@hasna/terminal 1.1.0 → 1.2.0

package/dist/ai.js

@@ -36,6 +36,10 @@ const IRREVERSIBLE_PATTERNS = [
     /\brm\s/, /\brmdir\b/, /\btruncate\b/, /\bdrop\s+table\b/i,
     /\bdelete\s+from\b/i, /\bmv\b.*\/dev\/null/, /\b>\s*[^>]/,
     /\bdd\b/, /\bmkfs\b/, /\bformat\b/, /\bshred\b/,
+    // Process/service killing
+    /\bkill\b/, /\bkillall\b/, /\bpkill\b/,
+    // Git push/force operations
+    /\bgit\s+push\b/, /\bgit\s+reset\s+--hard\b/, /\bgit\s+force\b/,
 ];
 export function isIrreversible(command) {
     return IRREVERSIBLE_PATTERNS.some((r) => r.test(command));
@@ -96,6 +100,23 @@ function detectProjectContext() {
     if (existsSync(join(cwd, "build.gradle")) || existsSync(join(cwd, "build.gradle.kts"))) {
         parts.push("Project: Java/Gradle. Use gradle commands.");
     }
+    // Directory structure — so AI knows actual paths (not guessed ones)
+    try {
+        const { execSync } = require("child_process");
+        // Top-level dirs
+        const topLevel = execSync("ls -1", { cwd, encoding: "utf8", timeout: 2000 }).trim();
+        parts.push(`Top-level: ${topLevel.split("\n").join(", ")}`);
+        // src/ structure (2 levels deep, most important for path resolution)
+        for (const srcDir of ["src", "lib", "app"]) {
+            if (existsSync(join(cwd, srcDir))) {
+                const tree = execSync(`find ${srcDir} -maxdepth 2 -type d -not -path '*/node_modules/*' 2>/dev/null | head -30`, { cwd, encoding: "utf8", timeout: 2000 }).trim();
+                if (tree)
+                    parts.push(`Directories in ${srcDir}/:\n${tree}`);
+                break;
+            }
+        }
+    }
+    catch { /* timeout or no exec — skip */ }
     return parts.length > 0 ? `\n\nPROJECT CONTEXT:\n${parts.join("\n")}` : "";
 }
 // ── system prompt ─────────────────────────────────────────────────────────────

package/dist/cli.js

@@ -4,32 +4,31 @@ import { render } from "ink";
 const args = process.argv.slice(2);
 // ── Help / Version ───────────────────────────────────────────────────────────
 if (args[0] === "--help" || args[0] === "-h" || args[0] === "help") {
-    console.log(`open-terminal v1.0.0 — Smart terminal for AI agents and humans
+    console.log(`open-terminal — Natural language shell for AI agents and humans
 
 USAGE:
+  terminal "your request"      NL → AI picks command → runs → smart output
   terminal                     Launch interactive NL terminal (TUI)
-  terminal <subcommand>        Run a specific command
+
+EXAMPLES:
+  terminal "list all typescript files"
+  terminal "run tests"
+  terminal "what changed in git"
+  terminal "show me the auth functions"
+  terminal "kill port 3000"
+  terminal "how many lines of code"
 
 SUBCOMMANDS:
-  mcp serve                    Start MCP server (stdio transport)
-  mcp install --claude|--codex|--gemini|--all
-                               Install as MCP server for AI agents
-  hook install --claude        Install Claude Code PostToolUse hook
-  hook uninstall               Remove hooks
-  recipe add <name> <cmd>      Save a reusable command recipe
-  recipe list                  List saved recipes
-  recipe run <name> [--var=X]  Run a recipe with variable substitution
-  recipe delete <name>         Delete a recipe
-  collection create <name>     Create a recipe collection
-  collection list              List collections
-  project init                 Initialize project-scoped recipes
-  repo                         Show git repo state (branch + status + log)
-  symbols <file>               Show file outline (functions, classes, exports)
-  stats                        Show token economy dashboard
-  sessions                     List recent terminal sessions
-  sessions stats               Show session analytics
-  sessions <id>                Show session details
-  snapshot                     Capture terminal state as JSON
+  repo                         Git repo state (branch + status + log)
+  symbols <file>               File outline (functions, classes, exports)
+  overview                     Project overview (deps, scripts, structure)
+  stats                        Token economy dashboard
+  sessions [stats|<id>]        Session history and analytics
+  recipe add|list|run|delete   Reusable command recipes
+  collection create|list       Recipe collections
+  mcp serve                    Start MCP server for AI agents
+  mcp install --claude|--codex Install MCP server
+  snapshot                     Terminal state as JSON
   --help                       Show this help
   --version                    Show version
 
