diff --git a/.forgejo/workflows/master.yml b/.forgejo/workflows/master.yml
index 7c959ba40..ccdede22d 100644
--- a/.forgejo/workflows/master.yml
+++ b/.forgejo/workflows/master.yml
@@ -3,8 +3,6 @@ name: CI
 on:
     push:
         branches: [master, develop]
-    pull_request:
-        branches: [master, develop]
 
 vars:
     docker_prefix: "http://harbor.cleverthis.com/docker/"
diff --git a/features/steps/tui_shell_safety_integration_steps.py b/features/steps/tui_shell_safety_integration_steps.py
new file mode 100644
index 000000000..0ba62b07b
--- /dev/null
+++ b/features/steps/tui_shell_safety_integration_steps.py
@@ -0,0 +1,296 @@
+"""Step definitions for tui_shell_safety_integration.feature.
+
+These steps exercise the end-to-end integration between :func:`run_shell_command`
+and :class:`~cleveragents.tui.shell_safety.ShellSafetyService`, verifying that:
+
+- Comprehensive pattern matching replaces the legacy substring-based
+  ``looks_dangerous`` check.
+- Both new-style (*warn_callback*) and legacy (*confirm_dangerous*) callbacks
+  are handled correctly.
+- Safety results include full warning details (danger level, matched pattern).
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+from unittest.mock import patch
+
+from behave import given, then, when
+from behave.runner import Context
+
+from cleveragents.tui.input.shell_exec import (
+    ShellResult,
+    looks_dangerous,
+    run_shell_command,
+)
+from cleveragents.tui.shell_safety import DangerousCommandWarning, ShellSafetyService
+
+
+# ---------------------------------------------------------------------------
+# Background & fixtures
+# ---------------------------------------------------------------------------
+
+
+@given("the shell safety integration module is imported")
+def step_import_module(context):  # noqa: ARG001
+    """Verify the integration module is importable."""
+    assert run_shell_command is not None
+    assert looks_dangerous is not None
+    assert ShellResult is not None
+
+
+# ---------------------------------------------------------------------------
+# Comprehensive danger detection (replaces legacy looks_dangerous)
+# ---------------------------------------------------------------------------
+
+
+@when(
+    "I check a MEDIUM-level command '{command}' through run_shell_command"
+)
+def step_run_medium_command(context, command):  # noqa: ARG001
+    """Run a MEDIUM-level dangerous command without any callback."""
+    context.shell_result = run_shell_command(command)
+
+
+@when(
+    "I check a HIGH-level command '{command}' through run_shell_command"
+)
+def step_run_high_command(context, command):  # noqa: ARG001
+    """Run a HIGH-level dangerous command without any callback."""
+    context.shell_result = run_shell_command(command)
+
+
+@when(
+    "I check a CRITICAL-level command '{command}' through run_shell_command"
+)
+def step_run_critical_command(context, command):  # noqa: ARG001
+    """Run a CRITICAL-level dangerous command without any callback."""
+    context.shell_result = run_shell_command(command)
+
+
+@when("I check a safe command '{command}' through run_shell_command")
+def step_run_safe_command(context, command):
+    """Run a safe command — should execute normally (mocked)."""
+    fake_proc = subprocess.CompletedProcess(
+        args=command, returncode=0, stdout="safe output", stderr=""
+    )
+    with patch(
+        "cleveragents.tui.input.shell_exec.subprocess.run", return_value=fake_proc
+    ):
+        context.shell_result = run_shell_command(command)
+
+
+@when("I check a LOW-level command '{command}' through run_shell_command")
+def step_run_low_command(context, command):
+    """Run a LOW-level dangerous command without any callback.
+
+    MEDIUM and above are auto-blocked; LOW should execute (mocked).
+    """
+    fake_proc = subprocess.CompletedProcess(
+        args=command, returncode=0, stdout="low-risk output", stderr=""
+    )
+    with patch(
+        "cleveragents.tui.input.shell_exec.subprocess.run", return_value=fake_proc
+    ):
+        context.shell_result = run_shell_command(command)
+
+
+# ---------------------------------------------------------------------------
+# Backward-compatible legacy confirm_dangerous callback
+# ---------------------------------------------------------------------------
+
+
+@given("a legacy confirm_dangerous callback")
+def step_create_legacy_callback(context):
+    """Create a legacy (command-string-based) confirmation callback."""
+    context.confirm_called = False
+
+    def old_style_confirm(command_string):  # noqa: ARG001
+        context.confirm_called = True
+        context.last_command_arg = command_string
+        return True
+
+    context.legacy_callback = old_style_confirm
+
+
+@given("a legacy confirm_dangerous callback that returns False")
+def step_create_legacy_callback_false(context):
+    """Create a legacy callback that always blocks."""
