kimiflare 0.29.1 → 0.31.0

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 || "";
@@ -1572,8 +1630,20 @@ ${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({
         sessionId: opts2.sessionId,
@@ -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 +1919,199 @@ 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.
+
+# Critical hand-off rule
+
+When your Brief is complete, you MUST call the hand_off tool to transfer control to the next agent. Simply saying you have handed off is NOT sufficient \u2014 the tool call is required. If you do not call hand_off, your work will be stranded and the next agent will never run.
+
+You MUST include the full Brief text in your final assistant message BEFORE calling the hand_off tool. The next agent receives your last assistant message in its entirety \u2014 no summarization, no truncation. If you produce the Brief in one message and then call hand_off in a separate message with only "Handing off now," the next agent will see only "Handing off now" and will not know what to implement.
+
+Correct: One assistant message containing the full Brief + the hand_off tool call.
+Incorrect: Brief in message N, then "Handing off" + hand_off in message N+1.
+Incorrect: Saying "I have handed off" without calling the hand_off tool.
+
+`;
+    case "coding":
+      return `You are the Coding Agent in kimiflare. You write, modify, debug, and reason about code. You receive tasks from the General Agent or research briefs from the Research Agent. Your audience is sometimes the user directly, sometimes another agent.
+
+# Your job
+
+Implement the task as scoped. Correctly, narrowly, and in a way that fits the codebase you're working in. Stop when it's done.
+
+# How to think
+
+1. Read before you write. Look at the existing code \u2014 patterns, utilities, conventions, naming. Match the codebase's style, don't impose your own. The repo should look like one author wrote it even after you've worked in it.
+
+2. Stay in scope. Touch what the task requires and nothing else. If you notice something else worth fixing, mention it \u2014 don't fix it uninvited. Scope creep is the most common way coding agents make things worse.
+
+3. Trust the runtime. When something doesn't work, run it, read the actual error, and update your understanding. Don't argue with reality based on what the docs or types said. The runtime is the source of truth.
+
+4. Be honest about uncertainty before acting, not after. "I'm going to try X \u2014 if it fails I'll try Y" is right. Confident execution followed by silent breakage is wrong.
+
+5. Ask only when ambiguity is load-bearing. If a choice would meaningfully change the result and you can't infer the user's intent, ask. If it's a trivial choice, make it and move on.
+
+6. Done means done. Working, fitting the codebase, tests passing where applicable, loose ends named. Not "the command exited zero." Don't claim done when you only have passing.
+
+# Working style
+
+- Small, verifiable steps over large speculative ones.
+- Run the code. Read the output. Believe the output.
+- Prefer existing utilities over new ones. Prefer the codebase's patterns over your defaults.
+- New dependencies are a real cost. Justify them or skip them.
+- Comments narrate why, not what. If the code needs a comment to explain what it does, the code is probably wrong.
+
+# Voice
+
+Direct. No throat-clearing, no narration of obvious steps, no celebration of completion. When you explain something, explain only what isn't already visible in the code or output.
+
+# Output
+
+Show the work \u2014 the diff, the file, the command output \u2014 and a one- or two-line summary of what you did and anything the next agent or the user should know. That's it. No "I hope this helps." No "let me know if you'd like me to..."
+
+If something didn't work or you couldn't finish cleanly, say so plainly with what you tried and what you'd try next.
+
+# Things that are not your job
+
+- Investigating broad questions (Research Agent's job).
+- Routing or chatting (General Agent's job).
+- Improving the codebase beyond the task at hand.
+- Producing long explanations of code the reader can read.
+
+# Receiving work from the Research Agent
+
+When you are activated after a Research Agent hand-off, the full Research Brief is included in the system message that precedes your turn. Read it carefully \u2014 it contains the decision, findings, recommendation, confidence levels, open questions, and risks. Do not ask the user to repeat what the Research Agent already determined.
+
+When your implementation is complete, you MUST call the hand_off tool to return to the General Agent. Simply saying you are done is NOT sufficient \u2014 the tool call is required. If you do not call hand_off, your work will be stranded and the General Agent will never run.
+
+`;
+    case "generalist":
+      return `You are the General Agent in kimiflare. You are the user's primary point of contact. Behind you are two specialists: the Research Agent (investigation, analysis, synthesis) and the Coding Agent (writing, modifying, and reasoning about code).
+
+# Your job
+
+Triage. Route. Stay out of the way. Handle small stuff. Present specialist work cleanly.
+
+You are fast and light by design. Substantive thinking is not your job \u2014 it's the specialists' job. Your job is to recognize what kind of help the user needs and get them to the right agent quickly, or to handle the request yourself if it's small enough that routing would be overkill.
+
+# How to think
+
+1. Default to routing. If a request involves real investigation, real synthesis, or real code work, call hand_off to the appropriate specialist. Do not try to answer it yourself just because you can produce something plausible-sounding.
+
+2. Route on partial information. You don't need to fully understand the request before routing \u2014 the specialist will ask follow-ups if needed. Spending three turns clarifying before handoff is worse than handing off now and letting the specialist clarify.
+
+3. Handle the small stuff yourself. Greetings, clarifications, "what can you do," confirming what just happened, one-line factual answers, formatting preferences, scope adjustments \u2014 these don't need a specialist. Be quick.
+
+4. Notice escalation. A conversation that started small can become a research or coding task. When it does, route. Don't keep answering out of inertia.
+
+5. Do not editorialize the specialists' output. When work comes back from Research or Coding, present it. Don't summarize it back at the user with your own framing on top. The user can read.
+
+# Routing rules
+
+Call hand_off to Research Agent when the user wants:
+- Information you don't already have, or that may have changed.
+- Comparison, evaluation, or recommendation between options.
+- Synthesis across multiple sources.
+- Investigation of an unfamiliar codebase or library.
+- Anything where being wrong has real cost.
+
+Call hand_off to Coding Agent when the user wants:
+- Code written, modified, debugged, or reviewed.
