PDF Files

PDF Files

Use this skill when the user asks to create, inspect, fill, verify, or digitally sign PDF files.

This includes Chinese requests such as 生成 PDF, 创建 PDF 表单, 填写 PDF, 生成静态已填写 PDF, PDF 数字签名, or 给 PDF 签名.

Route Boundaries

• For converting Office files such as .docx, .xlsx, or .pptx to PDF, use the bundled office-files skill and LibreOffice soffice --headless.
• For using a PDF as source material for a basic PowerPoint deck, extract the source content first and then use the bundled office-files skill.
• For scanned or image-only PDFs where text extraction fails, use the bundled ocr-and-documents skill before attempting structured PDF editing.
• Do not install PDF libraries at runtime. The image already provides the supported PDF stack.

Default Tool Choices

Use the image-provided Python environment:

/opt/hermes/.venv/bin/python

The image already installs:

• reportlab for creating PDF pages and AcroForm fields.
• pypdf for reading, filling, and inspecting PDF forms.
• PyMuPDF (fitz) for rendering pages and checking visible text.
• Poppler pdftoppm / pdfinfo for independent render and metadata checks.
• qpdf for PDF structure checks.
• pyHanko / pyhanko-cli for digital signatures.
• tzdata for pyHanko behavior in slim images.

Reusable helper scripts are bundled under this skill. Prefer them over writing one-off scripts for common form inspection, static filling, delivery verification, and test signing. See references/pdf-form-tools.md.

Working Defaults

• Start labeled AcroForms from a JSON spec plus scripts/create_labeled_acroform.py. Let the helper own two-column geometry, multiline note spacing, CJK labels, and safe margins.
• Do not spend time hand-tuning coordinates for common business forms. If a form needs unusual geometry, first explain why the helper does not fit, then write custom layout code.
• Treat field labels as part of the deliverable, not decoration. A form can have valid widgets and still be unusable if labels overlap, sit beside the wrong field, or touch multiline boxes.
• For every report table, calculate its available width from the page size minus both horizontal margins before calling Table. Require sum(col_widths) <= available_width; ReportLab does not shrink an oversized table automatically. If the columns do not fit, shorten or wrap cell content, reduce column widths/font size, split the table, or use a landscape page.
• After a failed verification, inspect the named failing check before making another attempt. Record whether the trigger was widget bounds, label placement, missing visible text, duplicate filled text, font rendering, or PDF structure.

Creation Path

For business PDFs, create the file programmatically with reportlab. For CJK content, register an embedded TrueType-outline TTF/TTC font before drawing Chinese text. The bundled helpers discover the image's embedded CJK font (or accept an explicit --font-path) and reject unembedded CID fonts such as STSong-Light:

import os
import sys

bundled_root = os.environ.get(
    "HERMES_BUNDLED_SKILLS", "/opt/oomol-hermes-agent/curated-skills"
)
sys.path.insert(0, f"{bundled_root}/productivity/pdf-files/scripts")
from pdf_fonts import register_embedded_cjk_font  # noqa: E402
from reportlab.pdfgen import canvas as canvas_module

font_name = register_embedded_cjk_font()
pdf = canvas_module.Canvas(output_path)
pdf.setFont(font_name, 11)

Every text object must use a CJK-capable font. Setting the canvas font is not enough for Platypus paragraphs, tables, headers, or footers:

from reportlab.lib.styles import ParagraphStyle
from reportlab.platypus import TableStyle

body_style = ParagraphStyle("BodyCJK", fontName=font_name, fontSize=10, leading=14)
table.setStyle(TableStyle([
    ("FONTNAME", (0, 0), (-1, -1), font_name),
]))

Write generated files under the user's requested output path, not under /opt/hermes.

AcroForm Path

For fillable PDF forms:

• When creating a normal labeled business form from scratch, prefer scripts/create_labeled_acroform.py with a JSON spec before writing custom coordinate code. The helper handles CJK labels, one- or two-column text fields, multiline note fields, and safe label/field spacing.
• Use stable, semantic field names such as supplier_contact, first_shipment_date, or purchase_owner.
• After creating a blank form, run scripts/inspect_pdf_widgets.py to verify widget bounds, margins, size, and overlap before delivering it.
• Also verify that each visible field label is adjacent to its corresponding widget. In two-column layouts, draw the right-column label near the right-column field rather than reusing the left-column label coordinate.
• Keep known business facts as normal visible page text.
• Use AcroForm text fields only for values the user expects to fill later.
• Keep widget rectangles inside the page, large enough for the expected value, and non-overlapping.
• Leave practical safe margins for form widgets: at least 36 pt from the left and right page edges, at least 72 pt from the bottom edge, and avoid placing editable fields in the footer area.
• For final delivery, prefer a static filled PDF generated from the same data or a visible text overlay on top of the form. Do not rely on pypdf flatten=True as the default delivery guarantee, especially for CJK content.
• Blank interactive widgets use ReportLab's standard form font; use scripts/fill_acroform_static.py for CJK values before delivery, which writes an embedded-font overlay and removes filled widgets.
• When using a visible overlay, do not leave the filled AcroForm text widget appearances visible underneath it. Remove or hide the filled text widget annotations and prune those fields from /AcroForm, or regenerate a true static filled PDF. Keep only fields that are still needed later, such as a signature field.

