@freesyntax/notch-cli 0.5.13 → 0.5.16

package/dist/chunk-6M6CXXWR.js

@@ -0,0 +1,213 @@
+// src/agent/compression.ts
+import { generateText } from "ai";
+function estimateTokens(messages) {
+  let chars = 0;
+  for (const msg of messages) {
+    if (typeof msg.content === "string") {
+      chars += msg.content.length;
+    } else if (Array.isArray(msg.content)) {
+      for (const part of msg.content) {
+        if ("text" in part) chars += part.text.length;
+        else if ("result" in part) chars += JSON.stringify(part.result).length;
+        else if ("args" in part) chars += JSON.stringify(part.args).length;
+      }
+    }
+  }
+  return Math.ceil(chars / 4);
+}
+var TEXT_BLOCK_MAX = 8e3;
+function microCompact(messages) {
+  const result = [];
+  for (let idx = 0; idx < messages.length; idx++) {
+    const msg = messages[idx];
+    if (idx === 0 || idx >= messages.length - 4) {
+      result.push(msg);
+      continue;
+    }
+    if (typeof msg.content === "string" && msg.content.length > TEXT_BLOCK_MAX && msg.role === "assistant") {
+      result.push({
+        ...msg,
+        content: msg.content.slice(0, TEXT_BLOCK_MAX) + "\n... [truncated]"
+      });
+      continue;
+    }
+    result.push(msg);
+  }
+  return result;
+}
+var RESERVE_BUFFER_TOKENS = 13e3;
+var MAX_SUMMARY_TOKENS = 2e4;
+var MAX_COMPRESSION_FAILURES = 3;
+var compressionFailures = 0;
+async function autoCompactSummarize(messages, opts) {
+  const threshold = opts.contextWindow - RESERVE_BUFFER_TOKENS;
+  const currentTokens = estimateTokens(messages);
+  if (currentTokens < threshold * 0.75 || messages.length < 6) {
+    return { messages, compressed: false };
+  }
+  if (compressionFailures >= MAX_COMPRESSION_FAILURES) {
+    return deterministicCompress(messages, opts.keepRecent ?? 4);
+  }
+  const keepRecent = opts.keepRecent ?? 4;
+  const keepStart = 1;
+  const head = messages.slice(0, keepStart);
+  const middle = messages.slice(keepStart, -keepRecent);
+  const tail = messages.slice(-keepRecent);
+  if (middle.length === 0) {
+    return { messages, compressed: false };
+  }
+  const middleText = summarizeMessages(middle);
+  let summaryText;
+  try {
+    const result = await generateText({
+      model: opts.model,
+      system: "You are a conversation summarizer. Condense the following conversation history into a brief summary. Preserve: files modified, key decisions, errors encountered, and the current task state. Be concise but thorough. Output only the summary, no preamble.",
+      messages: [{ role: "user", content: middleText }],
+      maxTokens: Math.min(1024, Math.floor(MAX_SUMMARY_TOKENS / 4))
+    });
+    summaryText = result.text;
+    compressionFailures = 0;
+  } catch {
+    compressionFailures++;
+    return deterministicCompress(messages, keepRecent);
+  }
+  return buildCompressedHistory(head, summaryText, tail);
+}
+async function fullCompact(messages, model) {
+  if (messages.length < 4) {
+    return { messages, compressed: false };
+  }
+  const allButLast2 = messages.slice(0, -2);
+  const last2 = messages.slice(-2);
+  const middleText = summarizeMessages(allButLast2);
+  let summaryText;
+  try {
+    const result = await generateText({
+      model,
+      system: "Compress this entire conversation into a dense summary. Include: the original task, all files created/modified, key decisions, current state, and any unresolved issues. Max 500 words.",
+      messages: [{ role: "user", content: middleText }],
+      maxTokens: 2048
+    });
+    summaryText = result.text;
+  } catch {
+    summaryText = buildDeterministicSummary(allButLast2);
+  }
+  return buildCompressedHistory([], summaryText, last2);
+}
+function deterministicCompress(messages, keepRecent) {
+  const head = messages.slice(0, 1);
+  const middle = messages.slice(1, -keepRecent);
+  const tail = messages.slice(-keepRecent);
+  if (middle.length === 0) return { messages, compressed: false };
+  const summaryText = buildDeterministicSummary(middle);
+  return buildCompressedHistory(head, summaryText, tail);
+}
+function buildDeterministicSummary(messages) {
+  const filesModified = /* @__PURE__ */ new Set();
+  const toolsUsed = /* @__PURE__ */ new Set();
+  const userRequests = [];
+  let errorCount = 0;
+  for (const msg of messages) {
+    if (msg.role === "user" && typeof msg.content === "string") {
+      userRequests.push(msg.content.slice(0, 100));
+    }
+    if (Array.isArray(msg.content)) {
+      for (const part of msg.content) {
+        if ("toolName" in part) {
+          const p = part;
+          toolsUsed.add(String(p.toolName));
+          const args = p.args;
+          if (args?.path) filesModified.add(String(args.path));
+        }
+        if ("result" in part) {
+          const r = part;
+          const res = r.result;
+          if (res?.isError) errorCount++;
+        }
+      }
+    }
+  }
+  const lines = ["Summary of previous conversation:"];
+  if (userRequests.length > 0) lines.push(`- User requests: ${userRequests.join("; ")}`);
+  if (toolsUsed.size > 0) lines.push(`- Tools used: ${[...toolsUsed].join(", ")}`);
+  if (filesModified.size > 0) lines.push(`- Files touched: ${[...filesModified].join(", ")}`);
+  if (errorCount > 0) lines.push(`- Errors encountered: ${errorCount}`);
+  lines.push(`- Total messages summarized: ${messages.length}`);
+  return lines.join("\n");
+}
+function summarizeMessages(messages) {
+  const lines = [];
+  for (const msg of messages) {
+    const role = msg.role.toUpperCase();
+    if (typeof msg.content === "string") {
+      lines.push(`${role}: ${msg.content.slice(0, 500)}`);
+    } else if (Array.isArray(msg.content)) {
+      const parts = [];
+      for (const part of msg.content) {
+        if ("text" in part) parts.push(part.text.slice(0, 200));
+        else if ("toolName" in part) parts.push(`[tool: ${part.toolName}]`);
+        else if ("result" in part) parts.push(`[result: ${JSON.stringify(part.result).slice(0, 100)}]`);
+      }
+      lines.push(`${role}: ${parts.join(" | ")}`);
+    }
+  }
+  return lines.join("\n");
+}
+function buildCompressedHistory(head, summaryText, tail) {
+  const compressed = [...head];
+  const summaryContent = `[Previous conversation context]
+${summaryText}
+[End of context]`;
+  if (tail.length > 0 && tail[0].role === "user") {
+    const firstContent = typeof tail[0].content === "string" ? tail[0].content : "";
+    compressed.push({
+      role: "user",
+      content: `${summaryContent}
+
+---
+
+${firstContent}`
+    });
+    compressed.push(...tail.slice(1));
+  } else {
+    compressed.push({ role: "user", content: summaryContent });
+    compressed.push({
+      role: "assistant",
+      content: "Understood. I have the context from our previous conversation. Continuing."
+    });
+    compressed.push(...tail);
+  }
+  return { messages: compressed, compressed: true };
+}
+async function autoCompress(messages, model, contextWindow, onCompress) {
+  let result = microCompact(messages);
+  const threshold = (contextWindow - RESERVE_BUFFER_TOKENS) * 0.75;
+  let tokens = estimateTokens(result);
+  if (tokens < threshold) return result;
+  const auto = await autoCompactSummarize(result, { model, contextWindow });
+  if (auto.compressed) {
+    onCompress?.();
+    result = auto.messages;
+    tokens = estimateTokens(result);
+  }
+  if (tokens < threshold) return result;
+  const full = await fullCompact(result, model);
+  if (full.compressed) {
+    onCompress?.();
+    result = full.messages;
+  }
+  return result;
+}
+async function compressHistory(messages, opts) {
+  const result = await autoCompactSummarize(messages, opts);
+  return result;
+}
+
+export {
+  estimateTokens,
+  microCompact,
+  autoCompactSummarize,
+  fullCompact,
+  autoCompress,
+  compressHistory
+};