+- A file created, edited, or restructured.
+- A concrete build/run/test action taken.
+
+Handle yourself when:
+- The user is making conversation.
+- The user is asking what you (collectively) can do.
+- The answer is one line and you're confident.
+- The user is correcting or adjusting a previous handoff.
+- Work has come back from a specialist and you're presenting it to the user.
+
+When in doubt, route. The cost of an unnecessary handoff is small. The cost of you confidently producing wrong work is large.
+
+# Voice
+
+Warm, quick, natural. Short sentences. No corporate softeners, no "I'd be happy to," no "great question." Talk like a competent person who respects the user's time.
+
+# Handoff style
+
+When you route, say so plainly in one line. "Handing this to the research agent \u2014 back in a moment." or "Coding agent will take this one." Then stop. Don't fill the wait with chatter.
+
+# Things that are not your job
+
+- Producing research findings.
+- Writing or analyzing code.
+- Synthesizing across many sources.
+- Long explanations of anything.
+
+If you find yourself drafting a long response, stop and ask whether this should have been routed. Usually it should have been.
+
+`;
+    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 +2900,47 @@ 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 +3429,7 @@ var init_executor = __esm({
     init_web_fetch();
     init_tasks();
     init_memory();
+    init_hand_off();
     init_artifact_store();
     init_reducer();
     init_expand_artifact();
@@ -3139,7 +3444,8 @@ var init_executor = __esm({
       tasksSetTool,
       memoryRememberTool,
       memoryRecallTool,
-      memoryForgetTool
+      memoryForgetTool,
+      handOffTool
     ];
     ToolExecutor = class {
       sessionAllowed = /* @__PURE__ */ new Set();
@@ -5654,6 +5960,7 @@ var init_agent_session = __esm({
       "lsp_rename",
       "lsp_typeDefinition",
       "lsp_workspaceSymbol",
+      "hand_off",
       "memory_recall",
       "read",
       "tasks_set",
@@ -5662,6 +5969,7 @@ var init_agent_session = __esm({
     CODING_TOOL_NAMES = [
       "bash",
       "edit",
+      "hand_off",
       "lsp_codeAction",
       "lsp_definition",
       "lsp_diagnostics",
@@ -5678,6 +5986,7 @@ var init_agent_session = __esm({
       "write"
     ].sort((a, b) => a.localeCompare(b));
     GENERALIST_TOOL_NAMES = [
+      "hand_off",
       "memory_forget",
       "memory_recall",
       "memory_remember",
@@ -5885,6 +6194,38 @@ 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;
+      }
+      /** Extract the last assistant message content from a session.
+       *  Returns the full text in its entirety — no truncation, no summarization.
+       *  If the agent has no text content, falls back to synthesis. */
+      extractDeliverable(session) {
+        for (let i = session.messages.length - 1; i >= 0; i--) {
+          const m = session.messages[i];
+          if (m.role === "assistant" && typeof m.content === "string" && m.content.trim().length > 0) {
+            return m.content;
+          }
+        }
+        return null;
+      }
       getToolsForRole(role) {
         const base = getAgentTools(role, this.opts.customAgents);
         if (this.opts.lspTools.length > 0) {
@@ -6040,6 +6381,119 @@ ${summary}`
           agentRole: this.activeRole
         });
         session.recentToolCalls = session.recentToolCalls.slice(-8);
+        const handOffTarget = this.detectHandOff(session.messages);
+        if (handOffTarget && handOffTarget !== this.activeRole) {
+          const fromRole = this.activeRole;
+          const fromSession = this.sessions.get(fromRole);
+          const deliverable = this.extractDeliverable(fromSession);
+          let handoffContent = "";
+          if (deliverable) {
+            handoffContent = `[hand-off from ${fromRole} agent \u2014 agent requested hand-off]
+
+The ${fromRole} agent completed its work. Here is their full deliverable:
+
+${deliverable}`;
+          } else {
+            const summary = await this.synthesizeHandoff(fromRole, handOffTarget);
+            handoffContent = `[hand-off from ${fromRole} agent \u2014 agent requested hand-off]
+${summary}`;
+          }
+          this.activeRole = handOffTarget;
+          const newSession = this.getActiveSession();
+          if (handoffContent) {
+            newSession.messages.push({
+              role: "system",
+              content: handoffContent
+            });
+          }
+          newSession.messages.push(userMessage);
+          this.turnCounts.set(handOffTarget, 0);
+          const targetTools = this.getToolsForRole(this.activeRole);
+          const targetCustomAgent = this.opts.customAgents?.find((a) => a.name === this.activeRole);
+          const targetConfig = this.resolveAgentConfig(this.activeRole);
+          const targetModel = targetCustomAgent?.model ?? targetConfig.model ?? this.opts.model;
+          const targetReasoning = targetCustomAgent?.reasoningEffort ?? targetConfig.reasoningEffort ?? this.opts.reasoningEffort;
+          if (targetCustomAgent?.systemPrompt && !newSession.messages.some((m) => m.role === "system" && m.content === targetCustomAgent.systemPrompt)) {
+            newSession.messages.unshift({
+              role: "system",
+              content: targetCustomAgent.systemPrompt
+            });
+          }
+          await runAgentTurn({
+            accountId: this.opts.accountId,
+            apiToken: this.opts.apiToken,
+            model: targetModel,
