feat(docx-mcp): add read-only get_document_outline tool

[stevenobiajulu]

What

Adds a read-only get_document_outline MCP tool that returns a compact structural map of a document's headings instead of full text — a projection over the existing docx-core document-view heading detection.

Each outline entry carries the heading text, outline level, heading source, and the stable _bk_* paragraph_id, so an agent can read the cheap outline first (hundreds of tokens), spot that §4.2 holds the indemnity clause, then scope a targeted read_file/replace_text to that region — instead of scanning a 60-page body and burning context window.

• Detection: style-based (Word HeadingN) headings only by default to keep the map low-noise; heuristic titles/run-in/centered-caps headers are opt-in via include_heuristic_headings.
• Output: format: "json" (default; structured outline array plus total_headings/total_paragraphs) or format: "markdown" (indented ATX outline under content).
• Scope: DOCX sessions only. ODF/Google Docs have no equivalent heading projection yet (flagged as follow-up in the proposal). Read-only annotation.

Why

docx-core already has the DOM traversal, heading-style detection, and stable _bk_* ids; this is a thin projection layer plus a tool_catalog.ts entry. Map first, then targeted read.

OpenSpec

New MCP capability, so this follows the OpenSpec workflow: change proposal add-document-outline-tool (proposal + tasks + mcp-server spec delta), validated with openspec validate --strict.

Files

• packages/docx-mcp/src/tools/get_document_outline.ts (new projection)
• packages/docx-mcp/src/tool_catalog.ts, src/server.ts (catalog entry + dispatch)
• packages/docx-mcp/src/tools/get_document_outline.test.ts (traceability tests, TEST_FEATURE='add-document-outline-tool')
• packages/safe-docx-mcpb/manifest.json (canonical-tool contract)
• packages/docx-mcp/docs/tool-reference.generated.md (regenerated)
• openspec/changes/add-document-outline-tool/**

Gates run (all green by exit code)

• npm run build — pass
• npm run lint:workspaces — pass (0 errors; pre-existing warnings only, none in touched files)
• node vitest run (full docx-mcp suite) — 849/849 pass, incl. new 3 outline tests
• npm run check:spec-coverage — pass (add-document-outline-tool: 3 scenarios covered by 3 story mappings)
• npm run check:tool-docs — pass (regenerated doc committed)
• npm run check:mcpb-manifest — pass
• npm run check:allure-labels / check:allure-filenames / check:allure-quality — pass
• npm run check:conformance-citations / check:conformance-doc — pass
• openspec validate add-document-outline-tool --strict — valid

Reviewer notes

• file_path reuses the shared optional field whose description says "DOCX or ODT"; the tool itself is DOCX-only (made explicit in the tool description and spec). Left as-is to avoid a one-off field variant — flagging in case you want it tightened.
• The proposal lists section-break / table anchors and ODF support as out-of-scope follow-ups, matching the issue's "start with style-based detection" guidance.

Ref: #493

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
site	Ready	Preview, Comment	Jun 26, 2026 5:44am

[stevenobiajulu]

Verdict: approve-with-nits — clean, well-scoped projection layer over existing docx-core heading detection. I built the PR branch in an isolated worktree and re-ran the claimed gates: the 3 new tests pass, openspec validate add-document-outline-tool --strict is valid, and check:spec-coverage reports add-document-outline-tool: 3 scenarios covered by 3 story mappings (THEN-keyword check passes too). Convention hygiene is good: Ref: #493 (no auto-close footer), conventional feat(docx-mcp): subject, TEST_FEATURE='add-document-outline-tool' declared, single-line .openspec(...) tags whose text matches the spec scenarios exactly. Scope matches the issue's "start with style-based detection" guidance (section-break/table anchors + ODF correctly deferred in the proposal's Out of scope). Tests assert real THEN observables (level, source, _bk_ id, default-vs-opt-in heuristic gating) — no tag-stuffing.

No blocking findings.

