kimiflare 0.14.0 → 0.15.0

package/dist/index.js

@@ -35,6 +35,8 @@ async function loadConfig() {
   const envEffort = readReasoningEffortEnv();
   const envTheme = process.env.KIMI_THEME;
   const envCoauthor = readCoauthorEnv();
+  const envCacheStable = process.env.KIMIFLARE_CACHE_STABLE_PROMPTS;
+  const cacheStablePrompts = envCacheStable === "0" || envCacheStable === "false" ? false : true;
   if (envAccount && envToken) {
     return {
       accountId: envAccount,
@@ -44,7 +46,8 @@ async function loadConfig() {
       reasoningEffort: envEffort,
       coauthor: envCoauthor?.enabled ?? true,
       coauthorName: envCoauthor?.name,
-      coauthorEmail: envCoauthor?.email
+      coauthorEmail: envCoauthor?.email,
+      cacheStablePrompts
     };
   }
   try {
@@ -60,7 +63,8 @@ async function loadConfig() {
         coauthor: envCoauthor?.enabled ?? parsed.coauthor ?? true,
         coauthorName: envCoauthor?.name ?? parsed.coauthorName,
         coauthorEmail: envCoauthor?.email ?? parsed.coauthorEmail,
-        mcpServers: parsed.mcpServers
+        mcpServers: parsed.mcpServers,
+        cacheStablePrompts: parsed.cacheStablePrompts ?? cacheStablePrompts
       };
     }
   } catch {
@@ -153,6 +157,20 @@ function jsonReplacer(_key, value) {
   }
   return value;
 }
+function stableStringify(value, replacer, space) {
+  function sortKeys(obj) {
+    if (obj === null || typeof obj !== "object") return obj;
+    if (Array.isArray(obj)) return obj.map(sortKeys);
+    const sorted2 = {};
+    const keys = Object.keys(obj).sort();
+    for (const k of keys) {
+      sorted2[k] = sortKeys(obj[k]);
+    }
+    return sorted2;
+  }
+  const sorted = sortKeys(value);
+  return JSON.stringify(sorted, replacer, space);
+}
 var init_messages = __esm({
   "src/agent/messages.ts"() {
     "use strict";
@@ -185,13 +203,17 @@ async function* runKimi(opts2) {
   for (let attempt = 0; attempt < MAX_ATTEMPTS; attempt++) {
     let res;
     try {
+      const headers = {
+        Authorization: `Bearer ${opts2.apiToken}`,
+        "Content-Type": "application/json"
+      };
+      if (opts2.sessionId) {
+        headers["X-Session-ID"] = opts2.sessionId;
+      }
       res = await fetch(url, {
         method: "POST",
-        headers: {
-          Authorization: `Bearer ${opts2.apiToken}`,
-          "Content-Type": "application/json"
-        },
-        body: JSON.stringify(body, jsonReplacer),
+        headers,
+        body: stableStringify(body, jsonReplacer),
         signal: opts2.signal
       });
     } catch (fetchErr) {
@@ -451,12 +473,76 @@ async function logCostDebug(entry) {
   await mkdir2(debugDir(), { recursive: true });
   await appendFile(debugPath(), JSON.stringify(entry) + "\n", "utf8");
 }
+function serializePrefix(messages) {
+  let end = 0;
+  while (end < messages.length && messages[end].role === "system") {
+    end++;
+  }
+  return messages.slice(0, end).map((m) => typeof m.content === "string" ? m.content : JSON.stringify(m.content)).join("\n---\n");
+}
+function comparePromptPrefixes(prev, curr) {
+  const prevPrefix = prev ? serializePrefix(prev) : "";
+  const currPrefix = serializePrefix(curr);
+  const totalChars = curr.reduce((sum, m) => {
+    if (typeof m.content === "string") return sum + m.content.length;
+    if (Array.isArray(m.content)) return sum + m.content.map((p) => p.type === "text" ? p.text.length : 0).reduce((a, b) => a + b, 0);
+    return sum;
+  }, 0);
+  let firstDiffByte = null;
+  let changedSegment = null;
