[codex] Strengthen IPython bash and edit prompts

src/rlm/prompt.py

@@ -47,9 +47,41 @@
     "benchmark, or API may have its own environment and normal interface. "
     "Evaluate external systems through their own interface, then use IPython "
     "to coordinate the process and analyze what comes back.\n\n"
-    "When running shell commands from IPython, use `%%bash` cells. Avoid "
-    "`!cmd` shell escapes for project commands so shell behavior is explicit "
-    "and multi-line commands share one shell context.\n\n"
+    "The IPython `code` argument is executed literally as one code cell. It "
+    "must contain only raw Python/IPython syntax, not chat-template wrappers, "
+    "native tool-call tags, JSON fields, or prose prefixes. Wrong: "
+    "`<arg_value>%%bash\\npytest -q` or `<arg_value>Now inspect the file`. "
+    "Correct: `%%bash\\npytest -q` or `print(path.read_text())`.\n\n"
+    "When running shell commands from IPython, use `%%bash` cells. This is a "
+    "hard syntax rule: if an IPython tool call contains `%%bash`, the `code` "
+    "argument must begin with exactly `%%bash\n` as its first seven "
+    "characters. No leading blank line, no leading spaces, no comments, no "
+    "imports, and no Python statements may appear before it. If a cell starts "
+    "with `# note\n%%bash`, IPython parses the whole cell as Python and the "
+    "shell command fails with `SyntaxError`.\n\n"
+    "Before every IPython tool call, silently check: does `code` contain "
+    "`%%bash`? If yes, the first line must be exactly `%%bash`. If not, "
+    "rewrite the tool call before sending it.\n\n"
+    "Correct:\n"
+    "```python\n"
+    "%%bash\n"
+    "# explanation belongs here, after the magic\n"
+    "cd /workspace/project && pytest -q\n"
+    "```\n"
+    "Wrong:\n"
+    "```python\n"
+    "# explanation before the magic breaks IPython\n"
+    "%%bash\n"
+    "cd /workspace/project && pytest -q\n"
+    "```\n"
+    "Wrong:\n"
+    "```python\n"
+    "\n"
+    "%%bash\n"
+    "cd /workspace/project && pytest -q\n"
+    "```\n\n"
+    "Avoid `!cmd` shell escapes for project commands so shell behavior is "
+    "explicit and multi-line commands share one shell context.\n\n"
     "Important: do not install dependencies into the IPython kernel just to "
     "make an external project import or run there. If a project import, test, "
     "script, CLI, or dependency check is needed, run it through that project's "
@@ -58,10 +90,61 @@
     "or the active project interpreter from the repo root. Treat failures from "
     "that native environment as the relevant result."
     "\n\n"
-    "Use Python for reading, searching, and editing files — it gives you "
-    "reusable variables you can slice, filter, and act on without re-reading. "
-    "Always assign read/search results to named variables so you can revisit "
-    "them later."
+    "Use Python for reading and searching files — it gives you reusable "
+    "variables you can slice, filter, and act on without re-reading. Always "
+    "assign read/search results to named variables so you can revisit them "
+    "later."
+)
+EDIT_SKILL_PROMPT = (
+    "For targeted modifications to existing files, you must use the "
+    "pre-imported `edit` skill from IPython instead of manual Python file "
+    "writes. Read and inspect files with normal Python, then make the change "
+    'with `await edit(path="relative/file.py", old_str=old, new_str=new)`, '
+    "where `old` and `new` are exact strings. Inline string literals are fine "
+    "when they are valid Python. The target `old_str` must appear exactly "
+    "once. Inside IPython `code`, never write native tool-call markup like "
+    "`<tool_call>`, `<arg_key>`, or `<arg_value>`; use Python "
+    'instead, for example: `old = "..."; new = "..."; await '
+    'edit(path="pkg/module.py", old_str=old, new_str=new)`. '
+    "The supported keyword arguments are `path`, `old_str`, `new_str`, "
+    "and optional `cwd`; do not use `file`, `old`, `new`, line numbers, "
+    "`after`, or `insert`. If an edit call fails because the string is not "
+    "found or not unique, inspect the file and retry with a smaller exact "
+    "snippet before falling back. Only use normal Python file I/O for creating "
+    "new files or for broad generated rewrites that cannot be expressed as "
+    "one or more exact replacements.\n\n"
+    "Before calling `edit`, make sure the Python string syntax is valid. Do "
+    'not wrap text that contains `"""` inside a `"""..."""` string; '
+    "that creates a SyntaxError before `edit` runs. If the target text "
+    "contains triple double quotes, use triple single quotes (`'''...'''`) or "
+    "assign `old`/`new` from inspected file slices. If the text contains both "
+    "triple quote styles or quoting becomes complex, build `old` and `new` as "
+    "variables first, then pass those variables to `await edit(...)`.\n\n"
+    "Good targeted edit pattern:\n"
+    "```python\n"
+    "from pathlib import Path\n"
+    'text = Path("pkg/module.py").read_text()\n'
+    'print(text[text.index("def broken") : text.index("def next_func")])\n'
+    "old = '''def broken():\n"
+    '    """Docstring with triple double quotes."""\n'
+    "    return False\n"
+    "'''\n"
+    "new = '''def broken():\n"
+    '    """Docstring with triple double quotes."""\n'
+    "    return True\n"
+    "'''\n"
+    'await edit(path="pkg/module.py", old_str=old, new_str=new)\n'
+    "```\n"
+    "Do not do this for targeted edits:\n"
+    "```python\n"
+    'await edit(path="pkg/module.py", old_str="""def broken():\n'
+    '    """Docstring closes the outer string early."""\n'
+    "    return False\n"
+    '""", new_str="""...\n'
+    '""")\n'
+    'Path("pkg/module.py").write_text(text.replace(old, new))\n'
+    'open("pkg/module.py", "w").write(new_contents)\n'
+    "```"
 )
 
