linkedin-live-validation

name	linkedin-live-validation
description	Use when a LinkedIn selector, scrape, or automation flow may have drifted, or before/after changing Selenium selectors — how to run the read-only live probe inside the selenium worker and read its JSON report.

Live LinkedIn selector grounding

Who may RUN the probe: the owner, and — since #1301 — a pipeline agent, under the three conditions below. The old rule was "owner only", because "read-only" was a property of the prose and a headless lane launches with --dangerously-skip-permissions, so nothing but the code stood between a probe and a post on the owner's real account. All three reasons are now mechanisms, not intentions:

1. It cannot write. install_read_only_guard() patches Selenium the moment the session is up (signing in is the one step that legitimately types — credentials and the emailed PIN — through production's own login path), so for every probe that follows: no printable character can be typed anywhere (every LinkedIn write starts with typed content) and no control whose own label commits something — Post, Send, Publish, Connect, Invite, Withdraw, Follow, Like, Join — can be clicked, via WebElement.click, ActionChains or arguments[0].click(). There is no override flag. Escape still closes a composer.
2. It cannot spend rate budget behind the breaker's back. The 429 breaker and the manual automation pause are read BEFORE Chrome opens and fail CLOSED — an unreadable breaker refuses too. There is NO override flag. A breaker stuck open is cleared at the breaker itself — the state this reads — never by teaching the probe to ignore it.
3. It cannot take a production Chrome slot. --require-debug-node pins the session to the watchable Grid node and refuses if it is unavailable. Agents must always pass it. That node offers 2 sessions (one probe + one MCP browser); a third is refused in about a second rather than queued, so a refusal is fast and unambiguous. The Selenium MCP browser (selenium-lem in .mcp.json) is pinned there too, with no opt-out.

Being off the pool is not being safe. The debug node reaches the same live LinkedIn on the same real account — it protects production CAPACITY, nothing else. Conditions 1 and 2 protect the probe; the Selenium MCP browser (selenium-lem) is pinned to the same node for capacity isolation but is interactive and NOT read-only-guarded, so it must be treated as a live browser you drive by hand. Do not use it to post, comment, send invites/DMs, or change settings.

A refusal is a WAIT, not a failure. Exit code 75 with a fenced JSON refusal block naming rate_limit_breaker_open / rate_limit_breaker_unreadable / debug_node_unavailable / debug_node_pin_unsupported. Re-run later; do NOT open a needs-human on the first one.

Still an escalation (needs-human + a Decision Comment): anything needing a WRITE on LinkedIn (posting, commenting, sending an invite or DM, changing a setting) — no flag makes that possible; a breaker that stays open across repeated attempts; and any probe run that needs a credential or an account decision. The Fix invariants below apply to every agent and need no probe run at all.

The agent form of the command (note the flag, and that 75 means wait):

sudo docker exec -i celery_worker_selenium python - --user-id 1 --require-debug-node --feed-sort \
    < scripts/linkedin_live_validation.py

The probe (scripts/linkedin_live_validation.py) is read-only — it navigates and reads, posts/comments/changes nothing. scripts/ is not baked into the image, so pipe it in on stdin — from the checkout you are validating, not whatever directory the shell landed in (/home/lem/linkedin_engagement_manager tracks main, so a branch's new capture is absent there and its report key never appears). Every report names the piped copy's optional captures in probe_script.capabilities: a capability listed there but missing from a surface means the PAGE had nothing, a capability absent from that list means the SCRIPT had nothing.

sudo docker exec -i celery_worker_selenium python - --user-id 1 \
    --post-url 'https://www.linkedin.com/feed/update/urn:li:activity:<id>/' \
    < scripts/linkedin_live_validation.py

