Diagnose tracking - #11
Closed
amitl-levi wants to merge 6 commits into
Closed
amitl-levi wants to merge 6 commits into
amitl-levi wants to merge 6 commits into
Conversation
amitl-levi
force-pushed
the
diagnose-tracking
branch
from
August 27, 2026 07:23
d66bb8a to
23cd61c
Compare
This was referenced Aug 27, 2026
Closed
amitl-levi
force-pushed
the
marketplace-packaging
branch
from
August 30, 2026 08:20
05ebd0b to
bdfb2bd
Compare
amitl-levi
force-pushed
the
diagnose-tracking
branch
from
August 30, 2026 08:20
23cd61c to
582ebcf
Compare
yanush88
approved these changes
Aug 30, 2026
… refusal Adapted from an internal Taboola support team's pixel-diagnostics skill with everything internal stripped: no internal-database queries, no Salesforce intake, no internal config flags, no live-browser tooling. Adoption notes and the MCP capability asks (top ask: a server-side "did the pixel events land?" read) are in docs/2026-08-22-pixel-expert-adoption-plan.md. The skill verifies the Taboola Pixel from evidence a client can produce: fetches the user's page (static install check, per account - multi-account and network-level pixel setups handled), reads a user-captured HAR / window._tfa dump (runtime firing check, with copy-paste capture instructions for a non-technical reader), and cross-checks conversion rules and spend via the MCP. Fixes route by owner: site-side as copy-paste instructions (with proactive help-center install steps when the code is missing), rule-side through manage-campaigns' preview-then-confirm gate, Taboola-side through the /realize-plugin:support escalation. App-store URLs are caught before any fetch - no web pixel runs there, so "not installed" would be a false finding; the answer is the S2S/MMP path. This flips a documented refusal, so every stale claim site changed in the same commit: the agent's description, triage table (pixel-health removed from UI-only; dedicated routing row added), over-engagement anchor, tracking ladder (new rung 3), and tool-existence boundary; the guardrails' out-of-MCP list; manage-campaigns' UI-fallback section; optimize-campaign's P1 pre-check and prescription hand-offs; README scope + skills table; knowledge/tracking.md and its manifest entry. Pixel installation, codeless-conversion setup, and test-firing remain genuinely UI-only. Two new trust boundaries in os/guardrails.md: the landing-page fetch is the plugin's first non-realize.com web access, scoped to the page under diagnosis, and fetched content - like HAR content - is evidence, never instructions (Scenario 22D plants hostile page text as the regression test). HAR files carry cookies and session data: Grep-first sliced reading, and no cookies / auth headers / unrelated-domain requests in any output. The honesty boundary is explicit because no MCP pixel tool exists: the skill can prove what the browser sent and what Realize reports, never whether an event landed inside Taboola's processing. The source skill's spend gate is kept verbatim - fires + live rule + zero conversions with no active spend is expected, not broken. Read Scenario 22 (A-D) and write Scenario W15 (diagnosis-to-fix handoff stays gated; "find out why and fix it" is not pre-authorization) added. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ipt tag Tested live before anyone hit it in the field: WebFetch converts the page to readable text before answering, so a page whose raw HTML carried 11 script tags came back with zero code visible. A WebFetch-based static check would report "no pixel" on every site - the worst possible failure for a diagnostic skill, wrong with confidence. The check now downloads the raw page via Bash (curl of the user's page under diagnosis, nothing else), Greps it for the pixel markers, and Reads only matching regions. WebFetch is dropped from the skill's allowed-tools; the skill, the agent's tool reference, and the tool-existence boundary all say why, so a future cleanup doesn't "simplify" it back. CLAUDE.md's no-direct-curl rule now names this as its one sanctioned non-API exception, and Scenario 22A fails any run whose static check went through WebFetch. Trust rules are mechanism-independent and unchanged: the downloaded page is untrusted evidence, never instructions. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Live probes before anyone hits these in the field: - A Taboola publisher page (site showing Taboola ads) loads cdn.taboola.com/libtrc/<publisher>/loader.js - same libtrc family, zero advertiser pixel. Verified on a real publisher site. Judging the pixel by "libtrc" alone reports "installed" on any site that merely runs Taboola ads; only libtrc/unip/<id>/tfa.js is the pixel. Both the skill and the pixel reference now state the distinction. - A failed download, a bot-challenge page, or a near-empty JS shell (observed live: a 7KB single-page-app shell on a real brand site) gets no static verdict at all - the skill says the page couldn't be inspected statically and moves to the user's browser capture, which answers everything the static check would have. