kimiflare 0.19.0 → 0.20.0

package/dist/index.js

@@ -41,6 +41,12 @@ async function loadConfig() {
   const compiledContext = envCompiled === "1" || envCompiled === "true" ? true : false;
   const envImageTurns = process.env.KIMIFLARE_IMAGE_HISTORY_TURNS;
   const imageHistoryTurns = envImageTurns ? parseInt(envImageTurns, 10) : void 0;
+  const envMaxToolIterations = process.env.KIMIFLARE_MAX_TOOL_ITERATIONS;
+  const maxToolIterations = envMaxToolIterations ? parseInt(envMaxToolIterations, 10) : void 0;
+  const envMaxInputTokens = process.env.KIMIFLARE_MAX_INPUT_TOKENS;
+  const maxInputTokens = envMaxInputTokens ? parseInt(envMaxInputTokens, 10) : void 0;
+  const envMaxCompletionTokens = process.env.KIMIFLARE_MAX_COMPLETION_TOKENS;
+  const maxCompletionTokens = envMaxCompletionTokens ? parseInt(envMaxCompletionTokens, 10) : void 0;
   if (envAccount && envToken) {
     return {
       accountId: envAccount,
@@ -53,7 +59,10 @@ async function loadConfig() {
       coauthorEmail: envCoauthor?.email,
       cacheStablePrompts,
       compiledContext,
-      imageHistoryTurns: Number.isNaN(imageHistoryTurns) ? void 0 : imageHistoryTurns
+      imageHistoryTurns: Number.isNaN(imageHistoryTurns) ? void 0 : imageHistoryTurns,
+      maxToolIterations: Number.isNaN(maxToolIterations) ? void 0 : maxToolIterations,
+      maxInputTokens: Number.isNaN(maxInputTokens) ? void 0 : maxInputTokens,
+      maxCompletionTokens: Number.isNaN(maxCompletionTokens) ? void 0 : maxCompletionTokens
     };
   }
   try {
@@ -72,7 +81,10 @@ async function loadConfig() {
         mcpServers: parsed.mcpServers,
         cacheStablePrompts: parsed.cacheStablePrompts ?? cacheStablePrompts,
         compiledContext: parsed.compiledContext ?? compiledContext,
-        imageHistoryTurns: Number.isNaN(imageHistoryTurns) ? parsed.imageHistoryTurns : imageHistoryTurns
+        imageHistoryTurns: Number.isNaN(imageHistoryTurns) ? parsed.imageHistoryTurns : imageHistoryTurns,
+        maxToolIterations: Number.isNaN(maxToolIterations) ? parsed.maxToolIterations : maxToolIterations,
+        maxInputTokens: Number.isNaN(maxInputTokens) ? parsed.maxInputTokens : maxInputTokens,
+        maxCompletionTokens: Number.isNaN(maxCompletionTokens) ? parsed.maxCompletionTokens : maxCompletionTokens
       };
     }
   } catch {
@@ -179,14 +191,14 @@ function stableStringify(value, replacer, space) {
   const sorted = sortKeys(value);
   return JSON.stringify(sorted, replacer, space);
 }
-function stripOldImages(messages, keepLastTurns) {
-  if (keepLastTurns < 0) return messages;
+function stripOldImages(messages, keepLastTurns2) {
+  if (keepLastTurns2 < 0) return messages;
   let userCount = 0;
   let cutoffIndex = messages.length;
   for (let i = messages.length - 1; i >= 0; i--) {
     if (messages[i].role === "user") {
       userCount++;
-      if (userCount === keepLastTurns) {
+      if (userCount === keepLastTurns2) {
         cutoffIndex = i;
         break;
       }
@@ -227,7 +239,7 @@ async function* runKimi(opts2) {
     ...opts2.tools && opts2.tools.length ? { tools: opts2.tools, tool_choice: "auto", parallel_tool_calls: true } : {},
     stream: true,
     temperature: opts2.temperature ?? 0.2,
-    max_completion_tokens: opts2.maxCompletionTokens ?? 16384
+    max_completion_tokens: opts2.maxCompletionTokens ?? 4096
   };
   if (opts2.reasoningEffort) {
     body.reasoning_effort = opts2.reasoningEffort;
@@ -417,7 +429,7 @@ var init_client = __esm({
     init_errors();
     init_messages();
     RETRYABLE_CODES = /* @__PURE__ */ new Set([3040]);
-    MAX_ATTEMPTS = 5;
+    MAX_ATTEMPTS = 2;
   }
 });
 
@@ -597,6 +609,37 @@ async function logCostDebug(entry) {
   await rotateJsonl(debugPath(), RETENTION.costDebugMaxBytes, RETENTION.costDebugRotations);
   await appendFile(debugPath(), JSON.stringify(entry) + "\n", "utf8");
 }
+function usageDir() {
+  return join3(homedir2(), ".kimiflare");
+}
+function usagePath() {
+  return join3(usageDir(), "usage.jsonl");
+}
+async function logTurnTokenMetrics(metrics) {
+  await mkdir2(usageDir(), { recursive: true });
+  await rotateJsonl(usagePath(), RETENTION.costDebugMaxBytes, RETENTION.costDebugRotations);
+  await appendFile(usagePath(), JSON.stringify(metrics) + "\n", "utf8");
+}
+function buildTurnTokenMetrics(sessionId, turn, breakdown, estimatedOutputTokenCap, wasCompacted, removedCount, exceedsLimit) {
+  return {
+    v: LOG_VERSION,
+    ts: now(),
+    sessionId,
+    turn,
+    estimatedInputTokens: breakdown.total,
+    estimatedOutputTokenCap,
+    messageCount: breakdown.messageCount,
+    toolOutputCount: breakdown.toolOutputCount,
+    tokensFromSystem: breakdown.fromSystem,
+    tokensFromSession: breakdown.fromSession,
+    tokensFromTools: breakdown.fromTools,
+    tokensFromHistory: breakdown.fromHistory,
+    tokensFromUserInput: breakdown.fromUserInput,
+    wasCompacted,
+    removedCount,
+    exceedsLimit
+  };
+}
 function serializePrefix(messages) {
   let end = 0;
