@hasna/terminal 3.0.1 → 3.2.0

package/dist/ai.js

@@ -177,24 +177,43 @@ function buildSystemPrompt(perms, sessionEntries, currentPrompt) {
     const compoundBlock = wantsMultiple ? `\nCOMPOUND QUESTIONS: Prefer ONE command that captures all info. NEVER split into separate expensive commands.` : "";
     const wantsAnalysis = /\b(quality|lint|coverage|complexity|unused|dead code|security|audit|scan|dependency)\b/i.test(nl);
     const blockedAltBlock = wantsAnalysis ? `\nBLOCKED ALTERNATIVES: If your preferred command needs installing packages, try READ-ONLY alternatives (grep, cat, wc, awk). NEVER give up on analysis questions.` : "";
-    return `You are a terminal assistant. Output ONLY the exact shell command — no explanation, no markdown, no backticks.
+    return `Translate to bash. One command. Simplest form. No explanation.
 
-RULES:
-- SIMPLICITY FIRST: Use the simplest command. Prefer grep | sort | head over 10-pipe chains.
-- ALWAYS use grep -rn when searching directories. NEVER grep without -r on a directory.
-- When user refers to items from previous output, use EXACT names shown.
-- For text search use grep -rn, NOT nm or objdump.
-- macOS/BSD tools: du -d 1 (not --max-depth), NEVER grep -P, use grep -E for extended regex.
-- NEVER invent commands. Stick to standard Unix/macOS.
-- NEVER install packages. READ-ONLY terminal.
-- NEVER modify source code. Only observe.
-- Search src/ not dist/ or node_modules/.
-- Use exact file paths from project context. Do NOT guess paths.
-- For DESTRUCTIVE requests: output BLOCKED: <reason>.
-- ACTION vs CONCEPTUAL: "run/test/build/check" → executable command. "explain/what does X mean" → read docs.
-- EXISTENCE CHECKS: "is there/does X exist" → use ls/find/test, NEVER run/launch.${astBlock}${compoundBlock}${blockedAltBlock}
+list files in current directory → ls
+list all files including hidden → ls -a
+show open files → lsof
+create copy of a.txt as b.txt → cp a.txt b.txt
+create file test.txt → touch test.txt
+make directory testdir → mkdir testdir
+display routing table → route
+show last logged in users → last
+show file stats → stat file
+print directory tree 2 levels → tree -L 2
+count word occurrences in file → grep -c "word" file
+print number of files in dir → ls -1 | wc -l
+print first line of file → head -1 file
+print last line of file → tail -1 file
+print lines 3 to 5 of file → sed -n '3,5p' file
+print every other line → awk 'NR%2==1' file
+count words in file → wc -w file
+find empty files not in subdirs → find . -maxdepth 1 -type f -empty
+show system load → w
+system utilization stats → vmstat
+DNS servers → cat /etc/resolv.conf | grep nameserver
+long integer size → getconf LONG_BIT
+base64 decode string → echo 'str' | base64 -d
+change owner to nobody → chown nobody file
+unique lines in file → uniq file
+max cpu time → ulimit -t
+memory info → lsmem
+process priority → nice
+bash profile → cat ~/.bashrc
+search recursively → grep -rn "pattern" src/
+${astBlock}${compoundBlock}${blockedAltBlock}
 cwd: ${process.cwd()}
-shell: zsh / macOS${projectContext}${safetyBlock}${restrictionBlock}${contextBlock}${currentPrompt ? loadCorrectionHints(currentPrompt) : ""}`;
+shell: zsh / macOS${projectContext}${safetyBlock}${restrictionBlock}${contextBlock}${currentPrompt ? loadCorrectionHints(currentPrompt) : ""}
+
+Q:`;
 }
 // ── streaming translate ───────────────────────────────────────────────────────
 export async function translateToCommand(nl, perms, sessionEntries, onToken) {