+    context.confirm_called = False
+
+    def old_style_confirm(command_string):  # noqa: ARG001
+        context.confirm_called = True
+        context.last_command_arg = command_string
+        return False
+
+    context.legacy_callback = old_style_confirm
+
+
+@when(
+    "I run a CRITICAL-level dangerous command '{command}' with the legacy callback"
+)
+def step_run_dangerous_with_legacy(context, command):
+    """Run a dangerous command using the legacy confirm_dangerous parameter."""
+    fake_proc = subprocess.CompletedProcess(
+        args=command, returncode=0, stdout="confirmed output", stderr=""
+    )
+
+    with patch(
+        "cleveragents.tui.input.shell_exec.subprocess.run", return_value=fake_proc
+    ):
+        context.shell_result = run_shell_command(
+            command, confirm_dangerous=context.legacy_callback  # type: ignore[arg-type]
+        )
+
+
+# ---------------------------------------------------------------------------
+# New-style warn_callback with full warning details
+# ---------------------------------------------------------------------------
+
+
+@given("a new-style warn_callback that inspects the warning")
+def step_create_warn_callback(context):
+    """Create a warn_callback that checks danger_level."""
+    context.warn_called = False
+    context.recorded_warning: DangerousCommandWarning | None = None
+
+    def new_style_warn(warning):
+        context.warn_called = True
+        context.recorded_warning = warning
+        return True
+
+    context.warn_callback = new_style_warn
+
+
+@when(
+    "I run a CRITICAL-level dangerous command '{command}' with the warn_callback"
+)
+def step_run_dangerous_with_warn(context, command):
+    """Run a dangerous command using the new-style warn_callback parameter."""
+    fake_proc = subprocess.CompletedProcess(
+        args=command, returncode=0, stdout="warn-allowed output", stderr=""
+    )
+
+    with patch(
+        "cleveragents.tui.input.shell_exec.subprocess.run", return_value=fake_proc
+    ):
+        context.shell_result = run_shell_command(
+            command, warn_callback=context.warn_callback  # type: ignore[arg-type]
+        )
+
+
+# ---------------------------------------------------------------------------
+# Deprecated looks_dangerous backward compatibility
+# ---------------------------------------------------------------------------
+
+
+@when("I call looks_dangerous on '{command}'")
+def step_check_legacy_is_safe(context, command):
+    """Assert legacy looks_dangerous uses the new service internally."""
+    context.is_dangerous = looks_dangerous(command)
+
+
+# ---------------------------------------------------------------------------
+# Common assertions for shell result
+# ---------------------------------------------------------------------------
+
+
+@then("the danger check should block the command (exit code 1)")
+def step_verify_blocked(context):  # noqa: ARG001
+    """Verify that dangerous commands without callback are blocked."""
+    assert context.shell_result.exit_code == 1, (
+        f"Expected exit_code=1, got {context.shell_result.exit_code}"
+    )
+    assert context.shell_result.stderr == "blocked dangerous shell command", (
+        f"Expected blocked message, got {context.shell_result.stderr!r}"
+    )
+
+
+@then("the danger check should allow the command")
+def step_verify_allowed(context):  # noqa: ARG001
+    """Verify that safe/low-risk commands proceed to execution."""
+    assert context.shell_result.exit_code == 0, (
+        f"Expected exit_code=0 after successful run, got {context.shell_result.exit_code}"
+    )
+
+
+@then(
+    "the legacy callback should have been invoked with a command string"
+)
+def step_verify_legacy_callback_invoked(context):
+    """Verify the legacy confirm_dangerous was called with just a string."""
+    assert context.confirm_called is True, "Legacy callback was not invoked"
+    assert isinstance(
+        context.last_command_arg, str  # type: ignore[attr-defined]
+    ), f"Expected string argument, got {type(context.last_command_arg)}"  # type: ignore[attr-defined]
+
+
+# ---------------------------------------------------------------------------
+# Common assertions for warn_callback invocations
+# ---------------------------------------------------------------------------
+
+
+@then("the warn callback should have been invoked with a warning")
+def step_verify_warn_callback_invoked(context):  # noqa: ARG001
+    """Verify the new-style warn_callback was called."""
+    assert context.warn_called is True, "warn_callback was not invoked"
+
+
+@then(
+    "the recorded warning should have danger level {level}"
+)
+def step_verify_warning_level(context, level):
+    """Verify the warning's danger level matches expectations."""