   while (end < messages.length && messages[end].role === "system") {
@@ -741,12 +784,282 @@ var init_strip_reasoning = __esm({
   }
 });
 
+// src/agent/token-limits.ts
+function loadSafetyLimits() {
+  return {
+    maxInputTokensPerRequest: parseIntEnv("KIMIFLARE_MAX_INPUT_TOKENS", DEFAULT_SAFETY_LIMITS.maxInputTokensPerRequest),
+    warningThreshold: parseIntEnv("KIMIFLARE_WARNING_TOKENS", DEFAULT_SAFETY_LIMITS.warningThreshold),
+    maxLlmCallsPerUserAction: parseIntEnv("KIMIFLARE_MAX_LLM_CALLS", DEFAULT_SAFETY_LIMITS.maxLlmCallsPerUserAction),
+    maxRetriesPerLlmCall: parseIntEnv("KIMIFLARE_MAX_RETRIES", DEFAULT_SAFETY_LIMITS.maxRetriesPerLlmCall),
+    maxCompletionTokens: parseIntEnv("KIMIFLARE_MAX_COMPLETION_TOKENS", DEFAULT_SAFETY_LIMITS.maxCompletionTokens),
+    maxToolIterations: parseIntEnv("KIMIFLARE_MAX_TOOL_ITERATIONS", DEFAULT_SAFETY_LIMITS.maxToolIterations),
+    maxRecentMessages: parseIntEnv("KIMIFLARE_MAX_RECENT_MESSAGES", DEFAULT_SAFETY_LIMITS.maxRecentMessages),
+    maxToolOutputChars: parseIntEnv("KIMIFLARE_MAX_TOOL_OUTPUT_CHARS", DEFAULT_SAFETY_LIMITS.maxToolOutputChars)
+  };
+}
+function parseIntEnv(name, fallback) {
+  const raw = process.env[name];
+  if (!raw) return fallback;
+  const n = parseInt(raw, 10);
+  return Number.isNaN(n) ? fallback : n;
+}
+function estimateTokens(text) {
+  return Math.ceil(text.length / 4);
+}
+function estimateMessageTokens(m) {
+  let chars = 0;
+  if (typeof m.content === "string") {
+    chars = m.content.length;
+  } else if (Array.isArray(m.content)) {
+    for (const part of m.content) {
+      if (part.type === "text") chars += part.text.length;
+      else if (part.type === "image_url") chars += 1e3;
+    }
+  }
+  if (m.reasoning_content) chars += m.reasoning_content.length;
+  if (m.tool_calls) {
+    for (const tc of m.tool_calls) {
+      chars += tc.function.name.length;
+      chars += tc.function.arguments.length;
+    }
+  }
+  return Math.ceil(chars / 4) + 4;
+}
+function estimateMessagesTokens(messages) {
+  return messages.reduce((sum, m) => sum + estimateMessageTokens(m), 0);
+}
+function estimateToolDefsTokens(tools) {
+  return estimateTokens(JSON.stringify(tools));
+}
+function breakdownTokens(systemMessages, sessionMessages, toolDefs, historyMessages, userMessage) {
+  const fromSystem = estimateMessagesTokens(systemMessages);
+  const fromSession = estimateMessagesTokens(sessionMessages);
+  const fromTools = estimateToolDefsTokens(toolDefs);
+  const fromHistory = estimateMessagesTokens(historyMessages);
+  const fromUserInput = userMessage ? estimateMessageTokens(userMessage) : 0;
+  return {
+    total: fromSystem + fromSession + fromTools + fromHistory + fromUserInput,
+    fromSystem,
+    fromSession,
+    fromTools,
+    fromHistory,
+    fromUserInput,
+    messageCount: systemMessages.length + sessionMessages.length + historyMessages.length + (userMessage ? 1 : 0),
+    toolOutputCount: historyMessages.filter((m) => m.role === "tool").length
+  };
+}
+function compactHistoryForSafety(messages, targetTokens) {
+  let compacted = messages.map((m) => {
+    if (m.role === "tool" && typeof m.content === "string" && m.content.length > 200) {
+      const lines = m.content.split("\n");
+      const firstLine2 = lines[0] ?? "";
+      const truncated = lines.length > 3 || m.content.length > 200;
+      return {
+        ...m,
+        content: `[${m.name ?? "tool"} result${truncated ? " (truncated)" : ""}] ${firstLine2.slice(0, 120)}`
+      };
+    }
+    return m;
+  });
+  let removedCount = 0;
+  while (estimateMessagesTokens(compacted) > targetTokens && compacted.length > 2) {
+    const dropIndex = compacted.findIndex((m, i) => i > 0 && m.role !== "system" && m.role !== "user");
+    if (dropIndex === -1) break;
+    compacted.splice(dropIndex, 1);
+    removedCount++;
+  }
+  return { messages: compacted, removedCount };
+}
+var DEFAULT_SAFETY_LIMITS;
+var init_token_limits = __esm({
+  "src/agent/token-limits.ts"() {
+    "use strict";
+    DEFAULT_SAFETY_LIMITS = {
+      maxInputTokensPerRequest: 3e4,
+      warningThreshold: 15e3,
+      maxLlmCallsPerUserAction: 10,
+      maxRetriesPerLlmCall: 2,
+      maxCompletionTokens: 4096,
+      maxToolIterations: 10,
+      maxRecentMessages: 4,
+      maxToolOutputChars: 800
+    };
+  }
+});
+
+// src/agent/tool-output-summarizer.ts
+import { createHash } from "crypto";
+function normalizeForHash(text) {
+  return text.toLowerCase().replace(/\s+/g, " ").trim().slice(0, 5e3);
+}
+function stableHash(text) {
+  return createHash("sha256").update(text).digest("hex").slice(0, 16);
+}
+function clearOutputHashCache() {
+  outputHashCache.clear();
+}
+function summarizeToolOutput(toolCallId, name, rawContent, maxChars = DEFAULT_MAX_CHARS) {
+  const normalized = normalizeForHash(rawContent);
+  const hash = stableHash(normalized);
+  const cached = outputHashCache.get(hash);
+  if (cached && cached.firstSeenId !== toolCallId) {
+    const ref = `same as previous ${cached.name ?? "tool"} call (result_id=${hash})`;