package/dist/cli.js

@@ -482,15 +482,24 @@ else if (args.length > 0) {
     const config = loadConfig();
     const perms = config.permissions;
     const sessionCtx = formatContext();
+    // ── Direct command detection ──
+    // If input looks like a shell command (starts with known binary), skip AI translation entirely.
+    // This saves one AI call ($0.0008) per invocation for agents that already know the command.
+    const KNOWN_BINARIES = /^(ls|cd|cat|head|tail|grep|rg|find|wc|du|df|git|bun|npm|pnpm|yarn|node|python3?|pip|curl|wget|ssh|scp|chmod|chown|chgrp|mkdir|rmdir|rm|cp|mv|touch|ln|tar|gzip|gunzip|zip|unzip|sed|awk|sort|uniq|cut|tr|tee|xargs|echo|printf|env|export|source|which|whereis|whatis|man|date|cal|uptime|whoami|hostname|uname|ps|top|htop|kill|killall|lsof|netstat|ss|ifconfig|ip|ping|dig|nslookup|docker|kubectl|make|cmake|cargo|go|rustc|gcc|g\+\+|clang|java|javac|mvn|gradle|npx|bunx|tsx|deno|tree|file|stat|readlink|realpath|basename|dirname|pwd|test|true|false|sleep|timeout|time|watch|diff|patch|rsync|lsblk|mount|umount|fdisk|free|vmstat|iostat|sar|strace|ltrace|gdb|lldb|sqlite3|psql|mysql|redis-cli|mongosh|jq|yq|bat|fd|exa|fzf|gh|hub|terraform|ansible|helm|k9s|lazygit|tmux|screen|nc|nmap|openssl|base64|md5|shasum|xxd|od|hexdump|strings|nm|objdump|readelf|ldd|ldconfig|pkg-config|brew|apt|yum|dnf|pacman|snap|flatpak|systemctl|service|journalctl|dmesg|crontab|at|nohup|bg|fg|jobs|disown|wait|nice|renice|ionice|chrt|taskset|ulimit|sysctl|getconf|locale|iconv|perl|ruby|php|lua|R|julia|swift|kotlin|scala|elixir|mix|rebar3|tsc|eslint|prettier|biome|ruff|black|isort|mypy|pyright|pylint|flake8|pytest|vitest|jest|mocha|ava|tap|phpunit|rspec|minitest|unittest2|nose2|coverage|nyc|c8|v8-profiler)(\s|$)/;
+    const isDirectCommand = KNOWN_BINARIES.test(prompt.trim()) || /^[.\/~]/.test(prompt.trim()) || /\|/.test(prompt);
     // Check usage learning cache first (zero AI cost for repeated queries)
     const learned = getLearned(prompt);
     if (learned && !offlineMode) {
         console.error(`[open-terminal] cached: $ ${learned}`);
     }
-    // Step 1: AI translates NL → shell command (with session context for follow-ups)
+    // Step 1: Determine command — either direct passthrough or AI translation
     let command;
-    if (offlineMode) {
-        // Offline: treat prompt as literal command, apply noise filter only
+    if (isDirectCommand) {
+        // Direct command — skip AI translation entirely (saves 1 AI call)
+        command = prompt;
+    }
+    else if (offlineMode) {
+        // Offline: treat prompt as literal command
         console.error("[open-terminal] offline mode (no API key) — running as literal command");
         command = prompt;
     }