+    from cleveragents.tui.shell_safety import ShellDangerLevel
+
+    expected = _level_from_name(level)
+    assert context.recorded_warning is not None, "Expected a recorded warning"
+    assert (
+        context.recorded_warning.danger_level == expected  # type: ignore[attr-defined]
+    ), f"Expected danger_level={expected}, got {context.recorded_warning.danger_level}"  # type: ignore[attr-defined]
+
+
+@then(
+    'the recorded warning message should contain "{fragment}"'
+)
+def step_verify_warning_message(context, fragment):
+    """Verify the warning's message contains an expected fragment."""
+    assert context.recorded_warning is not None, "Expected a recorded warning"
+    assert fragment in context.recorded_warning.message, (
+        f"Expected message to contain {fragment!r}, "
+        f"got {context.recorded_warning.message!r}"  # type: ignore[attr-defined]
+    )
+
+
+# ---------------------------------------------------------------------------
+# Deprecated looks_dangerous assertions
+# ---------------------------------------------------------------------------
+
+
+@then("looks_dangerous should detect the command as dangerous")
+def step_legacy_detects_danger(context):  # noqa: ARG001
+    """Verify deprecated looks_dangerous now uses full service patterns."""
+    assert context.is_dangerous is True, (
+        f"Expected looks_dangerous to flag dangerous"
+    )
+
+
+@then("looks_dangerous should detect the command as safe")
+def step_legacy_detects_safe(context):  # noqa: ARG001
+    """Verify deprecated looks_dangerous allows safe commands."""
+    assert context.is_dangerous is False, (
+        f"Expected looks_dangerous to allow safe command"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Helper
+# ---------------------------------------------------------------------------
+
+
+def _level_from_name(name: str) -> ShellDangerLevel:
+    """Convert a name string to the corresponding ShellDangerLevel."""
+    from cleveragents.tui.shell_safety import ShellDangerLevel
+
+    return ShellDangerLevel[name.upper()]
diff --git a/features/tui_shell_safety_integration.feature b/features/tui_shell_safety_integration.feature
new file mode 100644
index 000000000..2a5590179
--- /dev/null
+++ b/features/tui_shell_safety_integration.feature
@@ -0,0 +1,114 @@
+Feature: TUI Shell Safety Integration via ShellSafetyService
+  Replace legacy substring-based looks_dangerous with comprehensive,
+  regex-driven ShellSafetyService in run_shell_command.
+
+  As a CleverAgents TUI user
+  I want dangerous shell commands to be detected with full context
+  So that I can make informed decisions about risky operations
+
+  Background:
+    Given the shell safety integration module is imported
+
+
+  # ── Comprehensive pattern matching (replaces legacy looks_dangerous) ───
+
+
+  Scenario: MEDIUM-level command is auto-blocked without callback
+    When I check a MEDIUM-level command "chmod 777 /etc/passwd" through run_shell_command
+    Then the danger check should block the command (exit code 1)
+
+
+  Scenario: HIGH-level command is auto-blocked without callback
+    When I check a HIGH-level command "dd if=/dev/zero of=/dev/sda" through run_shell_command
+    Then the danger check should block the command (exit code 1)
+
+
+  Scenario: CRITICAL-level command is auto-blocked without callback
+    When I check a CRITICAL-level command "rm -rf /" through run_shell_command
+    Then the danger check should block the command (exit code 1)
+
+
+  Scenario: Low-risk commands are allowed to execute
+    When I check a LOW-level command "git push --force origin main" through run_shell_command
+    Then the danger check should allow the command
+
+
+  Scenario: Safe commands proceed normally
+    When I check a safe command "echo hello world" through run_shell_command
+    Then the danger check should allow the command
+
+
+  # ── Backward-compatible legacy confirm_dangerous callback ─────────────
+
+  # The deprecated looks_dangerous function is replaced internally but
+  # existing callers that pass a (command: str) -> bool callback continue to work.
+
+
+  Scenario: Legacy confirm_dangerous allows dangerous commands when confirmed
+    Given a legacy confirm_dangerous callback
+    When I run a CRITICAL-level dangerous command "rm -rf /" with the legacy callback
+    Then the danger check should allow the command
+    And the legacy callback should have been invoked with a command string
+
+
+  Scenario: Legacy confirm_dangerous blocks dangerous commands when declined
+    Given a legacy confirm_dangerous callback that returns False
+    When I run a CRITICAL-level dangerous command "rm -rf /" with the legacy callback
+    Then the danger check should block the command (exit code 1)
+
+
+  # ── New-style warn_callback with full warning context ─────────────────
+
+  # Users can provide a more informative callback that receives a
+  # DangerousCommandWarning containing matched pattern and danger level.