Flags (combine as needed): --user-id (default 1) · --surfaces (print the surface → probe coverage matrix as JSON; no browser, no network) · --sweep [+ --sweep-profile-url] (every target-free probe in ONE session — what the weekly drift cron runs) · --post-url (own-post render + stats) · --probe-composer (composer controls, opens then Escapes) · --comment-outcome-url + --our-slug + --comment-text (comment-outcome readability) · --dm-thread-url + --dm-thread-name (thread resolution) · --article-editor-url (newsletter editor; never clicks Next, so publish grades UNKNOWN not MISSING) · --feed-sort (Recent-sort control) · --reaction-probe + --reaction-cards N + --reaction-open-menu · --roster-follow <profile-url> (top-card Follow/Following control on a roster target's activity page, #962 — resolves and describes it, clicks nothing) · --roster-connect <profile-url> (what that same top card says about our CONNECTION — Pending / 1st-degree / Message, #979 — read-only, sends no invite) · --appreciation-sources (Recommendations Received + mentions notification feed, #968 — reports per card what the appreciation-DM triggers would do; messages nobody, claims no ledger row) · --sent-invites [+ --sent-invite-days N] (pending sent invites and how their "Sent … ago" stamps parse, #969 — withdraws nothing; zero rows is graded against the page's own empty state, so a clean account and a rotated anchor read differently) · --connect-dialog <profile-url> (custom-invite URL renders the dialog? plus every "Invite … to connect" control naming SOMEONE ELSE, #1012 — never clicks Send or an Invite control) · --profile-scrape <profile-url> (name/headline/degree badge vs the page's own text; use a 2nd/3rd-degree profile or the degree half reads degree_grounded: false) · --profile-experiences <profile-url> (what the rebuilt experience parser reads off /details/experience/, beside the raw lines and every rung's hit count, #970) · --catchup-cards (#964) · --group-composer <group-id> (opens the share box, never clicks Post) · --group-membership [<group-id>] (#1052 — the groups directory as the shipped sync reads it, plus one group page's own join/leave controls; defaults to the first DB-enabled group, clicks nothing, joins/leaves nobody) · [+ ] (#928 — walks a GROUP feed and reports per card whether the comment composer needs resolves, with hit counts and whether #916's widening is what found the box; runs the home feed as a control, defaults to the first DB-enabled group, opens each composer and Escapes it, types and submits nothing) · [+ ] (credits/invitee rows/checkboxes; ticks nothing, never clicks Invite) · (card → Comment action → composer on a post permalink, #966 — opens the composer and Escapes it, types and submits nothing) · (the noVNC debug node — the default since #1301, so this flag now only says so out loud) · (refuse rather than fall back to a production Chrome slot — ).

Reading the report: a run that refused carries refusal and NO surfaces — read reason and wait_seconds, then re-run; it grounds nothing. A run that started carries read_only_guard (what the guard refused, plus unlabelled_clicks — presses on controls with no readable label, the one thing the click half cannot classify) and breaker. Otherwise: every probe carries a three-state state next to its prose verdict — ok (chain resolved), drift (the PAGE shows content the locator can't see — the only state the weekly cron files an issue for), unknown (page didn't render; grounds nothing, never filed). A post_stats signal sourced from none while the analytics page plainly shows a number = layout drift — the *_lines arrays hold the new layout; update the selector rows in docs/sdui-selenium-notes.md. Paste the JSON into the issue.

Fix invariants: locators are data-testid / aria-label / href / TEXT — never class names. Composer lookups scoped to their OWN card (_post_composer_for_card / _reply_composer_for_comment), no page-wide fallback. Success is the OUTCOME being present, never a click having landed; never click a control whose label names a different entity than the target; zero items is not "nothing to do" until the page agrees (_report_zero_walk). A selector miss is a DEBUG no-op, never log_warning (a repeated warning files a defect). Strip non-BMP emoji before send_keys (_strip_non_bmp()). No-browser alternative for post stats: scripts/linkedin_post_stats_api_probe.py --post-id <lem id>.

Authoritative: docs/sdui-probe-coverage.md (surface → probe matrix + the weekly drift cron), docs/LIVE_VALIDATION_FORMAT_AND_STATS.md §4, docs/sdui-selenium-notes.md, docs/SELENIUM_DEBUGGING.md.