Auto-detect CLI commands — skip AI agent, save tokens

code_puppy/cli_runner.py

@@ -395,12 +395,17 @@ async def interactive_mode(message_renderer, initial_command: str = None) -> Non
     if initial_command:
         from code_puppy.command_line.shell_passthrough import (
             execute_shell_passthrough,
+            is_known_cli_command,
             is_shell_passthrough,
         )
 
         if is_shell_passthrough(initial_command):
             execute_shell_passthrough(initial_command)
             initial_command = None
+        elif is_known_cli_command(initial_command):
+            # Auto-detect known CLI commands — bypass AI agent, zero tokens
+            execute_shell_passthrough(f"!{initial_command.strip()}")
+            initial_command = None
 
     # Initialize the runtime agent manager
     if initial_command:
@@ -597,13 +602,20 @@ async def interactive_mode(message_renderer, initial_command: str = None) -> Non
         # Shell pass-through: !<command> executes directly, bypassing the agent
         from code_puppy.command_line.shell_passthrough import (
             execute_shell_passthrough,
+            is_known_cli_command,
             is_shell_passthrough,
         )
 
         if is_shell_passthrough(task):
             execute_shell_passthrough(task)
             continue
 
+        # Auto-detect known CLI commands (e.g. `ls`, `git status`, `grep …`)
+        # and route them directly to the shell — zero tokens consumed.
+        if is_known_cli_command(task):
+            execute_shell_passthrough(f"!{task.strip()}")
+            continue
+
         # Check for exit commands (plain text or command form)
         if task.strip().lower() in ["exit", "quit"] or task.strip().lower() in [
             "/exit",
@@ -1014,13 +1026,19 @@ async def execute_single_prompt(prompt: str, message_renderer) -> None:
     # Shell pass-through: !<cmd> bypasses the agent even in -p mode
     from code_puppy.command_line.shell_passthrough import (
         execute_shell_passthrough,
+        is_known_cli_command,
         is_shell_passthrough,
     )
 
     if is_shell_passthrough(prompt):
         execute_shell_passthrough(prompt)
         return
 
+    # Auto-detect known CLI commands — bypass AI agent, zero tokens consumed
+    if is_known_cli_command(prompt):
+        execute_shell_passthrough(f"!{prompt.strip()}")
+        return
+
     from code_puppy.messaging import emit_info
 
     emit_info(f"Executing prompt: {prompt}")

code_puppy/command_line/shell_passthrough.py

@@ -3,13 +3,22 @@
 Prepend a prompt with `!` to execute it as a shell command directly,
 bypassing the agent entirely. Inspired by Claude Code's `!` prefix.
 
+Also auto-detects well-known CLI commands (e.g. ``ls``, ``git``, ``grep``)
+so users can type them without the ``!`` prefix and still bypass the AI agent
+— consuming zero tokens.
+
 Examples:
     !ls -la
     !git status
     !python --version
+    ls -la          ← auto-detected, no tokens used
+    git status      ← auto-detected, no tokens used
+    ls | grep test  ← auto-detected, no tokens used
 """
 
 import os
+import re
+import shutil
 import subprocess
 import sys
 import time
@@ -25,6 +34,247 @@
 # Banner identifier — matches the key in DEFAULT_BANNER_COLORS
 _BANNER_NAME = "shell_passthrough"
 