+  if (prevPrefix !== currPrefix) {
+    const minLen = Math.min(prevPrefix.length, currPrefix.length);
+    for (let i = 0; i < minLen; i++) {
+      if (prevPrefix[i] !== currPrefix[i]) {
+        firstDiffByte = i;
+        break;
+      }
+    }
+    if (firstDiffByte === null && prevPrefix.length !== currPrefix.length) {
+      firstDiffByte = minLen;
+    }
+    if (curr.length >= 1 && curr[0].role === "system") {
+      const staticLen = typeof curr[0].content === "string" ? curr[0].content.length : JSON.stringify(curr[0].content).length;
+      if (firstDiffByte !== null && firstDiffByte < staticLen) {
+        changedSegment = "static";
+      } else if (curr.length >= 2 && curr[1].role === "system") {
+        const sessionLen = typeof curr[1].content === "string" ? curr[1].content.length : JSON.stringify(curr[1].content).length;
+        if (firstDiffByte !== null && firstDiffByte < staticLen + 5 + sessionLen) {
+          changedSegment = "session";
+        } else {
+          changedSegment = "dynamic";
+        }
+      } else {
+        changedSegment = "dynamic";
+      }
+    } else {
+      changedSegment = "dynamic";
+    }
+  } else {
+    changedSegment = "none";
+  }
+  const staticPrefixChars = curr.length > 0 && curr[0].role === "system" && typeof curr[0].content === "string" ? curr[0].content.length : 0;
+  const sessionPrefixChars = curr.length > 1 && curr[1].role === "system" && typeof curr[1].content === "string" ? curr[1].content.length : 0;
+  const dynamicSuffixChars = totalChars - staticPrefixChars - sessionPrefixChars;
+  return {
+    staticPrefixChars,
+    sessionPrefixChars,
+    dynamicSuffixChars,
+    firstDiffByte,
+    changedSegment,
+    cacheHitRatio: 0
+    // populated by caller with actual usage data
+  };
+}
 async function logTurnDebug(ctx) {
   const promptSections = analyzePrompt(ctx.messages);
   const promptTotalChars = promptSections.reduce((sum, s) => sum + s.chars, 0);
   const toolStats = buildToolStats(ctx.toolResults);
   const toolTotalRaw = toolStats.reduce((sum, t) => sum + t.rawBytes, 0);
   const toolTotalReduced = toolStats.reduce((sum, t) => sum + t.reducedBytes, 0);
+  const cacheDiagnostics = comparePromptPrefixes(ctx.previousMessages, ctx.messages);
+  const cachedTokens = ctx.usage.prompt_tokens_details?.cached_tokens ?? 0;
+  cacheDiagnostics.cacheHitRatio = ctx.usage.prompt_tokens > 0 ? cachedTokens / ctx.usage.prompt_tokens : 0;
   await logCostDebug({
     v: LOG_VERSION,
     ts: now(),
@@ -469,7 +555,8 @@ async function logTurnDebug(ctx) {
     toolStats,
     toolTotalRawBytes: toolTotalRaw,
     toolTotalReducedBytes: toolTotalReduced,
-    toolSavingsPct: toolTotalRaw > 0 ? Math.round((toolTotalRaw - toolTotalReduced) / toolTotalRaw * 100) : 0
+    toolSavingsPct: toolTotalRaw > 0 ? Math.round((toolTotalRaw - toolTotalReduced) / toolTotalRaw * 100) : 0,
+    cacheDiagnostics
   });
 }
 var LOG_VERSION;
@@ -488,6 +575,7 @@ async function runAgentTurn(opts2) {
   let lastUsage = null;
   for (let iter = 0; iter < max; iter++) {
     turn++;
+    const previousMessages = opts2.messages.slice();
     const toolCalls = [];
     const toolResults = [];
     let content = "";
@@ -502,7 +590,8 @@ async function runAgentTurn(opts2) {
       signal: opts2.signal,
       temperature: opts2.temperature,
       maxCompletionTokens: opts2.maxCompletionTokens,
-      reasoningEffort: opts2.reasoningEffort
+      reasoningEffort: opts2.reasoningEffort,
+      sessionId: opts2.sessionId