+    return {
+      tool_call_id: toolCallId,
+      name,
+      content: ref,
+      truncated: false
+    };
+  }
+  if (!cached) {
+    const preview = rawContent.slice(0, 120).replace(/\s+/g, " ");
+    outputHashCache.set(hash, { name, firstSeenId: toolCallId, preview });
+  }
+  const isFailure = rawContent.startsWith("Error:") || rawContent.startsWith("error:") || rawContent.includes("exit code") || rawContent.includes("not found") || rawContent.includes("No such file");
+  const isNoisy = rawContent.length > 0 && (rawContent.split("\n").length > 100 || rawContent.length > maxChars * 2);
+  if (isFailure && rawContent.length > 200) {
+    const firstLine2 = rawContent.split("\n")[0] ?? "";
+    return {
+      tool_call_id: toolCallId,
+      name,
+      content: `[${name ?? "tool"} failed] ${firstLine2.slice(0, 160)}`,
+      truncated: true
+    };
+  }
+  if (rawContent.length <= maxChars) {
+    return {
+      tool_call_id: toolCallId,
+      name,
+      content: rawContent,
+      truncated: false
+    };
+  }
+  const truncated = rawContent.slice(0, maxChars);
+  const lastNewline = truncated.lastIndexOf("\n");
+  const clean = lastNewline > maxChars * 0.5 ? truncated.slice(0, lastNewline) : truncated;
+  return {
+    tool_call_id: toolCallId,
+    name,
+    content: `${clean}
+... (${rawContent.length - clean.length} more chars truncated)`,
+    truncated: true
+  };
+}
+function summarizeToolMessage(msg, maxChars = DEFAULT_MAX_CHARS) {
+  if (msg.role !== "tool" || typeof msg.content !== "string") {
+    return msg;
+  }
+  const summary = summarizeToolOutput(msg.tool_call_id ?? "", msg.name, msg.content, maxChars);
+  return {
+    ...msg,
+    content: summary.content
+  };
+}
+function summarizeToolMessages(messages, maxChars = DEFAULT_MAX_CHARS) {
+  return messages.map((m) => m.role === "tool" ? summarizeToolMessage(m, maxChars) : m);
+}
+var DEFAULT_MAX_CHARS, outputHashCache;
+var init_tool_output_summarizer = __esm({
+  "src/agent/tool-output-summarizer.ts"() {
+    "use strict";
+    DEFAULT_MAX_CHARS = 800;
+    outputHashCache = /* @__PURE__ */ new Map();
+  }
+});
+
+// src/agent/context-builder.ts
+function buildContext(opts2) {
+  const { allMessages, systemMessages, sessionMessages, toolDefs, limits, currentUserMessage } = opts2;
+  const prefixLength = systemMessages.length + sessionMessages.length;
+  let history = allMessages.slice(prefixLength);
+  history = summarizeToolMessages(history, limits.maxToolOutputChars);
+  const recentHistory = keepLastTurns(history, limits.maxRecentMessages);
+  const contextMessages = [
+    ...systemMessages,
+    ...sessionMessages,
+    ...recentHistory
+  ];
+  if (currentUserMessage) {
+    contextMessages.push(currentUserMessage);
+  }
+  let breakdown = breakdownTokens(
+    systemMessages,
+    sessionMessages,
+    toolDefs,
+    recentHistory,
+    currentUserMessage ?? null
+  );
+  let wasCompacted = false;
+  let removedCount = 0;
+  if (breakdown.total > limits.maxInputTokensPerRequest) {
+    const target = limits.maxInputTokensPerRequest;
+    const compacted = compactHistoryForSafety(recentHistory, target - breakdown.fromSystem - breakdown.fromSession - breakdown.fromTools - breakdown.fromUserInput);
+    if (compacted.removedCount > 0) {
+      wasCompacted = true;
+      removedCount = compacted.removedCount;
+      const newContext = [
+        ...systemMessages,
+        ...sessionMessages,
+        ...compacted.messages
+      ];
+      if (currentUserMessage) {
+        newContext.push(currentUserMessage);
+      }
+      breakdown = breakdownTokens(
+        systemMessages,
+        sessionMessages,
+        toolDefs,
+        compacted.messages,
+        currentUserMessage ?? null
+      );
+      return {
+        messages: newContext,
+        breakdown,
+        wasCompacted,
+        removedCount,
+        exceedsLimit: breakdown.total > limits.maxInputTokensPerRequest
+      };
+    }
+  }
+  return {
+    messages: contextMessages,
+    breakdown,
+    wasCompacted,
+    removedCount,
+    exceedsLimit: breakdown.total > limits.maxInputTokensPerRequest
+  };
+}
+function keepLastTurns(messages, maxTurns) {
+  if (maxTurns <= 0) return [];
+  const turnStarts = [];
+  for (let i = 0; i < messages.length; i++) {
+    if (messages[i].role === "user") {
+      turnStarts.push(i);
+    }
+  }
+  if (turnStarts.length <= maxTurns) {
+    return messages;
+  }
+  const startIndex = turnStarts[turnStarts.length - maxTurns] ?? 0;
+  return messages.slice(startIndex);
+}
+var init_context_builder = __esm({
+  "src/agent/context-builder.ts"() {
+    "use strict";
+    init_token_limits();
+    init_tool_output_summarizer();
+  }
+});
+
 // src/agent/loop.ts
 async function runAgentTurn(opts2) {
-  const max = opts2.maxToolIterations ?? 50;
+  const limits = loadSafetyLimits();
+  const max = opts2.maxToolIterations ?? limits.maxToolIterations;
   const toolDefs = toOpenAIToolDefs(opts2.tools);
   let turn = 0;
   let lastUsage = null;
