feat(skills): add shell command execution in frontmatter for dynamic context

openhands-sdk/openhands/sdk/context/skills/execute.py

@@ -0,0 +1,113 @@
+"""Command execution for dynamic skill context injection.
+
+Supports inline !`command` syntax in skill content. Commands are executed
+at render time and their output replaces the placeholder.
+
+Safety rules:
+- Fenced (```) and inline (`) code blocks are preserved, never executed.
+- An unclosed fenced block (odd number of ```) extends to EOF, protecting
+  any trailing content from accidental execution.
+- Use \\!`cmd` to produce the literal text !`cmd` without execution.
+
+**Security Warning**: Commands are executed via shell with full process
+privileges. Only use with trusted skill sources.
+"""
+
+from __future__ import annotations
+
+import re
+import subprocess
+from pathlib import Path
+from typing import Final
+
+from openhands.sdk.logger import get_logger
+
+
+logger = get_logger(__name__)
+
+# 50KB per command output
+MAX_OUTPUT_SIZE: Final[int] = 50 * 1024
+
+# Default timeout per command in seconds
+DEFAULT_TIMEOUT: Final[float] = 10.0
+
+# Single-pass pattern: matches fenced code blocks, escaped commands, inline code,
+# or !`command`.  Order matters – earlier alternatives take priority.
+#
+# 1. Fenced blocks (``` ... ```).  An *unclosed* fence (odd number of ```)
+#    matches through to the end of the string so that content after the last
+#    opening ``` is never accidentally executed.
+# 2. Escaped commands (\!`...`) – the backslash is stripped and the rest is
+#    kept as a literal !`...` so authors can document the syntax itself.
+# 3. Inline code (`...`) not preceded by `!`.
+# 4. Executable commands (!`...`).
+_COMBINED_PATTERN: re.Pattern[str] = re.compile(
+    r"(?P<fenced>```[\s\S]*?(?:```|$))"  # fenced code block (unclosed → EOF)
+    r"|(?P<escaped>\\!`[^`]+`)"  # escaped \!`command` → literal
+    r"|(?P<inline>(?<!!)`[^`]+`)"  # inline code (not preceded by !)
+    r"|!`(?P<cmd>[^`]+)`"  # !`command`
+)
+
+
+def _execute_inline_command(
+    command: str,
+    working_dir: Path | None = None,
+    timeout: float = DEFAULT_TIMEOUT,
+) -> str:
+    """Execute a single inline shell command and return its output."""
+    cwd = str(working_dir) if working_dir else None
+    try:
+        result = subprocess.run(
+            command,
+            shell=True,
+            cwd=cwd,
+            capture_output=True,
+            text=True,
+            timeout=timeout,
+        )
+        if result.returncode != 0:
+            message = (
+                f"Command `{command}` exited with "
+                f"code {result.returncode}: {result.stderr}"
+            )
+            logger.warning("Skill command failed: %s", message)
+            return f"[Error: {message}]"
+
+        output = result.stdout.strip()
+        if len(output.encode()) > MAX_OUTPUT_SIZE:
+            output = output.encode()[:MAX_OUTPUT_SIZE].decode("utf-8", errors="ignore")
+            output += "\n... [output truncated]"
+        return output
+
+    except subprocess.TimeoutExpired:
+        message = f"Command `{command}` timed out after {timeout}s"
+        logger.warning("Skill command failed: %s", message)
+        return f"[Error: {message}]"
+    except Exception as e:
+        message = f"Failed to execute command `{command}`: {e}"
+        logger.warning("Skill command failed: %s", message)
+        return f"[Error: {message}]"
+
+
+def render_content_with_commands(
+    content: str,
+    working_dir: Path | None = None,
+    timeout: float = DEFAULT_TIMEOUT,
+) -> str:
+    """Execute inline !`command` patterns in content and replace with output.
+
+    Code blocks (fenced ``` and inline `) are preserved and not executed.
+    Unclosed fenced blocks (odd number of ```) are treated as extending to
+    EOF so that trailing content is never accidentally executed.
+    Use \\!`cmd` to produce the literal text !`cmd` without execution.
+    """
+
+    def _replace(match: re.Match[str]) -> str:
+        if match.group("fenced") or match.group("inline"):
+            return match.group(0)
+        if match.group("escaped"):
+            # Strip leading backslash: \!`cmd` → !`cmd`
+            return match.group("escaped")[1:]
+        return _execute_inline_command(match.group("cmd"), working_dir, timeout)
+
+    return _COMBINED_PATTERN.sub(_replace, content)

