kimiflare 0.29.1 → 0.30.0

package/dist/index.js

@@ -1359,6 +1359,13 @@ Use console.log() to return results. Only console.log output will be sent back t
   const recentToolCalls = opts2.recentToolCalls ?? [];
   const LOOP_WINDOW = 8;
   const LOOP_THRESHOLD = 2;
+  const webFetchHistory = [];
+  const MAX_WEB_FETCH_PER_TURN = 5;
+  const WEB_FETCH_DOMAIN_THRESHOLD = 2;
+  let totalToolCallsThisTurn = 0;
+  const BUDGET_CHECK_INTERVAL = 3;
+  const SOFT_BUDGET = 5;
+  const HARD_BUDGET = 15;
   for (let iter = 0; iter < max; iter++) {
     turn++;
     const previousMessages = opts2.messages.slice();
@@ -1513,6 +1520,57 @@ Use console.log() to return results. Only console.log output will be sent back t
         if (recentToolCalls.length > LOOP_WINDOW) recentToolCalls.shift();
         continue;
       }
+      if (tc.function.name === "web_fetch") {
+        const args = JSON.parse(tc.function.arguments || "{}");
+        const url = args.url || "";
+        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 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,
+              name: "web_fetch",
+              content: warning,
+              ok: false
+            };
+            toolResults.push(budgetResult);
+            opts2.messages.push({
+              role: "tool",
+              tool_call_id: tc.id,
+              content: sanitizeString(warning),
+              name: "web_fetch"
+            });
+            opts2.callbacks.onToolResult?.(budgetResult);
+            recentToolCalls.push(loopSignature);
+            if (recentToolCalls.length > LOOP_WINDOW) recentToolCalls.shift();
+            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 = {
+              tool_call_id: tc.id,
+              name: "web_fetch",
+              content: warning,
+              ok: false
+            };
+            toolResults.push(loopResult);
+            opts2.messages.push({
+              role: "tool",
+              tool_call_id: tc.id,
+              content: sanitizeString(warning),
+              name: "web_fetch"
+            });
+            opts2.callbacks.onToolResult?.(loopResult);
+            recentToolCalls.push(loopSignature);
+            if (recentToolCalls.length > LOOP_WINDOW) recentToolCalls.shift();
+            continue;
+          }
+          webFetchHistory.push({ url, domain });
+        } catch {
+        }
+      }
       if (codeMode && tc.function.name === "execute_code") {
         const args = JSON.parse(tc.function.arguments || "{}");
         const code = args.code || "";
@@ -1572,7 +1630,19 @@ ${sandboxResult.output}` : sandboxResult.output;
         opts2.callbacks.onToolResult?.(result);
         recentToolCalls.push(loopSignature);
         if (recentToolCalls.length > LOOP_WINDOW) recentToolCalls.shift();
+        totalToolCallsThisTurn++;
+      }
+    }
+    if (totalToolCallsThisTurn > 0 && totalToolCallsThisTurn % BUDGET_CHECK_INTERVAL === 0) {
+      let budgetMsg = "";
+      if (totalToolCallsThisTurn >= HARD_BUDGET) {
+        budgetMsg = `BUDGET CHECK: You have made ${totalToolCallsThisTurn} tool calls. This is the substantial-question budget. Produce your deliverable now unless you can name a specific gap that would change the recommendation.`;
+      } else if (totalToolCallsThisTurn >= SOFT_BUDGET) {
+        budgetMsg = `BUDGET CHECK: You have made ${totalToolCallsThisTurn} tool calls. If this is a routine question, produce your deliverable now. If substantial, justify the next call in one sentence.`;
+      } else {
+        budgetMsg = `BUDGET CHECK: You have made ${totalToolCallsThisTurn} tool calls. Assess: is the next call worth more than what you already have? If yes, justify in one sentence. If no, produce your deliverable.`;
       }
+      opts2.messages.push({ role: "system", content: budgetMsg });
     }
     if (opts2.sessionId && lastUsage) {
       void logTurnDebug({
@@ -1587,7 +1657,9 @@ ${sandboxResult.output}` : sandboxResult.output;
       });
     }
   }
-  throw new Error(`kimiflare: tool iteration limit reached (${opts2.maxToolIterations ?? 50})`);
+  const pauseMsg = `Paused after ${opts2.maxToolIterations ?? 50} tool calls. The user may say "go on" to continue. If you have a partial deliverable (Research Brief, Implementation Notes, etc.), include it in your next response.`;
+  opts2.messages.push({ role: "system", content: pauseMsg });
+  throw new Error(`kimiflare: tool iteration limit reached (${opts2.maxToolIterations ?? 50}). Say "go on" to continue, or ask me to focus on a specific area.`);
 }
 function validateToolArguments(raw) {
   if (!raw || !raw.trim()) return "{}";
@@ -1847,8 +1919,82 @@ function loadContextFile(cwd) {
   }
   return null;
 }
+function buildRolePrefix(role) {
+  switch (role) {
+    case "research":
+      return `You are the Research Agent in kimiflare. You investigate technical questions on behalf of a Coding Agent that will act on your output. You are not talking to a human. The Coding Agent is your reader.