+  const systemMessages = opts2.systemMessages ?? extractSystemMessages(opts2.messages);
+  const sessionMessages = opts2.sessionMessages ?? [];
   for (let iter = 0; iter < max; iter++) {
     turn++;
     const previousMessages = opts2.messages.slice();
@@ -766,8 +1079,8 @@ async function runAgentTurn(opts2) {
         keepLast: Number.isNaN(keepLast) ? 1 : keepLast
       });
       if (shadowStrip) {
-        const originalSections = analyzePrompt(opts2.messages);
-        const strippedSections = analyzePrompt(stripped);
+        const originalSections = analyzePromptSections(opts2.messages);
+        const strippedSections = analyzePromptSections(stripped);
         const originalApproxTokens = originalSections.reduce(
           (sum, s) => sum + s.approxTokens,
           0
@@ -791,15 +1104,46 @@ async function runAgentTurn(opts2) {
     if (opts2.keepLastImageTurns !== void 0) {
       apiMessages = stripOldImages(apiMessages, opts2.keepLastImageTurns);
     }
+    const currentUserMessage = findCurrentUserMessage(apiMessages);
+    const context = buildContext({
+      allMessages: apiMessages,
+      systemMessages,
+      sessionMessages,
+      toolDefs,
+      limits,
+      currentUserMessage
+    });
+    if (opts2.sessionId) {
+      void logTurnTokenMetrics(
+        buildTurnTokenMetrics(
+          opts2.sessionId,
+          turn,
+          context.breakdown,
+          opts2.maxCompletionTokens ?? limits.maxCompletionTokens,
+          context.wasCompacted,
+          context.removedCount,
+          context.exceedsLimit
+        )
+      );
+    }
+    if (context.exceedsLimit) {
+      const assistantMsg3 = {
+        role: "assistant",
+        content: `I cannot continue: the conversation context exceeds the safety limit of ${limits.maxInputTokensPerRequest} tokens. Try running /compact or /clear to reduce context size.`
+      };
+      opts2.messages.push(assistantMsg3);
+      opts2.callbacks.onAssistantFinal?.(assistantMsg3);
+      return;
+    }
     const events = runKimi({
       accountId: opts2.accountId,
       apiToken: opts2.apiToken,
       model: opts2.model,
-      messages: apiMessages,
+      messages: context.messages,
       tools: toolDefs,
       signal: opts2.signal,
       temperature: opts2.temperature,
-      maxCompletionTokens: opts2.maxCompletionTokens,
+      maxCompletionTokens: opts2.maxCompletionTokens ?? limits.maxCompletionTokens,
       reasoningEffort: opts2.reasoningEffort,
       sessionId: opts2.sessionId
     });
@@ -838,7 +1182,7 @@ async function runAgentTurn(opts2) {
           break;
       }
     }
-    const assistantMsg = {
+    const assistantMsg2 = {
       role: "assistant",
       content: content ? sanitizeString(content) : null,
       ...reasoning ? { reasoning_content: sanitizeString(reasoning) } : {},
@@ -852,8 +1196,8 @@ async function runAgentTurn(opts2) {
         }))
       } : {}
     };
-    opts2.messages.push(assistantMsg);
-    opts2.callbacks.onAssistantFinal?.(assistantMsg);
+    opts2.messages.push(assistantMsg2);
+    opts2.callbacks.onAssistantFinal?.(assistantMsg2);
     if (toolCalls.length === 0) {
       if (opts2.sessionId && lastUsage) {
         void logTurnDebug({
@@ -896,7 +1240,36 @@ async function runAgentTurn(opts2) {
       });
     }
   }
-  throw new Error(`kimiflare: tool iteration limit reached (${opts2.maxToolIterations ?? 50})`);
+  const remaining = toolCallsFromMessages(opts2.messages);
+  const assistantMsg = {
+    role: "assistant",
+    content: `I reached the tool iteration limit (${max}). There ${remaining === 1 ? "is" : "are"} ${remaining} pending tool call${remaining === 1 ? "" : "s"} that could not be executed. Run /compact or /clear to reset context, or rephrase your request.`
+  };
+  opts2.messages.push(assistantMsg);
+  opts2.callbacks.onAssistantFinal?.(assistantMsg);
+}
+function extractSystemMessages(messages) {
+  const end = messages.findIndex((m) => m.role !== "system");
+  return end === -1 ? messages.slice() : messages.slice(0, end);
+}
+function findCurrentUserMessage(messages) {
+  const prefixEnd = messages.findIndex((m) => m.role !== "system");
+  const history = prefixEnd === -1 ? [] : messages.slice(prefixEnd);
+  for (let i = history.length - 1; i >= 0; i--) {
+    if (history[i].role === "user") {
+      return history[i];
+    }
+  }
+  return null;
+}
+function toolCallsFromMessages(messages) {
+  let count = 0;
+  for (const m of messages) {
+    if (m.role === "assistant" && m.tool_calls) {
+      count += m.tool_calls.length;
+    }
+  }
+  return count;
 }
 function validateToolArguments(raw) {
   if (!raw || !raw.trim()) return "{}";
@@ -907,6 +1280,25 @@ function validateToolArguments(raw) {
     return "{}";
   }
 }
+function analyzePromptSections(messages) {
+  return messages.map((m) => {
+    let chars = 0;
+    if (typeof m.content === "string") {
+      chars = m.content.length;
+    } else if (Array.isArray(m.content)) {
+      for (const p of m.content) {