+            gateway: this.opts.gateway,
+            messages: newSession.messages,
+            tools: [...targetTools, ...this.opts.mcpTools],
+            executor: this.opts.executor,
+            cwd: this.opts.cwd,
+            signal: this.opts.signal,
+            reasoningEffort: targetReasoning,
+            coauthor: this.opts.coauthor,
+            sessionId: this.opts.sessionId,
+            memoryManager: this.opts.memoryManager,
+            keepLastImageTurns: this.opts.keepLastImageTurns,
+            codeMode: this.opts.codeMode,
+            onFileChange: this.opts.onFileChange,
+            callbacks: this.opts.callbacks,
+            recentToolCalls: newSession.recentToolCalls,
+            agentRole: this.activeRole
+          });
+          newSession.recentToolCalls = newSession.recentToolCalls.slice(-8);
+        }
+        if (!handOffTarget && this.activeRole !== "generalist") {
+          const deliverable = this.extractDeliverable(session);
+          if (deliverable && deliverable.length >= 300) {
+            const fromRole = this.activeRole;
+            const handoffContent = `[hand-off from ${fromRole} agent \u2014 auto-detected completion (hand_off tool was not called)]
+
+The ${fromRole} agent completed its work. Here is their full deliverable:
+
+${deliverable}`;
+            this.activeRole = "generalist";
+            const newSession = this.getActiveSession();
+            newSession.messages.push({
+              role: "system",
+              content: handoffContent
+            });
+            newSession.messages.push(userMessage);
+            this.turnCounts.set("generalist", 0);
+            const targetTools = this.getToolsForRole(this.activeRole);
+            const targetCustomAgent = this.opts.customAgents?.find((a) => a.name === this.activeRole);
+            const targetConfig = this.resolveAgentConfig(this.activeRole);
+            const targetModel = targetCustomAgent?.model ?? targetConfig.model ?? this.opts.model;
+            const targetReasoning = targetCustomAgent?.reasoningEffort ?? targetConfig.reasoningEffort ?? this.opts.reasoningEffort;
+            if (targetCustomAgent?.systemPrompt && !newSession.messages.some((m) => m.role === "system" && m.content === targetCustomAgent.systemPrompt)) {
+              newSession.messages.unshift({
+                role: "system",
+                content: targetCustomAgent.systemPrompt
+              });
+            }
+            await runAgentTurn({
+              accountId: this.opts.accountId,
+              apiToken: this.opts.apiToken,
+              model: targetModel,
+              gateway: this.opts.gateway,
+              messages: newSession.messages,
+              tools: [...targetTools, ...this.opts.mcpTools],
+              executor: this.opts.executor,
+              cwd: this.opts.cwd,
+              signal: this.opts.signal,
+              reasoningEffort: targetReasoning,
+              coauthor: this.opts.coauthor,
+              sessionId: this.opts.sessionId,
+              memoryManager: this.opts.memoryManager,
+              keepLastImageTurns: this.opts.keepLastImageTurns,
+              codeMode: this.opts.codeMode,
+              onFileChange: this.opts.onFileChange,
+              callbacks: this.opts.callbacks,
+              recentToolCalls: newSession.recentToolCalls,
+              agentRole: this.activeRole
+            });
+            newSession.recentToolCalls = newSession.recentToolCalls.slice(-8);
+          }
+        }
         this.turnCounts.set(this.activeRole, (this.turnCounts.get(this.activeRole) ?? 0) + 1);
       }
       async handOff(toRole) {
@@ -7499,6 +7953,11 @@ 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 +9591,7 @@ 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 +9787,70 @@ 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 +9863,11 @@ 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";
   }
@@ -10107,6 +10634,45 @@ var init_renderer2 = __esm({
   }
 });
 
+// src/commands/builtins.ts
+var BUILTIN_COMMANDS, BUILTIN_COMMAND_NAMES;
+var init_builtins = __esm({
+  "src/commands/builtins.ts"() {
+    "use strict";
+    BUILTIN_COMMANDS = [
+      { name: "help", description: "Show keybindings and command list", source: "builtin" },
+      { name: "model", description: "Show current model", source: "builtin" },
+      { name: "mode", argHint: "edit|plan|auto", description: "Switch agent mode", source: "builtin" },
+      { name: "plan", description: "Switch to plan mode", source: "builtin" },
+      { name: "auto", description: "Switch to auto mode", source: "builtin" },
+      { name: "edit", description: "Switch to edit mode", source: "builtin" },
+      { name: "thinking", argHint: "low|medium|high", description: "Set reasoning effort", source: "builtin" },
+      { name: "effort", argHint: "low|medium|high", description: "Alias for /thinking", source: "builtin" },
+      { name: "reasoning", description: "Toggle reasoning visibility", source: "builtin" },
+      { name: "theme", description: "Open theme picker", source: "builtin" },
+      { name: "agent", argHint: "[on|off|status]", description: "Multi-agent mode controls", source: "builtin" },
+      { name: "memory", argHint: "[on|off|clear|search ...]", description: "Manage memory", source: "builtin" },
+      { name: "cost", argHint: "[on|off]", description: "Show cost report or toggle attribution", source: "builtin" },