package/dist/compression-LPFNGAV6.js

@@ -0,0 +1,17 @@
+import {
+  autoCompactSummarize,
+  autoCompress,
+  compressHistory,
+  estimateTokens,
+  fullCompact,
+  microCompact
+} from "./chunk-6M6CXXWR.js";
+import "./chunk-3RG5ZIWI.js";
+export {
+  autoCompactSummarize,
+  autoCompress,
+  compressHistory,
+  estimateTokens,
+  fullCompact,
+  microCompact
+};

package/dist/index.js

@@ -7,7 +7,7 @@ import {
 import {
   autoCompress,
   estimateTokens
-} from "./chunk-MWM5TFY4.js";
+} from "./chunk-6M6CXXWR.js";
 import {
   __require
 } from "./chunk-3RG5ZIWI.js";
@@ -77,15 +77,6 @@ var MODEL_CATALOG = {
     maxOutputTokens: 16384,
     baseUrl: "https://cutmob--notch-serve-solace-notchsolaceserver-serve.modal.run/v1"
   },
-  "notch-forge-lite": {
-    id: "notch-forge-lite",
-    label: "Forge Lite",
-    size: "9B",
-    gpu: "L4",
-    contextWindow: 131072,
-    maxOutputTokens: 16384,
-    baseUrl: "https://cutmob--notch-serve-forge-lite-notchforgeliteserver-serve.modal.run/v1"
-  },
   "notch-solace-lite": {
     id: "notch-solace-lite",
     label: "Solace Lite",
@@ -417,36 +408,94 @@ var editTool = {
 };
 