+        if (p.type === "text") chars += p.text.length;
+      }
+    }
+    if (m.reasoning_content) chars += m.reasoning_content.length;
+    if (m.tool_calls) {
+      for (const tc of m.tool_calls) {
+        chars += tc.function.name.length + tc.function.arguments.length;
+      }
+    }
+    return { role: m.role, chars, approxTokens: Math.ceil(chars / 4) };
+  });
+}
 var init_loop = __esm({
   "src/agent/loop.ts"() {
     "use strict";
@@ -915,6 +1307,8 @@ var init_loop = __esm({
     init_messages();
     init_cost_debug();
     init_strip_reasoning();
+    init_token_limits();
+    init_context_builder();
   }
 });
 
@@ -2399,6 +2793,77 @@ var init_update_check = __esm({
   }
 });
 
+// src/usage-cli.ts
+var usage_cli_exports = {};
+__export(usage_cli_exports, {
+  showUsageLog: () => showUsageLog
+});
+import { readFile as readFile7 } from "fs/promises";
+import { homedir as homedir6 } from "os";
+import { join as join7 } from "path";
+function usagePath2() {
+  return join7(homedir6(), ".kimiflare", "usage.jsonl");
+}
+function fmt(n) {
+  return n.toLocaleString();
+}
+async function showUsageLog() {
+  const path = usagePath2();
+  let raw;
+  try {
+    raw = await readFile7(path, "utf8");
+  } catch {
+    console.log("No usage log found at " + path);
+    return;
+  }
+  const lines = raw.trim().split("\n").filter(Boolean);
+  if (lines.length === 0) {
+    console.log("Usage log is empty.");
+    return;
+  }
+  const entries = [];
+  for (const line of lines) {
+    try {
+      entries.push(JSON.parse(line));
+    } catch {
+    }
+  }
+  if (entries.length === 0) {
+    console.log("No valid entries in usage log.");
+    return;
+  }
+  const bySession = /* @__PURE__ */ new Map();
+  for (const e of entries) {
+    const arr = bySession.get(e.sessionId) ?? [];
+    arr.push(e);
+    bySession.set(e.sessionId, arr);
+  }
+  console.log(`Usage log: ${lines.length} entries, ${bySession.size} session(s)
+`);
+  for (const [sessionId, sessEntries] of bySession) {
+    const last = sessEntries[sessEntries.length - 1];
+    const totalInput = sessEntries.reduce((s, e) => s + e.estimatedInputTokens, 0);
+    const avgInput = Math.round(totalInput / sessEntries.length);
+    console.log(`Session: ${sessionId.slice(0, 16)}\u2026  Turns: ${sessEntries.length}`);
+    console.log(`  Last turn:  ${fmt(last.estimatedInputTokens)} input tokens  /  ${fmt(last.estimatedOutputTokenCap)} output cap`);
+    console.log(`  Avg input:  ${fmt(avgInput)} tokens`);
+    console.log(`  Messages:   ${last.messageCount}  |  Tool outputs: ${last.toolOutputCount}`);
+    console.log(`  Breakdown:  system=${fmt(last.tokensFromSystem)}  session=${fmt(last.tokensFromSession)}  tools=${fmt(last.tokensFromTools)}  history=${fmt(last.tokensFromHistory)}  user=${fmt(last.tokensFromUserInput)}`);
+    if (last.wasCompacted) {
+      console.log(`  \u26A0\uFE0F  Compacted: removed ${last.removedCount} messages`);
+    }
+    if (last.exceedsLimit) {
+      console.log(`  \u274C  EXCEEDS LIMIT`);
+    }
+    console.log("");
+  }
+}
+var init_usage_cli = __esm({
+  "src/usage-cli.ts"() {
+    "use strict";
+  }
+});
+
 // src/agent/compact.ts
 function indexOfNthUserFromEnd(messages, n) {
   let seen = 0;
@@ -2605,7 +3070,7 @@ var init_session_state = __esm({
 function approxTokens2(n) {
   return Math.round(n / 4);
 }
-function estimateMessageTokens(m) {
+function estimateMessageTokens2(m) {
   let chars = 0;
   if (typeof m.content === "string") {
     chars = m.content.length;
@@ -2621,7 +3086,7 @@ function estimateMessageTokens(m) {
   return approxTokens2(chars);
 }
 function estimatePromptTokens(messages) {
-  return messages.reduce((sum, m) => sum + estimateMessageTokens(m), 0);
+  return messages.reduce((sum, m) => sum + estimateMessageTokens2(m), 0);
 }
 function groupIntoTurns(messages) {
   const prefix = [];
@@ -2801,10 +3266,10 @@ function shouldCompact(opts2) {
   return tokens > tokenThreshold || turns.length > turnThreshold;
 }
 function compactMessages2(opts2) {
-  const keepLastTurns = opts2.keepLastTurns ?? 4;
+  const keepLastTurns2 = opts2.keepLastTurns ?? 4;
   const { prefix, turns } = groupIntoTurns(opts2.messages);
   const tokensBefore = estimatePromptTokens(opts2.messages);
-  if (turns.length <= keepLastTurns) {
+  if (turns.length <= keepLastTurns2) {
     return {
       newMessages: opts2.messages,
       newState: opts2.state,
@@ -2818,8 +3283,8 @@ function compactMessages2(opts2) {
       }
     };
   }
-  const toCompact = turns.slice(0, turns.length - keepLastTurns);
-  const toKeep = turns.slice(turns.length - keepLastTurns);
+  const toCompact = turns.slice(0, turns.length - keepLastTurns2);
+  const toKeep = turns.slice(turns.length - keepLastTurns2);
   let newState = { ...opts2.state };
   let archivedCount = 0;
   for (let i = 0; i < toCompact.length; i++) {
@@ -4850,12 +5315,12 @@ __export(sessions_exports, {
   pruneSessions: () => pruneSessions,
   saveSession: () => saveSession
 });