@@ -59,160 +58,6 @@ if (args[0] === "--version" || args[0] === "-v") {
     }
     process.exit(0);
 }
-// ── Exec command — smart execution for agents ────────────────────────────────
-if (args[0] === "exec") {
-    // Parse flags: --json, --offset=N, --limit=N, --raw
-    const flags = {};
-    const cmdParts = [];
-    for (const arg of args.slice(1)) {
-        const flagMatch = arg.match(/^--(\w+)(?:=(.+))?$/);
-        if (flagMatch) {
-            flags[flagMatch[1]] = flagMatch[2] ?? "true";
-        }
-        else {
-            cmdParts.push(arg);
-        }
-    }
-    const command = cmdParts.join(" ");
-    const jsonMode = flags.json === "true";
-    const rawMode = flags.raw === "true";
-    const offset = flags.offset ? parseInt(flags.offset) : undefined;
-    const limit = flags.limit ? parseInt(flags.limit) : undefined;
-    if (!command) {
-        console.error("Usage: terminal exec <command> [--json] [--raw] [--offset=N] [--limit=N]");
-        process.exit(1);
-    }
-    const { execSync } = await import("child_process");
-    const { compress, stripAnsi } = await import("./compression.js");
-    const { stripNoise } = await import("./noise-filter.js");
-    const { processOutput, shouldProcess } = await import("./output-processor.js");
-    const { rewriteCommand } = await import("./command-rewriter.js");
-    const { shouldBeLazy, toLazy, getSlice } = await import("./lazy-executor.js");
-    const { parseOutput, estimateTokens } = await import("./parsers/index.js");
-    const { recordSaving, recordUsage } = await import("./economy.js");
-    const { isTestOutput, trackTests, formatWatchResult } = await import("./test-watchlist.js");
-    const { detectLoop } = await import("./loop-detector.js");
-    // Loop detection — suggest narrowing if running full test suite repeatedly
-    const loop = detectLoop(command);
-    if (loop.detected) {
-        console.error(`[open-terminal] loop detected: test run #${loop.iteration}${loop.suggestedNarrow ? ` — try: ${loop.suggestedNarrow}` : " — consider narrowing to specific test file"}`);
-    }
-    // Rewrite command if possible
-    const rw = rewriteCommand(command);
-    const actualCmd = rw.changed ? rw.rewritten : command;
-    if (rw.changed)
-        console.error(`[open-terminal] rewritten: ${actualCmd} (${rw.reason})`);
-    try {
-        const start = Date.now();
-        const raw = execSync(actualCmd, { encoding: "utf8", maxBuffer: 10 * 1024 * 1024, cwd: process.cwd() });
-        const duration = Date.now() - start;
-        const clean = stripNoise(stripAnsi(raw)).cleaned;
-        const rawTokens = estimateTokens(raw);
-        // Track usage
-        recordUsage(rawTokens);
-        // --raw flag: skip all processing
-        if (rawMode) {
-            console.log(clean);
-            process.exit(0);
-        }
-        // --json flag: always return structured JSON
-        if (jsonMode) {
-            const parsed = parseOutput(actualCmd, clean);
-            if (parsed) {
-                const saved = rawTokens - estimateTokens(JSON.stringify(parsed.data));
-                if (saved > 0)
-                    recordSaving("structured", saved);
-                console.log(JSON.stringify({ exitCode: 0, parser: parsed.parser, data: parsed.data, duration, tokensSaved: Math.max(0, saved) }));
-            }
-            else {
-                const compressed = compress(actualCmd, clean, { format: "json" });
-                console.log(JSON.stringify({ exitCode: 0, output: compressed.content, duration, tokensSaved: compressed.tokensSaved }));
-            }
-            process.exit(0);
-        }
-        // Pagination: --offset + --limit on a previous large result
-        if (offset !== undefined || limit !== undefined) {
-            const slice = getSlice(clean, offset ?? 0, limit ?? 50);
-            console.log(slice.lines.join("\n"));
-            if (slice.hasMore)
-                console.error(`[open-terminal] showing ${slice.lines.length}/${slice.total}, ${slice.total - (offset ?? 0) - slice.lines.length} remaining`);
-            process.exit(0);
-        }
-        // Test output detection — use watchlist for structured test tracking
-        if (isTestOutput(clean)) {
-            const result = trackTests(process.cwd(), clean);
-            const formatted = formatWatchResult(result);
-            const savedTokens = rawTokens - estimateTokens(formatted);
-            if (savedTokens > 20)
-                recordSaving("structured", savedTokens);
-            if (jsonMode) {
-                console.log(JSON.stringify({ exitCode: 0, type: "test-results", ...result, duration: Date.now() - start }));
-            }
-            else {