Nits (non-blocking):

• get_document_outline.test.ts:140 — the markdown test uses expect(content).toContain('# Introduction') / toContain('## Scope of Services'). Because '## Introduction'.includes('# Introduction') is true, this assertion is only a lower bound on # depth: a regression that over-indented H1 to ## would still pass. The scenario claims "depth SHALL reflect outline level," so consider anchoring to line starts (e.g. assert content.split('\n') equals the exact ['# Introduction','## Scope of Services']). Level is pinned exactly in the JSON test, so this is cosmetic.
• tasks.md 2.2 says "ODT via the shared session lane," but the dispatch in server.ts:163-164 calls getDocumentOutline directly with no isOdfRequest/isGDocsRequest guard; ODT is actually rejected with UNSUPPORTED_FOR_ODF inside resolveSessionForTool, and a bare google_doc_id yields MISSING_FILE_PATH. Behavior is correct and safe — just the task note is inaccurate.
• file_path reuses FILE_FIELD_OPTIONAL, so the generated doc renders "Path to the DOCX or ODT file." for a DOCX-only tool. You already flagged this; fine to leave, but the "or ODT" wording could nudge an agent into a guaranteed UNSUPPORTED_FOR_ODF round-trip.

[github-actions]

LLM-Based Quality Gate

Overall: ✅ PASS (8 pass · 0 warn · 9 skipped · 17 total)

	Check	Verdict
⏭️ SKIPPED	read_file response metadata parity	paths not touched by this PR
⏭️ SKIPPED	Live DOM namespace-safe OOXML writes	paths not touched by this PR
✅	Deleted field markup keeps w:fldChar outside w:del	The PR only implements the get_document_outline tool and does not touch field atomization, validateFieldStructure, w:fldChar, or other field-related logic.
⏭️ SKIPPED	Field validation per story, not global	paths not touched by this PR
⏭️ SKIPPED	Revision IDs seeded from all revision-bearing side parts	paths not touched by this PR
✅	Accept/reject sweep side parts and caches	The PR does not touch DocxDocument.acceptChanges, DocxDocument.rejectChanges, REVISION_STORY_PART_PATHS, accept_changes, reject_changes, or side-part revision markup, as it only implements the read-only get_document_outline MCP tool.
⏭️ SKIPPED	DocumentViewNode.heading stays canonical	paths not touched by this PR
✅	AI-author parity across entry points	All entry points instantiating SessionManager (packages/docx-mcp/src/server.ts:267, packages/docx-mcp/src/cli/tool_runner.ts:27, packages/docx-mcp/src/cli/commands/grep.ts:140, and packages/docx-mcp/src/cli/commands/edit.ts:132) correctly resolve SAFE_DOCX_AI_AUTHOR with identical three-way semantics.
⏭️ SKIPPED	Property-change wrapper discipline	paths not touched by this PR
⏭️ SKIPPED	SUPPORT.md Table A drift vs. implementation	paths not touched by this PR
⏭️ SKIPPED	Table A / Table B boundary on side-part revisions	paths not touched by this PR
✅	Canonical-emission surface completeness	The PR only adds a read-only document outline retrieval tool (get_document_outline) and does not modify or introduce any tracked-edit surfaces in packages/docx-core/src/primitives/** or packages/docx-mcp/src/tools/**, so the canonical-emission testing and documentation preconditions are not triggered.
⏭️ SKIPPED	Lean predicate drift against engine semantics (asymmetric)	paths not touched by this PR
✅	Unit-test quality (avoid tautological / change-detector tests)	The tests in packages/docx-mcp/src/tools/get_document_outline.test.ts:69-149 use first-principles XML inputs and assert exact expected semantic properties (such as text, level, source, and markdown formatting) directly without mocking the system under test or relying on implementation details.
✅	Re-derived facts vs canonical sources	The PR consumes the canonical document view in packages/docx-mcp/src/tools/get_document_outline.ts:70 by calling session.doc.buildDocumentView, leveraging pre-existing heading metadata rather than introducing any independent re-derivations.
✅	.openspec tag ↔ test-assertion drift	All three added .openspec tags in packages/docx-mcp/src/tools/get_document_outline.test.ts (lines 71-146) are accompanied by precise, dedicated assertions that fully exercise each scenario's preconditions and observables (a: ok; b: ok; c: ok).
✅	Library stays general (no downstream-domain leakage)	The PR adds get_document_outline in packages/docx-mcp/src/tools/get_document_outline.ts as a generic structural outline tool. It relies entirely on general OOXML heading/paragraph concepts and contains no downstream domain-specific vocabulary (such as signatures, parties, or cover terms).