-import { readFile as readFile7, writeFile as writeFile5, mkdir as mkdir5, readdir as readdir2, stat as stat3 } from "fs/promises";
-import { homedir as homedir6 } from "os";
-import { join as join7 } from "path";
+import { readFile as readFile8, writeFile as writeFile5, mkdir as mkdir5, readdir as readdir2, stat as stat3 } from "fs/promises";
+import { homedir as homedir7 } from "os";
+import { join as join8 } from "path";
 function sessionsDir() {
-  const xdg = process.env.XDG_DATA_HOME || join7(homedir6(), ".local", "share");
-  return join7(xdg, "kimiflare", "sessions");
+  const xdg = process.env.XDG_DATA_HOME || join8(homedir7(), ".local", "share");
+  return join8(xdg, "kimiflare", "sessions");
 }
 function sanitize(text) {
   return text.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "").slice(0, 40);
@@ -4868,7 +5333,7 @@ function makeSessionId(firstPrompt) {
 async function saveSession(file) {
   const dir = sessionsDir();
   await mkdir5(dir, { recursive: true });
-  const path = join7(dir, `${file.id}.json`);
+  const path = join8(dir, `${file.id}.json`);
   await writeFile5(path, JSON.stringify(file, null, 2), "utf8");
   return path;
 }
@@ -4888,9 +5353,9 @@ async function listSessions(limit = 30) {
   const summaries = [];
   for (const name of entries) {
     if (!name.endsWith(".json")) continue;
-    const path = join7(dir, name);
+    const path = join8(dir, name);
     try {
-      const [s, raw] = await Promise.all([stat3(path), readFile7(path, "utf8")]);
+      const [s, raw] = await Promise.all([stat3(path), readFile8(path, "utf8")]);
       const parsed = JSON.parse(raw);
       const firstUser = parsed.messages.find((m) => m.role === "user");
       const firstPrompt = typeof firstUser?.content === "string" ? firstUser.content : firstUser?.content ? firstUser.content.find((p) => p.type === "text")?.text ?? "(no prompt)" : "(no prompt)";
@@ -4909,7 +5374,7 @@ async function listSessions(limit = 30) {
   return summaries.slice(0, limit);
 }
 async function loadSession(filePath) {
-  const raw = await readFile7(filePath, "utf8");
+  const raw = await readFile8(filePath, "utf8");
   return JSON.parse(raw);
 }
 var init_sessions = __esm({
@@ -4920,10 +5385,10 @@ var init_sessions = __esm({
 });
 
 // src/util/image.ts
-import { readFile as readFile8 } from "fs/promises";
+import { readFile as readFile9 } from "fs/promises";
 import { basename as basename2 } from "path";
 async function encodeImageFile(filePath) {
-  const buf = await readFile8(filePath);
+  const buf = await readFile9(filePath);
   if (buf.byteLength > MAX_IMAGE_BYTES) {
     throw new Error(
       `image too large (${(buf.byteLength / 1024 / 1024).toFixed(1)} MB); max is ${MAX_IMAGE_BYTES / 1024 / 1024} MB`
@@ -4959,15 +5424,15 @@ var init_image = __esm({
 });
 
 // src/usage-tracker.ts
-import { readFile as readFile9, writeFile as writeFile6, mkdir as mkdir6 } from "fs/promises";
-import { homedir as homedir7 } from "os";
-import { join as join8 } from "path";
-function usageDir() {
-  const xdg = process.env.XDG_DATA_HOME || join8(homedir7(), ".local", "share");
-  return join8(xdg, "kimiflare");
+import { readFile as readFile10, writeFile as writeFile6, mkdir as mkdir6 } from "fs/promises";
+import { homedir as homedir8 } from "os";
+import { join as join9 } from "path";
+function usageDir2() {
+  const xdg = process.env.XDG_DATA_HOME || join9(homedir8(), ".local", "share");
+  return join9(xdg, "kimiflare");
 }
-function usagePath() {
-  return join8(usageDir(), "usage.json");
+function usagePath3() {
+  return join9(usageDir2(), "usage.json");
 }
 function today() {
   return (/* @__PURE__ */ new Date()).toISOString().slice(0, 10);
@@ -4978,7 +5443,7 @@ function cutoffDate(daysBack) {
 }
 async function loadLog() {
   try {
-    const raw = await readFile9(usagePath(), "utf8");
+    const raw = await readFile10(usagePath3(), "utf8");
     const parsed = JSON.parse(raw);
     if (parsed.version === LOG_VERSION2) return parsed;
   } catch {
@@ -4986,8 +5451,8 @@ async function loadLog() {
   return { version: LOG_VERSION2, days: [], sessions: [] };
 }
 async function saveLog(log) {
-  await mkdir6(usageDir(), { recursive: true });
-  await writeFile6(usagePath(), JSON.stringify(log, null, 2), "utf8");
+  await mkdir6(usageDir2(), { recursive: true });
+  await writeFile6(usagePath3(), JSON.stringify(log, null, 2), "utf8");
 }
 function getOrCreateDay(log, date) {
   let day = log.days.find((d) => d.date === date);
@@ -5104,7 +5569,7 @@ __export(app_exports, {
 import { useState as useState6, useRef as useRef3, useEffect as useEffect4, useCallback } from "react";
 import { Box as Box12, Text as Text13, useApp, useInput as useInput2, render } from "ink";
 import { existsSync } from "fs";
-import { join as join9 } from "path";
+import { join as join10 } from "path";
 import { unlink as unlink2 } from "fs/promises";
 import { jsx as jsx13, jsxs as jsxs12 } from "react/jsx-runtime";
 function capEvents(prev) {
@@ -5170,9 +5635,10 @@ function App({ initialCfg, initialUpdateResult }) {