+      { name: "gateway", argHint: "[status|off|<id>|cache-ttl|skip-cache|...]", description: "Manage AI Gateway", source: "builtin" },
+      { name: "mcp", argHint: "[list|reload]", description: "Manage MCP servers", source: "builtin" },
+      { name: "lsp", argHint: "[config|list|reload|scope]", description: "Manage language servers", source: "builtin" },
+      { name: "command", argHint: "[create|edit|delete|list]", description: "Manage custom slash commands", source: "builtin" },
+      { name: "resume", description: "Pick a past conversation to resume", source: "builtin" },
+      { name: "compact", description: "Summarize old turns to free context", source: "builtin" },
+      { name: "clear", description: "Clear current conversation", source: "builtin" },
+      { name: "init", description: "Scan repo and write KIMI.md", source: "builtin" },
+      { name: "update", description: "Check for updates", source: "builtin" },
+      { name: "hello", description: "Send a voice note to the creator", source: "builtin" },
+      { name: "logout", description: "Clear stored credentials", source: "builtin" },
+      { name: "exit", description: "Exit kimiflare", source: "builtin" },
+      { name: "quit", description: "Alias for /exit", source: "builtin" }
+    ];
+    BUILTIN_COMMAND_NAMES = new Set(
+      BUILTIN_COMMANDS.map((c) => c.name.toLowerCase())
+    );
+  }
+});
+
 // src/commands/save.ts
 import { mkdir as mkdir9, writeFile as writeFile9, unlink as unlink2 } from "fs/promises";
 import { dirname as dirname6 } from "path";
@@ -11268,6 +11834,125 @@ var init_file_picker = __esm({
   }
 });
 
+// src/ui/slash-picker.tsx
+import { Box as Box18, Text as Text19 } from "ink";
+import { jsx as jsx19, jsxs as jsxs18 } from "react/jsx-runtime";
+function sourceBadge(source) {
+  if (source === "builtin") return "";
+  if (source === "project") return "project";
+  return "global";
+}
+function commandLabel(item) {
+  return `/${item.name}${item.argHint ? ` ${item.argHint}` : ""}`;
+}
+function SlashPicker({ items, selectedIndex, theme, query }) {
+  let startIndex = 0;
+  if (selectedIndex >= VISIBLE_LIMIT2) {
+    startIndex = selectedIndex - VISIBLE_LIMIT2 + 1;
+  }
+  const visible = items.slice(startIndex, startIndex + VISIBLE_LIMIT2);
+  const hasMoreAbove = startIndex > 0;
+  const hasMoreBelow = items.length > startIndex + VISIBLE_LIMIT2;
+  const longestLabel = visible.reduce((m, it) => Math.max(m, commandLabel(it).length), 0);
+  const nameColWidth = Math.max(NAME_COL_MIN_WIDTH, longestLabel + NAME_DESC_GAP);
+  return /* @__PURE__ */ jsxs18(Box18, { flexDirection: "column", borderStyle: "round", borderColor: theme.accent, paddingX: 1, children: [
+    /* @__PURE__ */ jsx19(Text19, { color: theme.accent, bold: true, children: query ? `Commands matching "/${query}"` : "Slash commands" }),
+    /* @__PURE__ */ jsx19(Text19, { color: theme.info.color, children: "Arrow keys to navigate, Enter to select, Esc to cancel." }),
+    /* @__PURE__ */ jsxs18(Box18, { marginTop: 1, flexDirection: "column", children: [
+      visible.length === 0 && /* @__PURE__ */ jsx19(Text19, { color: theme.info.color, dimColor: true, children: "No matches" }),
+      hasMoreAbove && /* @__PURE__ */ jsxs18(Text19, { color: theme.info.color, dimColor: true, children: [
+        "\u2026 ",
+        startIndex,
+        " more above"
+      ] }),
+      visible.map((item, i) => {
+        const actualIndex = startIndex + i;
+        const isSelected = actualIndex === selectedIndex;
+        const nameCol = commandLabel(item).padEnd(nameColWidth);
+        const badge = sourceBadge(item.source);
+        return /* @__PURE__ */ jsxs18(Text19, { color: isSelected ? theme.accent : void 0, bold: isSelected, children: [
+          isSelected ? "\u203A " : "  ",
+          nameCol,
+          /* @__PURE__ */ jsxs18(Text19, { color: theme.info.color, dimColor: true, children: [
+            item.description,
+            badge && `  [${badge}]`
+          ] })
+        ] }, item.name);
+      }),
+      hasMoreBelow && /* @__PURE__ */ jsxs18(Text19, { color: theme.info.color, dimColor: true, children: [
+        "\u2026 ",
+        items.length - (startIndex + VISIBLE_LIMIT2),
+        " more below"
+      ] })
+    ] })
+  ] });
+}
+var VISIBLE_LIMIT2, NAME_COL_MIN_WIDTH, NAME_DESC_GAP;
+var init_slash_picker = __esm({
+  "src/ui/slash-picker.tsx"() {
+    "use strict";
+    VISIBLE_LIMIT2 = 7;
+    NAME_COL_MIN_WIDTH = 18;
+    NAME_DESC_GAP = 2;
+  }
+});
+
+// src/util/fuzzy.ts
+function fuzzyMatch(query, text) {
+  const q = query.toLowerCase();
+  const t = text.toLowerCase();
+  if (q.length === 0) return { matches: true, score: 0 };
+  if (q.length > t.length) return { matches: false, score: 0 };
+  let qi = 0;
+  let score = 0;
+  let lastMatch = -1;
+  let consecutive = 0;
+  for (let i = 0; i < t.length && qi < q.length; i++) {
+    if (t[i] !== q[qi]) continue;
+    const isWordBoundary = i === 0 || /[\s\-_./:]/.test(t[i - 1]);
+    if (lastMatch === i - 1) {
+      consecutive++;
+      score -= consecutive * 5;
+    } else {
+      consecutive = 0;
+      if (lastMatch >= 0) score += (i - lastMatch - 1) * 2;
+    }
+    if (isWordBoundary) score -= 10;
+    score += i * 0.1;
+    lastMatch = i;
+    qi++;
+  }
+  if (qi < q.length) return { matches: false, score: 0 };
+  return { matches: true, score };
+}
+function fuzzyFilter(items, query, getText) {
+  const trimmed = query.trim();
+  if (trimmed.length === 0) return items;
+  const tokens = trimmed.split(/\s+/);
+  const scored = [];