-                console.log(formatted);
-            }
-            if (savedTokens > 10)
-                console.error(`[open-terminal] test watchlist: saved ${savedTokens} tokens`);
-            process.exit(0);
-        }
-        // Lazy mode for huge output (threshold 200, skip cat/summary commands)
-        if (shouldBeLazy(clean, actualCmd)) {
-            const lazy = toLazy(clean, actualCmd);
-            const savedTokens = rawTokens - estimateTokens(JSON.stringify(lazy));
-            if (savedTokens > 0)
-                recordSaving("compressed", savedTokens);
-            console.log(JSON.stringify({ ...lazy, duration, tokensSaved: savedTokens }));
-            process.exit(0);
-        }
-        // AI summary for medium-large output (>15 lines)
-        if (shouldProcess(clean)) {
-            const processed = await processOutput(actualCmd, clean);
-            if (processed.aiProcessed && processed.tokensSaved > 30) {
-                recordSaving("compressed", processed.tokensSaved);
-                console.log(processed.summary);
-                console.error(`[open-terminal] ${rawTokens} → ${rawTokens - processed.tokensSaved} tokens (saved ${processed.tokensSaved}, ${Math.round(processed.tokensSaved / rawTokens * 100)}%)`);
-                process.exit(0);
-            }
-        }
-        // Small/medium output — just noise-strip and return
-        console.log(clean);
-        const savedTokens = rawTokens - estimateTokens(clean);
-        if (savedTokens > 10) {
-            recordSaving("compressed", savedTokens);
-            console.error(`[open-terminal] saved ${savedTokens} tokens (noise filter)`);
-        }
-    }
-    catch (e) {
-        // Command failed — parse error output for structured diagnosis
-        const stderr = e.stderr?.toString() ?? "";
-        const stdout = e.stdout?.toString() ?? "";
-        // Deduplicate: if stderr content appears in stdout, skip it
-        const combined = stderr && stdout.includes(stderr.trim()) ? stdout : stdout + stderr;
-        const errorOutput = stripNoise(stripAnsi(combined)).cleaned;
-        // Try structured error parsing
-        const { errorParser } = await import("./parsers/errors.js");
-        if (errorOutput.length > 200 && errorParser.detect(actualCmd, errorOutput)) {
-            const info = errorParser.parse(actualCmd, errorOutput);
-            if (jsonMode) {
-                console.log(JSON.stringify({ exitCode: e.status ?? 1, error: info }));
-            }
-            else {
-                console.log(`Error: ${info.type}`);
-                console.log(`  ${info.message}`);
-                if (info.file)
-                    console.log(`  File: ${info.file}${info.line ? `:${info.line}` : ""}`);
-                if (info.suggestion)
-                    console.log(`  Fix: ${info.suggestion}`);
-            }
-        }
-        else {
-            // Short error or no parser match — pass through cleaned
-            console.log(errorOutput);
-        }
-        process.exit(e.status ?? 1);
-    }
-    process.exit(0);
-}
 // ── MCP commands ─────────────────────────────────────────────────────────────
 if (args[0] === "mcp") {
     if (args[1] === "serve" || args.length === 1) {
@@ -585,6 +430,13 @@ else if (args.length > 0) {
         console.error(`blocked: ${blocked}`);
         process.exit(1);
     }
+    // Safety: warn about irreversible commands (kill, push, rm, etc.)
+    if (isIrreversible(command)) {
+        console.error(`⚠ IRREVERSIBLE: $ ${command}`);
+        console.error(`  This command may kill processes, push code, or delete data.`);
+        console.error(`  Run with terminal exec "${command}" to bypass, or use the TUI for confirmation.`);
+        process.exit(1);
+    }
     // Show what we're running
     console.error(`$ ${command}`);
     // Step 2: Rewrite for optimization
@@ -605,7 +457,7 @@ else if (args.length > 0) {
         const rawTokens = estimateTokens(raw);
         recordUsage(rawTokens);
         // Test output detection
-        if (isTestOutput(clean)) {
+        if (isTestOutput(clean, actualCmd)) {
             const result = trackTests(process.cwd(), clean);
             console.log(formatWatchResult(result));
             process.exit(0);
@@ -621,7 +473,7 @@ else if (args.length > 0) {
         }
         // AI summary for medium-large output
         if (shouldProcess(clean)) {
-            const processed = await processOutput(actualCmd, clean);
+            const processed = await processOutput(actualCmd, clean, prompt);
             if (processed.aiProcessed && processed.tokensSaved > 30) {
                 recordSaving("compressed", processed.tokensSaved);
                 console.log(processed.summary);