+
+
+  Scenario: Warn-callback allows dangerous commands with full details
+    Given a new-style warn_callback that inspects the warning
+    When I run a CRITICAL-level dangerous command "rm -rf /" with the warn_callback
+    Then the danger check should allow the command
+    And the warn callback should have been invoked with a warning
+    And the recorded warning should have danger level CRITICAL
+
+
+  Scenario: Warn-callback receives the matched pattern description in message
+    Given a new-style warn_callback that inspects the warning
+    When I run a CRITICAL-level dangerous command "rm -rf /" with the warn_callback
+    Then the danger check should allow the command
+    And the recorded warning message should contain "Critical"
+
+
+  # ── Deprecated looks_dangerous backward compatibility ─────────────────
+
+  # The standalone looks_dangerous() function now delegates to ShellSafetyService,
+  # gaining access to all 15 built-in patterns across four severity levels.
+
+
+  Scenario: Deprecated looks_dangerous detects CRITICAL-level commands
+    When I call looks_dangerous on "rm -rf /"
+    Then looks_dangerous should detect the command as dangerous
+
+
+  Scenario: Deprecated looks_dangerous detects HIGH-level commands
+    When I call looks_dangerous on "dd if=/dev/zero of=/dev/sda"
+    Then looks_dangerous should detect the command as dangerous
+
+
+  Scenario: Deprecated looks_dangerous detects MEDIUM-level commands
+    When I call looks_dangerous on "chmod 777 /etc/passwd"
+    Then looks_dangerous should detect the command as dangerous
+
+
+  Scenario: Deprecated looks_dangerous detects LOW-level commands
+    When I call looks_dangerous on "git push --force origin main"
+    Then looks_dangerous should detect the command as dangerous
+
+
+  Scenario: Deprecated looks_dangerous allows safe commands
+    When I call looks_dangerous on "ls -la /home"
+    Then looks_dangerous should detect the command as safe
+
+
+  Scenario: Deprecated looks_dangerous detects fork bombs
+    When I call looks_dangerous on ":(){ :|:& };:"
+    Then looks_dangerous should detect the command as dangerous
diff --git a/src/cleveragents/tui/input/shell_exec.py b/src/cleveragents/tui/input/shell_exec.py
index 015214695..faf72f23d 100644
--- a/src/cleveragents/tui/input/shell_exec.py
+++ b/src/cleveragents/tui/input/shell_exec.py
@@ -7,6 +7,21 @@ import subprocess
 from collections.abc import Callable
 from dataclasses import dataclass
 
+from cleveragents.tui.shell_safety import (
+    DangerousCommandWarning,
+    ShellSafetyService,
+)
+
+# ---------------------------------------------------------------------------
+# Module-level safety service — single instance shared by all callers
+# ---------------------------------------------------------------------------
+
+#: Primary shell-safety application service used by :func:`run_shell_command`.
+#: The default block level is :attr:`~cleveragents.tui.shell_safety.ShellDangerLevel.MEDIUM`,
+#: meaning commands at *MEDIUM* danger or above are blocked automatically when no
+#: callback is provided.
+_safety_service = ShellSafetyService()
+
 
 @dataclass(slots=True, frozen=True)
 class ShellResult:
@@ -18,31 +33,111 @@ class ShellResult:
     stderr: str
 
 
-def looks_dangerous(command: str) -> bool:
-    """Best-effort dangerous command detector."""
-    lowered = command.strip().lower()
-    patterns = (
-        "rm -rf /",
-        "git push --force",
-        "mkfs.",
-        "dd if=",
-        ":(){:|:&};:",
-    )
-    return any(pattern in lowered for pattern in patterns)
+# ---------------------------------------------------------------------------
+# Backward-compatible danger detection (deprecated)
+# ---------------------------------------------------------------------------
 
+def looks_dangerous(command: str) -> bool:
+    """Best-effort dangerous command detector.
+
+    .. deprecated::
+        This function is preserved for backward compatibility and delegates to
+        the internal :class:`~cleveragents.tui.shell_safety.ShellSafetyService`
+        which uses comprehensive regex-based pattern matching across four
+        severity levels (LOW, MEDIUM, HIGH, CRITICAL).  New callers should use
+        :func:`run_shell_command` which handles safety checks automatically.
+
+    The legacy implementation matched against a small set of substring-based
+    patterns (``rm -rf /``, ``git push --force``, ``mkfs.``, ``dd if=``, fork
+    bomb syntax).  The underlying service now detects over a dozen additional
+    dangerous shell operations such as ``chmod 777``, ``sudo rm``, piping to
+    ``sh``/``bash``, and recursive permissive permissions.