+  for (const item of items) {
+    const text = getText(item);
+    let total = 0;
+    let allMatch = true;
+    for (const token of tokens) {
+      const m = fuzzyMatch(token, text);
+      if (!m.matches) {
+        allMatch = false;
+        break;
+      }
+      total += m.score;
+    }
+    if (allMatch) scored.push({ item, score: total });
+  }
+  scored.sort((a, b) => a.score - b.score);
+  return scored.map((s) => s.item);
+}
+var init_fuzzy = __esm({
+  "src/util/fuzzy.ts"() {
+    "use strict";
+  }
+});
+
 // src/cost-attribution/tui-report.ts
 var tui_report_exports = {};
 __export(tui_report_exports, {
@@ -11360,11 +12045,13 @@ var app_exports = {};
 __export(app_exports, {
   buildFilePickerIgnoreList: () => buildFilePickerIgnoreList,
   filterPickerItems: () => filterPickerItems,
+  insertSlashCommand: () => insertSlashCommand,
   renderApp: () => renderApp,
-  shouldOpenMentionPicker: () => shouldOpenMentionPicker
+  shouldOpenMentionPicker: () => shouldOpenMentionPicker,
+  shouldOpenSlashPicker: () => shouldOpenSlashPicker
 });
 import React12, { useState as useState9, useRef as useRef3, useEffect as useEffect4, useCallback } from "react";
-import { Box as Box18, Text as Text19, useApp, useInput as useInput5, render } from "ink";
+import { Box as Box19, Text as Text20, useApp, useInput as useInput5, render } from "ink";
 import SelectInput8 from "ink-select-input";
 import { existsSync as existsSync2, statSync as statSync3 } from "fs";
 import { join as join17 } from "path";
@@ -11373,7 +12060,7 @@ import { spawn as spawn4 } from "child_process";
 import { platform as platform2 } from "os";
 import fg4 from "fast-glob";
 import { readFileSync as readFileSync3 } from "fs";
-import { jsx as jsx19, jsxs as jsxs18 } from "react/jsx-runtime";
+import { jsx as jsx20, jsxs as jsxs19 } from "react/jsx-runtime";
 function buildFilePickerIgnoreList(cwd) {
   const hardcoded = [
     // Dependencies
@@ -11482,6 +12169,18 @@ function shouldOpenMentionPicker(input, cursorOffset, pickerCancelOffset) {
   }
   return false;
 }
+function shouldOpenSlashPicker(input, cursorOffset, cancelOffset) {
+  if (cancelOffset === cursorOffset) return false;
+  if (cursorOffset === 0 || input[cursorOffset - 1] !== "/") return false;
+  return /^\s*$/.test(input.slice(0, cursorOffset - 1));
+}
+function insertSlashCommand(input, anchor, name) {
+  let tokenEnd = anchor + 1;
+  while (tokenEnd < input.length && !/\s/.test(input[tokenEnd])) tokenEnd++;
+  const head = input.slice(0, anchor + 1) + name;
+  const tail = " " + input.slice(tokenEnd).replace(/^\s+/, "");
+  return { value: head + tail, cursor: head.length + 1 };
+}
 function gatewayFromConfig(cfg) {
   if (!cfg.aiGatewayId) return void 0;
   return {
@@ -11529,14 +12228,14 @@ 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 })
     }
   ];
 }
@@ -11616,9 +12315,10 @@ function App({
   const [hasUpdate, setHasUpdate] = useState9(initialUpdateResult?.hasUpdate ?? false);
   const [latestVersion, setLatestVersion] = useState9(initialUpdateResult?.latestVersion ?? null);
   const [cursorOffset, setCursorOffset] = useState9(0);
-  const [pickerItems, setPickerItems] = useState9([]);
-  const [pickerSelected, setPickerSelected] = useState9(0);
-  const [pickerAnchor, setPickerAnchor] = useState9(null);
+  const [activePicker, setActivePicker] = useState9(null);
+  const [filePickerItems, setFilePickerItems] = useState9([]);
+  const filePickerLoadedRef = useRef3(false);
+  const [customCommandsVersion, setCustomCommandsVersion] = useState9(0);
   const cacheStableRef = useRef3(initialCfg?.cacheStablePrompts !== false);
   const messagesRef = useRef3(
     makePrefixMessages(cacheStableRef.current, cfg?.model ?? DEFAULT_MODEL, "edit", ALL_TOOLS)
@@ -11654,28 +12354,42 @@ function App({
   const flushTimeoutRef = useRef3(null);
   const customCommandsRef = useRef3([]);
   const pickerCancelRef = useRef3(null);
-  const mentionState = React12.useMemo(() => {
+  const pickerAnchor = activePicker?.anchor ?? null;
+  const pickerKind = activePicker?.kind ?? null;
+  const pickerQuery = React12.useMemo(() => {
     if (pickerAnchor === null) return null;
-    const query = input.slice(pickerAnchor + 1, cursorOffset);
-    return { query, anchor: pickerAnchor };
+    return input.slice(pickerAnchor + 1, cursorOffset);
   }, [input, cursorOffset, pickerAnchor]);
-  const filteredPickerItems = React12.useMemo(() => {
-    if (!mentionState) return [];
-    return filterPickerItems(pickerItems, mentionState.query);
-  }, [pickerItems, mentionState]);
+  const filteredFileItems = React12.useMemo(() => {
+    if (pickerKind !== "file" || pickerQuery === null) return [];
+    return filterPickerItems(filePickerItems, pickerQuery);
+  }, [pickerKind, filePickerItems, pickerQuery]);
+  const allSlashCommands = React12.useMemo(() => {
+    const customs = customCommandsRef.current.filter((c) => !BUILTIN_COMMAND_NAMES.has(c.name.toLowerCase())).map((c) => ({
+      name: c.name,
+      description: c.description ?? "",
+      source: c.source
+    }));
+    return [...BUILTIN_COMMANDS, ...customs];
+  }, [customCommandsVersion]);
+  const filteredSlashItems = React12.useMemo(() => {
+    if (pickerKind !== "slash" || pickerQuery === null) return [];
+    return fuzzyFilter(allSlashCommands, pickerQuery, (c) => c.name).slice(0, 50);
+  }, [pickerKind, allSlashCommands, pickerQuery]);
   useEffect4(() => {
-    if (pickerAnchor !== null) {
-      if (cursorOffset < pickerAnchor) {
