Interfaces d’outils pour coder (Read, Edit, Bash)

De quels outils un agent de code a-t-il besoin ?

read_file

def read_file(path: str, start_line: int = 0, end_line: int | None = None) -> str:
    with open(path) as f:
        lines = f.readlines()
    return ''.join(lines[start_line:end_line])

# --- demo ---
with open('demo_source.py', 'w', encoding='utf-8') as f:
    f.write('line0\nline1\nline2\nline3\nline4\n')

snippet = read_file('demo_source.py', start_line=1, end_line=3)
print('Lines 1-2 of demo_source.py:')
print(snippet)

Les lectures par intervalle de lignes économisent des jetons

edit_file (remplacement de chaîne)

def edit_file(path: str, old_string: str, new_string: str):
    contents = read_file(path)
    occurrences = contents.count(old_string)
    if occurrences == 0:
        raise ValueError('old_string not found')
    if occurrences > 1:
        raise ValueError('old_string is not unique')
    contents = contents.replace(old_string, new_string)
    with open(path, 'w') as f:
        f.write(contents)

Pourquoi remplacer des chaînes plutôt qu’utiliser des différences

write_file

def write_file(path: str, content: str):
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, 'w') as f:
        f.write(content)

bash

def bash(command: str, timeout: int = 60) -> dict:
    import subprocess
    try:
        result = subprocess.run(command, shell=True, capture_output=True, text=True, timeout=timeout)
        return {'stdout': result.stdout[-5000:], 'stderr': result.stderr[-5000:], 'exit_code': result.returncode}
    except subprocess.TimeoutExpired:
        return {'error': 'timeout', 'exit_code': -1}

# --- demo ---
result = bash("echo hello from the sandboxed shell")
print(f"exit_code={result['exit_code']}")
print(f"stdout={result['stdout'].strip()!r}")

Isoler l’outil bash

tronquer les sorties longues

stderr = 'error line\n' * 1000

if len(stderr) > 5000:
    stderr = '...[truncated]...' + stderr[-5000:]

print(f'Truncated stderr length: {len(stderr)}')

search / grep

def grep(pattern: str, path: str = '.', file_glob: str = '**/*') -> list[str]:
    matches = []
    for f in glob(file_glob, recursive=True):
        for i, line in enumerate(open(f)):
            if re.search(pattern, line):
                matches.append(f'{f}:{i+1}: {line.rstrip()}')
    return matches[:50]

list_files

def list_files(path: str = '.') -> list[str]:
    return [p for p in glob(f'{path}/**', recursive=True) if os.path.isfile(p)][:200]

Facultatif : outils AST

def find_function(name: str, path: str):
    tree = ast.parse(open(path).read())
    for node in ast.walk(tree):
        if isinstance(node, ast.FunctionDef) and node.name == name:
            return ast.unparse(node)

Les descriptions d’outils sont essentielles ici

description='Edit a file by replacing a unique string. The old_string MUST appear exactly once in the file. Reads the file first to validate. Use this for any code change in an existing file.'

C’est tout l’ensemble d’outils

Modèle de modification

Récapitulatif