Agent File Processor Subagent

File Processor Sub-Agent

Quand utiliser ce skill

Déléguer à ce sous-agent quand l'agent parent doit traiter des fichiers de formats variés sans polluer son contexte principal : ingestion de données, conversion de format, extraction de contenu (dont OCR), génération de rapports, traitement par lot d'un répertoire entier.

Ne pas utiliser si le fichier est < 5 Ko et que le format est trivial (JSON, YAML simple) : l'agent parent peut le lire directement.

Workflow en étapes

1. Validation des inputs

Avant toute opération, vérifier :

import os, pathlib

def validate_input(inp: dict) -> list[str]:
    errors = []
    fp = inp.get("file_path", "")
    if not fp:
        errors.append("file_path manquant")
    elif not pathlib.Path(fp).exists() and inp.get("operation") != "generate":
        errors.append(f"Fichier introuvable : {fp}")
    if pathlib.Path(fp).stat().st_size > 500 * 1024 * 1024:
        errors.append("Fichier > 500 Mo : utiliser batch + chunk_size")
    if inp.get("operation") not in ("read","transform","generate","convert","batch","extract"):
        errors.append("operation invalide")
    return errors

Retourner un output_schema avec errors rempli si la validation échoue — ne jamais lever d'exception non catchée vers l'agent parent.

2. Détection du type de fichier

Ne pas faire confiance à l'extension seule.

import magic  # python-magic
import chardet

def detect_file_type(path: str) -> tuple[str, str]:
    mime = magic.from_file(path, mime=True)  # ex: "application/pdf"
    encoding = "binary"
    if mime.startswith("text/"):
        with open(path, "rb") as f:
            raw = f.read(32_768)
        encoding = chardet.detect(raw)["encoding"] or "utf-8"
    return mime, encoding

Critères de sélection du parser :

3. Parsing par format

PDF (texte + tableaux) :

import pdfplumber

with pdfplumber.open(file_path) as pdf:
    text = "\n".join(p.extract_text() or "" for p in pdf.pages)
    tables = [p.extract_tables() for p in pdf.pages]

PDF scanné → OCR :

import fitz  # PyMuPDF
import pytesseract
from PIL import Image
import io

doc = fitz.open(file_path)
for page in doc:
    pix = page.get_pixmap(dpi=300)
    img = Image.open(io.BytesIO(pix.tobytes("png")))
    text = pytesseract.image_to_string(img, lang="fra+eng")

Excel avec multi-feuilles :

import openpyxl

wb = openpyxl.load_workbook(file_path, read_only=True, data_only=True)
for sheet_name in wb.sheetnames:
    ws = wb[sheet_name]
    rows = list(ws.values)

CSV gros fichier (streaming) :

import pandas as pd

for chunk in pd.read_csv(file_path, chunksize=10_000, encoding=encoding,
                          sep=None, engine="python"):  # sep auto-détecté
    process(chunk)

XML avec namespace :

from lxml import etree

tree = etree.parse(file_path)
ns = {"ns": "http://example.com/schema"}
nodes = tree.xpath("//ns:Record", namespaces=ns)

4. Transformation des données

import pandas as pd

def transform(df: pd.DataFrame, params: dict) -> pd.DataFrame:
    if cols := params.get("columns"):
        df = df[cols]
    if filters := params.get("filters"):
        for col, val in filters.items():
            df = df[df[col] == val]
    for t in params.get("transformations", []):
        if t["type"] == "rename":
            df = df.rename(columns=t["mapping"])
        elif t["type"] == "groupby":
            df = df.groupby(t["by"]).agg(t["agg"]).reset_index()
        elif t["type"] == "fillna":
            df = df.fillna(t["value"])
        elif t["type"] == "deduplicate":
            df = df.drop_duplicates(subset=t.get("subset"))
    return df

Pour JSON complexe, préférer jq via subprocess :

jq '.data[] | select(.status == "active") | {id, name}' input.json > output.json

5. Génération de fichiers

Rapport PDF avec ReportLab :

from reportlab.lib.pagesizes import A4
from reportlab.platypus import SimpleDocTemplate, Table, TableStyle

doc = SimpleDocTemplate(output_path, pagesize=A4)
data = [["Colonne A", "Colonne B"]] + rows
table = Table(data)
doc.build([table])

Excel avec styles :

from openpyxl.styles import Font, PatternFill

ws["A1"].font = Font(bold=True)
ws["A1"].fill = PatternFill("solid", fgColor="4472C4")
wb.save(output_path)

HTML via Jinja2 :

from jinja2 import Environment, FileSystemLoader

env = Environment(loader=FileSystemLoader("templates/"))
html = env.get_template("report.html.j2").render(data=rows, title="Rapport")
Path(output_path).write_text(html, encoding="utf-8")

6. Batch processing

from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path

def batch_process(directory: str, pattern: str, workers: int = 4) -> list[dict]:
    files = list(Path(directory).glob(pattern))
    results, errors = [], []
    with ThreadPoolExecutor(max_workers=workers) as pool:
        futures = {pool.submit(process_single, f): f for f in files}
        for future in as_completed(futures):
            try:
                results.append(future.result())
            except Exception as e:
                errors.append({"file": str(futures[future]), "error": str(e)})
    return results, errors

Règle : l'échec d'un fichier ne bloque jamais les autres. Logger chaque erreur avec nom de fichier + nature.

7. Output schema

{
  "result": dict | list | None,   # Données extraites, ou None si fichier généré
  "output_path": str | None,      # Chemin fichier généré
  "stats": {
    "rows_processed": int,
    "pages_processed": int,
    "files_processed": int,
    "bytes_read": int,
    "bytes_written": int,
    "detected_type": str,
    "detected_encoding": str
  },
  "errors": [{"file": str, "page": int, "error": str}],
  "warnings": ["string"],
  "execution_time_s": float
}

Garde-fous et anti-patterns

Librairies recommandées (2026)

pdfplumber>=0.11.0
PyMuPDF>=1.24.0
openpyxl>=3.2.0
pandas>=2.2.0
python-docx>=1.1.0
python-magic>=0.4.27
chardet>=5.2.0
pytesseract>=0.3.10
Pillow>=10.3.0
camelot-py[cv]>=0.11.0
reportlab>=4.2.0
jinja2>=3.1.4
PyYAML>=6.0.1
lxml>=5.2.0
orjson>=3.10.0

Installation OCR système requis :

# Ubuntu/Debian
apt-get install tesseract-ocr tesseract-ocr-fra poppler-utils libmagic1

# Windows
choco install tesseract poppler

Exemple d'intégration agent parent

# L'agent parent appelle le sous-agent via tool call
result = file_processor_subagent.run({
    "file_path": "/data/invoices/",
    "operation": "batch",
    "batch_pattern": "*.pdf",
    "params": {"ocr_language": "fra", "columns": ["date", "montant", "fournisseur"]},
    "output_format": "excel",
    "output_path": "/data/output/invoices_summary.xlsx",
    "parallel_workers": 4
})

if result["errors"]:
    # Signaler à l'agent parent les fichiers en échec sans bloquer
    log_errors(result["errors"])

df_summary = pd.read_excel(result["output_path"])

Communication Rules — MANDATORY

• Ultra-concise. No filler, no preamble, no pleasantries.
• Never say "happy to help", "sure!", "great question", "let me", or similar.
• Tool first, talk second. Act before explaining.
• Result first. Lead with outcome, not process.
• Stop when done. No summary, no recap, no trailing commentary.
• No politeness wrappers. Direct and blunt.
• Minimum words. If one word works, do not use ten.
• No unsolicited explanations.
• No emoji unless asked.