Datasets:

micro1-inc
/

ContractBench

"""LLM-facing view of a .docx contract.
A `DocumentView` is a frozen snapshot of a document for the LLM to reason
about: it carries paragraph IDs the model can cite, a Markdown serialization
that fits in a prompt, and enumerations of any pre-existing comments and
tracked changes (so the model can reply on threads and accept prior edits).
The view is built fresh each time the document is loaded; paragraph IDs are
assigned by load-order (`p-000`, `p-001`, ...) and are stable for the duration
of one redlining session. After we apply edits and save, reloading produces a
new view — the LLM must re-read in that case.
"""

from __future__ import annotations

from dataclasses import dataclass, field
from pathlib import Path
from typing import Iterable

from docx import Document
from docx.document import Document as DocxDocument
from docx.text.paragraph import Paragraph
from lxml import etree

from .ooxml_utils import NSMAP, qn


_PARAGRAPH_ID_FMT = "p-{:03d}"


def _iter_body_paragraphs(doc: DocxDocument) -> "list[Paragraph]":
    """Every paragraph in the document body **in document order, including
    paragraphs inside tables** (and nested tables).
    python-docx's `Document.paragraphs` only returns body-level `<w:p>` and
    silently skips `<w:tbl>//<w:p>`, so any edit inside a pricing schedule,
    SLA grid, or credit table was invisible to the grader (see the
    "table blindness" QC finding). We walk the body XML ourselves: yield each
    `<w:p>` where it sits, descending rows → cells → (recursively) nested
    tables, so paragraph IDs stay in reading order.
    """
    body = doc.element.body
    parent = doc._body  # a valid python-docx parent (resolves .part for styles)

    def walk(el):
        for child in el.iterchildren():
            if child.tag == qn("w:p"):
                yield Paragraph(child, parent)
            elif child.tag == qn("w:tbl"):
                for row in child.iterchildren(qn("w:tr")):
                    for cell in row.iterchildren(qn("w:tc")):
                        yield from walk(cell)

    return list(walk(body))


@dataclass
class ParagraphInfo:
    """A paragraph in the document, with the metadata the LLM and engine need."""

    id: str
    index: int
    text: str
    style: str
    is_empty: bool
    has_revisions: bool
    has_comments: bool
    # Word's auto-numbering is driven by w:numPr; we record the abstract
    # level (0 = top-level) when present so the serializer can indent.
    list_level: int | None = None


@dataclass
class ExistingComment:
    """A pre-existing comment in the document the LLM may need to reply to."""

    id: int
    author: str
    initials: str
    text: str
    paragraph_ids: list[str] = field(default_factory=list)
    parent_id: int | None = None  # if this comment is itself a reply
    # Set by annotated_view.tag_this_turn(): True if this comment is net-new by
    # the represented party THIS turn, False if carried-in (counterparty / prior
    # turn), None if turn attribution was not computed (no baseline). Never read
    # from the model-writable docx — derived from the parsed baseline diff only.
    is_this_turn: bool | None = None


@dataclass
class ExistingRevision:
    """A pre-existing tracked change the LLM may need to accept/reject."""

    id: int
    type: str  # "ins" | "del" | "moveFrom" | "moveTo"
    author: str
    text: str
    paragraph_ids: list[str] = field(default_factory=list)
    # w:date on the revision element (ISO string) if present, else None.
    date: str | None = None
    # Set by annotated_view.tag_this_turn(): True if this revision is net-new by
    # the represented party THIS turn, False if carried-in (counterparty / prior
    # turn), None if turn attribution was not computed (no baseline). Derived only
    # from the parsed OOXML/baseline diff — never model-controllable.
    is_this_turn: bool | None = None


@dataclass
class DocumentView:
    docx_path: Path
    paragraphs: list[ParagraphInfo]
    comments: list[ExistingComment]
    revisions: list[ExistingRevision]
    # B1 (Class B): raw model-controlled w:id -> grader-canonical sequential id
    # (the enumeration index) for comments and revisions. The annotated (judge) view
    # resolves every rendered id through these maps, so no raw, model-forgeable w:id
    # ever reaches the judge. Populated by load(); default empty for direct
    # construction / backward compatibility.
    comment_id_map: dict[int, int] = field(default_factory=dict, repr=False)
    revision_id_map: dict[int, int] = field(default_factory=dict, repr=False)
    # Map paragraph ID → underlying python-docx Paragraph object (not serialized
    # for the LLM; used internally when applying edits).
    _paragraph_objects: dict[str, Paragraph] = field(default_factory=dict, repr=False)
    _docx: DocxDocument | None = field(default=None, repr=False)