openhands-sdk/openhands/sdk/context/skills/skill.py

@@ -11,6 +11,7 @@
 from pydantic import BaseModel, Field, field_validator, model_validator
 
 from openhands.sdk.context.skills.exceptions import SkillError, SkillValidationError
+from openhands.sdk.context.skills.execute import render_content_with_commands
 from openhands.sdk.context.skills.trigger import (
     KeywordTrigger,
     TaskTrigger,
@@ -625,6 +626,25 @@ def to_skill_info(self) -> SkillInfo:
             is_agentskills_format=self.is_agentskills_format,
         )
 
+    def render_content(
+        self,
+        working_dir: Path | None = None,
+    ) -> str:
+        """Render skill content, executing inline !`command` blocks.
+
+        Inline !`command` patterns in the content are executed and
+        replaced with their stdout output. Code blocks (fenced and
+        inline) are preserved. Unclosed fenced blocks are treated as
+        extending to EOF. Use \\!`cmd` to produce literal !`cmd` text.
+
+        Args:
+            working_dir: Directory to run commands in.
+
+        Returns:
+            Processed content with command outputs substituted.
+        """
+        return render_content_with_commands(self.content, working_dir)
+
 
 def load_skills_from_dir(
     skill_dir: str | Path,

tests/sdk/context/skill/test_skill_commands.py

@@ -0,0 +1,210 @@
+"""Tests for inline !`command` execution in skill content.
+
+The !`command` syntax lets skill authors embed dynamic shell output in
+markdown.  These tests verify:
+
+  - Basic execution: !`echo hello` → hello
+  - Error / timeout handling
+  - Output truncation for large outputs
+  - Code-block safety: fenced (```) and inline (`) blocks are never executed
+  - Unclosed fenced blocks: an odd number of ``` delimiters must not leak
+    commands that follow the last unclosed fence
+  - Escape hatch: \\!`cmd` is preserved as the literal text !`cmd`
+  - Integration with the Skill model (load + render)
+"""
+
+from pathlib import Path
+
+import pytest
+
+from openhands.sdk.context.skills import Skill
+from openhands.sdk.context.skills.execute import (
+    MAX_OUTPUT_SIZE,
+    _execute_inline_command,
+    render_content_with_commands,
+)
+
+
+# ---------------------------------------------------------------------------
+# Low-level: _execute_inline_command
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("command", "timeout", "check_fn"),
+    [
+        pytest.param("echo hello", 10.0, lambda r: r == "hello", id="success"),
+        pytest.param(
+            "printf 'line1\nline2\nline3'",
+            10.0,
+            lambda r: r == "line1\nline2\nline3",
+            id="multiline_output",
+        ),
+        pytest.param("exit 1", 10.0, lambda r: "[Error:" in r, id="failure"),
+        pytest.param("sleep 5", 0.1, lambda r: "timed out" in r, id="timeout"),
+    ],
+)
+def test_execute_inline_command(command, timeout, check_fn):
+    assert check_fn(_execute_inline_command(command, timeout=timeout))
+
+
+def test_execute_inline_command_respects_working_dir(tmp_path: Path):
+    result = _execute_inline_command("pwd", working_dir=tmp_path)
+    assert result == str(tmp_path.resolve())
+
+
+def test_execute_inline_command_truncates_large_output():
+    size = MAX_OUTPUT_SIZE + 100
+    result = _execute_inline_command(f"python3 -c \"print('x' * {size})\"")
+    assert result.endswith("... [output truncated]")
+    assert len(result.encode()) <= MAX_OUTPUT_SIZE + 50  # small overhead ok
+
+
+# ---------------------------------------------------------------------------
+# Rendering: basic command substitution
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("content", "expected"),
+    [
+        pytest.param("Hello world", "Hello world", id="plain_text_unchanged"),
+        pytest.param("Branch: !`echo main`", "Branch: main", id="single_command"),
+        pytest.param(
+            "A: !`echo one` B: !`echo two`", "A: one B: two", id="multiple_commands"
+        ),
+        pytest.param("!``", "!``", id="empty_backticks_ignored"),
+    ],
+)
+def test_render_basic(content, expected):
+    assert render_content_with_commands(content) == expected
+
+
+# ---------------------------------------------------------------------------