 // src/tools/shell.ts
-import { execSync } from "child_process";
+import { execFile, exec } from "child_process";
+import { promisify } from "util";
 import path5 from "path";
 import { z as z4 } from "zod";
+var execFileAsync = promisify(execFile);
+var execAsync = promisify(exec);
 var BLOCKED_PATTERNS = [
   /rm\s+-rf\s+\/(?!\S)/,
   // rm -rf /
   /mkfs\./,
+  // format filesystem
   /dd\s+if=.*of=\/dev/,
+  // raw disk write
   /:\(\)\s*\{.*:\|:.*\}/,
-  // fork bomb variants
-  /chmod\s+-R\s+777\s+\//
+  // fork bomb
+  /chmod\s+-R\s+777\s+\//,
   // recursive chmod on root
+  /curl\s.*\|\s*(?:ba)?sh/,
+  // curl | sh (remote code execution)
+  /wget\s.*\|\s*(?:ba)?sh/,
+  // wget | sh
+  />\s*\/dev\/sd[a-z]/,
+  // overwrite disk device
+  /shutdown|reboot|init\s+[06]/,
+  // system shutdown/reboot
+  /rm\s+-rf\s+~\//
+  // rm -rf ~/
 ];
 var DESTRUCTIVE_PATTERNS = [
   /rm\s+-rf/,
   /rm\s+-r\s/,
   /git\s+push\s+--force(?!\s+--with-lease)/,
   /git\s+reset\s+--hard/,
+  /git\s+clean\s+-f/,
+  /git\s+checkout\s+--\s*\./,
   /DROP\s+(TABLE|DATABASE)/i,
   /TRUNCATE/i,
-  />\s*\/dev\/sd/
+  />\s*\/dev\/sd/,
+  /docker\s+(rm|rmi|system\s+prune)/,
+  /kubectl\s+delete/,
+  /npm\s+unpublish/
+];
+var BLOCKED_ENV_PATTERNS = [
+  /\bLD_PRELOAD=/,
+  /\bLD_LIBRARY_PATH=/,
+  /\bDYLD_INSERT_LIBRARIES=/,
+  /\bPATH=\//,
+  // Setting PATH to absolute (could shadow binaries)
+  /\bHOME=\//,
+  /\bSHELL=/
 ];
 var MAX_OUTPUT = 5e4;
 var DEFAULT_TIMEOUT = 3e4;
 var MAX_TIMEOUT = 6e5;
 var parameters4 = z4.object({
   command: z4.string().describe("Shell command to execute"),
-  timeout: z4.number().optional().describe("Timeout in ms (default 30s, max configurable up to 10m)")
+  timeout: z4.number().optional().describe("Timeout in ms (default 30s, max 10m)")
 });
 function validateCommand(command, cwd) {
+  for (const pattern of BLOCKED_PATTERNS) {
+    if (pattern.test(command)) {
+      return `Blocked: this command is too dangerous to execute.`;
+    }
+  }
+  for (const pattern of BLOCKED_ENV_PATTERNS) {
+    if (pattern.test(command)) {
+      return `Blocked: command attempts to override a protected environment variable.`;
+    }
+  }
+  const pipeSegments = command.split(/\s*\|\s*/);
+  for (const segment of pipeSegments) {
+    for (const pattern of BLOCKED_PATTERNS) {
+      if (pattern.test(segment.trim())) {
+        return `Blocked: a pipe segment contains a dangerous command.`;
+      }
+    }
+  }
+  const subCommands = [
+    ...command.matchAll(/\$\(([^)]+)\)/g),
+    ...command.matchAll(/`([^`]+)`/g)
+  ];
+  for (const match2 of subCommands) {
+    for (const pattern of BLOCKED_PATTERNS) {
+      if (pattern.test(match2[1])) {
+        return `Blocked: command substitution contains a dangerous command.`;
+      }
+    }
+  }
   const fileOpRegex = /(?:^|\s)(?:>|>>|cat|cp|mv|ln|tee|tar|zip|scp|rsync|chmod|chown|rm)\s+(\/(?!tmp\b|dev\/null\b)[^\s]+)/g;
   let match;
   while ((match = fileOpRegex.exec(command)) !== null) {
@@ -456,10 +505,19 @@ function validateCommand(command, cwd) {
     }
   }
   if (/(?:^|\s)(?:\.\.\/){3,}/.test(command)) {
-    return "Blocked: deep path traversal detected";
+    return "Blocked: deep path traversal detected.";
   }
   return null;
 }
+function isDestructive(command) {
+  const segments = [command, ...command.split(/\s*\|\s*/)];
+  for (const segment of segments) {
+    for (const pattern of DESTRUCTIVE_PATTERNS) {
+      if (pattern.test(segment)) return true;
+    }
+  }
+  return false;
+}
 var shellTool = {
   name: "shell",
   description: "Execute a shell command in the project directory. Dangerous commands (rm -rf, DROP TABLE, git push --force) require confirmation. Some destructive system commands are blocked entirely.",
@@ -468,50 +526,38 @@ var shellTool = {
     const { command } = params;
     const maxTimeout = ctx.shellTimeout ?? MAX_TIMEOUT;
     const timeout = Math.min(params.timeout ?? DEFAULT_TIMEOUT, maxTimeout);
-    for (const pattern of BLOCKED_PATTERNS) {
-      if (pattern.test(command)) {
-        return {
-          content: `Blocked: "${command}" is too dangerous to execute.`,
-          isError: true
-        };
-      }
+    const validationError = validateCommand(command, ctx.cwd);
+    if (validationError) {
+      return { content: validationError, isError: true };
     }
-    const pathError = validateCommand(command, ctx.cwd);
-    if (pathError) {
-      return { content: pathError, isError: true };
-    }
-    if (ctx.requireConfirm) {
-      for (const pattern of DESTRUCTIVE_PATTERNS) {
-        if (pattern.test(command)) {
-          const confirmed = await ctx.confirm(
-            `\u26A0 Destructive command: ${command}
+    if (ctx.requireConfirm && isDestructive(command)) {
+      const confirmed = await ctx.confirm(
+        `\u26A0 Destructive command: ${command}
 Proceed?`
-          );
-          if (!confirmed) {
-            return { content: "Command cancelled by user.", isError: true };
-          }
-          break;
-        }
+      );
+      if (!confirmed) {
+        return { content: "Command cancelled by user.", isError: true };
       }
     }
     try {
-      const output = execSync(command, {
+      const { stdout, stderr } = await execAsync(command, {
         cwd: ctx.cwd,
         encoding: "utf-8",
         timeout,
         maxBuffer: 10 * 1024 * 1024,
-        env: { ...process.env, FORCE_COLOR: "0" },
-        stdio: ["pipe", "pipe", "pipe"]
+        env: { ...process.env, FORCE_COLOR: "0" }
       });
-      const trimmed = output.length > MAX_OUTPUT ? output.slice(0, MAX_OUTPUT) + `
-... (truncated, ${output.length} chars total)` : output;
+      const combined = [stdout, stderr].filter(Boolean).join("\n");
+      const trimmed = combined.length > MAX_OUTPUT ? combined.slice(0, MAX_OUTPUT) + `
+... (truncated, ${combined.length} chars total)` : combined;
       return { content: trimmed || "(no output)" };
     } catch (err) {
-      const stderr = err.stderr?.toString() ?? "";
-      const stdout = err.stdout?.toString() ?? "";
+      const e = err;
+      const stderr = e.stderr ?? "";
+      const stdout = e.stdout ?? "";
       const combined = [stdout, stderr].filter(Boolean).join("\n");
       const trimmed = combined.length > MAX_OUTPUT ? combined.slice(0, MAX_OUTPUT) + "\n... (truncated)" : combined;
-      if (err.killed && err.signal === "SIGTERM") {
+      if (e.killed && e.signal === "SIGTERM") {
         return {
           content: `Command timed out after ${(timeout / 1e3).toFixed(0)}s: ${command}
 
@@ -521,8 +567,8 @@ ${trimmed || "(none)"}`,
         };
       }
       return {
-        content: `Command failed (exit ${err.status ?? "unknown"}):
-${trimmed || err.message}`,
+        content: `Command failed (exit ${e.status ?? e.code ?? "unknown"}):
+${trimmed || e.message || "Unknown error"}`,
         isError: true
       };
     }
@@ -666,7 +712,7 @@ Proceed?`
 };
 