package/dist/lazy-executor.js

@@ -39,7 +39,7 @@ export function toLazy(output, command) {
         count: lines.length,
         sample,
         categories: Object.keys(categories).length > 1 ? categories : undefined,
-        hint: `${lines.length} results. Showing first 20. Use terminal exec --offset=20 --limit=20 to paginate.`,
+        hint: `${lines.length} results. Showing first 20. Use a more specific query to narrow results.`,
     };
 }
 /** Get a slice of output */

package/dist/output-processor.js

@@ -5,26 +5,22 @@ import { estimateTokens } from "./parsers/index.js";
 import { recordSaving } from "./economy.js";
 const MIN_LINES_TO_PROCESS = 15;
 const MAX_OUTPUT_FOR_AI = 8000; // chars to send to AI (truncate if longer)
-const SUMMARIZE_PROMPT = `You are an output summarizer for a terminal. Given command output, return a CONCISE structured summary.
+const SUMMARIZE_PROMPT = `You are an intelligent terminal assistant. Given a user's original question and the command output, ANSWER THE QUESTION directly.
 
 RULES:
-- Return ONLY the summary, no explanations
-- For test output: show pass count, fail count, and ONLY the failing test names + errors
-- For build output: show status (ok/fail), error count, warning count
-- For install output: show package count, time, vulnerabilities
-- For file listings: show directory count, file count, notable files
-- For git output: show branch, status, key info
-- For logs: show line count, error count, latest error
-- For search results: show match count, top files
-- For ANY output: keep errors/failures/warnings, drop verbose/repetitive/progress lines
-- Use symbols: ✓ for success, ✗ for failure, ⚠ for warnings
-- Maximum 8 lines in your summary
-- If there are errors, ALWAYS include them verbatim`;
+- If the user asked a YES/NO question, start with Yes or No, then explain briefly
+- If the user asked "how many", give the number first, then context
+- If the user asked "show me X", show only X, not everything
+- ANSWER the question using the data — don't just summarize the raw output
+- Use symbols: ✓ for success/yes, ✗ for failure/no, ⚠ for warnings
+- Maximum 8 lines
+- Keep errors/failures verbatim
+- Be direct and concise — the user wants an ANSWER, not a data dump`;
 /**
  * Process command output through AI summarization.
  * Cheap AI call (~100 tokens) saves 1000+ tokens downstream.
  */
-export async function processOutput(command, output) {
+export async function processOutput(command, output, originalPrompt) {
     const lines = output.split("\n");
     // Short output — pass through, no AI needed
     if (lines.length <= MIN_LINES_TO_PROCESS) {
@@ -50,7 +46,7 @@ export async function processOutput(command, output) {
     }
     try {
         const provider = getProvider();
-        const summary = await provider.complete(`Command: ${command}\nOutput (${lines.length} lines):\n${toSummarize}`, {
+        const summary = await provider.complete(`${originalPrompt ? `User asked: ${originalPrompt}\n` : ""}Command: ${command}\nOutput (${lines.length} lines):\n${toSummarize}`, {
             system: SUMMARIZE_PROMPT,
             maxTokens: 300,
         });

package/dist/test-watchlist.js

@@ -37,10 +37,13 @@ function extractTests(output) {
     return tests;
 }
 /** Detect if output looks like test runner output */
-export function isTestOutput(output) {
-    // Must have a summary line with counts (not just words "pass"/"fail" in prose)
-    const summaryLine = /(?:\d+\s+pass|\d+\s+fail|Tests?:\s+\d+|Ran\s+\d+\s+tests?)/i;
-    const testMarkers = /(?:✓|✗|✔|✕|PASS\s+\S+\.test|FAIL\s+\S+\.test|bun test|jest|vitest|pytest)/;
+export function isTestOutput(output, command) {
+    // If the command is explicitly a test command, trust it
+    if (command && /\b(bun\s+test|npm\s+test|jest|vitest|pytest|cargo\s+test|go\s+test)\b/.test(command))
+        return true;
+    // Otherwise require BOTH a summary line AND a test runner marker in the output
+    const summaryLine = /(?:\d+\s+pass|\d+\s+fail|Tests?:\s+\d+|Ran\s+\d+\s+tests?)\s*$/im;
+    const testMarkers = /(?:✓|✗|✔|✕|PASS\s+\S+\.test|FAIL\s+\S+\.test|bun test v|jest|vitest|pytest)/;
     return summaryLine.test(output) && testMarkers.test(output);
 }
 /** Track test results and return only changes */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hasna/terminal",
-  "version": "1.1.0",
+  "version": "1.2.0",
   "description": "Smart terminal wrapper for AI agents and humans — structured output, token compression, MCP server, natural language",
   "type": "module",
   "bin": {

