kimiflare 0.66.0 → 0.68.0

package/dist/index.js

@@ -3150,6 +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;
 }
@@ -3304,7 +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;
@@ -3543,8 +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,
@@ -3565,6 +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 = {
@@ -3587,6 +3618,7 @@ Use console.log() to return results. Only console.log output will be sent back t
             continue;
           }
           webFetchHistory.push({ url, domain });
+          webFetchesThisTurn++;
         } catch {
         }
       }
@@ -3622,9 +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,
@@ -3653,9 +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);
@@ -3679,6 +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 () => {
@@ -3704,11 +3752,14 @@ ${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, []);
                       }
                     }
                   }
@@ -3740,12 +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");
@@ -3804,7 +3849,7 @@ function validateToolArguments(raw) {
     return "{}";
   }
 }
-var BudgetExhaustedError, AgentLoopError, codeModeApiCache, driftAccumulator, DRIFT_THRESHOLD, memoryExtractionErrorCounts, 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";
@@ -3832,9 +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;
   }
@@ -6274,6 +6322,12 @@ var init_renderer = __esm({
 });
 
 // src/cost-attribution/reconcile.ts
+var reconcile_exports = {};
+__export(reconcile_exports, {
+  aggregateByFeature: () => aggregateByFeature,
+  fetchGatewayLogs: () => fetchGatewayLogs,
+  reconcileWithCloudflare: () => reconcileWithCloudflare
+});
 function cacheKey(opts2) {
   return `${opts2.gatewayId ?? "none"}:${opts2.startDate}:${opts2.endDate}`;
 }
@@ -6312,6 +6366,28 @@ async function fetchGatewayLogs(accountId, apiToken, gatewayId, startDate, endDa
   }
   return out;
 }
+function aggregateByFeature(logs) {
+  const map = /* @__PURE__ */ new Map();
+  for (const log2 of logs) {
+    let feature = "unknown";
+    const m = log2.metadata;
+    if (m && typeof m === "object" && !Array.isArray(m)) {
+      const f = m.feature;
+      if (typeof f === "string") feature = f;
+    } else if (typeof m === "string") {
+      try {
+        const parsed = JSON.parse(m);
+        if (typeof parsed.feature === "string") feature = parsed.feature;
+      } catch {
+      }
+    }
+    const entry = map.get(feature) ?? { feature, cost: 0, requests: 0 };
+    entry.cost += typeof log2.cost === "number" ? log2.cost : 0;
+    entry.requests += 1;
+    map.set(feature, entry);
+  }
+  return Array.from(map.values()).sort((a, b) => b.cost - a.cost);
+}
 async function reconcileWithCloudflare(opts2) {
   if (!opts2.accountId || !opts2.apiToken) {
     return {
@@ -6351,7 +6427,8 @@ async function reconcileWithCloudflare(opts2) {
       localCost: opts2.localCost,
       cloudflareCost,
       driftPct: Math.round(driftPct * 1e3) / 10,
-      message: `Reconciled ${logs.length} Gateway log entries`
+      message: `Reconciled ${logs.length} Gateway log entries`,
+      featureBreakdown: aggregateByFeature(logs)
     };
     cache.set(key, { result, expires: Date.now() + 60 * 60 * 1e3 });
     return result;
@@ -9411,7 +9488,8 @@ async function getCostReport(sessionId) {
     gatewayRequests: rawSession.gatewayRequests,
     gatewayCachedRequests: rawSession.gatewayCachedRequests,
     gatewayCost: rawSession.gatewayCost,
-    reconcilePending: hasPendingReconcile(rawSession)
+    reconcilePending: hasPendingReconcile(rawSession),
+    lastTurnMs: latestConfirmedDurationMs(rawSession)
   } : { date, promptTokens: 0, completionTokens: 0, cachedTokens: 0, cost: 0 };
   const todayUsage = log2.days.find((d) => d.date === date) ?? { date, promptTokens: 0, completionTokens: 0, cachedTokens: 0, cost: 0 };
   const monthUsage = {
@@ -9456,6 +9534,14 @@ function hasPendingReconcile(session) {
     (t) => t.logId && t.confirmedCost === void 0 && !t.reconcileFailed
   );
 }
+function latestConfirmedDurationMs(session) {
+  if (!session.turns) return void 0;
+  for (let i = session.turns.length - 1; i >= 0; i--) {
+    const ms = session.turns[i]?.durationMs;
+    if (typeof ms === "number") return ms;