 // src/tools/grep.ts
-import { execSync as execSync2 } from "child_process";
+import { execSync } from "child_process";
 import fs5 from "fs/promises";
 import path6 from "path";
 import { z as z6 } from "zod";
@@ -696,7 +742,7 @@ var grepTool = {
         JSON.stringify(params.pattern),
         searchPath
       ].filter(Boolean).join(" ");
-      const output = execSync2(rgArgs, {
+      const output = execSync(rgArgs, {
         cwd: ctx.cwd,
         encoding: "utf-8",
         timeout: 15e3,
@@ -2404,6 +2450,25 @@ async function updateIndex() {
 }
 
 // src/agent/loop.ts
+var MAX_RETRIES = 3;
+var RETRY_DELAYS = [1e3, 3e3, 8e3];
+function isRetryableError(err) {
+  if (!(err instanceof Error)) return false;
+  const msg = err.message.toLowerCase();
+  return msg.includes("502") || msg.includes("503") || msg.includes("429") || msg.includes("rate limit") || msg.includes("timeout") || msg.includes("econnreset") || msg.includes("econnrefused") || msg.includes("fetch failed") || msg.includes("network") || msg.includes("aborted");
+}
+async function sleep(ms) {
+  return new Promise((resolve2) => setTimeout(resolve2, ms));
+}
+function getErrorSignature(toolName, result) {
+  return {
+    toolName,
+    errorPrefix: result.slice(0, 120)
+  };
+}
+function signaturesMatch(a, b) {
+  return a.toolName === b.toolName && a.errorPrefix === b.errorPrefix;
+}
 async function runAgentLoop(messages, config) {
   const readCache = /* @__PURE__ */ new Map();
   const toolCtxWithCache = {
@@ -2418,6 +2483,8 @@ async function runAgentLoop(messages, config) {
   let totalPromptTokens = 0;
   let totalCompletionTokens = 0;
   let wasCompressed = false;
+  const recentErrors = [];
+  const MAX_REPEATED_ERRORS = 3;
   let history = [...messages];
   await config.toolContext.runHook?.("pre-compact", { messageCount: history.length });
   history = await autoCompress(history, config.model, contextWindow, () => {
@@ -2429,53 +2496,104 @@ async function runAgentLoop(messages, config) {
   }
   while (iterations < maxIter) {
     iterations++;
-    const result = streamText({
-      model: config.model,
-      system: config.systemPrompt,
-      messages: history,
-      tools,
-      maxSteps: 1
-      // We manage the loop ourselves for better control
-    });
     let fullText = "";
     const toolCalls = [];
     const toolResults = [];
-    for await (const event of result.fullStream) {
-      if (event.type === "text-delta") {
-        fullText += event.textDelta;
-        config.onTextChunk?.(event.textDelta);
-      } else if (event.type === "tool-call") {
-        toolCalls.push({
-          toolCallId: event.toolCallId,
-          toolName: event.toolName,
-          args: event.args
+    let streamUsage = null;
+    let lastError = null;
+    for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+      if (attempt > 0) {
+        const delay = RETRY_DELAYS[Math.min(attempt - 1, RETRY_DELAYS.length - 1)];
+        config.onRetry?.(attempt, lastError instanceof Error ? lastError.message : "unknown");
+        await sleep(delay);
+      }
+      try {
+        fullText = "";
+        toolCalls.length = 0;
+        toolResults.length = 0;
+        const result = streamText({
+          model: config.model,
+          system: config.systemPrompt,
+          messages: history,
+          tools,
+          maxSteps: 1
         });
-        config.onToolCall?.(event.toolName, event.args);
-      }
-      const evt = event;
-      if (evt.type === "tool-result") {
-        const res = evt.result;
-        toolResults.push({
-          toolCallId: evt.toolCallId,
-          result: evt.result
+        for await (const event of result.fullStream) {
+          if (event.type === "text-delta") {
+            fullText += event.textDelta;
+            config.onTextChunk?.(event.textDelta);
+          } else if (event.type === "tool-call") {
+            toolCalls.push({
+              toolCallId: event.toolCallId,
+              toolName: event.toolName,
+              args: event.args
+            });
+            config.onToolCall?.(event.toolName, event.args);
+          }
+          const evt = event;
+          if (evt.type === "tool-result") {
+            const res = evt.result;
+            toolResults.push({
+              toolCallId: evt.toolCallId,
+              result: evt.result
+            });
+            config.onToolResult?.(
+              toolCalls.find((tc) => tc.toolCallId === evt.toolCallId)?.toolName ?? "unknown",
+              res?.content ?? String(evt.result),
+              res?.isError ?? false
+            );
+          }
+        }
+        try {
+          const u = await result.usage;
+          if (u) streamUsage = u;
+        } catch {
+        }
+        lastError = null;
+        break;
+      } catch (err) {
+        lastError = err;
+        if (attempt < MAX_RETRIES && isRetryableError(err)) {
+          continue;
+        }
+        const errMsg = err instanceof Error ? err.message : String(err);
+        history.push({
+          role: "assistant",
+          content: `[Error: ${errMsg}. The model endpoint may be unavailable. Try again or switch models with /model.]`
         });
-        config.onToolResult?.(
-          toolCalls.find((tc) => tc.toolCallId === evt.toolCallId)?.toolName ?? "unknown",
-          res?.content ?? String(evt.result),
-          res?.isError ?? false
-        );
+        return {
+          text: `[Error: ${errMsg}]`,
+          messages: history,
+          iterations,
+          toolCallCount: totalToolCalls,
+          compressed: wasCompressed,
+          usage: {
+            promptTokens: totalPromptTokens,
+            completionTokens: totalCompletionTokens,
+            totalTokens: totalPromptTokens + totalCompletionTokens
+          }
+        };
       }
     }
-    try {
-      const u = await result.usage;
-      if (u) {
-        totalPromptTokens += u.promptTokens ?? 0;
-        totalCompletionTokens += u.completionTokens ?? 0;
-      }
-    } catch {
+    if (streamUsage) {
+      totalPromptTokens += streamUsage.promptTokens ?? 0;
+      totalCompletionTokens += streamUsage.completionTokens ?? 0;
     }
     totalToolCalls += toolCalls.length;
     if (toolCalls.length > 0) {
+      let hasRepeatedError = false;
+      for (const tr of toolResults) {
+        const res = tr.result;
+        if (res?.isError) {
+          const toolName = toolCalls.find((tc) => tc.toolCallId === tr.toolCallId)?.toolName ?? "unknown";
+          const sig = getErrorSignature(toolName, res.content ?? "");
+          const repeated = recentErrors.filter((e) => signaturesMatch(e, sig)).length;
+          recentErrors.push(sig);
+          if (repeated >= MAX_REPEATED_ERRORS - 1) {
+            hasRepeatedError = true;
+          }
+        }
+      }
       history.push({
         role: "assistant",
         content: [
@@ -2497,6 +2615,13 @@ async function runAgentLoop(messages, config) {
           result: tr.result
         }))
       });
+      if (hasRepeatedError) {
+        history.push({
+          role: "user",
+          content: "[System: You have repeated the same failing tool call multiple times. Stop and try a different approach. If the tool is unavailable or the command keeps failing, explain the issue to the user instead of retrying.]"
+        });
+        recentErrors.length = 0;
+      }
       if (iterations % 5 === 0) {
         const prevLen = history.length;
         await config.toolContext.runHook?.("pre-compact", { messageCount: prevLen });
@@ -2566,7 +2691,8 @@ async function buildSystemPrompt(projectRoot, modelId) {
     "- Explain what you're doing before making changes.",
     "- If a task is complex, break it into steps.",
     "- When running shell commands, prefer non-destructive operations.",
-    "- If you encounter an error, analyze it and suggest a fix."
+    "- If you encounter an error, analyze it and suggest a fix.",
+    "- If the same tool call keeps failing, stop retrying and try a different approach."
   ];
   try {
     const instructions = await loadProjectInstructions(projectRoot);
@@ -2755,12 +2881,12 @@ async function withRetry(fn, options = {}) {
         delay += Math.random() * delay * 0.5;
       }
       options.onRetry?.(attempt, err, delay);
-      await sleep(delay);
+      await sleep2(delay);
     }
   }
   throw lastError;
 }
-function sleep(ms) {
+function sleep2(ms) {
   return new Promise((resolve2) => setTimeout(resolve2, ms));
 }
 