    @classmethod
    def load(cls, path: str | Path) -> "DocumentView":
        path = Path(path)
        doc = Document(str(path))
        paragraphs: list[ParagraphInfo] = []
        paragraph_objects: dict[str, Paragraph] = {}

        for i, p in enumerate(_iter_body_paragraphs(doc)):
            pid = _PARAGRAPH_ID_FMT.format(i)
            info = ParagraphInfo(
                id=pid,
                index=i,
                text=p.text,
                style=p.style.name if p.style else "",
                is_empty=not p.text.strip(),
                has_revisions=_paragraph_has_revisions(p),
                has_comments=_paragraph_has_comments(p),
                list_level=_paragraph_list_level(p),
            )
            paragraphs.append(info)
            paragraph_objects[pid] = p

        comments = _enumerate_comments(doc, paragraphs, paragraph_objects)
        revisions = _enumerate_revisions(doc, paragraphs, paragraph_objects)
        # B1: grader-canonical id = enumeration index; map the raw, model-set w:id
        # (.id) -> canonical so the render never surfaces a raw w:id.
        comment_id_map = {c.id: i for i, c in enumerate(comments)}
        revision_id_map = {r.id: i for i, r in enumerate(revisions)}

        return cls(
            docx_path=path,
            paragraphs=paragraphs,
            comments=comments,
            revisions=revisions,
            comment_id_map=comment_id_map,
            revision_id_map=revision_id_map,
            _paragraph_objects=paragraph_objects,
            _docx=doc,
        )

    def get(self, paragraph_id: str) -> ParagraphInfo | None:
        for p in self.paragraphs:
            if p.id == paragraph_id:
                return p
        return None

    def runtime_paragraph(self, paragraph_id: str) -> Paragraph | None:
        """Return the underlying python-docx `Paragraph` object for the given id."""
        return self._paragraph_objects.get(paragraph_id)

    def to_markdown(self, max_chars: int | None = None) -> str:
        """Serialize the document into an LLM-friendly Markdown view.
        Blank paragraphs are omitted from the visible output but their IDs are
        preserved (skipped) so paragraph IDs match the underlying document.
        If `max_chars` is set, truncates with a sentinel — useful for prompt
        budgeting. The LargeCo template is ~90 KB; most current frontier models
        comfortably fit the whole thing, but smaller models or budget runs may
        need windowing.
        """
        lines: list[str] = [f"# {self.docx_path.name}", ""]
        for p in self.paragraphs:
            if p.is_empty:
                continue
            flags = []
            if p.has_revisions:
                flags.append("REVS")
            if p.has_comments:
                flags.append("CMTS")
            flag_str = f" {{{','.join(flags)}}}" if flags else ""
            indent = "  " * (p.list_level or 0)
            lines.append(f"[{p.id}] ({p.style}){flag_str} {indent}{p.text}")

        if self.comments:
            lines.append("")
            lines.append("## Existing comments")
            for c in self.comments:
                head = f"[cmt-{c.id}] {c.author!r}"
                if c.parent_id is not None:
                    head += f" (reply to cmt-{c.parent_id})"
                spans = ",".join(c.paragraph_ids) or "?"
                lines.append(f"{head} on [{spans}]: {c.text}")

        if self.revisions:
            lines.append("")
            lines.append("## Existing tracked changes")
            for r in self.revisions:
                spans = ",".join(r.paragraph_ids) or "?"
                lines.append(
                    f"[rev-{r.id}] {r.type} by {r.author!r} on [{spans}]: {r.text!r}"
                )

        out = "\n".join(lines)
        if max_chars is not None and len(out) > max_chars:
            out = out[: max_chars - 100] + "\n\n[... document truncated ...]"
        return out

    def stats(self) -> dict[str, int]:
        return {
            "paragraphs": len(self.paragraphs),
            "non_empty_paragraphs": sum(1 for p in self.paragraphs if not p.is_empty),
            "existing_comments": len(self.comments),
            "existing_revisions": len(self.revisions),
        }