+# ---------------------------------------------------------------------------
+# Auto-detection: known CLI commands that should bypass the AI agent
+# ---------------------------------------------------------------------------
+# When a user types one of these commands directly (without the ``!`` prefix),
+# Code Puppy automatically routes the input to the shell — no tokens consumed.
+#
+# Extend this set to add more commands. Keep it sorted for readability.
+KNOWN_CLI_COMMANDS: frozenset[str] = frozenset(
+    {
+        # ── Archives ──────────────────────────────────────────────────────
+        "bzip2",
+        "gunzip",
+        "gzip",
+        "tar",
+        "unzip",
+        "xz",
+        "zip",
+        # ── Build / package managers ──────────────────────────────────────
+        "cargo",
+        "cmake",
+        "gradle",
+        "make",
+        "maven",
+        "mvn",
+        "npm",
+        "npx",
+        "pip",
+        "pip3",
+        "poetry",
+        "pnpm",
+        "yarn",
+        # ── Containers / orchestration ────────────────────────────────────
+        "docker",
+        "docker-compose",
+        "helm",
+        "kubectl",
+        "podman",
+        # ── File system ───────────────────────────────────────────────────
+        "cat",
+        "cd",
+        "chmod",
+        "chown",
+        "cp",
+        "dir",
+        "du",
+        "file",
+        "find",
+        "head",
+        "la",
+        "less",
+        "ll",
+        "ln",
+        "locate",
+        "ls",
+        "mkdir",
+        "more",
+        "mv",
+        "popd",
+        "pushd",
+        "pwd",
+        "rm",
+        "rmdir",
+        "tail",
+        "touch",
+        "tree",
+        "wc",
+        # ── Language runtimes ─────────────────────────────────────────────
+        "go",
+        "java",
+        "javac",
+        "node",
+        "python",
+        "python3",
+        "ruby",
+        "rustc",
+        # ── Misc utilities ────────────────────────────────────────────────
+        "alias",
+        "cal",
+        "date",
+        "df",
+        "echo",
+        "env",
+        "export",
+        "free",
+        "history",
+        "id",
+        "info",
+        "man",
+        "open",
+        "pbcopy",
+        "pbpaste",
+        "printf",
+        "type",
+        "uname",
+        "unalias",
+        "uptime",
+        "which",
+        "whereis",
+        "whoami",
+        "xclip",
+        "xsel",
+        # ── Network ───────────────────────────────────────────────────────
+        "curl",
+        "dig",
+        "host",
+        "ifconfig",
+        "ip",
+        "nc",
+        "netstat",
+        "nslookup",
+        "ping",
+        "rsync",
+        "scp",
+        "ssh",
+        "wget",
+        # ── Process / system ──────────────────────────────────────────────
+        "bg",
+        "fg",
+        "htop",
+        "jobs",
+        "kill",
+        "killall",
+        "ps",
+        "top",
+        "who",
+        # ── System package managers ───────────────────────────────────────
+        "apt",
+        "apt-get",
+        "dnf",
+        "pacman",
+        "snap",
+        "systemctl",
+        "yum",
+        # ── Text processing ───────────────────────────────────────────────
+        "awk",
+        "cut",
+        "diff",
+        "egrep",
+        "fgrep",
+        "grep",
+        "jq",
+        "patch",
+        "rg",
+        "ripgrep",
+        "sed",
+        "sort",
+        "tee",
+        "tr",
+        "uniq",
+        "xargs",
+        # ── Version control ───────────────────────────────────────────────
+        "git",
+        "hg",
+        "svn",
+    }
+)
+
+# Commands that double as common English words and frequently appear at the
+# start of natural-language prompts (e.g. "find memory leak in parser",
+# "open the config file", "type the password").  For these we require stronger
+# shell intent evidence before auto-routing.
+_AMBIGUOUS_CLI_COMMANDS: frozenset[str] = frozenset(
+    {
+        "date",
+        "find",
+        "history",
+        "id",
+        "info",
+        "man",
+        "open",
+        "type",
+        "which",
+        "who",
+    }
+)
+
+# Patterns that strongly indicate the user intends a real shell invocation:
+#   |  &  ;  <  >  `  $  ( )   — shell operators / substitution
+#   -flag                       — a CLI flag (-v, --verbose, -la …)
+#   ./path  ../path  /abs/path  — explicit filesystem paths
+_SHELL_INTENT_RE = re.compile(r"[|&;<>`$()]|(?:^|\s)-\w|(?:^|\s)(\.{1,2}/|/)")
+
+# Pre-compiled regex: first "word" of the input (handles leading whitespace)
+_FIRST_WORD_RE = re.compile(r"^\s*(\S+)")
+
+
+def is_known_cli_command(task: str) -> bool:
+    """Return True when *task* starts with a well-known CLI command name **and**
+    the overall input looks like a real shell invocation rather than natural
+    language.
+
+    Three-stage filter (conservative by design):
+
+    1. **Known list** — first token must be in ``KNOWN_CLI_COMMANDS``.
+    2. **PATH check** — the executable must actually exist on ``$PATH``
+       (via ``shutil.which``).  This rejects invented commands that happen to
+       share a name with a list entry on this machine.
+    3. **Ambiguity guard** — for commands that are also common English words
+       (``find``, ``open``, ``date`` …) the rest of the input must show at
+       least one strong shell-intent signal: a flag (``-v``), a path
+       (``./src``), or a shell operator (``|``, ``>``, etc.).
+
+    Single-word inputs (``pwd``, ``git``) skip stage 3 — they're unambiguous.
+
+    Args:
+        task: Raw user input string.
+
+    Returns:
+        True if the input should be routed directly to the shell.
+    """
+    stripped = task.strip()
+
+    # Already handled by other code paths — skip early.
+    if stripped.startswith(SHELL_PASSTHROUGH_PREFIX) or stripped.startswith("/"):
+        return False
+
+    match = _FIRST_WORD_RE.match(stripped)
+    if not match:
+        return False
+
+    first_word = match.group(1).lower()
+
+    # Stage 1: must be a known CLI command.
+    if first_word not in KNOWN_CLI_COMMANDS:
+        return False
+
+    # Stage 2: executable must exist on PATH (prevents false positives on
+    # machines where the command isn't installed, and catches typos).
+    if shutil.which(first_word) is None:
+        return False
+
+    # Single-word commands are unambiguous — accept immediately.
+    if " " not in stripped:
+        return True
+
+    # Stage 3: ambiguous English words require explicit shell-intent evidence.
+    if first_word in _AMBIGUOUS_CLI_COMMANDS and not _SHELL_INTENT_RE.search(stripped):
+        return False
+
+    return True
+
 
 def _get_console() -> Console:
     """Get a Rich console for direct output.