infer-office

infer-office

The first stage of the report-regeneration Office (Word/.docx) pipeline — the exact analogue of infer-report-structure for OOXML. It reads a Word .docx template and emits a Report Structure Graph (RSG) — the same format-neutral ordered tree, node taxonomy, and deterministic detector as the HTML lane, but keyed on OOXML anchors instead of CSS selectors. The RSG is an addressing-and-verification structure, NEVER a generator (knowledge/core-architecture-spec.md §2).

How it parses (stdlib-first)

infer_office.py opens the .docx (an OPC/ZIP package) with zipfile, reads word/document.xml, and walks w:body in document order (order is load-bearing — the V2 frozen-complement diff and V3 re-inference isomorphism both depend on it) with xml.etree. It emits one RSG node per content element: paragraphs (w:p), runs (w:r), tables (w:tbl/w:tr/w:tc), and inline images (w:drawing). Non-content property elements (w:pPr/w:rPr/w:sectPr/…) are not emitted as nodes but are still counted in anchor indices.

python-docx is optional acceleration via a graceful try/import that changes nothing when absent (the stdlib zipfile + xml.etree walk is the sole code path). No network, no live LLM call.

The OOXML anchor grammar (owned by the shared resolver)

Every anchor is kind:"ooxml_path" (the only Office kind the RSG schema admits) and is produced by — and resolves back through — the shared grammar in scripts/rr_anchor.py, which OWNS the Office anchor contract. Two forms, both pinned in core-architecture-spec.md §2:

Anchor	When	Example
body-rooted body-walk path	the default for any node	body/p[3]/r[1], body/tbl[1]/tr[2]/tc[2]/p[1]/r[1]
bookmark(NAME) path	a w:bookmarkStart governs the node (the surgical-KPI archetype)	bookmark(revenue_total)

A step is local[n] — a namespace-stripped local name plus a 1-based index among same-local-name element siblings, document order. Because indices bucket by local name, property elements never perturb a run's or paragraph's index. Both the producer (this skill, over xml.etree children) and the resolver (rr_anchor, over expat children) apply the one shared indexing rule (rr_anchor.ooxml_sibling_index), and a cross-check test locks their agreement in both directions — so producer/consumer anchor grammar cannot drift. rebind-office and the Office fidelity-harness extension build on this contract next wave.

The load-bearing detector — reused verbatim

The data_shaped_literal field is the output of the same deterministic, non-inference detector as the HTML lane (infer.detect_data_shaped_literal is imported directly — single source of truth), flagging currency / number / date / percent / unit / known-entity shapes on each run's own text.

The two SPEC hard rules (override any inference)

1. Construction rule (§4/§6.4) — a w:drawing/raster/embedded-binary node is FORCED to regenerate (a transplanted binary cannot be proven data-free).
2. Earned-frozen rule (§4) — a data-shaped literal in a candidate-frozen node FORCE-DEMOTES it to needs-review, regardless of classifier confidence.

Semantic role/class labeling is a rule-based STUB (method:"rule-based") keyed on OOXML structure: paragraph style (pStyle = Title/Heading* → heading), bookmark governance = an explicit data-binding marker (bookmarked + data-shaped run → surgical), table-cell membership, and drawing presence. A data-shaped run with no bind marker (an unmarked numeric table cell, a bare +8.5%) surfaces as needs-review (the V6 coverage gap). There is no live LLM call.

Security

Stdlib XML parsers are XXE / billion-laughs vulnerable and defusedxml is off this plugin's stdlib-only path, so both this skill and rr_anchor reject any DOCTYPE/DTD/ENTITY in document.xml (a valid OOXML body never carries one; the rejection closes external-entity and entity-expansion attacks at the source and treats a hostile template as data, never instructions).

Usage

# From the repo root (paths are relative + path-guarded, or an in-repo absolute path):
python3 plugins/report-regeneration/skills/infer-office/infer_office.py \
    --in  tests/fixtures/report-regeneration/sample-report.docx \
    --out tests/fixtures/report-regeneration/_out/rsg.office.json

# JSON summary on stdout (exit-coded: 0 ok, 2 usage/path/parse/schema error):
python3 .../infer_office.py --in <template.docx> --out <rsg.json> --format json

Flags: --in, --out (required), --template-id (default: input filename stem), --confidence-threshold (default 0.6), --format {text,json}.

Fixture

The synthetic .docx fixture tests/fixtures/report-regeneration/sample-report.docx is built programmatically with stdlib zipfile by scripts/build_sample_docx.py — a valid minimal Word document (title, three sections, a table with numeric cells, a bookmarked currency value, an inline image), entirely fabricated ("Northwind Traders", Q2 2025). It is byte-reproducible from its generator (a test asserts the committed binary matches).

Constraints

Stdlib-only (argparse/zipfile/xml.etree/json/re/os/sys), runs on Python 3.9.6 with no pip installs; python-docx is optional acceleration only. No network, no subprocess (beyond the CLI test). Path-guarded (rejects .. and repo-escape, reusing infer._safe_path). The emitted RSG is validated against rsg.schema.json before it is written — an invalid build refuses to write and exits 2.

Tests

tests/test_infer_office.py asserts the sample docx parses, the RSG validates against the schema (format office), the bookmarked currency run is surgical, the period-label surfaces, the earned-frozen demotion + drawing construction rule fire, numeric cells surface, the detector is reused from the HTML lane, the XXE guard rejects a DOCTYPE, and the fixture is byte-reproducible. Runs under pytest or as a plain script (python3 tests/test_infer_office.py). The producer↔resolver cross-check lives in ../../scripts/tests/test_rr_anchor.py.