# --- Paragraph metadata helpers -------------------------------------------- #


def _paragraph_has_revisions(p: Paragraph) -> bool:
    for tag in ("w:ins", "w:del", "w:moveFrom", "w:moveTo"):
        if next(p._p.iter(qn(tag)), None) is not None:
            return True
    return False


def _paragraph_has_comments(p: Paragraph) -> bool:
    for tag in ("w:commentRangeStart", "w:commentRangeEnd", "w:commentReference"):
        if next(p._p.iter(qn(tag)), None) is not None:
            return True
    return False


def _paragraph_list_level(p: Paragraph) -> int | None:
    """Return the list nesting level if the paragraph is part of a numbered/bulleted list."""
    numPr = p._p.find(f"{qn('w:pPr')}/{qn('w:numPr')}")
    if numPr is None:
        return None
    ilvl = numPr.find(qn("w:ilvl"))
    if ilvl is None:
        return 0
    val = ilvl.get(qn("w:val"))
    try:
        return int(val) if val is not None else 0
    except ValueError:
        return 0


# --- Comment & revision enumeration ---------------------------------------- #


def _paragraph_ids_for_element_range(
    start_id_attr: str,
    end_id_attr: str,
    el_id: int,
    paragraphs: list[ParagraphInfo],
    paragraph_objects: dict[str, Paragraph],
) -> list[str]:
    """Find paragraph IDs that the given range marker pair spans.
    Walks paragraphs in document order looking for `w:commentRangeStart`/`End`
    (or analogous markers) with matching IDs. Returns paragraph IDs from start
    to end inclusive.
    """
    started = False
    spans: list[str] = []
    target = str(el_id)
    for pi in paragraphs:
        p = paragraph_objects[pi.id]
        # Did this paragraph contain start?
        for el in p._p.iter(qn(start_id_attr)):
            if el.get(qn("w:id")) == target:
                started = True
                break
        if started:
            spans.append(pi.id)
        # Did this paragraph contain end?
        for el in p._p.iter(qn(end_id_attr)):
            if el.get(qn("w:id")) == target:
                return spans
    return spans


def _enumerate_comments(
    doc: DocxDocument,
    paragraphs: list[ParagraphInfo],
    paragraph_objects: dict[str, Paragraph],
) -> list[ExistingComment]:
    out: list[ExistingComment] = []
    # python-docx exposes `Document.comments` (CommentCollection) in 1.2.
    try:
        collection = doc.comments
    except AttributeError:
        return out

    # Build a paraId → parent_id lookup from commentsExtended.xml so we can
    # mark replies. The commentsExtended part is optional; absence just means
    # no thread metadata available.
    paraid_parent = _read_comments_extended(doc)

    for c in collection:
        cid = int(c._element.get(qn("w:id"), -1))
        if cid < 0:
            continue
        spans = _paragraph_ids_for_element_range(
            "w:commentRangeStart",
            "w:commentRangeEnd",
            cid,
            paragraphs,
            paragraph_objects,
        )
        # Text — join all w:t descendants of the w:comment element.
        text = "".join(t.text or "" for t in c._element.iter(qn("w:t")))
        # Determine if this is a reply: look up our first paragraph's paraId.
        parent_id = None
        first_para = c._element.find(qn("w:p"))
        if first_para is not None:
            pid_attr = first_para.get(qn("w14:paraId"))
            if pid_attr and pid_attr in paraid_parent:
                parent_id = paraid_parent[pid_attr]

        out.append(
            ExistingComment(
                id=cid,
                author=c._element.get(qn("w:author"), ""),
                initials=c._element.get(qn("w:initials"), ""),
                text=text,
                paragraph_ids=spans,
                parent_id=parent_id,
            )
        )
    # A threaded reply has no body comment-range, so its paragraph_ids are empty
    # and it would render `on [?]`. Inherit the anchor from its thread root so the
    # read view / judge appendix show the real paragraph.
    _by_id = {c.id: c for c in out}
    for c in out:
        if c.paragraph_ids or c.parent_id is None:
            continue
        seen: set[int] = set()
        anc = _by_id.get(c.parent_id)
        while anc is not None and anc.id not in seen:
            seen.add(anc.id)
            if anc.paragraph_ids:
                c.paragraph_ids = list(anc.paragraph_ids)
                break
            anc = _by_id.get(anc.parent_id) if anc.parent_id is not None else None
    return out