   const [hasUpdate, setHasUpdate] = useState6(initialUpdateResult?.hasUpdate ?? false);
   const [latestVersion, setLatestVersion] = useState6(initialUpdateResult?.latestVersion ?? null);
   const cacheStableRef = useRef3(initialCfg?.cacheStablePrompts !== false);
-  const messagesRef = useRef3(
+  const systemMessagesRef = useRef3(
     makePrefixMessages(cacheStableRef.current, cfg?.model ?? DEFAULT_MODEL, "edit", ALL_TOOLS)
   );
+  const messagesRef = useRef3(systemMessagesRef.current.slice());
   const executorRef = useRef3(new ToolExecutor(ALL_TOOLS));
   const activeAsstIdRef = useRef3(null);
   const activeControllerRef = useRef3(null);
@@ -5257,7 +5723,7 @@ function App({ initialCfg, initialUpdateResult }) {
   useEffect4(() => {
     modeRef.current = mode;
     if (cacheStableRef.current) {
-      messagesRef.current[1] = {
+      const sessionMsg = {
         role: "system",
         content: buildSessionPrefix({
           cwd: process.cwd(),
@@ -5266,8 +5732,10 @@ function App({ initialCfg, initialUpdateResult }) {
           mode
         })
       };
+      messagesRef.current[1] = sessionMsg;
+      systemMessagesRef.current[1] = sessionMsg;
     } else {
-      messagesRef.current[0] = {
+      const sysMsg = {
         role: "system",
         content: buildSystemPrompt({
           cwd: process.cwd(),
@@ -5276,6 +5744,8 @@ function App({ initialCfg, initialUpdateResult }) {
           mode
         })
       };
+      messagesRef.current[0] = sysMsg;
+      systemMessagesRef.current[0] = sysMsg;
     }
     if (mode === "plan") {
       executorRef.current.clearSessionPermissions();
@@ -5350,7 +5820,7 @@ function App({ initialCfg, initialUpdateResult }) {
     }
     if (totalTools > 0) {
       if (cacheStableRef.current) {
-        messagesRef.current[1] = {
+        const sessionMsg = {
           role: "system",
           content: buildSessionPrefix({
             cwd: process.cwd(),
@@ -5359,8 +5829,10 @@ function App({ initialCfg, initialUpdateResult }) {
             mode: modeRef.current
           })
         };
+        messagesRef.current[1] = sessionMsg;
+        systemMessagesRef.current[1] = sessionMsg;
       } else {
-        messagesRef.current[0] = {
+        const sysMsg = {
           role: "system",
           content: buildSystemPrompt({
             cwd: process.cwd(),
@@ -5369,6 +5841,8 @@ function App({ initialCfg, initialUpdateResult }) {
             mode: modeRef.current
           })
         };
+        messagesRef.current[0] = sysMsg;
+        systemMessagesRef.current[0] = sysMsg;
       }
       setEvents((e) => [
         ...e,
@@ -5545,13 +6019,13 @@ function App({ initialCfg, initialUpdateResult }) {
     }
     const cwd = process.cwd();
     for (const name of ["KIMI.md", "KIMIFLARE.md", "AGENT.md"]) {
-      if (existsSync(join9(cwd, name))) {
+      if (existsSync(join10(cwd, name))) {
         setEvents((e) => [
           ...e,
           {
             kind: "info",
             key: mkKey(),
-            text: `${name} already exists at ${join9(cwd, name)} \u2014 delete it first if you want to regenerate`
+            text: `${name} already exists at ${join10(cwd, name)} \u2014 delete it first if you want to regenerate`
           }
         ]);
         return;
@@ -5591,6 +6065,7 @@ function App({ initialCfg, initialUpdateResult }) {
         reasoningEffort: effortRef.current,
         coauthor: cfg.coauthor !== false ? { name: cfg.coauthorName || "kimiflare", email: cfg.coauthorEmail || "kimiflare@proton.me" } : void 0,
         sessionId: ensureSessionId(),
+        systemMessages: systemMessagesRef.current,
         callbacks: {
           onAssistantStart: () => {
             const id = nextAssistantId++;
@@ -5668,9 +6143,9 @@ function App({ initialCfg, initialUpdateResult }) {
           })
         }
       });
-      if (existsSync(join9(cwd, "KIMI.md"))) {
+      if (existsSync(join10(cwd, "KIMI.md"))) {
         if (cacheStableRef.current) {
-          messagesRef.current[1] = {
+          const sessionMsg = {
             role: "system",
             content: buildSessionPrefix({
               cwd,
@@ -5679,8 +6154,10 @@ function App({ initialCfg, initialUpdateResult }) {
               mode: modeRef.current
             })
           };
+          messagesRef.current[1] = sessionMsg;
+          systemMessagesRef.current[1] = sessionMsg;
         } else {
-          messagesRef.current[0] = {
+          const sysMsg = {
             role: "system",
             content: buildSystemPrompt({
               cwd,
@@ -5689,6 +6166,8 @@ function App({ initialCfg, initialUpdateResult }) {
               mode: modeRef.current
             })
           };
+          messagesRef.current[0] = sysMsg;
+          systemMessagesRef.current[0] = sysMsg;
         }
         setEvents((e) => [
           ...e,
@@ -5715,7 +6194,10 @@ function App({ initialCfg, initialUpdateResult }) {
       if (!picked) return;
       try {
         const file = await loadSession(picked.filePath);