+
+# Your job
+
+Produce the smallest research artifact that lets the Coding Agent act correctly and confidently on the task it has been given. Not the most thorough \u2014 the smallest sufficient one. Research exists to enable action. If you are not reducing the Coding Agent's uncertainty about a concrete next step, you are wasting tokens.
+
+# How to think
+
+1. Start by naming the decision. Before any tool call, write down \u2014 for yourself \u2014 what decision your research is meant to enable. "Pick a library." "Choose between approach A and B." "Determine if X is possible." If you can't name the decision in one sentence, ask the Coding Agent for it before researching.
+
+2. Surface area before depth. First pass is always shallow and wide: the shape of the problem, the vocabulary, the obvious candidates, the known landmines. Only then go deep, and only on what the decision actually hinges on.
+
+3. Hold hypotheses loosely and visibly. Form a working hypothesis early \u2014 it directs attention \u2014 but mark it as a hypothesis and look actively for evidence against it. Sycophantic research is useless research.
+
+4. Budget is real. You have a finite tool-call budget per task. Default to ~5 calls for routine questions, up to ~15 for substantial ones. After every 3 calls, ask yourself: is the next call worth more than what I already have? Usually it isn't. Stop earlier than feels comfortable.
+
+5. Separate finding from inference from recommendation. Sources said X. I infer Y. Therefore Z for our case. Keep these layers visible so the Coding Agent can audit any of them.
+
+6. Know when to recommend running the code instead. Sometimes the cheapest research is letting the runtime answer. Say so when true.
+
+# When to stop
+
+Stop when all of these are true:
+- The named decision can be made from what you have.
+- Remaining uncertainties are named, not hidden.
+- The next tool call would predictably add little.
+
+Do not stop just because you found something. Do not stop just because you ran out of patience. Stop on the criteria above, and only those.
+
+# Output format
+
+You are writing for an agent, not a person. No preamble, no narrative, no "in this report we will." Structure:
+
+- DECISION: one sentence \u2014 what this research enables.
+- FINDINGS: scannable facts, with source attribution. Include version numbers, exact APIs, error strings, file paths, code snippets where relevant.
+- RECOMMENDATION: what the Coding Agent should do, concretely.
+- CONFIDENCE: per claim where it varies. "High / Medium / Low" is fine.
+- OPEN QUESTIONS: things you couldn't resolve. Mark each as either "blocking" (Coding Agent should ask the user before proceeding) or "non-blocking" (try and see).
+- RISKS: what could go wrong if the Recommendation is followed, including the strongest counter-argument you found.
+
+# Voice
+
+Terse. Direct. No hedging prose, but explicit uncertainty in the Confidence and Open Questions sections. No apologies, no throat-clearing, no "I hope this helps."