-        setPickerAnchor(null);
+    if (activePicker !== null) {
+      const trigger = activePicker.kind === "file" ? "@" : "/";
+      if (cursorOffset < activePicker.anchor) {
+        setActivePicker(null);
         return;
       }
-      if (input[pickerAnchor] !== "@") {
-        setPickerAnchor(null);
+      if (input[activePicker.anchor] !== trigger) {
+        setActivePicker(null);
         return;
       }
-      const query = input.slice(pickerAnchor + 1, cursorOffset);
-      if (query.includes(" ")) {
-        setPickerAnchor(null);
+      const query = input.slice(activePicker.anchor + 1, cursorOffset);
+      if (/\s/.test(query)) {
+        setActivePicker(null);
         return;
       }
       return;
@@ -11685,9 +12399,9 @@ function App({
       return;
     }
     if (filePickerEnabled && shouldOpenMentionPicker(input, cursorOffset, pickerCancelRef.current)) {
-      setPickerAnchor(cursorOffset - 1);
-      setPickerSelected(0);
-      if (pickerItems.length === 0) {
+      setActivePicker({ kind: "file", anchor: cursorOffset - 1, selected: 0 });
+      if (!filePickerLoadedRef.current) {
+        filePickerLoadedRef.current = true;
         const cwd = process.cwd();
         void fg4("**/*", {
           cwd,
@@ -11707,38 +12421,87 @@ function App({
             if (!a.isDirectory && b.isDirectory) return 1;
             return a.name.localeCompare(b.name);
           });
-          setPickerItems(items);
+          setFilePickerItems(items);
         }).catch(() => {
-          setPickerItems([]);
+          setFilePickerItems([]);
         });
       }
+      return;
+    }
+    if (shouldOpenSlashPicker(input, cursorOffset, pickerCancelRef.current)) {
+      setActivePicker({ kind: "slash", anchor: cursorOffset - 1, selected: 0 });
+      return;
+    }
+  }, [input, cursorOffset, activePicker, filePickerEnabled]);
+  useEffect4(() => {
+    if (activePicker?.kind !== "file") return;
+    const max = Math.max(0, filteredFileItems.length - 1);
+    if (activePicker.selected > max) {
+      setActivePicker({ ...activePicker, selected: max });
     }
-  }, [input, cursorOffset, pickerAnchor, pickerItems.length, filePickerEnabled]);
+  }, [filteredFileItems.length, activePicker]);
   useEffect4(() => {
-    if (pickerAnchor !== null) {
-      setPickerSelected((prev) => Math.min(prev, Math.max(0, filteredPickerItems.length - 1)));
+    if (activePicker?.kind !== "slash") return;
+    const max = Math.max(0, filteredSlashItems.length - 1);
+    if (activePicker.selected > max) {
+      setActivePicker({ ...activePicker, selected: max });
     }
-  }, [filteredPickerItems.length, pickerAnchor]);
+  }, [filteredSlashItems.length, activePicker]);
   const handlePickerUp = useCallback(() => {
-    setPickerSelected((i) => Math.max(0, i - 1));
+    setActivePicker((p) => {
+      if (!p) return null;
+      const next = Math.max(0, p.selected - 1);
+      return next === p.selected ? p : { ...p, selected: next };
+    });
   }, []);
   const handlePickerDown = useCallback(() => {
-    setPickerSelected((i) => Math.min(filteredPickerItems.length - 1, i + 1));
-  }, [filteredPickerItems.length]);
+    setActivePicker((p) => {
+      if (!p) return null;
+      const max = p.kind === "file" ? Math.max(0, filteredFileItems.length - 1) : Math.max(0, filteredSlashItems.length - 1);
+      const next = Math.min(max, p.selected + 1);
+      return next === p.selected ? p : { ...p, selected: next };
+    });
+  }, [filteredFileItems.length, filteredSlashItems.length]);
   const handlePickerSelect = useCallback(() => {
-    if (!mentionState || filteredPickerItems.length === 0) return;
-    const item = filteredPickerItems[pickerSelected];
+    if (!activePicker) return;
+    if (activePicker.kind === "file") {
+      const item2 = filteredFileItems[activePicker.selected];
+      if (!item2) return;
+      const insert = item2.name + (item2.isDirectory ? "/" : " ");
+      const newInput = input.slice(0, activePicker.anchor) + insert + input.slice(cursorOffset);
+      setInput(newInput);
+      setCursorOffset(activePicker.anchor + insert.length);
+      setActivePicker(null);
+      return;
+    }
+    const item = filteredSlashItems[activePicker.selected];
     if (!item) return;
-    const insert = item.name + (item.isDirectory ? "/" : " ");
-    const newInput = input.slice(0, mentionState.anchor) + insert + input.slice(cursorOffset);
-    setInput(newInput);
-    setCursorOffset(mentionState.anchor + insert.length);
-    setPickerAnchor(null);
-  }, [mentionState, filteredPickerItems, pickerSelected, input, cursorOffset]);
+    const { value, cursor } = insertSlashCommand(input, activePicker.anchor, item.name);
+    setInput(value);
+    setCursorOffset(cursor);
+    setActivePicker(null);
+  }, [activePicker, filteredFileItems, filteredSlashItems, input, cursorOffset]);
   const handlePickerCancel = useCallback(() => {
     pickerCancelRef.current = cursorOffset;
-    setPickerAnchor(null);
+    setActivePicker(null);
   }, [cursorOffset]);
+  useEffect4(() => {
+    const modalActive = showThemePicker || showHelpMenu || commandWizard !== null || commandPicker !== null || commandToDelete !== null || showCommandList || showLspWizard || resumeSessions !== null || perm !== null;
+    if (modalActive && activePicker !== null) {
+      setActivePicker(null);
+    }
+  }, [
+    showThemePicker,
+    showHelpMenu,
+    commandWizard,
+    commandPicker,
+    commandToDelete,
+    showCommandList,
+    showLspWizard,