@@ -3054,20 +3180,18 @@ function isPlanComplete(plan) {
 // src/agent/cost.ts
 import chalk4 from "chalk";
 var MODEL_COSTS = {
-  "notch-cinder": { input: 0.05, output: 0.15 },
-  // L4 — cheapest
-  "notch-forge": { input: 0.1, output: 0.3 },
-  // L40S
-  "notch-pyre": { input: 0.3, output: 0.9 },
-  // A100
-  "notch-ignis": { input: 0.4, output: 1.2 },
-  // A100
-  "notch-solace": { input: 0.35, output: 1 },
-  // A100 — Gemma 4 31B
-  "notch-forge-lite": { input: 0.08, output: 0.25 },
-  // L4 — INT8
-  "notch-solace-lite": { input: 0.04, output: 0.12 }
-  // L4 — Gemma 4 E4B
+  "notch-cinder": { input: 1.59, output: 7.66 },
+  // L4 (benchmarked 2026-04-02)
+  "notch-forge": { input: 3.17, output: 12.32 },
+  // L40S (benchmarked 2026-04-02)
+  "notch-pyre": { input: 4.34, output: 15.42 },
+  // A100-80GB (benchmarked 2026-04-02)
+  "notch-ignis": { input: 2.86, output: 25.7 },
+  // A100-80GB (benchmarked 2026-04-02)
+  "notch-solace": { input: 4.63, output: 36.15 },
+  // A100-80GB (benchmarked 2026-04-06)
+  "notch-solace-lite": { input: 1.11, output: 10.28 }
+  // L4 (benchmarked 2026-04-05)
 };
 var CostTracker = class {
   entries = [];
@@ -4298,7 +4422,7 @@ function formatTokens(n) {
 import fs15 from "fs/promises";
 import path18 from "path";
 import os4 from "os";
-import { execSync as execSync3 } from "child_process";
+import { execSync as execSync2 } from "child_process";
 import chalk8 from "chalk";
 var CACHE_FILE = path18.join(os4.homedir(), ".notch", "update-check.json");
 var CHECK_INTERVAL = 0;
@@ -4337,7 +4461,7 @@ function autoUpdate(current, latest) {
   \u2B06 Updating Notch CLI: ${current} \u2192 ${latest}...
 `));
   try {
-    execSync3(`npm install -g ${PACKAGE_NAME}@${latest}`, {
+    execSync2(`npm install -g ${PACKAGE_NAME}@${latest}`, {
       stdio: "inherit",
       timeout: 6e4
     });
@@ -4456,7 +4580,7 @@ function mergePermissions(base, override) {
 }
 
 // src/hooks/index.ts
-import { execSync as execSync4 } from "child_process";
+import { execSync as execSync3 } from "child_process";
 import fs17 from "fs/promises";
 import { watch } from "fs";
 import path20 from "path";
@@ -4564,7 +4688,7 @@ async function executeHook(hook, context) {
     NOTCH_CWD: context.cwd
   };
   try {
-    const output = execSync4(hook.command, {
+    const output = execSync3(hook.command, {
       cwd: context.cwd,
       encoding: "utf-8",
       timeout: hook.timeout ?? 1e4,
@@ -4936,7 +5060,7 @@ function findSync(oldLines, newLines, oi, ni, lookAhead) {
 }
 