package/dist/output-processor.js

@@ -1,6 +1,6 @@
 // AI-powered output processor — uses cheap AI to intelligently summarize any output
 // NOTHING is hardcoded. The AI decides what's important, what's noise, what to keep.
-import { getProvider } from "./providers/index.js";
+import { getOutputProvider } from "./providers/index.js";
 import { estimateTokens } from "./tokens.js";
 import { recordSaving } from "./economy.js";
 import { discoverOutputHints } from "./context-hints.js";
@@ -8,9 +8,76 @@ import { formatProfileHints } from "./tool-profiles.js";
 import { stripAnsi } from "./compression.js";
 import { stripNoise } from "./noise-filter.js";
 const MIN_LINES_TO_PROCESS = 15;
-// Reserve ~2000 chars for system prompt + hints + profile + overhead
-const PROMPT_OVERHEAD_CHARS = 2000;
-const MAX_OUTPUT_FOR_AI = 6000; // chars of output to send to AI (leaves room for prompt overhead)
+const MAX_OUTPUT_FOR_AI = 6000;
+// ── Output fingerprinting — skip AI for outputs we can summarize instantly ──
+// These patterns match common terminal outputs that don't need AI interpretation.
+// Returns a short summary string, or null if AI should handle it.
+function fingerprint(command, output, exitCode) {
+    const trimmed = output.trim();
+    const lines = trimmed.split("\n").filter(l => l.trim());
+    // Empty output with success = command succeeded silently (build, lint, etc.)
+    if (lines.length === 0 && (exitCode === 0 || exitCode === undefined)) {
+        return "✓ Success (no output)";
+    }
+    // Single-line trivial outputs — pass through without AI
+    if (lines.length === 1 && trimmed.length < 80) {
+        return trimmed; // Already concise enough
+    }
+    // Git: common known patterns
+    if (/^Already up to date\.?$/i.test(trimmed))
+        return "✓ Already up to date";
+    if (/^nothing to commit, working tree clean$/i.test(trimmed))
+        return "✓ Clean working tree, nothing to commit";
+    if (/^On branch \S+\nnothing to commit/m.test(trimmed)) {
+        const branch = trimmed.match(/^On branch (\S+)/)?.[1];
+        return `✓ On branch ${branch}, clean working tree`;
+    }
+    if (/^Your branch is up to date/m.test(trimmed) && /nothing to commit/m.test(trimmed)) {
+        const branch = trimmed.match(/^On branch (\S+)/m)?.[1] ?? "?";
+        return `✓ Branch ${branch} up to date, clean`;
+    }
+    // Build/compile success with no errors
+    if (/^(tsc|bun|npm|yarn|pnpm)\s/.test(command)) {
+        if (lines.length <= 3 && (exitCode === 0 || exitCode === undefined) && !/error|Error|ERROR|fail|FAIL/.test(trimmed)) {
+            return `✓ Build succeeded${lines.length > 0 ? ` (${lines.length} lines)` : ""}`;
+        }
+    }
+    // npm/bun install success
+    if (/\binstall(ed)?\b.*\d+\s+packages?/i.test(trimmed) && !/error|Error|fail/i.test(trimmed)) {
+        const pkgMatch = trimmed.match(/(\d+)\s+packages?/);
+        return `✓ Installed ${pkgMatch?.[1] ?? "?"} packages`;
+    }
+    // Permission denied / not found — short errors pass through
+    if (lines.length <= 3 && /permission denied|command not found|No such file|ENOENT/i.test(trimmed)) {
+        return trimmed; // Already short enough, preserve error verbatim
+    }
+    // Hash-based dedup: if we've seen this exact output before, return cached summary
+    const hash = simpleHash(trimmed);
+    const cached = outputCache.get(hash);
+    if (cached)
+        return cached;
+    return null; // No fingerprint match — AI should handle this
+}
+// Simple string hash for output dedup
+function simpleHash(s) {
+    let hash = 0;
+    for (let i = 0; i < s.length; i++) {
+        hash = ((hash << 5) - hash + s.charCodeAt(i)) | 0;
+    }
+    return hash;
+}
+// LRU cache for output summaries (keyed by content hash)
+const OUTPUT_CACHE_MAX = 200;
+const outputCache = new Map();
+function cacheOutputSummary(output, summary) {
+    const hash = simpleHash(output.trim());
+    if (outputCache.size >= OUTPUT_CACHE_MAX) {
+        const oldest = outputCache.keys().next().value;
+        if (oldest !== undefined)
+            outputCache.delete(oldest);
+    }
+    outputCache.set(hash, 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:
@@ -30,6 +97,23 @@ RULES:
  */
 export async function processOutput(command, output, originalPrompt) {
     const lines = output.split("\n");
+    // Fingerprint check — skip AI entirely for known patterns (0ms, $0)
+    const fp = fingerprint(command, output);
+    if (fp && !originalPrompt) {
+        const saved = Math.max(0, estimateTokens(output) - estimateTokens(fp));
+        if (saved > 0)
+            recordSaving("compressed", saved);
+        return {
+            summary: fp,
+            full: output,
+            tokensSaved: saved,
+            aiTokensUsed: 0,
+            aiProcessed: false,
+            aiCostUsd: 0,
+            savingsValueUsd: 0,
+            netSavingsUsd: 0,
+        };
+    }
     // Short output — skip AI UNLESS we have an original prompt (NL mode needs answer framing)
     if (lines.length <= MIN_LINES_TO_PROCESS && !originalPrompt) {
         return {
@@ -62,8 +146,12 @@ export async function processOutput(command, output, originalPrompt) {
         // Inject tool-specific profile hints
         const profileBlock = formatProfileHints(command);
         const profileHints = profileBlock ? `\n\n${profileBlock}` : "";
-        const provider = getProvider();
+        // Use output-optimized provider (Groq llama-8b: fastest + best compression)
+        // Falls back to main provider if Groq unavailable
+        const provider = getOutputProvider();
+        const outputModel = provider.name === "groq" ? "llama-3.1-8b-instant" : undefined;
         const summary = await provider.complete(`${originalPrompt ? `User asked: ${originalPrompt}\n` : ""}Command: ${command}\nOutput (${lines.length} lines):\n${toSummarize}${hintsBlock}${profileHints}`, {
+            model: outputModel,
             system: SUMMARIZE_PROMPT,
             maxTokens: 300,
             temperature: 0.2,
@@ -94,6 +182,8 @@ export async function processOutput(command, output, originalPrompt) {
         if (netSavingsUsd > 0 && saved > 0) {
             recordSaving("compressed", saved);
         }
+        // Cache the AI summary for future identical outputs
+        cacheOutputSummary(output, summary);
         return {
             summary,
             full: output,