package/src/ai.ts

@@ -44,6 +44,10 @@ const IRREVERSIBLE_PATTERNS = [
   /\brm\s/, /\brmdir\b/, /\btruncate\b/, /\bdrop\s+table\b/i,
   /\bdelete\s+from\b/i, /\bmv\b.*\/dev\/null/, /\b>\s*[^>]/,
   /\bdd\b/, /\bmkfs\b/, /\bformat\b/, /\bshred\b/,
+  // Process/service killing
+  /\bkill\b/, /\bkillall\b/, /\bpkill\b/,
+  // Git push/force operations
+  /\bgit\s+push\b/, /\bgit\s+reset\s+--hard\b/, /\bgit\s+force\b/,
 ];
 
 export function isIrreversible(command: string): boolean {
@@ -124,6 +128,26 @@ function detectProjectContext(): string {
     parts.push("Project: Java/Gradle. Use gradle commands.");
   }
 
+  // Directory structure — so AI knows actual paths (not guessed ones)
+  try {
+    const { execSync } = require("child_process");
+    // Top-level dirs
+    const topLevel = execSync("ls -1", { cwd, encoding: "utf8", timeout: 2000 }).trim();
+    parts.push(`Top-level: ${topLevel.split("\n").join(", ")}`);
+
+    // src/ structure (2 levels deep, most important for path resolution)
+    for (const srcDir of ["src", "lib", "app"]) {
+      if (existsSync(join(cwd, srcDir))) {
+        const tree = execSync(
+          `find ${srcDir} -maxdepth 2 -type d -not -path '*/node_modules/*' 2>/dev/null | head -30`,
+          { cwd, encoding: "utf8", timeout: 2000 }
+        ).trim();
+        if (tree) parts.push(`Directories in ${srcDir}/:\n${tree}`);
+        break;
+      }
+    }
+  } catch { /* timeout or no exec — skip */ }
+
   return parts.length > 0 ? `\n\nPROJECT CONTEXT:\n${parts.join("\n")}` : "";
 }
 

package/src/cli.tsx

@@ -7,32 +7,31 @@ const args = process.argv.slice(2);
 // ── Help / Version ───────────────────────────────────────────────────────────
 
 if (args[0] === "--help" || args[0] === "-h" || args[0] === "help") {
-  console.log(`open-terminal v1.0.0 — Smart terminal for AI agents and humans
+  console.log(`open-terminal — Natural language shell for AI agents and humans
 
 USAGE:
+  terminal "your request"      NL → AI picks command → runs → smart output
   terminal                     Launch interactive NL terminal (TUI)
-  terminal <subcommand>        Run a specific command
+
+EXAMPLES:
+  terminal "list all typescript files"
+  terminal "run tests"
+  terminal "what changed in git"
+  terminal "show me the auth functions"
+  terminal "kill port 3000"
+  terminal "how many lines of code"
 
 SUBCOMMANDS:
-  mcp serve                    Start MCP server (stdio transport)
-  mcp install --claude|--codex|--gemini|--all
-                               Install as MCP server for AI agents
-  hook install --claude        Install Claude Code PostToolUse hook
-  hook uninstall               Remove hooks
-  recipe add <name> <cmd>      Save a reusable command recipe
-  recipe list                  List saved recipes
-  recipe run <name> [--var=X]  Run a recipe with variable substitution
-  recipe delete <name>         Delete a recipe
-  collection create <name>     Create a recipe collection
-  collection list              List collections
-  project init                 Initialize project-scoped recipes
-  repo                         Show git repo state (branch + status + log)
-  symbols <file>               Show file outline (functions, classes, exports)
-  stats                        Show token economy dashboard
-  sessions                     List recent terminal sessions
-  sessions stats               Show session analytics
-  sessions <id>                Show session details
-  snapshot                     Capture terminal state as JSON
+  repo                         Git repo state (branch + status + log)
+  symbols <file>               File outline (functions, classes, exports)
+  overview                     Project overview (deps, scripts, structure)
+  stats                        Token economy dashboard
+  sessions [stats|<id>]        Session history and analytics
+  recipe add|list|run|delete   Reusable command recipes
+  collection create|list       Recipe collections
+  mcp serve                    Start MCP server for AI agents
+  mcp install --claude|--codex Install MCP server
+  snapshot                     Terminal state as JSON
   --help                       Show this help
   --version                    Show version
 