 // src/commands/doctor.ts
-import { execSync as execSync5 } from "child_process";
+import { execSync as execSync4 } from "child_process";
 import fs20 from "fs/promises";
 import path23 from "path";
 import os8 from "os";
@@ -4953,7 +5077,7 @@ async function runDiagnostics(cwd) {
     results.push({ name: "Node.js", status: "fail", message: `v${nodeVersion} (requires >= 18)` });
   }
   try {
-    const gitVersion = execSync5("git --version", { encoding: "utf-8", timeout: 5e3 }).trim();
+    const gitVersion = execSync4("git --version", { encoding: "utf-8", timeout: 5e3 }).trim();
     results.push({ name: "Git", status: "ok", message: gitVersion });
   } catch {
     results.push({ name: "Git", status: "fail", message: "Not found. Install git to use git tools." });
@@ -5055,23 +5179,23 @@ registerCommand("/doctor", async (_args, ctx) => {
 });
 
 // src/commands/copy.ts
-import { execSync as execSync6 } from "child_process";
+import { execSync as execSync5 } from "child_process";
 import chalk11 from "chalk";
 function copyToClipboard(text) {
   try {
     const platform = process.platform;
     if (platform === "win32") {
-      execSync6("clip.exe", { input: text, timeout: 5e3 });
+      execSync5("clip.exe", { input: text, timeout: 5e3 });
     } else if (platform === "darwin") {
-      execSync6("pbcopy", { input: text, timeout: 5e3 });
+      execSync5("pbcopy", { input: text, timeout: 5e3 });
     } else {
       try {
-        execSync6("xclip -selection clipboard", { input: text, timeout: 5e3 });
+        execSync5("xclip -selection clipboard", { input: text, timeout: 5e3 });
       } catch {
         try {
-          execSync6("xsel --clipboard --input", { input: text, timeout: 5e3 });
+          execSync5("xsel --clipboard --input", { input: text, timeout: 5e3 });
         } catch {
-          execSync6("wl-copy", { input: text, timeout: 5e3 });
+          execSync5("wl-copy", { input: text, timeout: 5e3 });
         }
       }
     }
@@ -5117,7 +5241,7 @@ registerCommand("/btw", async (args, ctx) => {
 });
 
 // src/commands/security-review.ts
-import { execFileSync as execFileSync2, execSync as execSync7 } from "child_process";
+import { execFileSync as execFileSync2, execSync as execSync6 } from "child_process";
 import chalk13 from "chalk";
 function isValidGitRange(range) {
   return /^[a-zA-Z0-9._~^\/\-]+(\.\.[a-zA-Z0-9._~^\/\-]+)?$/.test(range);
@@ -5145,12 +5269,12 @@ registerCommand("/security-review", async (args, ctx) => {
     }).trim();
   } catch {
     try {
-      stat = execSync7("git diff --stat", {
+      stat = execSync6("git diff --stat", {
         cwd: ctx.cwd,
         encoding: "utf-8",
         timeout: 1e4
       }).trim();
-      diff = execSync7("git diff", {
+      diff = execSync6("git diff", {
         cwd: ctx.cwd,
         encoding: "utf-8",
         timeout: 1e4,
@@ -5422,7 +5546,7 @@ Read the file first, then make the change. Only modify this one file.`
 });
 
 // src/commands/plugin.ts
-import { execSync as execSync8, execFileSync as execFileSync3 } from "child_process";
+import { execSync as execSync7, execFileSync as execFileSync3 } from "child_process";
 import fs21 from "fs/promises";
 import path24 from "path";
 import os9 from "os";
@@ -5491,7 +5615,7 @@ registerCommand("/plugin", async (args, ctx) => {
         try {
           const pkgExists = await fs21.access(path24.join(pluginDir, "package.json")).then(() => true).catch(() => false);
           if (pkgExists) {
-            execSync8("npm install --production", {
+            execSync7("npm install --production", {
               cwd: pluginDir,
               encoding: "utf-8",
               timeout: 12e4,
@@ -5642,12 +5766,12 @@ Reply with ONLY the commit message, nothing else. No markdown, no explanation.`;
 });
 
 // src/commands/pr.ts
-import { execSync as execSync10, execFileSync as execFileSync5 } from "child_process";
+import { execSync as execSync9, execFileSync as execFileSync5 } from "child_process";
 import chalk18 from "chalk";
 import ora3 from "ora";
 function tryExec(cmd, cwd) {
   try {
-    return execSync10(cmd, { cwd, encoding: "utf-8", timeout: 15e3 }).trim();
+    return execSync9(cmd, { cwd, encoding: "utf-8", timeout: 15e3 }).trim();
   } catch {
     return null;
   }
@@ -5778,11 +5902,11 @@ BODY:
 });
 
