Render a DOCX to PDF (LibreOffice), extract per-page geometry, and check layout rules. Returns structured JSON — no images.

Use this to see whether a document is visually correct: margin/table overflow, low-resolution images, likely orphan headings, plus the measured size and effective margins of every page.

Args: document: Absolute or relative path to a .docx file. pages: Optional 1-based page numbers to report (e.g. [4,5,6]). Omit for all. rules: Optional path to a YAML rule pack. Omit to use rules/general.yaml.

Returns: A dict with summary.headline (read this first), summary.counts, and a pages[] list. Each page has size_mm, effective_margins_mm, its tables/ images, and a warnings[] list of {rule, severity, target, detail, engine_sensitivity}. engine_sensitivity=dependent means the finding is only valid for the LibreOffice layout and may differ in Word.

Health check: confirm the render + geometry stack is wired up.

Call this first from a new client to prove the server is alive and that LibreOffice / PyMuPDF are reachable, before pointing it at a real document.

List the addressable elements of a DOCX with ids AND stable anchors.

Two ways to address an element as a doc_apply_patch target:

• id = el-N (document order; stable only within one version), or

• anchor = the paragraph's w14:paraId (Word's persistent id; survives edits/reordering). Tables anchor via their first cell as t:{paraId}.

assign_anchors=True writes a normalized copy with anchors filled in and returns it as normalized_output — patch THAT to address by anchor. Default is read-only (existing anchors reported, missing ones as null).

Pass the returned version_hash back as expected_version_hash so a stale plan is rejected. Paragraphs include style/spacing/keep_with_next/ page_break_before; tables include width_mm and overflows_text_column.

Apply safe layout ops to a COPY of the DOCX (original untouched), validate by reopening the result, and save a revision.

Each op = {"op": , "target": "el-N" | anchor, ...params}. target may be a positional el-N id or a stable anchor (w14:paraId, or t:{paraId} for a table). Supported ops: set_spacing(before_pt?, after_pt?, line?) · set_keep_with_next(enabled) · set_page_break_before(enabled) · set_paragraph_style(style) · set_font(size_pt?, name?) · set_table_width(width_mm)

Call doc_get_structure first for ids/anchors + version_hash; pass that as expected_version_hash. The output always carries anchors so a chain of patches stays addressable. Returns {ok, output, version_hash, applied, results[], validation}. Each op reports applied/error independently.

Closed loop as one tool: report → fix → re-render → diff, until nothing is left to fix or max_iter. The original file is never modified.

mode (default 'safe'):

• 'review' plans the ops without applying (dry run),

• 'safe' applies renderer-invariant fixes (table wider than the text column → fit it),

• 'full' also pins headings with keep_with_next (prevents orphans) and clamps over-large paragraph spacing to the rule-pack caps.

Regression guard: if an iteration introduces a NEW warning, that revision is discarded and the loop stops with the previous good version kept.

Returns document_final (path to the fixed revision), baseline vs final counts, per-iteration history with before/after diffs (and reverted if rolled back), accepted (critical==0), and unresolved issues that need judgement.

Render two DOCX versions and diff their layout reports: what got resolved, what regressed (new), page-count delta, and per-severity count deltas.

OPTIONAL visual channel: render ONE page of the DOCX to a PNG for the agent to eyeball a specific concern.

Use sparingly — images are costly in model context (see docs/01-compatibility.md); the deterministic doc_render_report (no images) is the primary signal. dpi is clamped to 36..300. On failure returns a structured {ok:false, error} instead.