package/dist/providers/index.js

@@ -6,6 +6,7 @@ import { GroqProvider } from "./groq.js";
 import { XaiProvider } from "./xai.js";
 export { DEFAULT_PROVIDER_CONFIG } from "./base.js";
 let _provider = null;
+let _outputProvider = null;
 let _failedProviders = new Set();
 /** Get the active LLM provider. Auto-detects based on available API keys. */
 export function getProvider(config) {
@@ -18,8 +19,27 @@ export function getProvider(config) {
 /** Reset the cached provider (useful when config changes). */
 export function resetProvider() {
     _provider = null;
+    _outputProvider = null;
     _failedProviders.clear();
 }
+/**
+ * Get the provider optimized for output summarization.
+ * Priority: Groq (fastest, 234ms avg) > Cerebras > xAI > Anthropic.
+ * Falls back to the main provider if Groq is unavailable.
+ */
+export function getOutputProvider() {
+    if (_outputProvider)
+        return _outputProvider;
+    // Prefer Groq for output processing (fastest + best compression in evals)
+    const groq = new GroqProvider();
+    if (groq.isAvailable()) {
+        _outputProvider = groq;
+        return groq;
+    }
+    // Fall back to main provider
+    _outputProvider = getProvider();
+    return _outputProvider;
+}
 /** Get a fallback-wrapped provider that tries alternatives on failure */
 export function getProviderWithFallback(config) {
     const primary = getProvider(config);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hasna/terminal",
-  "version": "3.0.1",
+  "version": "3.2.0",
   "description": "Smart terminal wrapper for AI agents and humans — structured output, token compression, MCP server, natural language",
   "type": "module",
   "files": [