+    """
+    return _safety_service.is_safe(command) is False
+
+
+# ---------------------------------------------------------------------------
+# Internal callback adapters
+# ---------------------------------------------------------------------------
+
+def _adapt_confirm_dangerous(
+    callback: Callable[[str], bool] | None,
+) -> Callable[[DangerousCommandWarning], bool] | None:
+    """Convert an old-style command-based callback into a warning-aware one.
+
+    When *callback* is ``None`` we return ``None`` so that the default
+    block-level behaviour of :class:`ShellSafetyService` takes over.
+
+    When a callable is provided, we wrap it so that the service receives a
+    :class:`~cleveragents.tui.shell_safety.DangerousCommandWarning`.  The
+    original callback (which expects only a command string) is invoked with
+    ``warning.command`` to preserve legacy behaviour.
+
+    Args:
+        callback: Legacy ``(command: str) -> bool`` confirmation callable or
+            ``None``.
+
+    Returns:
+        A warning-aware callback suitable for
+        :class:`ShellSafetyService`, or ``None``.
+    """
+    if callback is None:
+        return None
+
+    def _wrapper(warning: DangerousCommandWarning) -> bool:
+        return callback(warning.command)  # type: ignore[arg-type]
+
+    return _wrapper
+
+
+# ---------------------------------------------------------------------------
+# Public execution API
+# ---------------------------------------------------------------------------
 
 def run_shell_command(
     command: str,
     *,
+    warn_callback: Callable[[DangerousCommandWarning], bool] | None = None,
     confirm_dangerous: Callable[[str], bool] | None = None,
     timeout_seconds: int = 30,
 ) -> ShellResult:
-    """Execute shell command with basic safeguards.
+    """Execute shell command with comprehensive safety checks.
 
     Threat model:
     - Shell mode is a convenience feature for local development, not a sandbox.
     - We block obviously dangerous command patterns unless explicitly confirmed.
     - We apply a timeout to avoid hanging the UI event loop indefinitely.
+
+    Safety enforcement is handled by
+    :class:`~cleveragents.tui.shell_safety.ShellSafetyService` which checks
+    commands against 15 built-in regex-based patterns spanning four severity
+    levels (LOW, MEDIUM, HIGH, CRITICAL).
+
+    Args:
+        command: Shell command to execute.
+        warn_callback: New-style callback ``(warning) -> bool`` that receives
+            a :class:`~cleveragents.tui.shell_safety.DangerousCommandWarning`
+            when a dangerous pattern matches.  Return ``True`` to allow the
+            command or ``False`` to block it.  Ignored when *confirm_dangerous*
+            is provided (see note below).
+        confirm_dangerous: Legacy ``(command: str) -> bool`` confirmation
+            callback, preserved for backward compatibility with callers that
+            predate :class:`~cleveragents.tui.shell_safety.ShellSafetyService`.
+            When provided it is adapted so the service invokes it with just the
+            command string.  *warn_callback* takes precedence over this when
+            both are supplied.
+
+        .. note::
+           If neither callback is set and the command triggers a MEDIUM-level
+           or higher danger pattern, execution is blocked automatically (exit
+           code 1).
+
+    Returns:
+        A :class:`ShellResult` describing the outcome.
+
+    Raises:
+        subprocess.TimeoutExpired: Caught internally; yields a result with
+            ``exit_code=124``.
     """
     if not command.strip():
         return ShellResult(
@@ -55,17 +150,30 @@ def run_shell_command(
             stdout="",
             stderr="shell mode is disabled",
         )
-    if looks_dangerous(command):
-        confirmed = False
-        if confirm_dangerous is not None:
-            confirmed = confirm_dangerous(command)
-        if not confirmed:
+
+    # ------------------------------------------------------------------
+    # Safety gate — delegate to the full shell-safety service
+    # ------------------------------------------------------------------
+    effective_callback = warn_callback
+    if effective_callback is None and confirm_dangerous is not None:
+        effective_callback = _adapt_confirm_dangerous(confirm_dangerous)
+
+    result = _safety_service.check_command(command)
+    if not result.allowed:
+        # Dangerous command detected — check callback override.
+        if effective_callback is not None:
+            allowed = effective_callback(result.warning)  # type: ignore[arg-type]
+        else:
+            allowed = False
+
+        if not allowed:
             return ShellResult(
                 command=command,
                 exit_code=1,
                 stdout="",
                 stderr="blocked dangerous shell command",
             )
+
     try:
         proc = subprocess.run(
             command,