kimiflare 0.65.0 → 0.67.0

package/dist/index.js

@@ -401,11 +401,12 @@ var init_logger = __esm({
 });
 
 // src/util/sse.ts
-async function* readSSE(stream, signal, idleTimeoutMs) {
+async function* readSSE(stream, signal, idleTimeoutMs, postFirstByteIdleTimeoutMs) {
   const reader = stream.getReader();
   const decoder = new TextDecoder("utf-8");
   let buffer = "";
   let lastDataAt = Date.now();
+  let gotFirstByte = false;
   const onAbort = () => {
     reader.cancel(new DOMException("aborted", "AbortError")).catch(() => {
     });
@@ -427,16 +428,18 @@ async function* readSSE(stream, signal, idleTimeoutMs) {
   try {
     while (true) {
       if (signal?.aborted) throw new DOMException("aborted", "AbortError");
-      if (idleTimeoutMs !== void 0 && Date.now() - lastDataAt > idleTimeoutMs) {
-        logger.warn("sse:idle_timeout", { idleTimeoutMs });
+      const activeIdleTimeoutMs = gotFirstByte && postFirstByteIdleTimeoutMs !== void 0 ? postFirstByteIdleTimeoutMs : idleTimeoutMs;
+      if (activeIdleTimeoutMs !== void 0 && Date.now() - lastDataAt > activeIdleTimeoutMs) {
+        logger.warn("sse:idle_timeout", { idleTimeoutMs: activeIdleTimeoutMs, gotFirstByte });
         throw new DOMException(
-          `kimiflare: stream idle for ${idleTimeoutMs}ms \u2014 no data received from API`,
+          `kimiflare: stream idle for ${activeIdleTimeoutMs}ms \u2014 no data received from API`,
           "TimeoutError"
         );
       }
       const { done, value } = await abortRace(reader.read());
       if (done) break;
       lastDataAt = Date.now();
+      gotFirstByte = true;
       buffer += decoder.decode(value, { stream: true });
       buffer = buffer.replace(/\r\n/g, "\n");
       let sep3;
@@ -717,7 +720,7 @@ async function* runKimi(opts2) {
     if (meta) yield { type: "gateway_meta", meta };
     let lastUsage = null;
     logger.debug("runKimi:stream_start", { requestId });
-    for await (const ev of parseStream(res.body, opts2.signal, opts2.idleTimeoutMs)) {
+    for await (const ev of parseStream(res.body, opts2.signal, opts2.idleTimeoutMs, opts2.postFirstByteIdleTimeoutMs)) {
       if (ev.type === "usage") lastUsage = ev.usage;
       yield ev;
     }
@@ -799,12 +802,11 @@ function readGatewayMeta(headers) {
   if (model) meta.model = model;
   return Object.keys(meta).length > 0 ? meta : null;
 }
-async function* parseStream(body, signal, idleTimeoutMs = DEFAULT_IDLE_TIMEOUT_MS) {
+async function* parseStream(body, signal, idleTimeoutMs = DEFAULT_IDLE_TIMEOUT_MS, postFirstByteIdleTimeoutMs = DEFAULT_POST_FIRST_BYTE_IDLE_TIMEOUT_MS) {
   const toolCalls = /* @__PURE__ */ new Map();
   let lastUsage = null;
   let finishReason = null;
-  let lastDataAt = Date.now();
-  for await (const dataStr of readSSE(body, signal, idleTimeoutMs)) {
+  for await (const dataStr of readSSE(body, signal, idleTimeoutMs, postFirstByteIdleTimeoutMs)) {
     if (dataStr === "[DONE]") break;
     let chunk = null;
     try {
@@ -942,7 +944,7 @@ function sleep(ms, signal) {
     signal?.addEventListener("abort", onAbort, { once: true });
   });
 }
-var RETRYABLE_CODES, MAX_ATTEMPTS, DEFAULT_IDLE_TIMEOUT_MS;
+var RETRYABLE_CODES, MAX_ATTEMPTS, DEFAULT_IDLE_TIMEOUT_MS, DEFAULT_POST_FIRST_BYTE_IDLE_TIMEOUT_MS;
 var init_client = __esm({
   "src/agent/client.ts"() {
     "use strict";
@@ -954,6 +956,7 @@ var init_client = __esm({
     RETRYABLE_CODES = /* @__PURE__ */ new Set([3040]);
     MAX_ATTEMPTS = 5;
     DEFAULT_IDLE_TIMEOUT_MS = 6e4;
+    DEFAULT_POST_FIRST_BYTE_IDLE_TIMEOUT_MS = 3e4;
   }
 });
 
@@ -2155,7 +2158,7 @@ var init_session_state = __esm({
       }
       add(a) {
         while (this.totalChars() + a.raw.length > this.maxTotalChars && this.artifacts.size > 0) {
-          this.evictOldest();
+          this.evictSizeWeighted();
         }
         while (this.artifacts.size >= this.maxArtifacts) {
           this.evictOldest();
@@ -2196,6 +2199,21 @@ var init_session_state = __esm({
         }
         if (oldest) this.artifacts.delete(oldest.id);
       }
+      /** Evict the largest artifact among the oldest quartile (by timestamp).