package/src/ai.ts

@@ -210,24 +210,43 @@ function buildSystemPrompt(perms: Permissions, sessionEntries: SessionEntry[], c
   const wantsAnalysis = /\b(quality|lint|coverage|complexity|unused|dead code|security|audit|scan|dependency)\b/i.test(nl);
   const blockedAltBlock = wantsAnalysis ? `\nBLOCKED ALTERNATIVES: If your preferred command needs installing packages, try READ-ONLY alternatives (grep, cat, wc, awk). NEVER give up on analysis questions.` : "";
 
-  return `You are a terminal assistant. Output ONLY the exact shell command — no explanation, no markdown, no backticks.
-
-RULES:
-- SIMPLICITY FIRST: Use the simplest command. Prefer grep | sort | head over 10-pipe chains.
-- ALWAYS use grep -rn when searching directories. NEVER grep without -r on a directory.
-- When user refers to items from previous output, use EXACT names shown.
-- For text search use grep -rn, NOT nm or objdump.
-- macOS/BSD tools: du -d 1 (not --max-depth), NEVER grep -P, use grep -E for extended regex.
-- NEVER invent commands. Stick to standard Unix/macOS.
-- NEVER install packages. READ-ONLY terminal.
-- NEVER modify source code. Only observe.
-- Search src/ not dist/ or node_modules/.
-- Use exact file paths from project context. Do NOT guess paths.
-- For DESTRUCTIVE requests: output BLOCKED: <reason>.
-- ACTION vs CONCEPTUAL: "run/test/build/check" → executable command. "explain/what does X mean" → read docs.
-- EXISTENCE CHECKS: "is there/does X exist" → use ls/find/test, NEVER run/launch.${astBlock}${compoundBlock}${blockedAltBlock}
+  return `Translate to bash. One command. Simplest form. No explanation.
+
+list files in current directory → ls
+list all files including hidden → ls -a
+show open files → lsof
+create copy of a.txt as b.txt → cp a.txt b.txt
+create file test.txt → touch test.txt
+make directory testdir → mkdir testdir
+display routing table → route
+show last logged in users → last
+show file stats → stat file
+print directory tree 2 levels → tree -L 2
+count word occurrences in file → grep -c "word" file
+print number of files in dir → ls -1 | wc -l
+print first line of file → head -1 file
+print last line of file → tail -1 file
+print lines 3 to 5 of file → sed -n '3,5p' file
+print every other line → awk 'NR%2==1' file
+count words in file → wc -w file
+find empty files not in subdirs → find . -maxdepth 1 -type f -empty
+show system load → w
+system utilization stats → vmstat
+DNS servers → cat /etc/resolv.conf | grep nameserver
+long integer size → getconf LONG_BIT
+base64 decode string → echo 'str' | base64 -d
+change owner to nobody → chown nobody file
+unique lines in file → uniq file
+max cpu time → ulimit -t
+memory info → lsmem
+process priority → nice
+bash profile → cat ~/.bashrc
+search recursively → grep -rn "pattern" src/
+${astBlock}${compoundBlock}${blockedAltBlock}
 cwd: ${process.cwd()}
-shell: zsh / macOS${projectContext}${safetyBlock}${restrictionBlock}${contextBlock}${currentPrompt ? loadCorrectionHints(currentPrompt) : ""}`;
+shell: zsh / macOS${projectContext}${safetyBlock}${restrictionBlock}${contextBlock}${currentPrompt ? loadCorrectionHints(currentPrompt) : ""}
+
+Q:`;
 }
 
 // ── streaming translate ───────────────────────────────────────────────────────

package/src/cli.tsx

@@ -459,17 +459,27 @@ else if (args.length > 0) {
   const perms = config.permissions;
   const sessionCtx = formatContext();
 