+
+# Addressing the user
+
+You do not address the user. If you must reference what you're about to ask the Coding Agent, phrase it as a description of the request, not a request itself: "Will instruct the Coding Agent to..." \u2014 never "please do X." The user is overhearing, not participating.
+
+# Things that are not research
+
+- Restating the task back at length.
+- Listing every option without ranking them.
+- Producing an essay when a table would do.
+- Continuing to search after the decision can already be made.
+- Hiding uncertainty inside confident prose.
+
+When in doubt, deliver the smaller artifact sooner. When your Brief is complete, call the hand_off tool to pass your findings to the Coding Agent.
+
+`;
+    case "coding":
+      return `You are the Coding Agent of kimiflare. You write code, edit files, and execute tools directly. Do not ask the user to do your work for you. Implement the changes yourself.
+
+`;
+    case "generalist":
+      return `You are the Generalist Agent of kimiflare. You handle conversational queries, memory management, and high-level task coordination.
+
+`;
+    default:
+      return "";
+  }
+}
 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.
+  return buildRolePrefix(opts2.role) + `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.
@@ -2637,6 +2783,47 @@ var init_memory = __esm({
   }
 });
 
+// src/tools/hand-off.ts
+var handOffTool;
+var init_hand_off = __esm({
+  "src/tools/hand-off.ts"() {
+    "use strict";
+    handOffTool = {
+      name: "hand_off",
+      needsPermission: false,
+      description: `Signal that your work is complete and request a hand-off to another agent. Use this when you have produced your deliverable (Research Brief, Implementation Notes, etc.) and the next agent should take over.
+
+Parameters:
+- target: the agent to hand off to (e.g., "coding", "generalist")
+- reason: optional one-sentence summary of what was completed`,
+      parameters: {
+        type: "object",
+        properties: {
+          target: {
+            type: "string",
+            description: 'Agent to hand off to: "coding", "generalist", or another agent role'
+          },
+          reason: {
+            type: "string",
+            description: "One-sentence summary of what was completed and why hand-off is appropriate"
+          }
+        },
+        required: ["target"],
+        additionalProperties: false
+      },
+      async run(args, _ctx) {
+        const target = args.target;
+        const reason = args.reason ?? "";
+        return {
+          content: `Hand-off requested to ${target} agent.${reason ? ` Reason: ${reason}` : ""}`,
+          rawBytes: 0,
+          reducedBytes: 0
+        };
+      }
+    };
+  }
+});
+
 // src/tools/artifact-store.ts
 var ToolArtifactStore;
 var init_artifact_store = __esm({
@@ -3125,6 +3312,7 @@ var init_executor = __esm({
     init_web_fetch();
     init_tasks();
     init_memory();
+    init_hand_off();
     init_artifact_store();
     init_reducer();
     init_expand_artifact();
@@ -3139,7 +3327,8 @@ var init_executor = __esm({
       tasksSetTool,
       memoryRememberTool,
       memoryRecallTool,
-      memoryForgetTool
+      memoryForgetTool,
+      handOffTool
     ];
     ToolExecutor = class {
       sessionAllowed = /* @__PURE__ */ new Set();
@@ -5654,6 +5843,7 @@ var init_agent_session = __esm({
       "lsp_rename",
       "lsp_typeDefinition",
       "lsp_workspaceSymbol",
+      "hand_off",
       "memory_recall",
       "read",
       "tasks_set",
@@ -5885,6 +6075,26 @@ var init_orchestrator = __esm({
       getAutoSwitch() {
         return this.autoSwitch;
       }
+      /** Scan the session's last assistant message for a hand_off tool call.