def _read_comments_extended(doc: DocxDocument) -> dict[str, int]:
    """Return a paraId → parent_comment_id map from commentsExtended.xml.
    Word stores comment-thread parentage by paraId (a w14 attribute on the
    paragraph inside the comment body) rather than by comment IDs directly.
    We resolve that mapping here.
    """
    out: dict[str, int] = {}
    package_part = doc.part.package
    # Walk for any part whose URI ends with /word/commentsExtended.xml
    for rel in doc.part.rels.values():
        if rel.target_ref.endswith("commentsExtended.xml"):
            blob = rel.target_part.blob
            root = etree.fromstring(blob)
            # Build paraId → parent_paraId from <w15:commentEx>
            paraid_to_parent = {}
            for ex in root.iter(qn("w15:commentEx")):
                pid = ex.get(qn("w15:paraId"))
                ppid = ex.get(qn("w15:paraIdParent"))
                if pid:
                    paraid_to_parent[pid] = ppid
            if not paraid_to_parent:
                return out
            # Now we need to resolve parent paraId → comment id by walking
            # comments.xml.
            paraid_to_commentid = {}
            for crel in doc.part.rels.values():
                if crel.target_ref.endswith("comments.xml"):
                    cblob = crel.target_part.blob
                    croot = etree.fromstring(cblob)
                    for c in croot.iter(qn("w:comment")):
                        first_p = c.find(qn("w:p"))
                        if first_p is None:
                            continue
                        pid = first_p.get(qn("w14:paraId"))
                        cid = c.get(qn("w:id"))
                        if pid and cid is not None:
                            try:
                                paraid_to_commentid[pid] = int(cid)
                            except ValueError:
                                continue
            for pid, ppid in paraid_to_parent.items():
                if ppid in paraid_to_commentid:
                    out[pid] = paraid_to_commentid[ppid]
            break
    return out


def _enumerate_revisions(
    doc: DocxDocument,
    paragraphs: list[ParagraphInfo],
    paragraph_objects: dict[str, Paragraph],
) -> list[ExistingRevision]:
    out: list[ExistingRevision] = []
    seen_ids: set[int] = set()
    for pi in paragraphs:
        p = paragraph_objects[pi.id]
        for tag in ("w:ins", "w:del", "w:moveFrom", "w:moveTo"):
            for el in p._p.iter(qn(tag)):
                raw = el.get(qn("w:id"))
                if raw is None:
                    continue
                try:
                    rid = int(raw)
                except ValueError:
                    continue
                if rid in seen_ids:
                    continue
                seen_ids.add(rid)
                out.append(
                    ExistingRevision(
                        id=rid,
                        type=etree.QName(el).localname,
                        author=el.get(qn("w:author"), ""),
                        # Text belonging to THIS wrapper only — NOT the text of a
                        # deeper nested revision (a model edit made inside a
                        # counterparty insertion is a separate revision, usually a
                        # different author). The old flat ``el.iter(w:t/w:delText)``
                        # folded that nested text into the outer revision, so the
                        # appendix mis-attributed the model's insert/strike to the
                        # counterparty (R4). See _own_revision_text.
                        text=_own_revision_text(el),
                        paragraph_ids=[pi.id],
                        date=el.get(qn("w:date")),
                    )
                )
    return out


# Wrapper tags whose text is a SEPARATE (nested) revision, and transparent
# wrappers we descend through without crossing a revision boundary.
_REVISION_WRAPPER_TAGS = frozenset(
    qn(t) for t in ("w:ins", "w:del", "w:moveFrom", "w:moveTo")
)
_TRANSPARENT_WRAPPER_TAGS = frozenset(
    qn(t) for t in ("w:smartTag", "w:hyperlink", "w:sdt", "w:sdtContent")
)


def _own_revision_text(el) -> str:
    """Text of a revision wrapper (`w:ins`/`w:del`/`w:moveFrom`/`w:moveTo`) that
    belongs to IT, excluding any text inside a DEEPER nested revision wrapper.
    A nested revision (e.g. a model `w:del` inside a counterparty `w:ins`) is a
    distinct tracked change with its own author; folding its text into the outer
    revision's ``.text`` mis-attributes it in the appendix and corrupts the
    signature used for this-turn diffing (R4). We walk children, gathering
    ``w:r`` run text (``w:t`` + ``w:delText``) and descending only through
    transparent wrappers, stopping at any nested revision wrapper. For a
    non-nested revision this returns exactly the old flat join.
    """
    out: list[str] = []

    def walk(node) -> None:
        for c in node.iterchildren():
            if not isinstance(c.tag, str):
                continue  # XML comment / PI node
            if c.tag in _REVISION_WRAPPER_TAGS:
                continue  # separate nested revision — its text is its own
            if c.tag in _TRANSPARENT_WRAPPER_TAGS:
                walk(c)
            elif c.tag == qn("w:r"):
                for t in c.iter(qn("w:t")):
                    out.append(t.text or "")
                for t in c.iter(qn("w:delText")):
                    out.append(t.text or "")
            # else: pPr / rPr / bookmarks / comment refs — no revision text

    walk(el)
    return "".join(out)

	"""LLM-facing view of a .docx contract.