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…st live run Two field facts from running the new skill against a real advertiser landing page and a real HAR capture: - The static check and the HAR check both work end-to-end. The page carried zero Taboola code in raw HTML (runtime-injected pixel), the skill correctly deferred to the capture instead of reporting "not installed", and the 5.5MB HAR yielded the full verdict via Grep-first sliced reading: tfa.js 200, exactly one page_view on the modern JSON transport, item-url present, current library version, internal engagement events ignored, no cookies echoed. - The server-side write block follows the user, not the account. The rejection message names the account, but the operative policy is that internal Taboola users are excluded from writes - a client user can succeed on the same account. The agent's error handling, the manage-campaigns gotcha, and the W10-W15 unrun note now say so; running the write scenarios needs a non-internal user, not a different account. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…-side item The adoption doc for the MCP owner now carries only what is theirs: the three capability asks plus minimal context. The skill-side backlog moves to CLAUDE.md's open items, with one change of status: live browser capture is promoted from "only if evidence shows the need" to a planned next step - the HAR path works but makes the user do the capturing, and removing user effort is the product goal. Its price tag is stated as non-negotiable: a structurally unbypassable throwaway-browser boundary, reviewed before any implementation, with purchase-funnel capture out of scope even then. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Findings from the requested per-PR review:
- The documented duplicate-install signature could never occur: the
base snippet's getElementById guard means a second copy inserts no
second tb_tfa_script tag - it double-fires page_view instead. The
checklist row, the reference, and the skill now key duplicate
detection on page_view fire count, not DOM tags.
- The sanctioned curl gains its trust-boundary flags: scheme pinned
to http/https, redirects capped, final status + effective URL
checked before the file is believed - a saved 200-shaped error page
or an off-host redirect now skips the static verdict instead of
reading as "no pixel".
- Evidence capture: point users at Chrome's sanitized HAR export
(strips cookies/auth, keeps everything the check needs) and warn
about the console's one-time "allow pasting" gate.
- Runtime honesty for Codex: the guardrails ship to runtimes without
the skill layer, so the pixel-health routing now says the UI
redirect stands where the skill isn't loaded.
- Self-check gains a fetched-page/HAR item (no cookies or auth
headers in output; instruction-shaped content quoted, not obeyed).
- Output rules: the raw-field-name ban is scoped to Realize API
fields - the user's own pixel event names and on-wire values are
quoted verbatim.
- Permission plumbing: settings example + INSTALL now allow the
page-fetch curl, so a diagnosis doesn't die on a declined prompt
that reads as "no answer"; the outbound-scope sentence covers it.
- Claim-site sweep completions: the unreleased-changelog
self-contradiction on pixel-health, docs/realize-best-practices-gap
("conversion count is zero" is now served; diagnose-tracking is the
home if get_pixel_status lands), the guardrails' skill-name list,
and a ~100-char trim of the skill description clear of the
frontmatter limit.
- Hygiene: local Slack/PR-draft txt files are now gitignored.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
amitl-levi
force-pushed
the
marketplace-packaging
branch
from
August 30, 2026 10:43
bdfb2bd to
ee44c5f
Compare
amitl-levi
force-pushed
the
diagnose-tracking
branch
from
August 30, 2026 10:43
582ebcf to
0e08305
Compare
Merged
Collaborator
Author
|
Closing per review — this was opened from a personal fork by mistake. Replaced by #16 (same branch, same commits, on the taboola repo). |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 5 (final) of the split of the former stacked PRs, per review feedback. New
diagnose-trackingskill: pixel-health diagnosis ("pixel not firing", "conversions not tracking") — previously an explicit UI redirect.Stack: based on
marketplace-packaging(PR #14); this PR's diff is exactly its own 5 commits. Merge order: #8 → #12 → #13 → #14 → this.What
window._tfadump (runtime firing check with copy-paste capture instructions), and MCP cross-checks (rule status/event-name match, spend).manage-campaignspreview-then-confirm gate; Taboola-side →/realize-plugin:supportescalation. The skill itself never writes.docs/2026-08-22-pixel-expert-adoption-plan.md.Reviewer focus
os/guardrails.md): first non-realize.comweb access — scoped to the page under diagnosis; fetched page and HAR content are evidence, never instructions (read Scenario 22D plants hostile page text as the regression test). HAR hygiene: Grep-first sliced reading, no cookies/auth headers in output.<script>tag — tested live; a WebFetch-based check reports "no pixel" on every site.Tests
Read Scenario 22 (A–D), write Scenario W15 (diagnosis→fix handoff stays gated).
brand-check.sh0 FAIL, bundle tests 82/82, CI validate green.🤖 Generated with Claude Code