+       *  Returns the target role if found, null otherwise. */
+      detectHandOff(messages) {
+        for (let i = messages.length - 1; i >= 0; i--) {
+          const m = messages[i];
+          if (m.role === "assistant" && m.tool_calls) {
+            for (const tc of m.tool_calls) {
+              if (tc.function.name === "hand_off") {
+                try {
+                  const args = JSON.parse(tc.function.arguments);
+                  if (args.target) return args.target;
+                } catch {
+                }
+              }
+            }
+            break;
+          }
+        }
+        return null;
+      }
       getToolsForRole(role) {
         const base = getAgentTools(role, this.opts.customAgents);
         if (this.opts.lspTools.length > 0) {
@@ -6040,6 +6250,20 @@ ${summary}`
           agentRole: this.activeRole
         });
         session.recentToolCalls = session.recentToolCalls.slice(-8);
+        const handOffTarget = this.detectHandOff(session.messages);
+        if (handOffTarget && handOffTarget !== this.activeRole) {
+          const summary = await this.synthesizeHandoff(this.activeRole, handOffTarget);
+          this.activeRole = handOffTarget;
+          const newSession = this.getActiveSession();
+          if (summary) {
+            newSession.messages.push({
+              role: "system",
+              content: `[hand-off from ${this.activeRole} agent \u2014 agent requested hand-off]
+${summary}`
+            });
+          }
+          this.turnCounts.set(handOffTarget, 0);
+        }
         this.turnCounts.set(this.activeRole, (this.turnCounts.get(this.activeRole) ?? 0) + 1);
       }
       async handOff(toRole) {
@@ -7499,6 +7723,11 @@ var init_chat = __esm({
       }
       if (evt.kind === "assistant") {
         return /* @__PURE__ */ jsxs4(Box4, { flexDirection: "column", paddingLeft: 2, children: [
+          evt.agentRole && /* @__PURE__ */ jsx4(Box4, { marginBottom: 1, children: /* @__PURE__ */ jsxs4(Text4, { color: theme.info.color, dimColor: theme.info.dim, children: [
+            "\u25C6 ",
+            evt.agentRole,
+            " agent"
+          ] }) }),
           showReasoning && evt.reasoning ? /* @__PURE__ */ jsx4(Box4, { flexDirection: "column", marginBottom: 1, children: /* @__PURE__ */ jsxs4(Text4, { color: theme.reasoning.color, dimColor: theme.reasoning.dim, children: [
             "thinking\u2026",
             " ",
@@ -9132,7 +9361,7 @@ function themeNames() {
 function themeList() {
   return Object.values(THEMES);
 }
-var dark, light, highContrast, dracula, nord, monokai, solarizedDark, solarizedLight, tokyoNight, gruvboxDark, catppuccinMocha, rosePine, THEMES, DEFAULT_THEME_NAME;
+var dark, light, highContrast, dracula, nord, monokai, solarizedDark, solarizedLight, tokyoNight, gruvboxDark, catppuccinMocha, rosePine, oneDark, ayu, nightOwl, palenight, THEMES, DEFAULT_THEME_NAME;
 var init_theme = __esm({
   "src/ui/theme.ts"() {
     "use strict";
@@ -9328,6 +9557,70 @@ var init_theme = __esm({
       accent: "#ebbcba",
       modeBadge: { plan: "#31748f", auto: "#9ccfd8", edit: "#ebbcba" }
     };
+    oneDark = {
+      name: "one-dark",
+      label: "one-dark (Atom's iconic dark \u2014 blue & purple)",
+      user: "#61afef",
+      assistant: void 0,
+      reasoning: { color: "#5c6370", dim: true },
+      info: { color: "#5c6370", dim: true },
+      error: "#e06c75",
+      warn: "#e5c07b",
+      tool: "#c678dd",
+      spinner: "#61afef",
+      permission: "#e5c07b",
+      queue: { color: "#5c6370", dim: true },
+      accent: "#c678dd",
+      modeBadge: { plan: "#61afef", auto: "#98c379", edit: "#c678dd" }
+    };
+    ayu = {
+      name: "ayu",
+      label: "ayu (clean modern \u2014 orange & cyan)",
+      user: "#39bae6",
+      assistant: void 0,
+      reasoning: { color: "#4d5566", dim: true },
+      info: { color: "#4d5566", dim: true },