 // src/commands/worktree.ts
-import { execSync as execSync11, execFileSync as execFileSync6 } from "child_process";
+import { execSync as execSync10, execFileSync as execFileSync6 } from "child_process";
 import chalk19 from "chalk";
 function tryExec2(cmd, cwd) {
   try {
-    return execSync11(cmd, { cwd, encoding: "utf-8", timeout: 15e3 }).trim();
+    return execSync10(cmd, { cwd, encoding: "utf-8", timeout: 15e3 }).trim();
   } catch {
     return null;
   }
@@ -5893,7 +6017,7 @@ registerCommand("/worktree", async (args, ctx) => {
     }
     case "prune": {
       try {
-        execSync11("git worktree prune", { cwd: ctx.cwd, encoding: "utf-8" });
+        execSync10("git worktree prune", { cwd: ctx.cwd, encoding: "utf-8" });
         console.log(chalk19.green("  \u2713 Pruned stale worktrees.\n"));
       } catch (err) {
         console.log(chalk19.red(`  Failed: ${err.message}
@@ -6390,8 +6514,8 @@ function clearMenu(state) {
   if (state.renderedLines === 0) return "";
   const lines = state.renderedLines;
   state.renderedLines = 0;
-  let ansi = "";
-  for (let i = 0; i < lines; i++) {
+  let ansi = "\x1B[2K";
+  for (let i = 0; i < lines + 1; i++) {
     ansi += "\x1B[1A\x1B[2K";
   }
   return ansi;
@@ -6456,16 +6580,15 @@ function attachSlashMenu(rl) {
         const shouldShow = updateMenu(state, line);
         if (wasVisible && !shouldShow) {
           process.stdout.write(clearMenu(state));
+          rewritePromptLine(rl);
         } else if (shouldShow) {
           if (wasVisible) {
             process.stdout.write(clearMenu(state));
           }
           const menuStr = renderMenu(state);
           if (menuStr) {
-            process.stdout.write("\x1B[s");
             process.stdout.write(menuStr + "\n");
             rewritePromptLine(rl);
-            process.stdout.write("\x1B[u");
           }
         }
       }
@@ -6516,19 +6639,60 @@ var modelChoices = MODEL_IDS.join(", ");
 var program = new Command().name("notch").description("Notch CLI \u2014 AI-powered coding assistant by Driftrail").version(VERSION).argument("[prompt...]", "One-shot prompt (runs once and exits)").option(`-m, --model <model>`, `Notch model (${modelChoices})`).option("--base-url <url>", "Override Notch API base URL").option("--api-key <key>", "Notch API key (prefer NOTCH_API_KEY env var)").option("--no-repo-map", "Disable automatic repository mapping").option("--no-markdown", "Disable markdown rendering in output").option("--max-iterations <n>", "Max tool-call rounds per turn", "25").option("-y, --yes", "Auto-confirm destructive actions").option("--trust", "Trust mode \u2014 auto-allow all tool calls").option("--theme <theme>", `UI color theme (${THEME_IDS.join(", ")})`).option("--resume", "Resume the last session for this project").option("--session <id>", "Resume a specific session by ID").option("--cwd <dir>", "Set working directory").parse(process.argv);
 var opts = program.opts();
 var promptArgs = program.args;
-function printModelTable(activeModel) {
-  const t = theme();
-  console.log(t.dim("\n  Available models:\n"));
-  for (const id of MODEL_IDS) {
-    const info = MODEL_CATALOG[id];
-    const active = id === activeModel ? t.success(" \u25CF") : "  ";
-    const label = id === activeModel ? t.bold(`${info.label}`) : t.dim(`${info.label}`);
-    const ctx = t.dim(`${(info.contextWindow / 1024).toFixed(0)}K ctx`);
-    console.log(`  ${active} ${t.brand(id.padEnd(14))} ${label}  ${ctx}`);
-  }
-  console.log(t.dim(`
-  Switch with: /model <name>
-`));
+function interactiveModelPicker(activeModel) {
+  return new Promise((resolve2) => {
+    const t = theme();
+    let cursor = MODEL_IDS.indexOf(activeModel);
+    if (cursor < 0) cursor = 0;
+    const render = () => {
+      process.stdout.write(`\x1B[${MODEL_IDS.length + 2}A\x1B[J`);
+      draw();
+    };
+    const draw = () => {
+      console.log(t.dim("  Select a model (\u2191\u2193 to move, Enter to select, Esc to cancel)\n"));
+      for (let i = 0; i < MODEL_IDS.length; i++) {
+        const id = MODEL_IDS[i];
+        const info = MODEL_CATALOG[id];
+        const isCurrent = id === activeModel;
+        const isSelected = i === cursor;
+        const pointer = isSelected ? t.brand("\u276F") : " ";
+        const dot = isCurrent ? t.success("\u25CF") : " ";
+        const label = isSelected ? t.bold(info.label) : t.dim(info.label);
+        const size = t.dim(info.size);
+        const gpu = t.dim(info.gpu);
+        const ctx = t.dim(`${(info.contextWindow / 1024).toFixed(0)}K`);
+        console.log(`  ${pointer} ${dot} ${t.brand(id.replace("notch-", "").padEnd(12))} ${label.padEnd(20)} ${size.padEnd(6)} ${gpu.padEnd(12)} ${ctx}`);
+      }
+    };
+    console.log("");
+    draw();
+    const stdin = process.stdin;
+    const wasRaw = stdin.isRaw;
+    stdin.setRawMode(true);
+    stdin.resume();
+    const onKey = (key) => {
+      const s = key.toString();
+      if (s === "\x1B[A") {
+        cursor = (cursor - 1 + MODEL_IDS.length) % MODEL_IDS.length;
+        render();
+      } else if (s === "\x1B[B") {
+        cursor = (cursor + 1) % MODEL_IDS.length;
+        render();
+      } else if (s === "\r" || s === "\n") {
+        cleanup();
+        resolve2(MODEL_IDS[cursor] ?? null);
+      } else if (s === "\x1B" || s === "") {
+        cleanup();
+        resolve2(null);
+      }
+    };
+    const cleanup = () => {
+      stdin.removeListener("data", onKey);
+      stdin.setRawMode(wasRaw ?? false);
+      process.stdout.write(`\x1B[${MODEL_IDS.length + 2}A\x1B[J`);
+    };
+    stdin.on("data", onKey);
+  });
 }
 function printHelp() {
   console.log(chalk27.gray(`
@@ -6974,7 +7138,23 @@ Analyze the above input.`;
       return;
     }
     if (input === "/model" || input === "/models") {
-      printModelTable(activeModelId);
+      rl.pause();
+      const picked = await interactiveModelPicker(activeModelId);
+      if (picked && picked !== activeModelId) {
+        activeModelId = picked;
+        config.models.chat.model = activeModelId;
+        model = resolveModel(config.models.chat);
+        const switchedInfo = MODEL_CATALOG[activeModelId];
+        console.log(chalk27.green(`  \u2713 Switched to ${switchedInfo.label} (${switchedInfo.id})
+`));
+      } else if (picked) {
+        console.log(chalk27.gray(`  Already using ${MODEL_CATALOG[activeModelId].label}
+`));
+      } else {
+        console.log(chalk27.gray(`  Cancelled
+`));
+      }
+      rl.resume();
       rl.prompt();
       return;
     }
@@ -7052,7 +7232,7 @@ Analyze the above input.`;
       return;
     }
     if (input === "/compact") {
-      const { autoCompress: autoCompress2 } = await import("./compression-CXJN2ZYN.js");
+      const { autoCompress: autoCompress2 } = await import("./compression-LPFNGAV6.js");
       const before = messages.length;
       const compressed = await autoCompress2(messages, model, MODEL_CATALOG[activeModelId].contextWindow);
       messages.length = 0;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@freesyntax/notch-cli",
-  "version": "0.5.13",
+  "version": "0.5.16",
   "description": "Notch CLI — AI-powered coding assistant by Driftrail",
   "type": "module",
   "bin": {