+# Rendering: code blocks are never executed
+# ---------------------------------------------------------------------------
+
+
+def test_render_preserves_inline_code():
+    """Regular `code` spans are left alone."""
+    content = "Use `git status` to check"
+    assert render_content_with_commands(content) == content
+
+
+def test_render_preserves_fenced_block():
+    """Commands inside ``` fences are not executed."""
+    content = "Real: !`echo yes`\n```\n!`echo no`\n```"
+    result = render_content_with_commands(content)
+    assert "yes" in result
+    assert "!`echo no`" in result
+
+
+def test_render_inline_code_next_to_command():
+    """`code` immediately followed by a real !`cmd` — both handled correctly."""
+    content = "Run `git status` then !`echo done`"
+    result = render_content_with_commands(content)
+    assert "`git status`" in result
+    assert "done" in result
+
+
+# ---------------------------------------------------------------------------
+# Rendering: unclosed fenced blocks
+#
+# When a fenced block is opened but never closed (odd number of ```),
+# everything after the opening ``` must be treated as inside the fence —
+# no commands should be executed there.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("content", "executed", "preserved"),
+    [
+        pytest.param(
+            "```\nblock1\n```\n!`echo mid`\n```\n!`echo sneaky`\n",
+            "mid",
+            "!`echo sneaky`",
+            id="odd_fences_protects_trailing_command",
+        ),
+        pytest.param(
+            "```\n!`echo nope`\n",
+            None,
+            "!`echo nope`",
+            id="single_unclosed_fence",
+        ),
+    ],
+)
+def test_render_unclosed_fenced_blocks(content, executed, preserved):
+    result = render_content_with_commands(content)
+    if executed is not None:
+        assert executed in result
+    assert preserved in result
+
+
+def test_render_properly_closed_fences():
+    content = "```\nblock1\n```\n!`echo between`\n```\nblock2\n```"
+    result = render_content_with_commands(content)
+    assert "between" in result
+    assert "!`echo between`" not in result
+
+
+# ---------------------------------------------------------------------------
+# Rendering: escape hatch — \!`cmd` produces the literal text !`cmd`
+#
+# This lets skill authors document the !`...` syntax itself, or show
+# examples of commands without them being run at render time.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("content", "expected_literal", "expected_executed"),
+    [
+        pytest.param(
+            r"\!`echo hello`",
+            "!`echo hello`",
+            None,
+            id="escaped_becomes_literal",
+        ),
+        pytest.param(
+            r"Docs: \!`echo no` Real: !`echo yes`",
+            "!`echo no`",
+            "yes",
+            id="escaped_and_real_coexist",
+        ),
+    ],
+)
+def test_render_escaped_commands(content, expected_literal, expected_executed):
+    result = render_content_with_commands(content)
+    assert expected_literal in result
+    if expected_executed is not None:
+        assert expected_executed in result
+
+
+def test_render_escape_inside_fenced_block_untouched():
+    r"""\\!`cmd` inside a fenced block is left completely as-is."""
+    content = "```\n\\!`echo hi`\n```"
+    result = render_content_with_commands(content)
+    assert result == content
+
+
+# ---------------------------------------------------------------------------
+# Integration: Skill.render_content
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("content", "expected"),
+    [
+        pytest.param("Plain text", "Plain text", id="no_commands"),
+        pytest.param("Out: !`echo hi`", "Out: hi", id="with_command"),
+    ],
+)
+def test_skill_render_content(content, expected):
+    assert Skill(name="t", content=content).render_content() == expected
+
+
+def test_skill_load_and_render(tmp_path: Path):
+    skill_md = tmp_path / "test-skill" / "SKILL.md"
+    skill_md.parent.mkdir()
+    skill_md.write_text("---\nname: test-skill\n---\nBranch: !`echo main`\n")
+    skill = Skill.load(skill_md)
+    assert skill.render_content() == "Branch: main"