     });
     for await (const ev of events) {
       switch (ev.type) {
@@ -561,6 +650,7 @@ async function runAgentTurn(opts2) {
           sessionId: opts2.sessionId,
           turn,
           messages: opts2.messages,
+          previousMessages,
           toolResults,
           usage: lastUsage
         });
@@ -588,6 +678,7 @@ async function runAgentTurn(opts2) {
         sessionId: opts2.sessionId,
         turn,
         messages: opts2.messages,
+        previousMessages,
         toolResults,
         usage: lastUsage
       });
@@ -847,25 +938,8 @@ function loadContextFile(cwd) {
   }
   return null;
 }
-function buildSystemPrompt(opts2) {
-  const now2 = opts2.now ?? /* @__PURE__ */ new Date();
-  const date = now2.toISOString().slice(0, 10);
-  const shell = process.env.SHELL ? basename(process.env.SHELL) : "sh";
-  const toolsBlock = opts2.tools.map((t) => {
-    const perm = t.needsPermission ? " [needs user permission]" : "";
-    return `- \`${t.name}\`${perm}: ${t.description.split("\n")[0]}`;
-  }).join("\n");
-  const base = `You are kimiflare, an interactive coding assistant running in the user's terminal. You act on the user's local filesystem through the tools listed below. You are powered by the ${opts2.model} model on Cloudflare Workers AI.
-
-Environment:
-- Working directory: ${opts2.cwd}
-- Platform: ${platform()} ${release()}
-- Shell: ${shell}
-- Home: ${homedir3()}
-- Today: ${date}
-
-Tools available:
-${toolsBlock}
+function buildStaticPrefix(opts2) {
+  return `You are kimiflare, an interactive coding assistant running in the user's terminal. You act on the user's local filesystem through the tools listed below. You are powered by the ${opts2.model} model on Cloudflare Workers AI.
 
 How to work:
 - Prefer calling tools over guessing. Read files before editing them. Use \`glob\` and \`grep\` to explore code before assuming structure.
@@ -878,13 +952,39 @@ How to work:
 - If a request is ambiguous, ask one focused question instead of making large assumptions.
 - When you finish a task, stop. Do not add a closing summary.
 - When creating git commits, you must include \`Co-authored-by: kimiflare <kimiflare@proton.me>\` in the commit message so kimiflare is credited as a contributor. The bash tool will also auto-append this trailer when it detects git commit-creating commands.`;
+}
+function buildSessionPrefix(opts2) {
+  const now2 = opts2.now ?? /* @__PURE__ */ new Date();
+  const date = now2.toISOString().slice(0, 10);
+  const shell = process.env.SHELL ? basename(process.env.SHELL) : "sh";
+  const toolsBlock = opts2.tools.map((t) => {
+    const perm = t.needsPermission ? " [needs user permission]" : "";
+    return `- \`${t.name}\`${perm}: ${t.description.split("\n")[0]}`;
+  }).join("\n");
+  const env2 = `Environment:
+- Working directory: ${opts2.cwd}
+- Platform: ${platform()} ${release()}
+- Shell: ${shell}
+- Home: ${homedir3()}
+- Today: ${date}`;
+  const tools = `Tools available:
+${toolsBlock}`;
   const ctx = loadContextFile(opts2.cwd);
   const contextBlock = ctx ? `
 
 Project context from ${ctx.name} (${ctx.lineCount} lines, treat as authoritative):
 ${ctx.content.trim()}` : "";
   const modeBlock = opts2.mode ? systemPromptForMode(opts2.mode) : "";
-  return base + contextBlock + modeBlock;
+  return env2 + "\n\n" + tools + contextBlock + modeBlock;
+}
+function buildSystemPrompt(opts2) {
+  return buildStaticPrefix(opts2) + "\n\n" + buildSessionPrefix(opts2);
+}
+function buildSystemMessages(opts2) {