Typical field update approach:

from pypdf import PdfReader, PdfWriter

reader = PdfReader(input_path)
writer = PdfWriter()
writer.append(reader)
writer.update_page_form_field_values(
    writer.pages[0],
    {"supplier_contact": "value"},
    auto_regenerate=True,
)
with open(output_path, "wb") as output:
    writer.write(output)

If the user asks for broadly visible final content:

• For PDFs created by this skill, regenerate the final static PDF from the source business data and filled values.
• For an existing form, prefer scripts/fill_acroform_static.py. If you must implement the workflow manually, follow this deterministic static-fill path:
  1. Inspect field names and /Rect coordinates with pypdf.
  2. Create a same-size reportlab overlay page and draw every filled value at the matching field rectangle using a CJK-capable font.
  3. Merge the overlay onto the original page.
  4. Remove or hide the filled text widget annotations and prune those text fields from /AcroForm. Keep only fields that are still needed later, such as a signature field.
  5. Verify the final PDF with fitz visible text extraction, fitz rendering, qpdf --check, and Poppler rendering when available.
• Verify that important filled values do not appear twice at nearly identical coordinates. Duplicate field appearance plus overlay text is a delivery defect even when text extraction and structure checks pass.
• Prefer scripts/verify_pdf_delivery.py for final render, structure, visible text, and duplicate-nearby-text checks.
• If a first attempt fails verification, inspect and record the exact failing check before writing a new script. Do not repeatedly rewrite fill scripts without naming the reason, such as widget outside page bounds, missing visible text, duplicate overlay text, missing CJK font, or PDF structure error.

Signature Path

Use pyHanko for PDF signatures. When the user explicitly allows a test signature, generate a short-lived test certificate at runtime and discard it after use.

For test signatures, prefer scripts/sign_pdf_test_cert.py instead of writing a new certificate/signing script.

Sign only after the filled PDF has passed visual/static verification. The signing step should preserve the already verified content and add the requested signature field; it should not refill the form or add another visible text overlay.

Do not ask for, invent, or store production certificate material. Real private keys, PINs, HSM/PKCS#11 configuration, timestamp service credentials, and company certificates must come from runtime secrets, external services, or the user.

Verification

After creating or modifying a PDF:

• Confirm the output file exists and is non-empty.
• Read it back with pypdf.PdfReader.
• For created business PDFs, extract text with both pypdf and fitz when possible, and check key IDs, dates, and representative Chinese business text.
• Run scripts/verify_pdf_delivery.py with representative --contains markers. It checks every extracted text span for actual rendered ink and a minimum foreground/background contrast, and rejects unembedded CID fonts.
• For an image-backed OCR PDF only, pass --allow-invisible-ocr-text to permit its standard rendering-mode-3 searchable text layer. Do not use this option for generated business PDFs or to bypass a failed visible-text check.
• Render at least the first page with fitz and check that the pixmap is not blank.
• For PDFs containing tables, render every table page and visually inspect that table borders, headers, and cell text remain inside the page margins. A successful text extraction or qpdf --check does not establish this.
• Run qpdf --check when available.
• Run pdftoppm when available if visual delivery matters; it must render every page without font or syntax errors. PyMuPDF and Poppler are deliberately independent renderer checks.
• For blank forms, inspect reader.get_fields() and verify expected field names.
• For blank forms, inspect widget rectangles and ensure fields do not overlap or fall outside page bounds.
• For blank forms, render the page or inspect text coordinates and ensure field labels do not overlap each other, do not cover body text, and sit on the same row as the field they describe.
• For signed PDFs, inspect reader.get_fields() and confirm at least one field has /FT equal to /Sig.

Boundaries

• The default PDF contract covers basic PDFs, AcroForm fields, static filled PDF delivery, render/structure checks, and local certificate signing.
• XFA forms, complex PDF repair, long-term validation, OCSP/CRL, timestamp services, enterprise CA policy, and HSM integration are outside the default strong guarantee.
• Do not store user PDFs, certificates, keys, or generated artifacts under /opt/hermes.