Full checklist questions

1. read_file response metadata parity: If this PR touches packages/docx-mcp/src/tools/read_file.ts, budgeted pagination returns, or additive response metadata like warnings / comment_load_error, do every successful return path (default budgeted early return, non-budget fallthrough, explicit limit/node_ids) preserve the same additive diagnostic fields? read_file has multiple success exits; diagnostics have already disappeared on one path before. Reference: fix(docx-core): declare xmlns:w14/w15 on comments root before writing prefixed attributes (#154) #180 surfaced comment_load_error, fix(docx-mcp): warn when read_file budget is exceeded by a single node (closes #184) #186 added an early budget return + warnings, fix(docx-mcp): surface comment_load_error on the default budgeted read path (closes #189) #191 fixed the missing comment_load_error on the default budgeted path.

2. Live DOM namespace-safe OOXML writes: If this PR touches packages/docx-core/src/primitives/comments.ts or writes prefixed OOXML attributes/elements (w14:*, w15:*, xmlns:*, comments.xml, commentsExtended.xml, people.xml), are prefixed OOXML names written with namespace-aware APIs — root aliases bound with setAttributeNS(XMLNS_NS, ...), prefixed attributes with setAttributeNS(W14_NS/W15_NS, ...), and is there a test that proves the live DOM works before serialization/reparse? String-prefixed attributes can serialize plausibly while the live DOM still throws namespace errors. Reference: fix(docx-core): declare xmlns:w14/w15 on comments root before writing prefixed attributes (#154) #180 (xmlns:w14/w15 declared on comments root before writing prefixed attrs).

3. Deleted field markup keeps w:fldChar outside w:del: If this PR touches field atomization, validateFieldStructure, hasFldCharInsideDel, w:fldChar, w:instrText, w:delInstrText, or collapsed field comparison logic, does deleted field output stay ECMA-376-conformant — w:fldChar sibling-level (never inside w:del), deleted instructions use w:delInstrText only inside valid delete wrappers, accept/reject safety checks still reject malformed combined output? Word treats deleted field-state markup in the wrong container as document-corrupting. References: fix(docx-core): validate w:delInstrText placement and reject w:fldChar inside <w:del> #211, fix(docx-core): partition field-closure validation by ECMA-376 story (#212) #225, fix(docx-core): fragment w:fldChar outside w:del per ECMA-376 Part 4 #228.

4. Field validation per story, not global: If this PR touches packages/docx-core/src/baselines/atomizer/pipeline.ts, splitStories, validateFieldStructure, side-part merge logic, or footnote/endnote field handling, is field validation run independently per ECMA story (document.xml, each footnote, each endnote), with sidecars from both original and revised archives considered, and global counter balance not treated as sufficient? A document can be globally balanced but have an invalid field sequence inside one story. References: fix(docx-core): partition field-closure validation by ECMA-376 story (#212) #225, fix(docx-core): fragment w:fldChar outside w:del per ECMA-376 Part 4 #228, feat(docx-core): sweep side-part revisions on accept/reject #218.

