feat(post): comment_on_post tool + activity URNs in posts references

[Dominien]

Summary

Adds the engagement-loop pair upstream is missing in v4.10.1: callers can chain get_person_profile(sections=["posts"]) → comment_on_post end-to-end without external URL discovery.

• comment_on_post — write tool mirroring send_message's confirmation gate. confirm_post=False does a dry run that locates the composer; confirm_post=True types and submits.
• Activity URNs in posts references — classify_link now recognises share-style /posts/<slug>-activity-<id>-<sig>/ permalinks, and a data-urn harvest in the is_activity branch picks up the URNs that the recent-activity DOM exposes only on post wrappers (timestamps render as <button>, not <a>, so anchor-only references missed them). Both surface as feed_post references with the canonical /feed/update/urn:li:activity:NNN/ URL.

End-to-end live-tested against a German LinkedIn account: get_person_profile(<user>, sections=posts) returned 12 feed_post URNs; comment_on_post(<urn>, <text>, confirm_post=true) posted, composer cleared, comment visible in thread.

Why bundled

comment_on_post is dead weight without a way to discover post URLs programmatically; the URN harvest is the missing other half. Happy to split into two PRs if preferred.

Design highlights

Locale-independence (per AGENTS.md Scraping Rules)

• Composer detection: [role=textbox][contenteditable=true] + aria-label hint with structural fallback.
• Submit-button detection: componentkey*="commentButtonSection" (LinkedIn's React component identity, locale-independent), with exact-match aria-label fallbacks ("Comment" / "Kommentieren" / "Commenter") for older renderings.
• URN harvest: [data-urn^="urn:li:activity:"] + [data-id^="urn:li:activity:"] — attribute presence + URN prefix.
• Share permalink regex: ^/posts/[^/?#]*-activity-(\d{15,})-[^/?#]+/?$.

Reliability lessons baked in

• The Tiptap composer hydrates async; wait_for_selector lets it mount before the focus/type evaluate.
• Submit click uses a full mousedown → mouseup → click MouseEvent sequence; bare btn.click() registers but skips React parent gates that gate the actual submit.
• Verify-after-action is two-stage: composer-empty (the reliable structural success signal — LinkedIn clears Tiptap only on server-accept) then body-text-visible. The composer-clear stage eliminates the false positive where typed-but-unsubmitted text matches body.innerText.includes(comment).
• Candidate buttons re-ranked so componentkey matches always win — an earlier draft used aria-label*="Post" and substring-matched "Open control menu for post by NAME", clicking the post overflow menu instead of the comment submit.
• Harvest entries prepended (not appended) so URNs win the per-section reference cap over generic post-attachment anchors.

Verification

• uv run pytest --cov — 486 passing (461 baseline + 25 new). New classes:
  • TestPostTool — happy path, dry-run, composer-unavailable, error.
  • TestNormalizeActivityUrl — URL normalisation across all four input shapes.
  • TestClassifyShareActivityPermalink — share-permalink regex + dedup with /feed/update/ form.
  • TestActivityUrnExtraction — wrapper harvest, non-activity-section skip, dedup with anchor-derived refs, helper-level filtering.
  • comment_on_post added to TestToolTimeouts.test_all_tools_have_global_timeout.
• uv run ruff check . && uv run ruff format --check . && uv run ty check && uv run pre-commit run --all-files — clean.
• Live E2E confirmed on a real post (German UI).

Synthetic prompt

Add a comment_on_post MCP tool to linkedin-mcp-server v4.10.1 that mirrors send_message's confirmation-gated, JS-driven write pattern, with locale-independent composer + componentkey-anchored submit detection. Surface activity URNs in get_person_profile(posts) section so callers can chain into comment_on_post — both via a /posts/-activity--/ regex in classify_link and a data-urn wrapper harvest inside the is_activity branch of _extract_page_once. Tests, docs, manifest.

Generated with Claude Opus 4.7 (1M context)

[greptile-apps]

Greptile Summary

This PR adds a comment_on_post tool and surfaces activity URNs in get_person_profile(sections=["posts"]) so callers can chain the two end-to-end. The implementation is well-structured — locale-independence requirements from AGENTS.md are met via a dedicated _COMMENT_SUBMIT_LABELS_BY_LOCALE table, structural selectors, and componentkey-anchored submit detection. The two previous review concerns (inconsistent digit floor in _normalize_activity_url and inline locale labels) appear to be addressed in this revision.