+  // ── Direct command detection ──
+  // If input looks like a shell command (starts with known binary), skip AI translation entirely.
+  // This saves one AI call ($0.0008) per invocation for agents that already know the command.
+  const KNOWN_BINARIES = /^(ls|cd|cat|head|tail|grep|rg|find|wc|du|df|git|bun|npm|pnpm|yarn|node|python3?|pip|curl|wget|ssh|scp|chmod|chown|chgrp|mkdir|rmdir|rm|cp|mv|touch|ln|tar|gzip|gunzip|zip|unzip|sed|awk|sort|uniq|cut|tr|tee|xargs|echo|printf|env|export|source|which|whereis|whatis|man|date|cal|uptime|whoami|hostname|uname|ps|top|htop|kill|killall|lsof|netstat|ss|ifconfig|ip|ping|dig|nslookup|docker|kubectl|make|cmake|cargo|go|rustc|gcc|g\+\+|clang|java|javac|mvn|gradle|npx|bunx|tsx|deno|tree|file|stat|readlink|realpath|basename|dirname|pwd|test|true|false|sleep|timeout|time|watch|diff|patch|rsync|lsblk|mount|umount|fdisk|free|vmstat|iostat|sar|strace|ltrace|gdb|lldb|sqlite3|psql|mysql|redis-cli|mongosh|jq|yq|bat|fd|exa|fzf|gh|hub|terraform|ansible|helm|k9s|lazygit|tmux|screen|nc|nmap|openssl|base64|md5|shasum|xxd|od|hexdump|strings|nm|objdump|readelf|ldd|ldconfig|pkg-config|brew|apt|yum|dnf|pacman|snap|flatpak|systemctl|service|journalctl|dmesg|crontab|at|nohup|bg|fg|jobs|disown|wait|nice|renice|ionice|chrt|taskset|ulimit|sysctl|getconf|locale|iconv|perl|ruby|php|lua|R|julia|swift|kotlin|scala|elixir|mix|rebar3|tsc|eslint|prettier|biome|ruff|black|isort|mypy|pyright|pylint|flake8|pytest|vitest|jest|mocha|ava|tap|phpunit|rspec|minitest|unittest2|nose2|coverage|nyc|c8|v8-profiler)(\s|$)/;
+
+  const isDirectCommand = KNOWN_BINARIES.test(prompt.trim()) || /^[.\/~]/.test(prompt.trim()) || /\|/.test(prompt);
+
   // Check usage learning cache first (zero AI cost for repeated queries)
   const learned = getLearned(prompt);
   if (learned && !offlineMode) {
     console.error(`[open-terminal] cached: $ ${learned}`);
   }
 
-  // Step 1: AI translates NL → shell command (with session context for follow-ups)
+  // Step 1: Determine command — either direct passthrough or AI translation
   let command: string;
 
-  if (offlineMode) {
-    // Offline: treat prompt as literal command, apply noise filter only
+  if (isDirectCommand) {
+    // Direct command — skip AI translation entirely (saves 1 AI call)
+    command = prompt;
+  } else if (offlineMode) {
+    // Offline: treat prompt as literal command
     console.error("[open-terminal] offline mode (no API key) — running as literal command");
     command = prompt;
   } else if (learned) {

package/src/output-processor.ts

@@ -1,7 +1,7 @@
 // AI-powered output processor — uses cheap AI to intelligently summarize any output
 // NOTHING is hardcoded. The AI decides what's important, what's noise, what to keep.
 