+       *  Bounded by the oldest quartile so we never evict freshly-added artifacts;
+       *  size-weighted within that window so one big artifact gets dropped instead
+       *  of many small ones. */
+      evictSizeWeighted() {
+        const sorted = [...this.artifacts.values()].sort((a, b) => a.ts < b.ts ? -1 : 1);
+        if (sorted.length === 0) return;
+        const quartile = Math.max(1, Math.ceil(sorted.length / 4));
+        const candidates = sorted.slice(0, quartile);
+        let pick3 = candidates[0];
+        for (const a of candidates) {
+          if (a.raw.length > pick3.raw.length) pick3 = a;
+        }
+        this.artifacts.delete(pick3.id);
+      }
     };
   }
 });
@@ -3132,6 +3150,15 @@ var init_system_prompt = __esm({
 });
 
 // src/agent/loop.ts
+function getSessionWebFetchHistory(sessionId) {
+  const key = sessionId ?? "default";
+  let arr = sessionWebFetchHistory.get(key);
+  if (!arr) {
+    arr = [];
+    sessionWebFetchHistory.set(key, arr);
+  }
+  return arr;
+}
 function isHighSignalMemory(memory) {
   return memory.topicKey === "project_dependencies" || memory.topicKey === "project_tsconfig" || memory.topicKey === "project_entry_point" || memory.category === "instruction" || memory.category === "preference" || memory.category === "event" && memory.importance >= 3;
 }
@@ -3286,7 +3313,8 @@ Use console.log() to return results. Only console.log output will be sent back t
   const recentToolCalls = [];
   const LOOP_WINDOW = 8;
   const LOOP_THRESHOLD = 2;
-  const webFetchHistory = [];
+  const webFetchHistory = getSessionWebFetchHistory(opts2.sessionId);
+  let webFetchesThisTurn = 0;
   const MAX_WEB_FETCH_PER_TURN = 5;
   const WEB_FETCH_DOMAIN_THRESHOLD = 2;
   let cumulativePromptTokens = 0;
@@ -3404,7 +3432,8 @@ Use console.log() to return results. Only console.log output will be sent back t
       cloudMode: opts2.cloudMode,
       cloudToken: opts2.cloudToken,
       cloudDeviceId: opts2.cloudDeviceId,
-      idleTimeoutMs: 6e4
+      idleTimeoutMs: opts2.idleTimeoutMs ?? 6e4,
+      postFirstByteIdleTimeoutMs: opts2.postFirstByteIdleTimeoutMs
     });
     let gotFirstChunk = false;
     for await (const ev of events) {
@@ -3454,7 +3483,7 @@ Use console.log() to return results. Only console.log output will be sent back t
     if (lastUsage) {
       opts2.callbacks.onUsageFinal?.(lastUsage, gatewayMeta);
       cumulativePromptTokens += lastUsage.prompt_tokens;
-      if (!budgetExhausted && opts2.maxInputTokens !== void 0 && opts2.maxInputTokens > 0 && cumulativePromptTokens >= opts2.maxInputTokens && toolCalls.length > 0) {
+      if (!budgetExhausted && opts2.maxInputTokens !== void 0 && opts2.maxInputTokens > 0 && cumulativePromptTokens >= opts2.maxInputTokens) {
         budgetExhausted = true;
       }
     }
@@ -3524,8 +3553,8 @@ Use console.log() to return results. Only console.log output will be sent back t
         try {
           const domain = new URL(url).hostname;
           const domainCount = webFetchHistory.filter((h) => h.domain === domain).length;
-          const totalWebFetches = webFetchHistory.length;
-          if (totalWebFetches >= MAX_WEB_FETCH_PER_TURN) {
+          const totalSessionFetches = webFetchHistory.length;
+          if (webFetchesThisTurn >= MAX_WEB_FETCH_PER_TURN) {
             const warning = `Research budget exceeded: you have already made ${MAX_WEB_FETCH_PER_TURN} web requests this turn. Synthesize what you have learned instead of fetching more pages.`;
             const budgetResult = {
               tool_call_id: tc.id,
@@ -3546,6 +3575,27 @@ Use console.log() to return results. Only console.log output will be sent back t
             blockedCount++;
             continue;
           }
+          if (totalSessionFetches >= SESSION_WEB_FETCH_CAP) {