+      error: "#f07178",
+      warn: "#ffb454",
+      tool: "#73b8ff",
+      spinner: "#39bae6",
+      permission: "#ffb454",
+      queue: { color: "#4d5566", dim: true },
+      accent: "#39bae6",
+      modeBadge: { plan: "#39bae6", auto: "#7ee787", edit: "#ffb454" }
+    };
+    nightOwl = {
+      name: "night-owl",
+      label: "night-owl (deep navy \u2014 cyan & red)",
+      user: "#82aaff",
+      assistant: void 0,
+      reasoning: { color: "#4d6885", dim: true },
+      info: { color: "#4d6885", dim: true },
+      error: "#ef5350",
+      warn: "#ffca28",
+      tool: "#c792ea",
+      spinner: "#82aaff",
+      permission: "#ffca28",
+      queue: { color: "#4d6885", dim: true },
+      accent: "#c792ea",
+      modeBadge: { plan: "#82aaff", auto: "#7ee787", edit: "#c792ea" }
+    };
+    palenight = {
+      name: "palenight",
+      label: "palenight (Material pale \u2014 purple & cyan)",
+      user: "#82b1ff",
+      assistant: void 0,
+      reasoning: { color: "#4c566a", dim: true },
+      info: { color: "#4c566a", dim: true },
+      error: "#f07178",
+      warn: "#ffcb6b",
+      tool: "#c792ea",
+      spinner: "#82b1ff",
+      permission: "#ffcb6b",
+      queue: { color: "#4c566a", dim: true },
+      accent: "#c792ea",
+      modeBadge: { plan: "#82b1ff", auto: "#c3e88d", edit: "#c792ea" }
+    };
     THEMES = {
       dark,
       light,
@@ -9340,7 +9633,11 @@ var init_theme = __esm({
       "tokyo-night": tokyoNight,
       "gruvbox-dark": gruvboxDark,
       "catppuccin-mocha": catppuccinMocha,
-      "rose-pine": rosePine
+      "rose-pine": rosePine,
+      "one-dark": oneDark,
+      ayu,
+      "night-owl": nightOwl,
+      palenight
     };
     DEFAULT_THEME_NAME = "dark";
   }
@@ -11529,14 +11826,14 @@ function compactEventsVisual(prev, keepLastTurns) {
     ...kept
   ];
 }
-function makePrefixMessages(cacheStable, model, mode, tools) {
+function makePrefixMessages(cacheStable, model, mode, tools, role) {
   if (cacheStable) {
-    return buildSystemMessages({ cwd: process.cwd(), tools, model, mode });
+    return buildSystemMessages({ cwd: process.cwd(), tools, model, mode, role });
   }
   return [
     {
       role: "system",
-      content: buildSystemPrompt({ cwd: process.cwd(), tools, model, mode })
+      content: buildSystemPrompt({ cwd: process.cwd(), tools, model, mode, role })
     }
   ];
 }
@@ -12407,9 +12704,10 @@ function App({
           onAssistantStart: () => {
             const id = nextAssistantId++;
             activeAsstIdRef.current = id;
+            const role = orchestratorRef.current?.getActiveRole();
             setEvents((e) => [
               ...e,
-              { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true }
+              { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true, agentRole: role }
             ]);
           },
           onReasoningDelta: (d) => {
@@ -13422,9 +13720,10 @@ ${lines.join("\n")}` }]);
         onAssistantStart: () => {
           const id = nextAssistantId++;
           activeAsstIdRef.current = id;
+          const role = orchestratorRef.current?.getActiveRole();
           setEvents((e) => [
             ...e,
-            { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true }
+            { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true, agentRole: role }
           ]);
         },
         onReasoningDelta: (d) => {
@@ -13570,7 +13869,8 @@ ${lines.join("\n")}` }]);
                 cacheStableRef.current,
                 overrideModel ?? cfg.model,
                 modeRef.current,
-                [...ALL_TOOLS, ...mcpToolsRef.current, ...lspToolsRef.current]
+                [...ALL_TOOLS, ...mcpToolsRef.current, ...lspToolsRef.current],
+                orchestratorRef.current.getActiveRole()
               );
               activeSession2.messages.unshift(...prefix);
             }