-import { getProvider } from "./providers/index.js";
+import { getProvider, getOutputProvider } from "./providers/index.js";
 import { estimateTokens } from "./tokens.js";
 import { recordSaving } from "./economy.js";
 import { discoverOutputHints } from "./context-hints.js";
@@ -31,9 +31,85 @@ export interface ProcessedOutput {
 }
 
 const MIN_LINES_TO_PROCESS = 15;
-// Reserve ~2000 chars for system prompt + hints + profile + overhead
-const PROMPT_OVERHEAD_CHARS = 2000;
-const MAX_OUTPUT_FOR_AI = 6000; // chars of output to send to AI (leaves room for prompt overhead)
+const MAX_OUTPUT_FOR_AI = 6000;
+
+// ── Output fingerprinting — skip AI for outputs we can summarize instantly ──
+// These patterns match common terminal outputs that don't need AI interpretation.
+// Returns a short summary string, or null if AI should handle it.
+
+function fingerprint(command: string, output: string, exitCode?: number): string | null {
+  const trimmed = output.trim();
+  const lines = trimmed.split("\n").filter(l => l.trim());
+
+  // Empty output with success = command succeeded silently (build, lint, etc.)
+  if (lines.length === 0 && (exitCode === 0 || exitCode === undefined)) {
+    return "✓ Success (no output)";
+  }
+
+  // Single-line trivial outputs — pass through without AI
+  if (lines.length === 1 && trimmed.length < 80) {
+    return trimmed; // Already concise enough
+  }
+
+  // Git: common known patterns
+  if (/^Already up to date\.?$/i.test(trimmed)) return "✓ Already up to date";
+  if (/^nothing to commit, working tree clean$/i.test(trimmed)) return "✓ Clean working tree, nothing to commit";
+  if (/^On branch \S+\nnothing to commit/m.test(trimmed)) {
+    const branch = trimmed.match(/^On branch (\S+)/)?.[1];
+    return `✓ On branch ${branch}, clean working tree`;
+  }
+  if (/^Your branch is up to date/m.test(trimmed) && /nothing to commit/m.test(trimmed)) {
+    const branch = trimmed.match(/^On branch (\S+)/m)?.[1] ?? "?";
+    return `✓ Branch ${branch} up to date, clean`;
+  }
+
+  // Build/compile success with no errors
+  if (/^(tsc|bun|npm|yarn|pnpm)\s/.test(command)) {
+    if (lines.length <= 3 && (exitCode === 0 || exitCode === undefined) && !/error|Error|ERROR|fail|FAIL/.test(trimmed)) {
+      return `✓ Build succeeded${lines.length > 0 ? ` (${lines.length} lines)` : ""}`;
+    }
+  }
+
+  // npm/bun install success
+  if (/\binstall(ed)?\b.*\d+\s+packages?/i.test(trimmed) && !/error|Error|fail/i.test(trimmed)) {
+    const pkgMatch = trimmed.match(/(\d+)\s+packages?/);
+    return `✓ Installed ${pkgMatch?.[1] ?? "?"} packages`;
+  }
+
+  // Permission denied / not found — short errors pass through
+  if (lines.length <= 3 && /permission denied|command not found|No such file|ENOENT/i.test(trimmed)) {
+    return trimmed; // Already short enough, preserve error verbatim
+  }
+
+  // Hash-based dedup: if we've seen this exact output before, return cached summary
+  const hash = simpleHash(trimmed);
+  const cached = outputCache.get(hash);
+  if (cached) return cached;
+
+  return null; // No fingerprint match — AI should handle this
+}
+
+// Simple string hash for output dedup
+function simpleHash(s: string): number {
+  let hash = 0;
+  for (let i = 0; i < s.length; i++) {
+    hash = ((hash << 5) - hash + s.charCodeAt(i)) | 0;
+  }
+  return hash;
+}
+
+// LRU cache for output summaries (keyed by content hash)
+const OUTPUT_CACHE_MAX = 200;
+const outputCache = new Map<number, string>();
+
+function cacheOutputSummary(output: string, summary: string): void {
+  const hash = simpleHash(output.trim());
+  if (outputCache.size >= OUTPUT_CACHE_MAX) {
+    const oldest = outputCache.keys().next().value;
+    if (oldest !== undefined) outputCache.delete(oldest);
+  }
+  outputCache.set(hash, summary);
+}
 
 const SUMMARIZE_PROMPT = `You are an intelligent terminal assistant. Given a user's original question and the command output, ANSWER THE QUESTION directly.
 
