Document Skills PDF

PDF workbench

PDF tasks fail when the agent mixes incompatible libraries or generates pages as screenshots. Route by job, use one tool per job, verify the artifact.

Step 0 - Check the toolchain

python -c "import reportlab, pypdf, fitz; print('reportlab', reportlab.Version, '| pypdf ok | pymupdf ok')"

• reportlab - creation (preferred: selectable text, controlled layout).
• pypdf - page-level transforms: split, merge, rotate, encrypt, form fields.
• pymupdf (fitz) - extraction (text with coordinates, tables, images), rasterization, page-level inspection.

Missing any -> say so and degrade to the ones present; do not substitute screenshot pipelines.

Step 1 - Pick exactly one tool for the job

Chaining tools on one artifact is fine (create -> inspect). Using two creation libraries on one file is not.

Step 2 - Shared rules

1. Text-first: a generated PDF must contain extractable text. Image-of-text is acceptable only when the user asked for a rasterized look or provides only images.
2. Page geometry is explicit: A4 = 595.27 x 841.89 pt, US Letter = 612 x 792 pt. Declare the target size and margins up front; re-check fit after generation.
3. Fonts: the standard 14 fonts cover only limited encodings; they do not automatically support arbitrary Unicode. Check that the selected face contains every requested character. If any glyph is unsupported - including CJK, Cyrillic, Arabic, Devanagari, or emoji - embed one or more licensed TTF/OTF fonts with the required coverage and use them for those runs. If no suitable embeddable font is available, report the limitation instead of emitting tofu.
4. Overflow is a defect: content that spills past the last page or the margin must be detected in postcheck and fixed (shrink, paginate, or cut), never shipped.
5. Write output to a new path; keep inputs untouched unless in-place was requested.
6. Accessibility claims require tagged output. ReportLab does not produce a tagged PDF/UA document, so do not describe ordinary ReportLab output as accessible or screen-reader-ready. When PDF/UA or assistive-technology compatibility is required, use a toolchain that can create and validate tags, reading order, structure roles, language, and alternative text. If that toolchain is unavailable, stop and report the limitation; offer an accessible DOCX or HTML deliverable instead of silently returning an untagged PDF.

Step 3 - Postcheck (mandatory)

import os
import fitz
import pypdf

output_path = "output.pdf"
expected_page_count = 1  # set this from the task; do not derive it from the output
password = os.environ.get("PDF_PASSWORD")
r = pypdf.PdfReader(output_path)
if r.is_encrypted:
    # Permission-encrypted PDFs commonly have an empty user password and open normally.
    if r.decrypt("") == 0:
        if not password:
            raise RuntimeError("set PDF_PASSWORD so the encrypted output can be postchecked")
        if r.decrypt(password) == 0:
            raise RuntimeError("PDF_PASSWORD could not decrypt the output")

def require(condition, message):
    # Mandatory verification must remain active under python -O.
    if not condition:
        raise ValueError(message)

page_count = len(r.pages)
require(
    page_count == expected_page_count,
    f"expected {expected_page_count} pages, got {page_count}",
)
render_doc = fitz.open(output_path)
if render_doc.needs_pass and render_doc.authenticate("") <= 0:
    if not password:
        raise RuntimeError("set PDF_PASSWORD so encrypted widgets can be rendered")
    if render_doc.authenticate(password) <= 0:
        raise RuntimeError("PDF_PASSWORD could not decrypt the output for widget rendering")
page_texts = {
    number: (page.extract_text() or "").strip()
    for number, page in enumerate(r.pages, start=1)
}

def normalized_box(box):
    x0, y0, x1, y1 = (float(value) for value in box)
    left, right = sorted((x0, x1))
    bottom, top = sorted((y0, y1))
    return left, bottom, right, top

def normalized_size(box):
    left, bottom, right, top = normalized_box(box)
    return right - left, top - bottom

def widget_count(page, rendered_page):
    # Interactive-only pages can be valid, but hidden metadata widgets are not visible content.
    non_viewable_flags = 1 | 2 | 32  # Invisible, Hidden, NoView annotation flags
    count = 0
    for ref in page.get("/Annots") or []:
        widget = ref.get_object()
        if widget.get("/Subtype") != "/Widget":
            continue
        flags = int(widget.get("/F", 0))
        rectangle = widget.get("/Rect")
        if flags & non_viewable_flags or rectangle is None:
            continue
        left, bottom, right, top = normalized_box(rectangle)
        crop_left, crop_bottom, crop_right, crop_top = normalized_box(page.cropbox)
        media_left, media_bottom, media_right, media_top = normalized_box(page.mediabox)
        visible_left = max(crop_left, media_left)
        visible_bottom = max(crop_bottom, media_bottom)
        visible_right = min(crop_right, media_right)
        visible_top = min(crop_top, media_top)
        intersects_visible_page = (
            min(right, visible_right) > max(left, visible_left)
            and min(top, visible_top) > max(bottom, visible_bottom)
        )
        if right <= left or top <= bottom or not intersects_visible_page:
            continue
        # Do not require /AP: viewers may synthesize it from field defaults. Instead,
        # render annotations on and off and require this widget region to change visibly.
        xref = getattr(ref, "idnum", None)
        if xref is None:
            continue
        try:
            rendered_widget = rendered_page.load_widget(xref)
            clip = (
                rendered_widget.rect * rendered_page.rotation_matrix
            ) & rendered_page.rect
            if clip.is_empty:
                continue
            with_widget = rendered_page.get_pixmap(
                matrix=fitz.Matrix(2, 2), clip=clip, alpha=False, annots=True,
            )
            without_widgets = rendered_page.get_pixmap(
                matrix=fitz.Matrix(2, 2), clip=clip, alpha=False, annots=False,
            )
        except (RuntimeError, ValueError):
            continue
        if with_widget.samples != without_widgets.samples:
            count += 1
    return count