	A `DocumentView` is a frozen snapshot of a document for the LLM to reason
	about: it carries paragraph IDs the model can cite, a Markdown serialization
	that fits in a prompt, and enumerations of any pre-existing comments and
	tracked changes (so the model can reply on threads and accept prior edits).

	The view is built fresh each time the document is loaded; paragraph IDs are
	assigned by load-order (`p-000`, `p-001`, ...) and are stable for the duration
	of one redlining session. After we apply edits and save, reloading produces a
	new view — the LLM must re-read in that case.
	"""

	from __future__ import annotations

	from dataclasses import dataclass, field
	from pathlib import Path
	from typing import Iterable

	from docx import Document
	from docx.document import Document as DocxDocument
	from docx.text.paragraph import Paragraph
	from lxml import etree

	from .ooxml_utils import NSMAP, qn


	_PARAGRAPH_ID_FMT = "p-{:03d}"


	def _iter_body_paragraphs(doc: DocxDocument) -> "list[Paragraph]":
	"""Every paragraph in the document body **in document order, including
	paragraphs inside tables** (and nested tables).

	python-docx's `Document.paragraphs` only returns body-level `<w:p>` and
	silently skips `<w:tbl>//<w:p>`, so any edit inside a pricing schedule,
	SLA grid, or credit table was invisible to the grader (see the
	"table blindness" QC finding). We walk the body XML ourselves: yield each
	`<w:p>` where it sits, descending rows → cells → (recursively) nested
	tables, so paragraph IDs stay in reading order.
	"""
	body = doc.element.body
	parent = doc._body # a valid python-docx parent (resolves .part for styles)

	def walk(el):
	for child in el.iterchildren():
	if child.tag == qn("w:p"):
	yield Paragraph(child, parent)
	elif child.tag == qn("w:tbl"):
	for row in child.iterchildren(qn("w:tr")):
	for cell in row.iterchildren(qn("w:tc")):
	yield from walk(cell)

	return list(walk(body))


	@dataclass
	class ParagraphInfo:
	"""A paragraph in the document, with the metadata the LLM and engine need."""

	id: str
	index: int
	text: str
	style: str
	is_empty: bool
	has_revisions: bool
	has_comments: bool
	# Word's auto-numbering is driven by w:numPr; we record the abstract
	# level (0 = top-level) when present so the serializer can indent.
	list_level: int \| None = None


	@dataclass
	class ExistingComment:
	"""A pre-existing comment in the document the LLM may need to reply to."""

	id: int
	author: str
	initials: str
	text: str
	paragraph_ids: list[str] = field(default_factory=list)
	parent_id: int \| None = None # if this comment is itself a reply
	# Set by annotated_view.tag_this_turn(): True if this comment is net-new by
	# the represented party THIS turn, False if carried-in (counterparty / prior
	# turn), None if turn attribution was not computed (no baseline). Never read
	# from the model-writable docx — derived from the parsed baseline diff only.
	is_this_turn: bool \| None = None


	@dataclass
	class ExistingRevision:
	"""A pre-existing tracked change the LLM may need to accept/reject."""

	id: int
	type: str # "ins" \| "del" \| "moveFrom" \| "moveTo"
	author: str
	text: str
	paragraph_ids: list[str] = field(default_factory=list)
	# w:date on the revision element (ISO string) if present, else None.
	date: str \| None = None
	# Set by annotated_view.tag_this_turn(): True if this revision is net-new by
	# the represented party THIS turn, False if carried-in (counterparty / prior
	# turn), None if turn attribution was not computed (no baseline). Derived only
	# from the parsed OOXML/baseline diff — never model-controllable.
	is_this_turn: bool \| None = None


	@dataclass
	class DocumentView:
	docx_path: Path
	paragraphs: list[ParagraphInfo]
	comments: list[ExistingComment]
	revisions: list[ExistingRevision]
	# B1 (Class B): raw model-controlled w:id -> grader-canonical sequential id
	# (the enumeration index) for comments and revisions. The annotated (judge) view
	# resolves every rendered id through these maps, so no raw, model-forgeable w:id
	# ever reaches the judge. Populated by load(); default empty for direct
	# construction / backward compatibility.
	comment_id_map: dict[int, int] = field(default_factory=dict, repr=False)
	revision_id_map: dict[int, int] = field(default_factory=dict, repr=False)
	# Map paragraph ID → underlying python-docx Paragraph object (not serialized
	# for the LLM; used internally when applying edits).
	_paragraph_objects: dict[str, Paragraph] = field(default_factory=dict, repr=False)
	_docx: DocxDocument \| None = field(default=None, repr=False)