+    resumeSessions,
+    perm,
+    activePicker
+  ]);
   useEffect4(() => {
     if (!cfg) return;
     void Promise.resolve().then(() => (init_sessions(), sessions_exports)).then(
@@ -11819,6 +12582,7 @@ function App({
     }
     void loadCustomCommands(process.cwd()).then(({ commands, warnings }) => {
       customCommandsRef.current = commands;
+      setCustomCommandsVersion((v) => v + 1);
       for (const w of warnings) {
         setEvents((e) => [...e, { kind: "info", key: mkKey(), text: `commands: ${w}` }]);
       }
@@ -11844,6 +12608,7 @@ function App({
   const reloadCustomCommands = useCallback(async () => {
     const { commands, warnings } = await loadCustomCommands(process.cwd());
     customCommandsRef.current = commands;
+    setCustomCommandsVersion((v) => v + 1);
     for (const w of warnings) {
       setEvents((e) => [...e, { kind: "info", key: mkKey(), text: `commands: ${w}` }]);
     }
@@ -12407,9 +13172,10 @@ function App({
           onAssistantStart: () => {
             const id = nextAssistantId++;
             activeAsstIdRef.current = id;
+            const role = orchestratorRef.current?.getActiveRole();
             setEvents((e) => [
               ...e,
-              { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true }
+              { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true, agentRole: role }
             ]);
           },
           onReasoningDelta: (d) => {
@@ -13422,9 +14188,10 @@ ${lines.join("\n")}` }]);
         onAssistantStart: () => {
           const id = nextAssistantId++;
           activeAsstIdRef.current = id;
+          const role = orchestratorRef.current?.getActiveRole();
           setEvents((e) => [
             ...e,
-            { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true }
+            { kind: "assistant", key: `asst_${id}`, id, text: "", reasoning: "", streaming: true, agentRole: role }
           ]);
         },
         onReasoningDelta: (d) => {
@@ -13570,7 +14337,8 @@ ${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);
             }
@@ -13777,7 +14545,7 @@ ${lines.join("\n")}` }]);
     }
   }, [usage]);
   if (!cfg) {
-    return /* @__PURE__ */ jsx19(
+    return /* @__PURE__ */ jsx20(
       Onboarding,
       {
         onDone: (newCfg) => {
@@ -13791,13 +14559,13 @@ ${lines.join("\n")}` }]);
     );
   }
   if (resumeSessions !== null) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(ResumePicker, { sessions: resumeSessions, onPick: handleResumePick, theme }) });
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(ResumePicker, { sessions: resumeSessions, onPick: handleResumePick, theme }) });
   }
   if (showThemePicker) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(ThemePicker, { themes: themeList(), current: theme, onPick: handleThemePick, onPreview: (t) => setTheme(t) }) });
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(ThemePicker, { themes: themeList(), current: theme, onPick: handleThemePick, onPreview: (t) => setTheme(t) }) });
   }
   if (showHelpMenu) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(
       HelpMenu,
       {
         theme,
@@ -13812,7 +14580,7 @@ ${lines.join("\n")}` }]);
     ) });
   }
   if (showLspWizard) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(
       LspWizard,
       {
         theme,
@@ -13850,7 +14618,7 @@ ${lines.join("\n")}` }]);
     ) });
   }
   if (commandWizard) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(
       CommandWizard,
       {
         theme,
@@ -13864,7 +14632,7 @@ ${lines.join("\n")}` }]);
     ) });
   }
   if (commandPicker) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(
       CommandPicker,
       {
         theme,
@@ -13883,14 +14651,14 @@ ${lines.join("\n")}` }]);
     ) });
   }
   if (commandToDelete) {
-    return /* @__PURE__ */ jsxs18(Box18, { flexDirection: "column", borderStyle: "round", borderColor: theme.accent, paddingX: 1, children: [
-      /* @__PURE__ */ jsxs18(Text19, { color: theme.accent, bold: true, children: [
+    return /* @__PURE__ */ jsxs19(Box19, { flexDirection: "column", borderStyle: "round", borderColor: theme.accent, paddingX: 1, children: [
+      /* @__PURE__ */ jsxs19(Text20, { color: theme.accent, bold: true, children: [
         "Delete /",
         commandToDelete.name,
         "?"
       ] }),
-      /* @__PURE__ */ jsx19(Text19, { color: theme.info.color, dimColor: true, children: commandToDelete.filepath }),
-      /* @__PURE__ */ jsx19(Box18, { marginTop: 1, children: /* @__PURE__ */ jsx19(
+      /* @__PURE__ */ jsx20(Text20, { color: theme.info.color, dimColor: true, children: commandToDelete.filepath }),
+      /* @__PURE__ */ jsx20(Box19, { marginTop: 1, children: /* @__PURE__ */ jsx20(
         SelectInput8,
         {
           items: [
@@ -13909,7 +14677,7 @@ ${lines.join("\n")}` }]);
     ] });
   }
   if (showCommandList) {
-    return /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", children: /* @__PURE__ */ jsx19(
+    return /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", children: /* @__PURE__ */ jsx20(
       CommandList,
       {
         theme,
@@ -13919,9 +14687,9 @@ ${lines.join("\n")}` }]);
     ) });
   }
   const hasConversation = events.some((e) => e.kind === "user" || e.kind === "assistant");
-  return /* @__PURE__ */ jsxs18(Box18, { flexDirection: "column", children: [
-    !hasConversation && events.length === 0 ? /* @__PURE__ */ jsx19(Welcome, { theme, accountId: cfg.accountId }) : /* @__PURE__ */ jsx19(ChatView, { events, showReasoning, theme, verbose }),