+            const warning = `Session research budget exceeded: ${totalSessionFetches} web fetches across this session. Synthesize what you have learned from prior fetches instead of starting another page.`;
+            const sessionCapResult = {
+              tool_call_id: tc.id,
+              name: "web_fetch",
+              content: warning,
+              ok: false
+            };
+            toolResults.push(sessionCapResult);
+            opts2.messages.push({
+              role: "tool",
+              tool_call_id: tc.id,
+              content: sanitizeString(warning),
+              name: "web_fetch"
+            });
+            opts2.callbacks.onToolResult?.(sessionCapResult);
+            recentToolCalls.push(loopSignature);
+            if (recentToolCalls.length > LOOP_WINDOW) recentToolCalls.shift();
+            blockedCount++;
+            continue;
+          }
           if (domainCount >= WEB_FETCH_DOMAIN_THRESHOLD) {
             const warning = `Loop detected: you have fetched from ${domain} multiple times. Consider a different approach or synthesize existing findings.`;
             const loopResult = {
@@ -3568,6 +3618,7 @@ Use console.log() to return results. Only console.log output will be sent back t
             continue;
           }
           webFetchHistory.push({ url, domain });
+          webFetchesThisTurn++;
         } catch {
         }
       }
@@ -3603,9 +3654,16 @@ Use console.log() to return results. Only console.log output will be sent back t
 Output:
 ${sandboxResult.output}` : sandboxResult.output;
         if (resultContent.length > MAX_TOOL_CONTENT_CHARS) {
+          const rawBytes = resultContent.length;
           resultContent = resultContent.slice(0, MAX_TOOL_CONTENT_CHARS) + `
 
-[truncated: ${resultContent.length - MAX_TOOL_CONTENT_CHARS} chars omitted]`;
+[truncated: ${rawBytes - MAX_TOOL_CONTENT_CHARS} chars omitted]`;
+          opts2.callbacks.onTruncation?.({
+            tool: "execute_code",
+            toolCallId: tc.id,
+            rawBytes,
+            reducedBytes: resultContent.length
+          });
         }
         const result = {
           tool_call_id: tc.id,
@@ -3634,9 +3692,17 @@ ${sandboxResult.output}` : sandboxResult.output;
         );
         let content2 = result.content;
         if (content2.length > MAX_TOOL_CONTENT_CHARS) {
+          const rawBytes = content2.length;
           content2 = content2.slice(0, MAX_TOOL_CONTENT_CHARS) + `
 
-[truncated: ${content2.length - MAX_TOOL_CONTENT_CHARS} chars omitted]`;
+[truncated: ${rawBytes - MAX_TOOL_CONTENT_CHARS} chars omitted]`;
+          opts2.callbacks.onTruncation?.({
+            tool: tc.function.name,
+            toolCallId: tc.id,
+            rawBytes,
+            reducedBytes: content2.length,
+            artifactId: result.artifactId
+          });
         }
         logger.debug("turn:tool_end", { sessionId: opts2.sessionId, tool: tc.function.name, toolCallId: tc.id, ok: result.ok });
         toolResults.push(result);
@@ -3660,6 +3726,7 @@ ${sandboxResult.output}` : sandboxResult.output;
           );
           const assistantMessage = lastAssistant?.content ?? "";
           const llmOpts = opts2.memoryManager.getExtractionLlmOpts();
+          const turnAtMemoryCommit = turn;
           for (const extractor of EXTRACTORS) {
             if (extractor.match(tc.function.name, filePath)) {
               void (async () => {
@@ -3685,15 +3752,33 @@ ${sandboxResult.output}` : sandboxResult.output;
                     );
                     if (isHighSignalMemory(memory)) {
                       const sid = opts2.sessionId ?? "default";
-                      const current = (driftAccumulator.get(sid) ?? 0) + 1;
-                      driftAccumulator.set(sid, current);
-                      if (current >= DRIFT_THRESHOLD) {
+                      const events2 = driftEvents.get(sid) ?? [];
+                      events2.push(turnAtMemoryCommit);