	@classmethod
	def load(cls, path: str \| Path) -> "DocumentView":
	path = Path(path)
	doc = Document(str(path))
	paragraphs: list[ParagraphInfo] = []
	paragraph_objects: dict[str, Paragraph] = {}

	for i, p in enumerate(_iter_body_paragraphs(doc)):
	pid = _PARAGRAPH_ID_FMT.format(i)
	info = ParagraphInfo(
	id=pid,
	index=i,
	text=p.text,
	style=p.style.name if p.style else "",
	is_empty=not p.text.strip(),
	has_revisions=_paragraph_has_revisions(p),
	has_comments=_paragraph_has_comments(p),
	list_level=_paragraph_list_level(p),
	)
	paragraphs.append(info)
	paragraph_objects[pid] = p

	comments = _enumerate_comments(doc, paragraphs, paragraph_objects)
	revisions = _enumerate_revisions(doc, paragraphs, paragraph_objects)
	# B1: grader-canonical id = enumeration index; map the raw, model-set w:id
	# (.id) -> canonical so the render never surfaces a raw w:id.
	comment_id_map = {c.id: i for i, c in enumerate(comments)}
	revision_id_map = {r.id: i for i, r in enumerate(revisions)}

	return cls(
	docx_path=path,
	paragraphs=paragraphs,
	comments=comments,
	revisions=revisions,
	comment_id_map=comment_id_map,
	revision_id_map=revision_id_map,
	_paragraph_objects=paragraph_objects,
	_docx=doc,
	)

	def get(self, paragraph_id: str) -> ParagraphInfo \| None:
	for p in self.paragraphs:
	if p.id == paragraph_id:
	return p
	return None

	def runtime_paragraph(self, paragraph_id: str) -> Paragraph \| None:
	"""Return the underlying python-docx `Paragraph` object for the given id."""
	return self._paragraph_objects.get(paragraph_id)

	def to_markdown(self, max_chars: int \| None = None) -> str:
	"""Serialize the document into an LLM-friendly Markdown view.

	Blank paragraphs are omitted from the visible output but their IDs are
	preserved (skipped) so paragraph IDs match the underlying document.

	If `max_chars` is set, truncates with a sentinel — useful for prompt
	budgeting. The LargeCo template is ~90 KB; most current frontier models
	comfortably fit the whole thing, but smaller models or budget runs may
	need windowing.
	"""
	lines: list[str] = [f"# {self.docx_path.name}", ""]
	for p in self.paragraphs:
	if p.is_empty:
	continue
	flags = []
	if p.has_revisions:
	flags.append("REVS")
	if p.has_comments:
	flags.append("CMTS")
	flag_str = f" {{{','.join(flags)}}}" if flags else ""
	indent = " " * (p.list_level or 0)
	lines.append(f"[{p.id}] ({p.style}){flag_str} {indent}{p.text}")

	if self.comments:
	lines.append("")
	lines.append("## Existing comments")
	for c in self.comments:
	head = f"[cmt-{c.id}] {c.author!r}"
	if c.parent_id is not None:
	head += f" (reply to cmt-{c.parent_id})"
	spans = ",".join(c.paragraph_ids) or "?"
	lines.append(f"{head} on [{spans}]: {c.text}")

	if self.revisions:
	lines.append("")
	lines.append("## Existing tracked changes")
	for r in self.revisions:
	spans = ",".join(r.paragraph_ids) or "?"
	lines.append(
	f"[rev-{r.id}] {r.type} by {r.author!r} on [{spans}]: {r.text!r}"
	)

	out = "\n".join(lines)
	if max_chars is not None and len(out) > max_chars:
	out = out[: max_chars - 100] + "\n\n[... document truncated ...]"
	return out

	def stats(self) -> dict[str, int]:
	return {
	"paragraphs": len(self.paragraphs),
	"non_empty_paragraphs": sum(1 for p in self.paragraphs if not p.is_empty),
	"existing_comments": len(self.comments),
	"existing_revisions": len(self.revisions),
	}