5. Revision IDs seeded from all revision-bearing side parts: If this PR touches packages/docx-mcp/src/session/manager.ts (especially getRevisionContextForSession or FIXED_REVISION_ID_SEED_PARTS), createRevisionContext, revision-ID allocation, or MCP tools that create tracked changes/comments/footnotes, does revision-ID allocation scan all relevant package parts before issuing new IDs — comments, footnotes, endnotes, glossary, headers, footers — ignore non-revision w:id values (comment IDs, bookmarks), and handle malformed optional parts gracefully? Revision IDs are package-wide; document-only seeding collides with existing side-part revisions. Reference: fix(docx-mcp): seed revision ids from side parts #216 (seed revision ids from side parts).

6. Accept/reject sweep side parts and caches: If this PR touches DocxDocument.acceptChanges, DocxDocument.rejectChanges, REVISION_STORY_PART_PATHS, accept_changes, reject_changes, or side-part revision markup, does accept/reject process every revision-bearing story — updating document.xml + footnotes.xml + endnotes.xml + comments.xml, writing back only changed side parts while refreshing cached XML, and pruning orphan footnotes without deleting reserved separator entries? Accepting only in the main document leaves stale revisions and dangling references in the package. References: feat(docx-core): sweep side-part revisions on accept/reject #218, fix(docx-mcp): seed revision ids from side parts #216, fix(docx-core): partition field-closure validation by ECMA-376 story (#212) #225.

7. DocumentViewNode.heading stays canonical: If this PR touches packages/docx-core/src/primitives/document_view.ts, HeadingValue, heading heuristics, ListMetadata.header_style, or Google Docs document-view heading normalization, does node.heading remain a structural heading signal — exact Word styles Heading1–Heading6 win, heuristic sources suppressed inside table cells while real Word heading styles still pass, ordinary body paragraphs omit the heading key? Consumers use node.heading != null as a structural test; heuristic false positives break downstream navigation. References: fix(docx-core): harden heading detection (#157 Phase 1) #178, fix(docx-core): suppress non-sectional false-positive headings (closes #187) #188, feat(docx-core): add derived heading object to DocumentViewNode (closes #179) #190.

8. AI-author parity across entry points: If this PR touches packages/docx-mcp/src/server.ts, packages/docx-mcp/src/cli/tool_runner.ts, packages/docx-mcp/src/cli/commands/**, or adds any new new SessionManager(...) call site in docx-mcp, does every entry point that constructs a SessionManager resolve SAFE_DOCX_AI_AUTHOR with the same three-way semantics (set → use it; empty string → opt out to untracked; unset → defaultAiAuthor), or has a new entry path silently bypassed tracked emission? Each entry path looks locally correct while diverging from another; tracked emission has gone dark in one path before anyone noticed. References: feat(docx-mcp): wire configurable AI author through MCP layer (#142) #172 (production MCP wiring would have kept tracked emission dark), fix(docx-mcp): honor SAFE_DOCX_AI_AUTHOR in CLI entry points (#181) #182 (CLI runners constructing bare SessionManager() silently produced untracked edits).

9. Property-change wrapper discipline: If this PR touches packages/docx-core/src/primitives/layout.ts, packages/docx-core/src/primitives/text.ts, packages/docx-mcp/src/tools/clear_formatting.ts, or packages/docx-core/src/primitives/track-changes-emitter.ts, do tracked formatting/property edits emit exactly one correct *PrChange wrapper (pPrChange / rPrChange / trPrChange / tcPrChange) carrying a snapshot of the prior live properties — not stacking stale wrappers, not stripping valid historical children (cellIns/cellDel/cellMerge), and not omitting the snapshot when the operation is formatting-aware? Emitted OOXML is visually plausible but subtle snapshot mistakes only surface during later accept/reject or in Word's tracked-changes UI. References: feat(docx-core): emit pPrChange/trPrChange/tcPrChange from layout setters (#140) #167 (duplicate pPrChange/trPrChange/tcPrChange stacking + over-broad tcPr exclusion), feat(docx-mcp): emit rPrChange from clear_formatting MCP tool (#141) #170 (clear_formatting failing to strip stale rPrChange), feat(docx-core): emit rPrChange for formatted paragraph replacements #215 (rPrChange for formatted paragraph replacements + filtering nested stale records).