+                      const cutoff = turnAtMemoryCommit - DRIFT_WINDOW + 1;
+                      const recent = events2.filter((t) => t >= cutoff);
+                      driftEvents.set(sid, recent);
+                      if (recent.length >= DRIFT_THRESHOLD) {
                         opts2.callbacks.onKimiMdStale?.();
-                        driftAccumulator.set(sid, 0);
+                        driftEvents.set(sid, []);
                       }
                     }
                   }
-                } catch {
+                } catch (err) {
+                  const sid = opts2.sessionId ?? "default";
+                  const next = (memoryExtractionErrorCounts.get(sid) ?? 0) + 1;
+                  memoryExtractionErrorCounts.set(sid, next);
+                  const msg = err instanceof Error ? err.message : String(err);
+                  logger.debug("memory:extract_error", {
+                    sessionId: opts2.sessionId,
+                    tool: tc.function.name,
+                    count: next,
+                    error: msg
+                  });
+                  if (next === 1) {
+                    opts2.callbacks.onWarning?.(
+                      `[memory] auto-extraction failed (${msg}). Subsequent failures will be counted silently; check /memory health.`
+                    );
+                  }
                 }
               })();
             }
@@ -3706,12 +3791,6 @@ ${sandboxResult.output}` : sandboxResult.output;
     if (blockedCount === toolCalls.length && toolCalls.length > 0) {
       loopExhausted = true;
     }
-    if (opts2.sessionId) {
-      const current = driftAccumulator.get(opts2.sessionId) ?? 0;
-      if (current > 0) {
-        driftAccumulator.set(opts2.sessionId, Math.max(0, current - 1));
-      }
-    }
     if (opts2.onIterationEnd) {
       opts2.messages = await opts2.onIterationEnd(opts2.messages, opts2.signal);
       if (opts2.signal.aborted) throw new DOMException("aborted", "AbortError");
@@ -3770,7 +3849,7 @@ function validateToolArguments(raw) {
     return "{}";
   }
 }
-var BudgetExhaustedError, AgentLoopError, codeModeApiCache, driftAccumulator, DRIFT_THRESHOLD, MAX_PROMPT_TOKENS, MAX_TOOL_CONTENT_CHARS;
+var BudgetExhaustedError, AgentLoopError, codeModeApiCache, driftEvents, DRIFT_WINDOW, DRIFT_THRESHOLD, memoryExtractionErrorCounts, sessionWebFetchHistory, SESSION_WEB_FETCH_CAP, MAX_PROMPT_TOKENS, MAX_TOOL_CONTENT_CHARS;
 var init_loop = __esm({
   "src/agent/loop.ts"() {
     "use strict";
@@ -3798,8 +3877,12 @@ var init_loop = __esm({
       }
     };
     codeModeApiCache = /* @__PURE__ */ new Map();
-    driftAccumulator = /* @__PURE__ */ new Map();
-    DRIFT_THRESHOLD = 5;
+    driftEvents = /* @__PURE__ */ new Map();
+    DRIFT_WINDOW = 10;
+    DRIFT_THRESHOLD = 3;
+    memoryExtractionErrorCounts = /* @__PURE__ */ new Map();
+    sessionWebFetchHistory = /* @__PURE__ */ new Map();
+    SESSION_WEB_FETCH_CAP = 25;
     MAX_PROMPT_TOKENS = 24e4;
     MAX_TOOL_CONTENT_CHARS = 1e4;
   }
@@ -3867,10 +3950,13 @@ var init_read = __esm({
       needsPermission: false,
       render: ({ path }) => ({ title: `read ${collapsePath(path, process.cwd())}` }),
       async run(args, ctx) {
+        if (ctx.signal?.aborted) throw new DOMException("aborted", "AbortError");
         const abs = resolvePath(ctx.cwd, args.path);
         const st = await stat2(abs);
         if (st.size > MAX_BYTES) throw new Error(`file too large: ${st.size} bytes (max ${MAX_BYTES})`);
-        const text = await readFile3(abs, "utf8");
+        if (ctx.signal?.aborted) throw new DOMException("aborted", "AbortError");
+        const text = await readFile3(abs, { encoding: "utf8", signal: ctx.signal });
+        if (ctx.signal?.aborted) throw new DOMException("aborted", "AbortError");
         const lines = text.split("\n");
         const start = Math.max(0, (args.offset ?? 1) - 1);