@@ -61,156 +60,6 @@ if (args[0] === "--version" || args[0] === "-v") {
   process.exit(0);
 }
 
-// ── Exec command — smart execution for agents ────────────────────────────────
-
-if (args[0] === "exec") {
-  // Parse flags: --json, --offset=N, --limit=N, --raw
-  const flags: Record<string, string> = {};
-  const cmdParts: string[] = [];
-  for (const arg of args.slice(1)) {
-    const flagMatch = arg.match(/^--(\w+)(?:=(.+))?$/);
-    if (flagMatch) { flags[flagMatch[1]] = flagMatch[2] ?? "true"; }
-    else { cmdParts.push(arg); }
-  }
-  const command = cmdParts.join(" ");
-  const jsonMode = flags.json === "true";
-  const rawMode = flags.raw === "true";
-  const offset = flags.offset ? parseInt(flags.offset) : undefined;
-  const limit = flags.limit ? parseInt(flags.limit) : undefined;
-
-  if (!command) {
-    console.error("Usage: terminal exec <command> [--json] [--raw] [--offset=N] [--limit=N]");
-    process.exit(1);
-  }
-
-  const { execSync } = await import("child_process");
-  const { compress, stripAnsi } = await import("./compression.js");
-  const { stripNoise } = await import("./noise-filter.js");
-  const { processOutput, shouldProcess } = await import("./output-processor.js");
-  const { rewriteCommand } = await import("./command-rewriter.js");
-  const { shouldBeLazy, toLazy, getSlice } = await import("./lazy-executor.js");
-  const { parseOutput, estimateTokens } = await import("./parsers/index.js");
-  const { recordSaving, recordUsage } = await import("./economy.js");
-  const { isTestOutput, trackTests, formatWatchResult } = await import("./test-watchlist.js");
-  const { detectLoop } = await import("./loop-detector.js");
-
-  // Loop detection — suggest narrowing if running full test suite repeatedly
-  const loop = detectLoop(command);
-  if (loop.detected) {
-    console.error(`[open-terminal] loop detected: test run #${loop.iteration}${loop.suggestedNarrow ? ` — try: ${loop.suggestedNarrow}` : " — consider narrowing to specific test file"}`);
-  }
-
-  // Rewrite command if possible
-  const rw = rewriteCommand(command);
-  const actualCmd = rw.changed ? rw.rewritten : command;
-  if (rw.changed) console.error(`[open-terminal] rewritten: ${actualCmd} (${rw.reason})`);
-
-  try {
-    const start = Date.now();
-    const raw = execSync(actualCmd, { encoding: "utf8", maxBuffer: 10 * 1024 * 1024, cwd: process.cwd() });
-    const duration = Date.now() - start;
-    const clean = stripNoise(stripAnsi(raw)).cleaned;
-    const rawTokens = estimateTokens(raw);
-
-    // Track usage
-    recordUsage(rawTokens);
-
-    // --raw flag: skip all processing
-    if (rawMode) { console.log(clean); process.exit(0); }
-
-    // --json flag: always return structured JSON
-    if (jsonMode) {
-      const parsed = parseOutput(actualCmd, clean);
-      if (parsed) {
-        const saved = rawTokens - estimateTokens(JSON.stringify(parsed.data));
-        if (saved > 0) recordSaving("structured", saved);
-        console.log(JSON.stringify({ exitCode: 0, parser: parsed.parser, data: parsed.data, duration, tokensSaved: Math.max(0, saved) }));
-      } else {
-        const compressed = compress(actualCmd, clean, { format: "json" });
-        console.log(JSON.stringify({ exitCode: 0, output: compressed.content, duration, tokensSaved: compressed.tokensSaved }));
-      }
-      process.exit(0);
-    }
-
-    // Pagination: --offset + --limit on a previous large result
-    if (offset !== undefined || limit !== undefined) {
-      const slice = getSlice(clean, offset ?? 0, limit ?? 50);
