feat: built-in bash skill (bash -c, shell parity with a native tool)

src/rlm/skills/bash.py

@@ -0,0 +1,43 @@
+"""Built-in ``bash`` skill — run a shell command from the REPL.
+
+Enabled via ``RLM_SKILLS``; pre-imported into the IPython kernel so the agent calls
+``await bash(command="...")`` (or ``await bash("...")``). One command per call, fresh
+subshell, like a plain bash-agent tool but surfaced as a skill.
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+
+async def run(command: str, timeout: int = 120) -> str:
+    """Run a shell command and return its combined output.
+
+    Args:
+        command: The shell command to run.
+        timeout: Seconds before the command is killed.
+
+    Returns:
+        stdout and stderr of the command (with the exit code when nonzero).
+    """
+    proc = await asyncio.create_subprocess_exec(
+        "bash",
+        "-c",
+        command,
+        stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.PIPE,
+    )
+    try:
+        out_b, err_b = await asyncio.wait_for(proc.communicate(), timeout=timeout)
+    except asyncio.TimeoutError:
+        proc.kill()
+        return f"Error: command timed out after {timeout}s"
+    out = out_b.decode(errors="replace")
+    err = err_b.decode(errors="replace")
+    result = out + (("\n" + err) if err else "")
+    if proc.returncode != 0:
+        result += f"\n[exit code {proc.returncode}]"
+    result = result.strip() or "(no output)"
+    if len(result) > 30_000:
+        result = result[:30_000] + f"\n... [truncated {len(result) - 30_000} chars]"
+    return result

tests/test_builtin_skills.py

@@ -16,6 +16,7 @@
 )
 from rlm.engine import RLMEngine
 from rlm.skills import available_builtin_skills, enable_builtin_skills
+from rlm.skills.bash import run as bash
 from rlm.skills.edit import run as edit
 from rlm.skills.search import format_results
 from rlm.skills.search import run as run_search
@@ -155,3 +156,23 @@ async def post(self, url, **kwargs):
     assert secret not in source
     meta = json.loads((session.dir / "meta.json").read_text())
     assert meta["programmatic_tool_call_stats"]["by_tool_python"] == {"search": 1}
+
+
+async def test_bash_returns_output():
+    assert await bash("echo hello") == "hello"
+
+
+async def test_bash_runs_real_bash_not_sh():
+    # process substitution is a bashism that /bin/sh (dash) rejects
+    out = await bash("cat <(echo bashism-works)")
+    assert "bashism-works" in out
+
+
+async def test_bash_nonzero_exit_reported():
+    out = await bash("exit 3")
+    assert "[exit code 3]" in out
+
+
+async def test_bash_combines_stderr():
+    out = await bash("echo out && echo err >&2")
+    assert "out" in out and "err" in out