intentionally_raster_only_pages = set()
missing_text_pages = [
    number for number, text in page_texts.items()
    if number not in intentionally_raster_only_pages
    and not text
    and widget_count(r.pages[number - 1], render_doc[number - 1]) == 0
]
require(not missing_text_pages, f"pages without extractable text: {missing_text_pages}")
# Add task-specific checks when exact copy matters, for example
# {1: ("Report title",), 2: ("Conclusion",)}. Per-page text presence is enforced above.
expected_strings_by_page = {}
for page_number, expected_strings in expected_strings_by_page.items():
    missing = [value for value in expected_strings if value not in page_texts[page_number]]
    require(not missing, f"page {page_number} is missing {missing}")
# Compare width/height to the exact size used at creation with a small point tolerance.
# ReportLab A4 is about (595.2756, 841.8898), not the rounded prose value (595.27, 841.89).
expected_page_size = (595.2756, 841.8898)  # replace for Letter or a task-specific size
page_size_tolerance = 0.5
page_sizes = [
    normalized_size(page.mediabox) for page in r.pages
]
size_mismatches = [
    (number, actual)
    for number, actual in enumerate(page_sizes, start=1)
    if any(abs(value - expected) > page_size_tolerance
           for value, expected in zip(actual, expected_page_size))
]
require(not size_mismatches, f"unexpected page sizes: {size_mismatches}")

# Overflow is a defect (shared rule 4): text, images, or vector drawings that run
# past the page box are clipped even though every check above still passes. Plain
# text extraction silently drops fully off-page text, so use an enlarged clip for
# text and inspect the placement boxes reported for every image and drawing.
overflow_doc = render_doc
overflow_pages = []
for page in overflow_doc:
    # These APIs report unrotated coordinates even when /Rotate is 90/270;
    # page.rect uses rotated dimensions. Compare against an unrotated crop-box
    # extent so valid high-y portrait content is not flagged on a rotated page.
    crop_left, crop_bottom, crop_right, crop_top = normalized_box(page.cropbox)
    page_box = fitz.Rect(0, 0, crop_right - crop_left, crop_top - crop_bottom)
    text_rects = [
        fitz.Rect(block[:4])
        for block in page.get_text("blocks", clip=fitz.INFINITE_RECT())
        if block[6] == 0
    ]
    image_rects = [
        fitz.Rect(0, 0, 1, 1) * fitz.Matrix(*image["transform"])
        for image in page.get_image_info()
    ]
    drawing_rects = []
    for drawing in page.get_drawings():
        rect = fitz.Rect(drawing["rect"])
        # Path rectangles exclude stroke thickness. Expand stroked paths so a
        # line centered on the page edge cannot hide half its width off-page.
        stroke_pad = (
            float(drawing.get("width") or 0) / 2
            if "s" in drawing.get("type", "") else 0
        )
        drawing_rects.append(fitz.Rect(
            rect.x0 - stroke_pad, rect.y0 - stroke_pad,
            rect.x1 + stroke_pad, rect.y1 + stroke_pad,
        ))
    beyond_box = any(
        rect.x0 < page_box.x0 - 0.5 or rect.y0 < page_box.y0 - 0.5
        or rect.x1 > page_box.x1 + 0.5 or rect.y1 > page_box.y1 + 0.5
        for rect in text_rects + image_rects + drawing_rects
    )
    if beyond_box:
        overflow_pages.append(page.number + 1)
overflow_doc.close()
require(not overflow_pages, f"content extends past the page box on pages: {overflow_pages}")
# The page box is the hard bound. When the task declares specific margins,
# additionally check key blocks against them (or render and inspect visually) -
# content inside the box but past a declared margin is a softer, task-specific
# defect to report.

Confirm: page count matches the request; every page except those explicitly listed in intentionally_raster_only_pages has extractable text or at least one form widget (a pure interactive page); each requested key string is listed in expected_strings_by_page and extracts on the correct page; every value in page_sizes is the declared size within page_size_tolerance; and no page's text, image placements, or vector drawings extend past the page box (the overflow check). Report all five. For pixel-sensitive work, render every applicable page with PyMuPDF at 100 dpi and check that each render is non-blank (mean pixel value).