+  return [
+    { role: "system", content: buildStaticPrefix(opts2) },
+    { role: "system", content: buildSessionPrefix(opts2) }
+  ];
 }
 var CONTEXT_FILENAMES, MAX_CONTEXT_BYTES;
 var init_system_prompt = __esm({
@@ -1712,11 +1812,15 @@ function indexOfNthUserFromEnd(messages, n) {
 async function compactMessages(opts2) {
   const keep = opts2.keepLastTurns ?? 4;
   const messages = opts2.messages;
-  const systemMsg = messages.find((m) => m.role === "system");
-  if (!systemMsg) throw new Error("compact: no system message found");
+  let prefixEnd = 0;
+  while (prefixEnd < messages.length && messages[prefixEnd].role === "system") {
+    prefixEnd++;
+  }
+  const prefix = messages.slice(0, prefixEnd);
+  if (prefix.length === 0) throw new Error("compact: no system message found");
   const cutoffUserIdx = indexOfNthUserFromEnd(messages, keep);
   const firstKeepIdx = cutoffUserIdx >= 0 ? cutoffUserIdx : messages.length;
-  const toSummarize = messages.slice(1, firstKeepIdx);
+  const toSummarize = messages.slice(prefixEnd, firstKeepIdx);
   const toKeep = messages.slice(firstKeepIdx);
   if (toSummarize.length === 0) {
     return { summary: "", newMessages: messages, replacedCount: 0 };
@@ -1758,7 +1862,7 @@ ${summary.trim()}`
   };
   return {
     summary: summary.trim(),
-    newMessages: [systemMsg, summaryMsg, ...toKeep],
+    newMessages: [...prefix, summaryMsg, ...toKeep],
     replacedCount: toSummarize.length
   };
 }
@@ -3970,6 +4074,17 @@ function capEvents(prev) {
   if (prev.length <= MAX_EVENTS) return prev;
   return prev.slice(prev.length - MAX_EVENTS);
 }
+function makePrefixMessages(cacheStable, model, mode, tools) {
+  if (cacheStable) {
+    return buildSystemMessages({ cwd: process.cwd(), tools, model, mode });
+  }
+  return [
+    {
+      role: "system",
+      content: buildSystemPrompt({ cwd: process.cwd(), tools, model, mode })
+    }
+  ];
+}
 function findImagePaths(text) {
   const paths = [];
   for (const token of text.split(/\s+/)) {
@@ -4017,17 +4132,10 @@ function App({ initialCfg, initialUpdateResult }) {
   const [verbose, setVerbose] = useState6(false);
   const [hasUpdate, setHasUpdate] = useState6(initialUpdateResult?.hasUpdate ?? false);
   const [latestVersion, setLatestVersion] = useState6(initialUpdateResult?.latestVersion ?? null);
-  const messagesRef = useRef3([
-    {
-      role: "system",
-      content: buildSystemPrompt({
-        cwd: process.cwd(),
-        tools: ALL_TOOLS,
-        model: cfg?.model ?? DEFAULT_MODEL,
-        mode: "edit"
-      })
-    }
-  ]);
+  const cacheStableRef = useRef3(initialCfg?.cacheStablePrompts !== false);
+  const messagesRef = useRef3(
+    makePrefixMessages(cacheStableRef.current, cfg?.model ?? DEFAULT_MODEL, "edit", ALL_TOOLS)
+  );
   const executorRef = useRef3(new ToolExecutor(ALL_TOOLS));
   const activeAsstIdRef = useRef3(null);
   const activeControllerRef = useRef3(null);
@@ -4095,15 +4203,27 @@ function App({ initialCfg, initialUpdateResult }) {
   }, [cfg, initialUpdateResult]);
   useEffect4(() => {
     modeRef.current = mode;
-    messagesRef.current[0] = {
-      role: "system",
-      content: buildSystemPrompt({
-        cwd: process.cwd(),
-        tools: [...ALL_TOOLS, ...mcpToolsRef.current],
-        model: cfg?.model ?? DEFAULT_MODEL,
-        mode
-      })
-    };
+    if (cacheStableRef.current) {
+      messagesRef.current[1] = {
+        role: "system",
+        content: buildSessionPrefix({
+          cwd: process.cwd(),