package/dist/chunk-MWM5TFY4.js

@@ -1,142 +0,0 @@
-// src/agent/compression.ts
-import { generateText } from "ai";
-function estimateTokens(messages) {
-  let chars = 0;
-  for (const msg of messages) {
-    if (typeof msg.content === "string") {
-      chars += msg.content.length;
-    } else if (Array.isArray(msg.content)) {
-      for (const part of msg.content) {
-        if ("text" in part) chars += part.text.length;
-        else if ("result" in part) chars += JSON.stringify(part.result).length;
-        else if ("args" in part) chars += JSON.stringify(part.args).length;
-      }
-    }
-  }
-  return Math.ceil(chars / 4);
-}
-async function compressHistory(messages, opts) {
-  const threshold = opts.contextWindow * 0.75;
-  const currentTokens = estimateTokens(messages);
-  if (currentTokens < threshold || messages.length < 6) {
-    return { messages, compressed: false };
-  }
-  const keepRecent = opts.keepRecent ?? 4;
-  const keepStart = 1;
-  const head = messages.slice(0, keepStart);
-  const middle = messages.slice(keepStart, -keepRecent);
-  const tail = messages.slice(-keepRecent);
-  if (middle.length === 0) {
-    return { messages, compressed: false };
-  }
-  const middleSummary = summarizeMessages(middle);
-  let summaryText;
-  try {
-    const result = await generateText({
-      model: opts.model,
-      system: "You are a conversation summarizer. Condense the following conversation history into a brief summary that preserves all important context, decisions made, files modified, and any errors encountered. Be concise but thorough. Output only the summary.",
-      messages: [{ role: "user", content: middleSummary }],
-      maxTokens: 1024
-    });
-    summaryText = result.text;
-  } catch {
-    summaryText = buildDeterministicSummary(middle);
-  }
-  const compressedMessages = [...head];
-  const summaryContent = `[Previous conversation context]
-${summaryText}
-[End of context]`;
-  if (tail.length > 0 && tail[0].role === "user") {
-    const firstContent = typeof tail[0].content === "string" ? tail[0].content : "";
-    compressedMessages.push({
-      role: "user",
-      content: `${summaryContent}
-
----
-
-${firstContent}`
-    });
-    compressedMessages.push(...tail.slice(1));
-  } else {
-    compressedMessages.push({ role: "user", content: summaryContent });
-    compressedMessages.push({
-      role: "assistant",
-      content: "Understood. I have the context from our previous conversation. Continuing."
-    });
-    compressedMessages.push(...tail);
-  }
-  return { messages: compressedMessages, compressed: true };
-}
-function summarizeMessages(messages) {
-  const lines = [];
-  for (const msg of messages) {
-    const role = msg.role.toUpperCase();
-    if (typeof msg.content === "string") {
-      lines.push(`${role}: ${msg.content.slice(0, 500)}`);
-    } else if (Array.isArray(msg.content)) {
-      const parts = [];
-      for (const part of msg.content) {
-        if ("text" in part) parts.push(part.text.slice(0, 200));
-        else if ("toolName" in part) parts.push(`[tool: ${part.toolName}]`);
-        else if ("result" in part) parts.push(`[result: ${JSON.stringify(part.result).slice(0, 100)}]`);
-      }
-      lines.push(`${role}: ${parts.join(" | ")}`);
-    }
-  }
-  return lines.join("\n");
-}
-function buildDeterministicSummary(messages) {
-  const filesModified = /* @__PURE__ */ new Set();
-  const toolsUsed = /* @__PURE__ */ new Set();
-  const userRequests = [];
-  let errorCount = 0;
-  for (const msg of messages) {
-    if (msg.role === "user" && typeof msg.content === "string") {
-      userRequests.push(msg.content.slice(0, 100));
-    }
-    if (Array.isArray(msg.content)) {
-      for (const part of msg.content) {
-        if ("toolName" in part) {
-          const p = part;
-          toolsUsed.add(p.toolName);
-          if (p.args?.path) filesModified.add(String(p.args.path));
-        }
-        if ("result" in part) {
-          const r = part;
-          if (r.result?.isError) errorCount++;
-        }
-      }
-    }
-  }
-  const lines = ["Summary of previous conversation:"];
-  if (userRequests.length > 0) {
-    lines.push(`- User requests: ${userRequests.join("; ")}`);
-  }
-  if (toolsUsed.size > 0) {
-    lines.push(`- Tools used: ${[...toolsUsed].join(", ")}`);
-  }
-  if (filesModified.size > 0) {
-    lines.push(`- Files touched: ${[...filesModified].join(", ")}`);
-  }
-  if (errorCount > 0) {
-    lines.push(`- Errors encountered: ${errorCount}`);
-  }
-  lines.push(`- Total messages summarized: ${messages.length}`);
-  return lines.join("\n");
-}
-async function autoCompress(messages, model, contextWindow, onCompress) {
-  const result = await compressHistory(messages, {
-    model,
-    contextWindow
-  });
-  if (result.compressed) {
-    onCompress?.();
-  }
-  return result.messages;
-}
-
-export {
-  estimateTokens,
-  compressHistory,
-  autoCompress
-};

package/dist/compression-CXJN2ZYN.js

@@ -1,11 +0,0 @@
-import {
-  autoCompress,
-  compressHistory,
-  estimateTokens
-} from "./chunk-MWM5TFY4.js";
-import "./chunk-3RG5ZIWI.js";
-export {
-  autoCompress,
-  compressHistory,
-  estimateTokens
-};