+        const prefixEnd = file.messages.findIndex((m) => m.role !== "system");
+        systemMessagesRef.current = prefixEnd === -1 ? file.messages.slice() : file.messages.slice(0, prefixEnd);
         messagesRef.current = file.messages;
+        clearOutputHashCache();
         sessionIdRef.current = file.id;
         if (file.sessionState && compiledContextRef.current) {
           sessionStateRef.current = file.sessionState;
@@ -5778,10 +6260,12 @@ function App({ initialCfg, initialUpdateResult }) {
       }
       if (c === "/clear") {
         if (cacheStableRef.current && messagesRef.current.length >= 2) {
-          messagesRef.current = [messagesRef.current[0], messagesRef.current[1]];
+          systemMessagesRef.current = [messagesRef.current[0], messagesRef.current[1]];
         } else {
-          messagesRef.current = [messagesRef.current[0]];
+          systemMessagesRef.current = [messagesRef.current[0]];
         }
+        messagesRef.current = systemMessagesRef.current.slice();
+        clearOutputHashCache();
         sessionIdRef.current = null;
         sessionStateRef.current = emptySessionState();
         artifactStoreRef.current = new ArtifactStore();
@@ -6103,6 +6587,7 @@ use: /thinking low | medium | high`
           coauthor: cfg.coauthor !== false ? { name: cfg.coauthorName || "kimiflare", email: cfg.coauthorEmail || "kimiflare@proton.me" } : void 0,
           sessionId: ensureSessionId(),
           keepLastImageTurns: cfg.imageHistoryTurns ?? 2,
+          systemMessages: systemMessagesRef.current,
           callbacks: {
             onAssistantStart: () => {
               const id = nextAssistantId++;
@@ -6433,6 +6918,7 @@ var init_app = __esm({
     init_sessions();
     init_image();
     init_usage_tracker();
+    init_tool_output_summarizer();
     CONTEXT_LIMIT = 262e3;
     AUTO_COMPACT_SUGGEST_PCT = 0.8;
     MAX_EVENTS = 500;
@@ -6457,22 +6943,30 @@ init_update_check();
 import { Command } from "commander";
 import { readFileSync as readFileSync2 } from "fs";
 import { fileURLToPath as fileURLToPath2 } from "url";
-import { dirname as dirname3, join as join10 } from "path";
+import { dirname as dirname3, join as join11 } from "path";
 function readPackageVersion() {
   try {
     const here = dirname3(fileURLToPath2(import.meta.url));
-    const pkg = JSON.parse(readFileSync2(join10(here, "..", "package.json"), "utf8"));
+    const pkg = JSON.parse(readFileSync2(join11(here, "..", "package.json"), "utf8"));
     return pkg.version ?? "0.0.0";
   } catch {
     return "0.0.0";
   }
 }
 var program = new Command();
-program.name("kimiflare").description("Terminal coding agent powered by Kimi-K2.6 on Cloudflare Workers AI.").version(readPackageVersion()).option("-p, --print <prompt>", "one-shot mode: send prompt, stream reply to stdout, exit").option("-m, --model <id>", "model id (defaults to @cf/moonshotai/kimi-k2.6)").option("--dangerously-allow-all", "auto-approve every permission prompt (print mode only)").option("--reasoning", "include reasoning in stdout (print mode only)").parse();
+program.name("kimiflare").description("Terminal coding agent powered by Kimi-K2.6 on Cloudflare Workers AI.").version(readPackageVersion()).option("-p, --print <prompt>", "one-shot mode: send prompt, stream reply to stdout, exit").option("-m, --model <id>", "model id (defaults to @cf/moonshotai/kimi-k2.6)").option("--dangerously-allow-all", "auto-approve every permission prompt (print mode only)").option("--reasoning", "include reasoning in stdout (print mode only)").command("usage").description("show per-turn token usage log from ~/.kimiflare/usage.jsonl").action(async () => {
+  const { showUsageLog: showUsageLog2 } = await Promise.resolve().then(() => (init_usage_cli(), usage_cli_exports));
+  await showUsageLog2();
+});
+program.parse();
 var opts = program.opts();
 async function main() {
   const cfg = await loadConfig();
   const updateResult = await checkForUpdate();
+  const args = program.args;
+  if (args[0] === "usage") {
+    return;
+  }
   if (opts.print !== void 0) {
     if (!cfg) {
       console.error(
@@ -6516,8 +7010,11 @@ async function runPrintMode(opts2) {
   }
   const cwd = process.cwd();
   const executor = new ToolExecutor(ALL_TOOLS);
+  const systemMessages = [
+    { role: "system", content: buildSystemPrompt({ cwd, tools: ALL_TOOLS, model: opts2.model }) }
+  ];
   const messages = [
-    { role: "system", content: buildSystemPrompt({ cwd, tools: ALL_TOOLS, model: opts2.model }) },
+    ...systemMessages,
     { role: "user", content: opts2.prompt }
   ];
   const controller = new AbortController();
@@ -6533,6 +7030,7 @@ async function runPrintMode(opts2) {
     executor,
     cwd,
     signal: controller.signal,
+    systemMessages,
     coauthor: opts2.coauthor !== false ? { name: opts2.coauthorName || "kimiflare", email: opts2.coauthorEmail || "kimiflare@proton.me" } : void 0,
     callbacks: {
       onReasoningDelta: opts2.showReasoning ? (delta) => {