 
@@ -108,6 +191,8 @@ def build_system_prompt(
             "Each skill is also available as a shell command by the same name: `<skill> ...`. "
             "Discover its CLI usage with `<skill> --help`."
         )
+        if "edit" in installed_skills:
+            skill_lines.append(EDIT_SKILL_PROMPT)
     if skill_lines:
         parts.extend(["", *skill_lines])
 

src/rlm/tools/ipython.py

@@ -4,11 +4,11 @@
 
 import copy
 import os
-from queue import Empty
 import re
 import sys
 import threading
 import time
+from queue import Empty
 from typing import TYPE_CHECKING, Any
 
 from rlm.tools.base import ToolContext, ToolOutcome
@@ -30,7 +30,10 @@
             "Use !command for shell commands (e.g. !ls -la, !cat file.py, !pip install foo). "
             "Use !python3 to run code with the project's own packages "
             "(e.g. !python3 -m pytest, !python3 -c 'import numpy'). "
-            "Use %%bash for multi-line shell scripts."
+            "Use %%bash for multi-line shell scripts. If code contains %%bash, "
+            "the code string must begin exactly with %%bash followed by a "
+            "newline. Do not put a blank line, spaces, comments, imports, or "
+            "Python code before %%bash."
         ),
         "parameters": {
             "type": "object",
@@ -65,8 +68,7 @@ def schema(self) -> dict[str, Any]:
         )
         schema = copy.deepcopy(IPYTHON_SCHEMA)
         schema["function"]["parameters"]["properties"]["timeout"]["description"] = (
-            "Optional timeout in seconds. "
-            f"Default: {timeout}s. Maximum: {IPYTHON_TIMEOUT_MAX_SECONDS}s."
+            f"Optional timeout in seconds. Default: {timeout}s. Maximum: {IPYTHON_TIMEOUT_MAX_SECONDS}s."
         )
         return schema
 

tests/test_prompt.py

@@ -5,6 +5,7 @@
 from dataclasses import dataclass
 
 from rlm.prompt import (
+    EDIT_SKILL_PROMPT,
     GIT_HISTORY_GUARD_PROMPT,
     IPYTHON_CONTROL_PROMPT,
     build_system_prompt,
@@ -16,11 +17,15 @@ class _Tool:
     name: str
 
 
-def _prompt(active_tools: list[_Tool]) -> str:
+def _prompt(
+    active_tools: list[_Tool],
+    *,
+    installed_skills: list[str] | None = None,
+) -> str:
     return build_system_prompt(
         "/repo",
         None,
-        [],
+        installed_skills or [],
         "/repo/.rlm/messages.jsonl",
         allow_recursion=False,
         active_tools=active_tools,
@@ -57,8 +62,33 @@ def test_ipython_control_prompt_included_for_ipython_tool():
     assert "long-lived notebook" in prompt
     assert "native runtime" in prompt
     assert "use `%%bash` cells" in prompt
+    assert "first seven characters" in prompt
+    assert "Before every IPython tool call" in prompt
     assert "do not install dependencies into the IPython kernel" in prompt
+    assert "Use Python for reading and searching files" in prompt
+    assert "Use Python for reading, searching, and editing files" not in prompt
 
 
 def test_ipython_control_prompt_omitted_without_ipython_tool():
     assert IPYTHON_CONTROL_PROMPT not in _prompt([_Tool("bash")])
+
+
+def test_edit_skill_prompt_included_when_edit_is_installed():
+    prompt = _prompt([_Tool("ipython")], installed_skills=["edit"])
+
+    assert EDIT_SKILL_PROMPT in prompt
+    assert 'await edit(path="relative/file.py", old_str=old, new_str=new)' in prompt
+    assert "pre-imported `edit` skill from IPython" in prompt
+    assert "`old_str` must appear exactly once" in prompt
+    assert "do not use `file`, `old`, `new`, line numbers" in prompt
+    assert "inspect the file and retry with a smaller exact snippet" in prompt
+    assert "Only use normal Python file I/O for creating new files" in prompt
+    assert 'Do not wrap text that contains `"""`' in prompt
+    assert "use triple single quotes" in prompt
+    assert "Good targeted edit pattern" in prompt
+
+
+def test_edit_skill_prompt_omitted_without_edit_skill():
+    prompt = _prompt([_Tool("ipython")], installed_skills=["search_docs"])
+
+    assert EDIT_SKILL_PROMPT not in prompt