Confidence Score: 5/5

Safe to merge — no P0 or P1 findings; all previous review concerns are addressed.

Only P2-or-lower observations remain. The implementation correctly applies the ≥15-digit floor consistently across _normalize_activity_url and _SHARE_POST_PATH_RE, the locale-label concern from the prior review cycle is resolved with _COMMENT_SUBMIT_LABELS_BY_LOCALE, the test naming ambiguity is resolved by the updated docstring, and the overall pattern (exception handling, confirmation gate, two-stage verify) faithfully mirrors the send_message precedent. 486 tests pass including 25 new ones covering all new code paths.

No files require special attention.

Important Files Changed

Filename	Overview
linkedin_mcp_server/scraping/extractor.py	Adds post_comment, _normalize_activity_url, _harvest_activity_urns, and _wait_for_composer_clear; _normalize_activity_url consistently uses the ≥15-digit floor; activity URN harvest correctly prepends refs before anchor refs so URNs win the section cap.
linkedin_mcp_server/tools/post.py	New tool registration file mirroring send_message's confirmation-gated pattern; exception handling matches the established codebase pattern used by every other tool.
linkedin_mcp_server/scraping/link_metadata.py	Adds _SHARE_POST_PATH_RE with ≥15-digit guard and a classify_link branch that canonicalizes /posts/slug-activity-id-sig/ share permalinks to /feed/update/urn:li:activity:id/.
tests/test_scraping.py	Adds TestNormalizeActivityUrl and TestActivityUrnExtraction with thorough coverage; test_harvest_helper_filters_invalid_urns docstring correctly documents that dedup is intentionally deferred to build_references, resolving the naming ambiguity from the previous review cycle.
tests/test_tools.py	Adds TestPostTool with happy-path, dry-run, composer-unavailable, and error cases; comment_on_post added to TestToolTimeouts.
tests/test_link_metadata.py	Adds TestClassifyShareActivityPermalink with 7 cases including short-id rejection, missing-sig rejection, and dedup with /feed/update/ anchors.
linkedin_mcp_server/server.py	Imports and registers register_post_tools alongside the other tool registration calls.
manifest.json	Adds comment_on_post entry to the manifest tools array.

Sequence Diagram

sequenceDiagram
    participant Caller as MCP Caller
    participant Tool as comment_on_post tool
    participant Ext as LinkedInExtractor
    participant Page as LinkedIn Page

    Caller->>Tool: comment_on_post(post_url, text, confirm_post)
    Tool->>Ext: post_comment(post_url, text, confirm_post=...)
    Ext->>Ext: _normalize_activity_url(post_url)
    Note over Ext: Accepts /feed/update/URN, /posts/slug-activity-id-sig/, bare URN, bare id
    Ext->>Page: navigate to canonical /feed/update/ URL
    Ext->>Page: wait_for_selector(main) + handle_modal_close
    Ext->>Page: wait_for_selector(div[role=textbox][contenteditable])
    Ext->>Page: evaluate() — focus first visible composer
    alt composer not found
        Ext-->>Tool: status: composer_unavailable
    else confirm_post=False
        Ext-->>Tool: status: confirmation_required
    else confirm_post=True
        Ext->>Page: keyboard.type(comment_text, delay=15)
        Ext->>Page: evaluate() — click submit (componentkey → aria-label → Ctrl+Enter fallback)
        Ext->>Page: _wait_for_composer_clear (poll until innerText empty)
        Ext->>Page: _message_text_visible(comment_text)
        alt composer cleared + text visible
            Ext-->>Tool: status: posted, posted: true, comment_visible: true
        else composer not cleared
            Ext-->>Tool: status: post_unverified, posted: false
        else text not visible
            Ext-->>Tool: status: post_unverified, posted: true, comment_visible: false
        end
    end
    Tool-->>Caller: result dict

Loading

_{Reviews (2): Last reviewed commit: "fix(post): address Greptile review on #4..." | Re-trigger Greptile}