         const end = args.limit ? Math.min(lines.length, start + args.limit) : lines.length;
@@ -4162,14 +4248,31 @@ var init_glob = __esm({
       needsPermission: false,
       render: (args) => ({ title: `glob ${args.pattern ?? ""}${args.path ? ` in ${collapsePath(String(args.path), process.cwd())}` : ""}` }),
       async run(args, ctx) {
+        if (ctx.signal?.aborted) throw new DOMException("aborted", "AbortError");
         const root = args.path ? resolvePath(ctx.cwd, args.path) : ctx.cwd;
-        const entries = await fg(args.pattern, {
+        const stream = fg.stream(args.pattern, {
           cwd: root,
           absolute: true,
           dot: false,
           onlyFiles: false,
           stats: true
         });
+        const entries = [];
+        const onAbort = () => {
+          try {
+            stream.destroy(new DOMException("aborted", "AbortError"));
+          } catch {
+          }
+        };
+        ctx.signal?.addEventListener("abort", onAbort, { once: true });
+        try {
+          for await (const entry of stream) {
+            if (ctx.signal?.aborted) throw new DOMException("aborted", "AbortError");
+            entries.push(entry);
+          }
+        } finally {
+          ctx.signal?.removeEventListener("abort", onAbort);
+        }
         entries.sort((a, b) => (b.stats?.mtimeMs ?? 0) - (a.stats?.mtimeMs ?? 0));
         const paths = entries.slice(0, 200).map((e) => e.path);
         return paths.length ? paths.join("\n") : "(no matches)";
@@ -4193,14 +4296,14 @@ async function hasRipgrep() {
   }
   return cachedHasRg;
 }
-async function runRipgrep(args, root, mode) {
+async function runRipgrep(args, root, mode, signal) {
   const rgArgs = ["--no-heading", "--color=never", "--line-number"];
   if (args.case_insensitive) rgArgs.push("-i");
   if (args.glob) rgArgs.push("--glob", args.glob);
   if (mode === "files") rgArgs.push("-l");
   rgArgs.push("--", args.pattern, root);
   try {
-    const { stdout } = await pExecFile("rg", rgArgs, { maxBuffer: 10 * 1024 * 1024 });
+    const { stdout } = await pExecFile("rg", rgArgs, { maxBuffer: 10 * 1024 * 1024, signal });
     const trimmed = stdout.trim();
     if (!trimmed) return { content: "(no matches)", rawBytes: 0, reducedBytes: 0 };
     return {
@@ -4214,7 +4317,7 @@ async function runRipgrep(args, root, mode) {
     throw new Error(err.stderr || String(e));
   }
 }
-async function runJsFallback(args, root, mode) {
+async function runJsFallback(args, root, mode, signal) {
   const re = new RegExp(args.pattern, args.case_insensitive ? "i" : "");
   const globPattern = args.glob ? `**/${args.glob}` : "**/*";
   const files = await fg2(globPattern, {
@@ -4225,7 +4328,9 @@ async function runJsFallback(args, root, mode) {
     ignore: ["**/node_modules/**", "**/.git/**", "**/dist/**"]
   });
   const out = [];
-  for (const file of files.slice(0, 5e3)) {
+  for (let fi = 0; fi < Math.min(files.length, 5e3); fi++) {
+    if (signal?.aborted) throw new DOMException("aborted", "AbortError");
+    const file = files[fi];
     try {
       const content = await readFile6(file, "utf8");
       if (mode === "files") {
@@ -4280,10 +4385,11 @@ var init_grep = __esm({
       needsPermission: false,
       render: (args) => ({ title: `grep ${args.pattern ?? ""}${args.glob ? ` (${args.glob})` : ""}` }),
       async run(args, ctx) {
+        if (ctx.signal?.aborted) throw new DOMException("aborted", "AbortError");
         const root = args.path ? resolvePath(ctx.cwd, args.path) : ctx.cwd;
         const mode = args.output_mode ?? "content";
-        if (await hasRipgrep()) return runRipgrep(args, root, mode);
-        return runJsFallback(args, root, mode);
+        if (await hasRipgrep()) return runRipgrep(args, root, mode, ctx.signal);
+        return runJsFallback(args, root, mode, ctx.signal);
       }
     };
   }