-      console.log(slice.lines.join("\n"));
-      if (slice.hasMore) console.error(`[open-terminal] showing ${slice.lines.length}/${slice.total}, ${slice.total - (offset ?? 0) - slice.lines.length} remaining`);
-      process.exit(0);
-    }
-
-    // Test output detection — use watchlist for structured test tracking
-    if (isTestOutput(clean)) {
-      const result = trackTests(process.cwd(), clean);
-      const formatted = formatWatchResult(result);
-      const savedTokens = rawTokens - estimateTokens(formatted);
-      if (savedTokens > 20) recordSaving("structured", savedTokens);
-      if (jsonMode) {
-        console.log(JSON.stringify({ exitCode: 0, type: "test-results", ...result, duration: Date.now() - start }));
-      } else {
-        console.log(formatted);
-      }
-      if (savedTokens > 10) console.error(`[open-terminal] test watchlist: saved ${savedTokens} tokens`);
-      process.exit(0);
-    }
-
-    // Lazy mode for huge output (threshold 200, skip cat/summary commands)
-    if (shouldBeLazy(clean, actualCmd)) {
-      const lazy = toLazy(clean, actualCmd);
-      const savedTokens = rawTokens - estimateTokens(JSON.stringify(lazy));
-      if (savedTokens > 0) recordSaving("compressed", savedTokens);
-      console.log(JSON.stringify({ ...lazy, duration, tokensSaved: savedTokens }));
-      process.exit(0);
-    }
-
-    // AI summary for medium-large output (>15 lines)
-    if (shouldProcess(clean)) {
-      const processed = await processOutput(actualCmd, clean);
-      if (processed.aiProcessed && processed.tokensSaved > 30) {
-        recordSaving("compressed", processed.tokensSaved);
-        console.log(processed.summary);
-        console.error(`[open-terminal] ${rawTokens} → ${rawTokens - processed.tokensSaved} tokens (saved ${processed.tokensSaved}, ${Math.round(processed.tokensSaved/rawTokens*100)}%)`);
-        process.exit(0);
-      }
-    }
-
-    // Small/medium output — just noise-strip and return
-    console.log(clean);
-    const savedTokens = rawTokens - estimateTokens(clean);
-    if (savedTokens > 10) {
-      recordSaving("compressed", savedTokens);
-      console.error(`[open-terminal] saved ${savedTokens} tokens (noise filter)`);
-    }
-  } catch (e: any) {
-    // Command failed — parse error output for structured diagnosis
-    const stderr = e.stderr?.toString() ?? "";
-    const stdout = e.stdout?.toString() ?? "";
-    // Deduplicate: if stderr content appears in stdout, skip it
-    const combined = stderr && stdout.includes(stderr.trim()) ? stdout : stdout + stderr;
-    const errorOutput = stripNoise(stripAnsi(combined)).cleaned;
-
-    // Try structured error parsing
-    const { errorParser } = await import("./parsers/errors.js");
-    if (errorOutput.length > 200 && errorParser.detect(actualCmd, errorOutput)) {
-      const info = errorParser.parse(actualCmd, errorOutput);
-      if (jsonMode) {
-        console.log(JSON.stringify({ exitCode: e.status ?? 1, error: info }));
-      } else {
-        console.log(`Error: ${info.type}`);
-        console.log(`  ${info.message}`);
-        if (info.file) console.log(`  File: ${info.file}${info.line ? `:${info.line}` : ""}`);
-        if (info.suggestion) console.log(`  Fix: ${info.suggestion}`);
-      }
-    } else {
-      // Short error or no parser match — pass through cleaned
-      console.log(errorOutput);
-    }
-    process.exit(e.status ?? 1);
-  }
-  process.exit(0);
-}
-
 // ── MCP commands ─────────────────────────────────────────────────────────────
 