@@ -60,6 +136,23 @@ export async function processOutput(
 ): Promise<ProcessedOutput> {
   const lines = output.split("\n");
 
+  // Fingerprint check — skip AI entirely for known patterns (0ms, $0)
+  const fp = fingerprint(command, output);
+  if (fp && !originalPrompt) {
+    const saved = Math.max(0, estimateTokens(output) - estimateTokens(fp));
+    if (saved > 0) recordSaving("compressed", saved);
+    return {
+      summary: fp,
+      full: output,
+      tokensSaved: saved,
+      aiTokensUsed: 0,
+      aiProcessed: false,
+      aiCostUsd: 0,
+      savingsValueUsd: 0,
+      netSavingsUsd: 0,
+    };
+  }
+
   // Short output — skip AI UNLESS we have an original prompt (NL mode needs answer framing)
   if (lines.length <= MIN_LINES_TO_PROCESS && !originalPrompt) {
     return {
@@ -97,10 +190,14 @@ export async function processOutput(
     const profileBlock = formatProfileHints(command);
     const profileHints = profileBlock ? `\n\n${profileBlock}` : "";
 
-    const provider = getProvider();
+    // Use output-optimized provider (Groq llama-8b: fastest + best compression)
+    // Falls back to main provider if Groq unavailable
+    const provider = getOutputProvider();
+    const outputModel = provider.name === "groq" ? "llama-3.1-8b-instant" : undefined;
     const summary = await provider.complete(
       `${originalPrompt ? `User asked: ${originalPrompt}\n` : ""}Command: ${command}\nOutput (${lines.length} lines):\n${toSummarize}${hintsBlock}${profileHints}`,
       {
+        model: outputModel,
         system: SUMMARIZE_PROMPT,
         maxTokens: 300,
         temperature: 0.2,
@@ -138,6 +235,9 @@ export async function processOutput(
       recordSaving("compressed", saved);
     }
 
+    // Cache the AI summary for future identical outputs
+    cacheOutputSummary(output, summary);
+
     return {
       summary,
       full: output,

package/src/providers/index.ts

@@ -11,6 +11,7 @@ export type { LLMProvider, ProviderOptions, StreamCallbacks, ProviderConfig } fr
 export { DEFAULT_PROVIDER_CONFIG } from "./base.js";
 
 let _provider: LLMProvider | null = null;
+let _outputProvider: LLMProvider | null = null;
 let _failedProviders: Set<string> = new Set();
 
 /** Get the active LLM provider. Auto-detects based on available API keys. */
@@ -25,9 +26,30 @@ export function getProvider(config?: ProviderConfig): LLMProvider {
 /** Reset the cached provider (useful when config changes). */
 export function resetProvider() {
   _provider = null;
+  _outputProvider = null;
   _failedProviders.clear();
 }
 
+/**
+ * Get the provider optimized for output summarization.
+ * Priority: Groq (fastest, 234ms avg) > Cerebras > xAI > Anthropic.
+ * Falls back to the main provider if Groq is unavailable.
+ */
+export function getOutputProvider(): LLMProvider {
+  if (_outputProvider) return _outputProvider;
+
+  // Prefer Groq for output processing (fastest + best compression in evals)
+  const groq = new GroqProvider();
+  if (groq.isAvailable()) {
+    _outputProvider = groq;
+    return groq;
+  }
+
+  // Fall back to main provider
+  _outputProvider = getProvider();
+  return _outputProvider;
+}
+
 /** Get a fallback-wrapped provider that tries alternatives on failure */
 export function getProviderWithFallback(config?: ProviderConfig): LLMProvider {
   const primary = getProvider(config);