	# --- Paragraph metadata helpers -------------------------------------------- #


	def _paragraph_has_revisions(p: Paragraph) -> bool:
	for tag in ("w:ins", "w:del", "w:moveFrom", "w:moveTo"):
	if next(p._p.iter(qn(tag)), None) is not None:
	return True
	return False


	def _paragraph_has_comments(p: Paragraph) -> bool:
	for tag in ("w:commentRangeStart", "w:commentRangeEnd", "w:commentReference"):
	if next(p._p.iter(qn(tag)), None) is not None:
	return True
	return False


	def _paragraph_list_level(p: Paragraph) -> int \| None:
	"""Return the list nesting level if the paragraph is part of a numbered/bulleted list."""
	numPr = p._p.find(f"{qn('w:pPr')}/{qn('w:numPr')}")
	if numPr is None:
	return None
	ilvl = numPr.find(qn("w:ilvl"))
	if ilvl is None:
	return 0
	val = ilvl.get(qn("w:val"))
	try:
	return int(val) if val is not None else 0
	except ValueError:
	return 0


	# --- Comment & revision enumeration ---------------------------------------- #


	def _paragraph_ids_for_element_range(
	start_id_attr: str,
	end_id_attr: str,
	el_id: int,
	paragraphs: list[ParagraphInfo],
	paragraph_objects: dict[str, Paragraph],
	) -> list[str]:
	"""Find paragraph IDs that the given range marker pair spans.

	Walks paragraphs in document order looking for `w:commentRangeStart`/`End`
	(or analogous markers) with matching IDs. Returns paragraph IDs from start
	to end inclusive.
	"""
	started = False
	spans: list[str] = []
	target = str(el_id)
	for pi in paragraphs:
	p = paragraph_objects[pi.id]
	# Did this paragraph contain start?
	for el in p._p.iter(qn(start_id_attr)):
	if el.get(qn("w:id")) == target:
	started = True
	break
	if started:
	spans.append(pi.id)
	# Did this paragraph contain end?
	for el in p._p.iter(qn(end_id_attr)):
	if el.get(qn("w:id")) == target:
	return spans
	return spans


	def _enumerate_comments(
	doc: DocxDocument,
	paragraphs: list[ParagraphInfo],
	paragraph_objects: dict[str, Paragraph],
	) -> list[ExistingComment]:
	out: list[ExistingComment] = []
	# python-docx exposes `Document.comments` (CommentCollection) in 1.2.
	try:
	collection = doc.comments
	except AttributeError:
	return out

	# Build a paraId → parent_id lookup from commentsExtended.xml so we can
	# mark replies. The commentsExtended part is optional; absence just means
	# no thread metadata available.
	paraid_parent = _read_comments_extended(doc)

	for c in collection:
	cid = int(c._element.get(qn("w:id"), -1))
	if cid < 0:
	continue
	spans = _paragraph_ids_for_element_range(
	"w:commentRangeStart",
	"w:commentRangeEnd",
	cid,
	paragraphs,
	paragraph_objects,
	)
	# Text — join all w:t descendants of the w:comment element.
	text = "".join(t.text or "" for t in c._element.iter(qn("w:t")))
	# Determine if this is a reply: look up our first paragraph's paraId.
	parent_id = None
	first_para = c._element.find(qn("w:p"))
	if first_para is not None:
	pid_attr = first_para.get(qn("w14:paraId"))
	if pid_attr and pid_attr in paraid_parent:
	parent_id = paraid_parent[pid_attr]

	out.append(
	ExistingComment(
	id=cid,
	author=c._element.get(qn("w:author"), ""),
	initials=c._element.get(qn("w:initials"), ""),
	text=text,
	paragraph_ids=spans,
	parent_id=parent_id,
	)
	)
	# A threaded reply has no body comment-range, so its paragraph_ids are empty
	# and it would render `on [?]`. Inherit the anchor from its thread root so the
	# read view / judge appendix show the real paragraph.
	_by_id = {c.id: c for c in out}
	for c in out:
	if c.paragraph_ids or c.parent_id is None:
	continue
	seen: set[int] = set()
	anc = _by_id.get(c.parent_id)
	while anc is not None and anc.id not in seen:
	seen.add(anc.id)
	if anc.paragraph_ids:
	c.paragraph_ids = list(anc.paragraph_ids)
	break
	anc = _by_id.get(anc.parent_id) if anc.parent_id is not None else None
	return out


	def _read_comments_extended(doc: DocxDocument) -> dict[str, int]:
	"""Return a paraId → parent_comment_id map from commentsExtended.xml.