 if (args[0] === "mcp") {
@@ -561,6 +410,14 @@ else if (args.length > 0) {
   const blocked = checkPermissions(command, perms);
   if (blocked) { console.error(`blocked: ${blocked}`); process.exit(1); }
 
+  // Safety: warn about irreversible commands (kill, push, rm, etc.)
+  if (isIrreversible(command)) {
+    console.error(`⚠ IRREVERSIBLE: $ ${command}`);
+    console.error(`  This command may kill processes, push code, or delete data.`);
+    console.error(`  Run with terminal exec "${command}" to bypass, or use the TUI for confirmation.`);
+    process.exit(1);
+  }
+
   // Show what we're running
   console.error(`$ ${command}`);
 
@@ -583,7 +440,7 @@ else if (args.length > 0) {
     recordUsage(rawTokens);
 
     // Test output detection
-    if (isTestOutput(clean)) {
+    if (isTestOutput(clean, actualCmd)) {
       const result = trackTests(process.cwd(), clean);
       console.log(formatWatchResult(result));
       process.exit(0);
@@ -600,7 +457,7 @@ else if (args.length > 0) {
 
     // AI summary for medium-large output
     if (shouldProcess(clean)) {
-      const processed = await processOutput(actualCmd, clean);
+      const processed = await processOutput(actualCmd, clean, prompt);
       if (processed.aiProcessed && processed.tokensSaved > 30) {
         recordSaving("compressed", processed.tokensSaved);
         console.log(processed.summary);

package/src/lazy-executor.ts

@@ -54,7 +54,7 @@ export function toLazy(output: string, command: string): LazyResult {
     count: lines.length,
     sample,
     categories: Object.keys(categories).length > 1 ? categories : undefined,
-    hint: `${lines.length} results. Showing first 20. Use terminal exec --offset=20 --limit=20 to paginate.`,
+    hint: `${lines.length} results. Showing first 20. Use a more specific query to narrow results.`,
   };
 }
 

package/src/output-processor.ts

@@ -29,21 +29,17 @@ export interface ProcessedOutput {
 const MIN_LINES_TO_PROCESS = 15;
 const MAX_OUTPUT_FOR_AI = 8000; // chars to send to AI (truncate if longer)
 
-const SUMMARIZE_PROMPT = `You are an output summarizer for a terminal. Given command output, return a CONCISE structured summary.
+const SUMMARIZE_PROMPT = `You are an intelligent terminal assistant. Given a user's original question and the command output, ANSWER THE QUESTION directly.
 
 RULES:
-- Return ONLY the summary, no explanations
-- For test output: show pass count, fail count, and ONLY the failing test names + errors
-- For build output: show status (ok/fail), error count, warning count
-- For install output: show package count, time, vulnerabilities
-- For file listings: show directory count, file count, notable files
-- For git output: show branch, status, key info
-- For logs: show line count, error count, latest error
-- For search results: show match count, top files
-- For ANY output: keep errors/failures/warnings, drop verbose/repetitive/progress lines
-- Use symbols: ✓ for success, ✗ for failure, ⚠ for warnings
-- Maximum 8 lines in your summary
-- If there are errors, ALWAYS include them verbatim`;
+- If the user asked a YES/NO question, start with Yes or No, then explain briefly
+- If the user asked "how many", give the number first, then context
+- If the user asked "show me X", show only X, not everything
+- ANSWER the question using the data — don't just summarize the raw output
+- Use symbols: ✓ for success/yes, ✗ for failure/no, ⚠ for warnings
+- Maximum 8 lines
+- Keep errors/failures verbatim
+- Be direct and concise — the user wants an ANSWER, not a data dump`;
 
 /**
  * Process command output through AI summarization.
@@ -52,6 +48,7 @@ RULES:
 export async function processOutput(
   command: string,
   output: string,
+  originalPrompt?: string,
 ): Promise<ProcessedOutput> {
   const lines = output.split("\n");
 
@@ -82,7 +79,7 @@ export async function processOutput(
   try {
     const provider = getProvider();
     const summary = await provider.complete(
-      `Command: ${command}\nOutput (${lines.length} lines):\n${toSummarize}`,
+      `${originalPrompt ? `User asked: ${originalPrompt}\n` : ""}Command: ${command}\nOutput (${lines.length} lines):\n${toSummarize}`,
       {
         system: SUMMARIZE_PROMPT,
         maxTokens: 300,

package/src/test-watchlist.ts

@@ -63,10 +63,12 @@ function extractTests(output: string): TestStatus[] {
 }
 
 /** Detect if output looks like test runner output */
-export function isTestOutput(output: string): boolean {
-  // Must have a summary line with counts (not just words "pass"/"fail" in prose)
-  const summaryLine = /(?:\d+\s+pass|\d+\s+fail|Tests?:\s+\d+|Ran\s+\d+\s+tests?)/i;
-  const testMarkers = /(?:✓|✗|✔|✕|PASS\s+\S+\.test|FAIL\s+\S+\.test|bun test|jest|vitest|pytest)/;
+export function isTestOutput(output: string, command?: string): boolean {
+  // If the command is explicitly a test command, trust it
+  if (command && /\b(bun\s+test|npm\s+test|jest|vitest|pytest|cargo\s+test|go\s+test)\b/.test(command)) return true;
+  // Otherwise require BOTH a summary line AND a test runner marker in the output
+  const summaryLine = /(?:\d+\s+pass|\d+\s+fail|Tests?:\s+\d+|Ran\s+\d+\s+tests?)\s*$/im;
+  const testMarkers = /(?:✓|✗|✔|✕|PASS\s+\S+\.test|FAIL\s+\S+\.test|bun test v|jest|vitest|pytest)/;
   return summaryLine.test(output) && testMarkers.test(output);
 }