+          tools: [...ALL_TOOLS, ...mcpToolsRef.current],
+          model: cfg?.model ?? DEFAULT_MODEL,
+          mode
+        })
+      };
+    } else {
+      messagesRef.current[0] = {
+        role: "system",
+        content: buildSystemPrompt({
+          cwd: process.cwd(),
+          tools: [...ALL_TOOLS, ...mcpToolsRef.current],
+          model: cfg?.model ?? DEFAULT_MODEL,
+          mode
+        })
+      };
+    }
     if (mode === "plan") {
       executorRef.current.clearSessionPermissions();
     }
@@ -4176,15 +4296,27 @@ function App({ initialCfg, initialUpdateResult }) {
       }
     }
     if (totalTools > 0) {
-      messagesRef.current[0] = {
-        role: "system",
-        content: buildSystemPrompt({
-          cwd: process.cwd(),
-          tools: [...ALL_TOOLS, ...mcpToolsRef.current],
-          model: cfg.model ?? DEFAULT_MODEL,
-          mode: modeRef.current
-        })
-      };
+      if (cacheStableRef.current) {
+        messagesRef.current[1] = {
+          role: "system",
+          content: buildSessionPrefix({
+            cwd: process.cwd(),
+            tools: [...ALL_TOOLS, ...mcpToolsRef.current],
+            model: cfg.model ?? DEFAULT_MODEL,
+            mode: modeRef.current
+          })
+        };
+      } else {
+        messagesRef.current[0] = {
+          role: "system",
+          content: buildSystemPrompt({
+            cwd: process.cwd(),
+            tools: [...ALL_TOOLS, ...mcpToolsRef.current],
+            model: cfg.model ?? DEFAULT_MODEL,
+            mode: modeRef.current
+          })
+        };
+      }
       setEvents((e) => [
         ...e,
         { kind: "info", key: mkKey(), text: `MCP connected \u2014 ${totalTools} external tool${totalTools === 1 ? "" : "s"} available` }
@@ -4457,15 +4589,27 @@ function App({ initialCfg, initialUpdateResult }) {
         }
       });
       if (existsSync(join8(cwd, "KIMI.md"))) {
-        messagesRef.current[0] = {
-          role: "system",
-          content: buildSystemPrompt({
-            cwd,
-            tools: [...ALL_TOOLS, ...mcpToolsRef.current],
-            model: cfg.model,
-            mode: modeRef.current
-          })
-        };
+        if (cacheStableRef.current) {
+          messagesRef.current[1] = {
+            role: "system",
+            content: buildSessionPrefix({
+              cwd,
+              tools: [...ALL_TOOLS, ...mcpToolsRef.current],
+              model: cfg.model,
+              mode: modeRef.current
+            })
+          };
+        } else {
+          messagesRef.current[0] = {
+            role: "system",
+            content: buildSystemPrompt({
+              cwd,
+              tools: [...ALL_TOOLS, ...mcpToolsRef.current],
+              model: cfg.model,
+              mode: modeRef.current
+            })
+          };
+        }
         setEvents((e) => [
           ...e,
           { kind: "info", key: mkKey(), text: "KIMI.md generated; context loaded for future turns" }
@@ -4549,7 +4693,11 @@ function App({ initialCfg, initialUpdateResult }) {
         return true;
       }
       if (c === "/clear") {
-        messagesRef.current = [messagesRef.current[0]];
+        if (cacheStableRef.current && messagesRef.current.length >= 2) {
+          messagesRef.current = [messagesRef.current[0], messagesRef.current[1]];
+        } else {
+          messagesRef.current = [messagesRef.current[0]];
+        }
         sessionIdRef.current = null;
         setEvents([]);
         setUsage(null);