10. SUPPORT.md Table A drift vs. implementation: If this PR modifies OOXML revision emission behavior (w:ins, w:del, w:rPrChange, etc.) in packages/docx-core/src/primitives/**, or touches packages/docx-core/SUPPORT.md, does the PR symmetrically update Table A in SUPPORT.md when the supported revision-emission surface in primitives changed — added, removed, or weakened — or is the documented contract now lying about what's supported? Reviewers focus on TS AST correctness and golden tests; Markdown contract tables get treated as an afterthought, so the documented surface drifts from the actual surface. Reference: [120.8] Regression suite for canonical revision emission across the surface #143 review caught replaceParagraphTextRange should emit w:rPrChange when run formatting changes #173 (formatting mismatch in Table A) and addCommentReply should emit body revision markup OR SUPPORT.md should be softened #174 (comment body revision omission forcing a Table A softening) late in peer review.

11. Table A / Table B boundary on side-part revisions: If this PR touches packages/docx-core/src/primitives/comments.ts, packages/docx-core/src/primitives/footnotes.ts, or other side-part primitives, and adds/changes revision markup (w:ins, w:del), does tracked-change revision logic stay scoped to Table A (document-body content inside the side part) without leaking revision markup into Table B (the side-part package bootstrap — comments.xml/footnotes.xml element registration itself)? Body runs and side-part package elements share nearly identical XML namespace schemas; revisions emitted in the wrong table corrupt the package contract while looking plausible. References: [120.3] Emit w:ins/w:del for comment body anchors #138 (comment-body straddle constraints), [120.4] Emit w:ins/w:del for footnote reference and text #139 (footnote-reference straddle constraints).

12. Canonical-emission surface completeness: If this PR adds or changes a tracked-edit surface in packages/docx-core/src/primitives/** or packages/docx-mcp/src/tools/**, are the paired artifacts updated together — packages/docx-core/src/integration/canonical-emission-regression.test.ts, packages/docx-mcp/src/integration/canonical-emission-mcp.test.ts, and the documented emitter surface (Table A) — or is the rollout only partially wired? The primitive change looks done before the MCP path, regression matrix, and documented contract are wired through; partial rollouts ship undocumented surface that drifts. References: feat(docx-mcp): wire configurable AI author through MCP layer (#142) #172 (RevisionContext threaded through every Table A MCP tool), test(docx-core,docx-mcp): final regression suite for canonical emission (#143) #175 (24-test regression suite + verified write-time emitter rows), feat(docx-core): emit rPrChange for formatted paragraph replacements #215 (re-enabled rPrChange regression + updated support surface for replaceParagraphTextRange).

13. Lean predicate drift against engine semantics (asymmetric): If this PR changes field-wrapper semantics, the proof boundary, or atomizer behavior — packages/docx-core/src/baselines/atomizer/**, verification/lean/LeanSpike/Spec.lean, verification/lean/Tier2/**, or packages/docx-core/src/integration/lean-spec-bridge.test.ts — and if the TS engine semantics shifted, did the PR also update the Lean residual predicate and bridge tests, or is the proof now pinned to a stale stronger/weaker assumption? Asymmetric: a TS change without a corresponding Lean update is WARN; a Lean-only change without a TS update should not fire. The Lean side can still compile while the abstraction boundary is subtly wrong for the next engine refactor. References: feat(verification): close inv_field_001 with Tier 2 OoxmlDoc subset #208 (closed inv_field_001 using stronger recursivelyWellformed), refactor(verification): weaken inv_field_001 axiom to document-level preservationFriendly (rebased follow-up to #208) #220 (weakened the axiom to document-level preservationFriendly to avoid breakage when field fragmentation lands).

