write(chunk);
+          if (res.writableNeedDrain) {
+            await Promise.race([
+              new Promise((r) => res.once("drain", () => r())),
+              new Promise((r) => res.once("close", () => r()))
+            ]);
           }
-          const out = u2.completion_tokens ?? u2.output_tokens;
-          if (typeof out === "number") prepared.session.stats.outputTokens += out;
+          if (res.destroyed || res.writableEnded) break;
         }
-        if (prepared.protocol === "openai") {
-          rewriteOpenaiJsonResponse(json, ctx);
-        } else if (prepared.protocol === "responses") {
-          rewriteResponsesJsonResponse(json, ctx);
-        } else {
-          rewriteJsonResponse(json, ctx);
+        res.end();
+      } catch (e) {
+        emitStreamError(res, prepared.protocol, e?.message ?? String(e), (m2) => log2("error", `[${prepared.session.id}] ${m2}`));
+      } finally {
+        clearUpstreamTimer();
+        if (dumpRaw) await dumpRaw;
+        markDirty(prepared.session);
+      }
+    } else {
+      try {
+        const buf = await upstream.arrayBuffer();
+        const text = Buffer.from(buf).toString("utf8");
+        try {
+          let json = JSON.parse(text);
+          if (prepared.protocol === "responses" && prepared.responsesTextProtocol) {
+            const requestBody = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
+            const requestHeaders = buildForwardHeaders(headers);
+            json = await compressLoopResponsesJson(
+              json,
+              { core, config, messages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, textProtocol: true },
+              requestBody,
+              { url: upstreamUrl, headers: requestHeaders }
+            );
+          }
+          const u2 = json.usage ?? {};
+          const { total, cached } = usageTotals(prepared.protocol, u2);
+          if (typeof total === "number") {
+            prepared.session.stats.inputTokens += total;
+            prepared.session.stats.lastInputTokens = Math.max(
+              0,
+              total - (prepared.session.stats.compressCreditTokens ?? 0)
+            );
+            if (typeof cached === "number") {
+              prepared.session.stats.cachedTokens += cached;
+              prepared.session.stats.cacheSamples += 1;
+            }
+            const out = u2.completion_tokens ?? u2.output_tokens;
+            if (typeof out === "number") prepared.session.stats.outputTokens += out;
+          }
+          if (prepared.protocol === "openai") {
+            rewriteOpenaiJsonResponse(json, ctx);
+          } else if (prepared.protocol === "responses") {
+            rewriteResponsesJsonResponse(json, ctx);
+          } else {
+            rewriteJsonResponse(json, ctx);
+          }
+          res.end(JSON.stringify(json));
+        } catch {
+          res.end(text);
         }
-        res.end(JSON.stringify(json));
-      } catch {
-        res.end(text);
+      } finally {
+        clearUpstreamTimer();
+        markDirty(prepared.session);
       }
-    } finally {
-      clearUpstreamTimer();
     }
+  } finally {
+    clearUpstreamTimer();
   }
-  markDirty(prepared.session);
 }
 async function readStreamToBuffer(stream2, maxBytes = 1 << 20) {
   const reader = stream2.getReader();