+  }
+  return void 0;
+}
 async function getSessionGatewayLogs(sessionId) {
   const log2 = await loadLog2();
   const session = log2.sessions.find((s) => s.id === sessionId);
@@ -9501,6 +9587,17 @@ function formatGatewaySection(report, accountId, gatewayId, recentLogs2 = []) {
   );
   return lines.join("\n");
 }
+function formatFeatureBreakdown(breakdown) {
+  if (!breakdown || breakdown.length === 0) return "";
+  if (breakdown.length === 1 && breakdown[0].feature === "unknown") return "";
+  const lines = ["\u2500\u2500\u2500 By feature (Gateway-confirmed) \u2500\u2500\u2500"];
+  for (const row of breakdown) {
+    lines.push(
+      `  ${row.feature.padEnd(20)} $${row.cost.toFixed(4)}  (${row.requests} req)`
+    );
+  }
+  return lines.join("\n");
+}
 function formatCostReport(report) {
   const lines = [];
   const add = (label, u) => {
@@ -11704,6 +11801,9 @@ function buildRightParts(usage, contextLimit, sessionUsage, gatewayMeta, cloudMo
     } else {
       parts.push(`${prefix}${sessionUsage.cost.toFixed(2)}`);
     }
+    if (typeof sessionUsage.lastTurnMs === "number") {
+      parts.push(formatDuration(sessionUsage.lastTurnMs));
+    }
   } else {
     const cached = usage.prompt_tokens_details?.cached_tokens ?? 0;
     const cost = calculateCost(usage.prompt_tokens, usage.completion_tokens, cached);
@@ -11732,6 +11832,10 @@ function formatGatewayCacheStatus(gatewayMeta) {
   const status = gatewayMeta?.cacheStatus?.trim();
   return status ? `AI Gateway \xB7 cache ${status.toLowerCase()}` : null;
 }
+function formatDuration(ms) {
+  if (ms < 1e3) return `${Math.round(ms)}ms`;
+  return `${(ms / 1e3).toFixed(1)}s`;
+}
 function formatElapsed2(ms) {
   const total = Math.floor(ms / 1e3);
   const m = Math.floor(total / 60);
@@ -20031,6 +20135,22 @@ ${wcagWarnings.join("\n")}` }
             const logs = sid ? await getSessionGatewayLogs(sid).catch(() => []) : [];
             const gwSection = formatGatewaySection(report, cfg.accountId, cfg.aiGatewayId, logs);
             if (gwSection) lines.push("", gwSection);
+            try {
+              const { reconcileWithCloudflare: reconcileWithCloudflare2 } = await Promise.resolve().then(() => (init_reconcile(), reconcile_exports));
+              const today4 = (/* @__PURE__ */ new Date()).toISOString().slice(0, 10);
+              const sevenDaysAgo = new Date(Date.now() - 7 * 24 * 60 * 60 * 1e3).toISOString().slice(0, 10);
+              const recon = await reconcileWithCloudflare2({
+                localCost: report.month.cost,
+                accountId: cfg.accountId,
+                apiToken: cfg.apiToken,
+                gatewayId: cfg.aiGatewayId,
+                startDate: sevenDaysAgo,
+                endDate: today4
+              });
+              const breakdown = formatFeatureBreakdown(recon.featureBreakdown);
+              if (breakdown) lines.push("", breakdown);
+            } catch {
+            }
           }
           if (cfg?.costAttribution) {
             const { getCategoryReportText: getCategoryReportText2 } = await Promise.resolve().then(() => (init_tui_report(), tui_report_exports));
@@ -20108,6 +20228,24 @@ ${wcagWarnings.join("\n")}` }
             lines.push(`collect-logs: ${cfg.aiGatewayCollectLogPayload ?? false}`);
             const meta = cfg.aiGatewayMetadata;
             lines.push(`metadata: ${meta && Object.keys(meta).length > 0 ? JSON.stringify(meta) : "none"}`);
+            const sid = sessionIdRef.current;
+            if (sid) {
+              void getCostReport(sid).then((report) => {
+                const req = report.session.gatewayRequests ?? 0;
+                if (req === 0) return;
+                const cached = report.session.gatewayCachedRequests ?? 0;
+                const pct = (cached / req * 100).toFixed(1);
+                setEvents((e) => [
+                  ...e,
+                  {
+                    kind: "info",
+                    key: mkKey(),
+                    text: `cache hits (session): ${cached}/${req} (${pct}%)`
+                  }
+                ]);
+              }).catch(() => {
+              });
+            }
           } else {
             lines.push("gateway: off (direct Workers AI)");
           }