14. Unit-test quality (avoid tautological / change-detector tests): If this PR adds or modifies any **/*.test.ts (or other test files), are the test assertions independent of the system under test — expected values constructed from first principles rather than re-derived from the function under test, mocks limited to external boundaries (filesystem, network, clocks) rather than mocking the SUT itself, assertions making concrete semantic claims rather than just snapshotting current behavior or asserting non-null, and any test added alongside a bug fix actually exercising the bug? Tests that re-implement the production code as the "expected" value, or mock out the system under test, pass green while providing no regression protection.

15. Re-derived facts vs canonical sources: If this PR adds logic that re-computes a fact the codebase already derives elsewhere — another fldChar begin/separate/end visible-content walk (canonical: getParagraphRuns in packages/docx-core/src/primitives/text.ts), a second footnote display-number map (buildFootnoteDisplayMap vs getFootnotes().displayNumber), or any sibling re-implementation of an existing helper instead of importing or exporting it — does the PR consume the canonical source, or explicitly justify the new derivation and add a test pinning the copies in agreement? Independent derivations of the same fact drift apart; read_file renders every footnote marker twice in text/tagged_text ([^N][^N]) #382's doubled footnote markers were two passes disagreeing about one fact, and the fix initially added a fifth copy of the field-state walk. References: read_file renders every footnote marker twice in text/tagged_text ([^N][^N]) #382, fix(docx-mcp): render each footnote marker exactly once in read_file output #386, refactor: expose footnote references as document-view node metadata — one derivation of "visible footnote refs", not five fldChar walks #393.

16. .openspec tag ↔ test-assertion drift: If this PR adds, moves, or changes any .openspec('[ID] …') tag on a test(...)/it(...), does the tagged test body actually exercise that scenario's GIVEN/WHEN/THEN — establishing the scenario's precondition and asserting its THEN observable — or is the tag attached without a covering assertion (tag-stuffing)? Judge each added/changed [ID] independently against the body of the test it sits on. Pay special attention to (a) tags moved onto generic clean/round-trip/property tests that never set up the scenario's specific precondition, (b) a single test absorbing several distinct scenario IDs it does not separately assert, and (c) "empirical proxy" / repo-convention justifications standing in for a real covering assertion. Deterministic coverage checks confirm the tag exists and (with the THEN-keyword guard) that a scenario token is mentioned; they cannot judge whether the assertion actually covers the scenario — this asymmetric, oversells-coverage drift is exactly what the gate is for. References: test(docx-core): raise docx-comparison spec coverage to 68/69 and split coverage scripts #513 (first pass tag-stuffed multiple docx-comparison scenarios onto tests that didn't assert them; only peer review caught it), check:spec-coverage never passes --strict to validate_openspec_coverage.mjs, so the docx-core matrix gate is WARN-only in CI #469.

17. Library stays general (no downstream-domain leakage): If this PR adds or renames public API, types, recipes, functions, or identifiers in packages/*/src/**, does it avoid baking in names or concepts specific to a single downstream consumer's domain — agreements especially (signature, signatory, party, coverTerms, print-name/title/date signing semantics, or other agreement field names)? safe-docx is a general OOXML library: its surface should be Word/OOXML vocabulary (table, row, cell, border, run) plus a small allowlist of genuinely LLM-driven affordances that are not in Word but exist because LLMs are the primary consumers (e.g. an outline/TOC view sized for a context window). Domain-shaped composition (a "signature block" is just a w:tbl built from TableSpec/BorderSpec/RunProps) belongs in the consumer, not here. WARN if a new symbol is named for, or shaped around, one product rather than the general OOXML/LLM surface. See CONTRIBUTING "Library scope & domain boundaries". Reference: remove-agreement-domain-recipes removed coverTermsTable/signatureBlock for this reason.

Estimated cost (this run): $0.0146 — 43,864 input + 515 output tokens (≈4 chars/token) on gemini-3.5-flash. Char-count estimate, not provider telemetry.

[codecov]

Codecov Report

❌ Patch coverage is 76.92308% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...ackages/docx-mcp/src/tools/get_document_outline.ts	80.00%	2 Missing and 3 partials ⚠️
packages/docx-mcp/src/server.ts	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!