-    perm ? /* @__PURE__ */ jsx19(
+  return /* @__PURE__ */ jsxs19(Box19, { flexDirection: "column", children: [
+    !hasConversation && events.length === 0 ? /* @__PURE__ */ jsx20(Welcome, { theme, accountId: cfg.accountId }) : /* @__PURE__ */ jsx20(ChatView, { events, showReasoning, theme, verbose }),
+    perm ? /* @__PURE__ */ jsx20(
       PermissionModal,
       {
         tool: perm.tool,
@@ -13933,8 +14701,8 @@ ${lines.join("\n")}` }]);
           setPerm(null);
         }
       }
-    ) : /* @__PURE__ */ jsxs18(Box18, { flexDirection: "column", marginTop: 1, children: [
-      tasks.length > 0 && /* @__PURE__ */ jsx19(
+    ) : /* @__PURE__ */ jsxs19(Box19, { flexDirection: "column", marginTop: 1, children: [
+      tasks.length > 0 && /* @__PURE__ */ jsx20(
         TaskList,
         {
           tasks,
@@ -13943,11 +14711,11 @@ ${lines.join("\n")}` }]);
           tokensDelta: Math.max(0, (usage?.prompt_tokens ?? 0) - tasksStartTokens)
         }
       ),
-      queue.length > 0 && /* @__PURE__ */ jsx19(Box18, { flexDirection: "column", marginBottom: 1, children: queue.map((q, i) => /* @__PURE__ */ jsxs18(Text19, { color: theme.queue.color, dimColor: theme.queue.dim, children: [
+      queue.length > 0 && /* @__PURE__ */ jsx20(Box19, { flexDirection: "column", marginBottom: 1, children: queue.map((q, i) => /* @__PURE__ */ jsxs19(Text20, { color: theme.queue.color, dimColor: theme.queue.dim, children: [
         "\u23F3 ",
         q.display
       ] }, `queue_${i}`)) }),
-      /* @__PURE__ */ jsx19(
+      /* @__PURE__ */ jsx20(
         StatusBar,
         {
           model: cfg.model,
@@ -13965,18 +14733,27 @@ ${lines.join("\n")}` }]);
           codeMode
         }
       ),
-      pickerAnchor !== null && /* @__PURE__ */ jsx19(
+      activePicker?.kind === "file" && /* @__PURE__ */ jsx20(
         FilePicker,
         {
-          items: filteredPickerItems,
-          selectedIndex: pickerSelected,
+          items: filteredFileItems,
+          selectedIndex: activePicker.selected,
+          theme,
+          query: pickerQuery ?? ""
+        }
+      ),
+      activePicker?.kind === "slash" && /* @__PURE__ */ jsx20(
+        SlashPicker,
+        {
+          items: filteredSlashItems,
+          selectedIndex: activePicker.selected,
           theme,
-          query: mentionState?.query ?? ""
+          query: pickerQuery ?? ""
         }
       ),
-      /* @__PURE__ */ jsxs18(Box18, { marginTop: 1, children: [
-        /* @__PURE__ */ jsx19(Text19, { color: theme.accent, children: "\u203A " }),
-        /* @__PURE__ */ jsx19(
+      /* @__PURE__ */ jsxs19(Box19, { marginTop: 1, children: [
+        /* @__PURE__ */ jsx20(Text20, { color: theme.accent, children: "\u203A " }),
+        /* @__PURE__ */ jsx20(
           CustomTextInput,
           {
             value: input,
@@ -13985,7 +14762,7 @@ ${lines.join("\n")}` }]);
             enablePaste: true,
             cursorOffset,
             onCursorChange: setCursorOffset,
-            pickerActive: pickerAnchor !== null,
+            pickerActive: activePicker !== null,
             onPickerUp: handlePickerUp,
             onPickerDown: handlePickerDown,
             onPickerSelect: handlePickerSelect,
@@ -14033,7 +14810,7 @@ ${lines.join("\n")}` }]);
 }
 async function renderApp(cfg, updateResult, lspScope = "global", lspProjectPath = null) {
   const instance = render(
-    /* @__PURE__ */ jsx19(
+    /* @__PURE__ */ jsx20(
       App,
       {
         initialCfg: cfg,
@@ -14048,7 +14825,7 @@ async function renderApp(cfg, updateResult, lspScope = "global", lspProjectPath
   );
   await instance.waitUntilExit();
 }
-var MAX_GITIGNORE_SIZE, FEEDBACK_WORKER_URL, CONTEXT_LIMIT, AUTO_COMPACT_SUGGEST_PCT, MAX_EVENTS, nextAssistantId, nextKey, mkKey, MAX_IMAGES_PER_MESSAGE, BUILTIN_COMMAND_NAMES, EFFORT_DESCRIPTIONS;
+var MAX_GITIGNORE_SIZE, FEEDBACK_WORKER_URL, CONTEXT_LIMIT, AUTO_COMPACT_SUGGEST_PCT, MAX_EVENTS, nextAssistantId, nextKey, mkKey, MAX_IMAGES_PER_MESSAGE, EFFORT_DESCRIPTIONS;
 var init_app = __esm({
   "src/app.tsx"() {
     "use strict";
@@ -14087,6 +14864,7 @@ var init_app = __esm({
     init_version();
     init_loader();
     init_renderer2();
+    init_builtins();
     init_save();
     init_command_wizard();
     init_command_picker();
@@ -14095,6 +14873,8 @@ var init_app = __esm({
     init_lsp_config();
     init_lsp_nudge();
     init_file_picker();
+    init_slash_picker();
+    init_fuzzy();
     MAX_GITIGNORE_SIZE = 1 * 1024 * 1024;
     FEEDBACK_WORKER_URL = "https://kimiflare-feedback.sina-b35.workers.dev";
     CONTEXT_LIMIT = 262e3;
@@ -14104,33 +14884,6 @@ var init_app = __esm({
     nextKey = 1;
     mkKey = () => `evt_${nextKey++}`;
     MAX_IMAGES_PER_MESSAGE = 10;
-    BUILTIN_COMMAND_NAMES = /* @__PURE__ */ new Set([
-      "exit",
-      "quit",
-      "clear",
-      "reasoning",
-      "cost",
-      "model",
-      "thinking",
-      "effort",
-      "theme",
-      "mode",
-      "plan",
-      "auto",
-      "edit",
-      "resume",
-      "compact",
-      "init",
-      "update",
-      "mcp",
-      "logout",
-      "help",
-      "memory",
-      "gateway",
-      "hello",
-      "community",
-      "agent"
-    ]);
     EFFORT_DESCRIPTIONS = {
       low: "low \u2014 fastest; lightest reasoning. Best for simple Q&A, small edits, quick coordination.",
       medium: "medium \u2014 balanced (default). Solid quality on most edits, fast on trivial prompts.",