	Word stores comment-thread parentage by paraId (a w14 attribute on the
	paragraph inside the comment body) rather than by comment IDs directly.
	We resolve that mapping here.
	"""
	out: dict[str, int] = {}
	package_part = doc.part.package
	# Walk for any part whose URI ends with /word/commentsExtended.xml
	for rel in doc.part.rels.values():
	if rel.target_ref.endswith("commentsExtended.xml"):
	blob = rel.target_part.blob
	root = etree.fromstring(blob)
	# Build paraId → parent_paraId from <w15:commentEx>
	paraid_to_parent = {}
	for ex in root.iter(qn("w15:commentEx")):
	pid = ex.get(qn("w15:paraId"))
	ppid = ex.get(qn("w15:paraIdParent"))
	if pid:
	paraid_to_parent[pid] = ppid
	if not paraid_to_parent:
	return out
	# Now we need to resolve parent paraId → comment id by walking
	# comments.xml.
	paraid_to_commentid = {}
	for crel in doc.part.rels.values():
	if crel.target_ref.endswith("comments.xml"):
	cblob = crel.target_part.blob
	croot = etree.fromstring(cblob)
	for c in croot.iter(qn("w:comment")):
	first_p = c.find(qn("w:p"))
	if first_p is None:
	continue
	pid = first_p.get(qn("w14:paraId"))
	cid = c.get(qn("w:id"))
	if pid and cid is not None:
	try:
	paraid_to_commentid[pid] = int(cid)
	except ValueError:
	continue
	for pid, ppid in paraid_to_parent.items():
	if ppid in paraid_to_commentid:
	out[pid] = paraid_to_commentid[ppid]
	break
	return out


	def _enumerate_revisions(
	doc: DocxDocument,
	paragraphs: list[ParagraphInfo],
	paragraph_objects: dict[str, Paragraph],
	) -> list[ExistingRevision]:
	out: list[ExistingRevision] = []
	seen_ids: set[int] = set()
	for pi in paragraphs:
	p = paragraph_objects[pi.id]
	for tag in ("w:ins", "w:del", "w:moveFrom", "w:moveTo"):
	for el in p._p.iter(qn(tag)):
	raw = el.get(qn("w:id"))
	if raw is None:
	continue
	try:
	rid = int(raw)
	except ValueError:
	continue
	if rid in seen_ids:
	continue
	seen_ids.add(rid)
	out.append(
	ExistingRevision(
	id=rid,
	type=etree.QName(el).localname,
	author=el.get(qn("w:author"), ""),
	# Text belonging to THIS wrapper only — NOT the text of a
	# deeper nested revision (a model edit made inside a
	# counterparty insertion is a separate revision, usually a
	# different author). The old flat ``el.iter(w:t/w:delText)``
	# folded that nested text into the outer revision, so the
	# appendix mis-attributed the model's insert/strike to the
	# counterparty (R4). See _own_revision_text.
	text=_own_revision_text(el),
	paragraph_ids=[pi.id],
	date=el.get(qn("w:date")),
	)
	)
	return out


	# Wrapper tags whose text is a SEPARATE (nested) revision, and transparent
	# wrappers we descend through without crossing a revision boundary.
	_REVISION_WRAPPER_TAGS = frozenset(
	qn(t) for t in ("w:ins", "w:del", "w:moveFrom", "w:moveTo")
	)
	_TRANSPARENT_WRAPPER_TAGS = frozenset(
	qn(t) for t in ("w:smartTag", "w:hyperlink", "w:sdt", "w:sdtContent")
	)


	def _own_revision_text(el) -> str:
	"""Text of a revision wrapper (`w:ins`/`w:del`/`w:moveFrom`/`w:moveTo`) that
	belongs to IT, excluding any text inside a DEEPER nested revision wrapper.

	A nested revision (e.g. a model `w:del` inside a counterparty `w:ins`) is a
	distinct tracked change with its own author; folding its text into the outer
	revision's ``.text`` mis-attributes it in the appendix and corrupts the
	signature used for this-turn diffing (R4). We walk children, gathering
	``w:r`` run text (``w:t`` + ``w:delText``) and descending only through
	transparent wrappers, stopping at any nested revision wrapper. For a
	non-nested revision this returns exactly the old flat join.
	"""
	out: list[str] = []

	def walk(node) -> None:
	for c in node.iterchildren():
	if not isinstance(c.tag, str):
	continue # XML comment / PI node
	if c.tag in _REVISION_WRAPPER_TAGS:
	continue # separate nested revision — its text is its own
	if c.tag in _TRANSPARENT_WRAPPER_TAGS:
	walk(c)
	elif c.tag == qn("w:r"):
	for t in c.iter(qn("w:t")):
	out.append(t.text or "")
	for t in c.iter(qn("w:delText")):
	out.append(t.text or "")
	# else: pPr / rPr / bookmarks / comment refs — no revision text

	walk(el)
	return "".join(out)