The tracker-pattern catalog was binary (linked to a vendor or not), so
generic and first-party artifacts (loglevel keys, wallet-extension keys,
an org's own trackers) were retried forever and, once one row was wrongly
attributed, re-propagated to every organization with no re-check.
Give catalog rows a terminal attribution verdict (UNDETERMINED,
THIRD_PARTY, FIRST_PARTY): FIRST_PARTY short-circuits the whole mapping
pipeline so the artifact is never attributed again. Gate deterministic
vendor adoption behind a trust bar so only curated/operator rows
auto-propagate; lower-confidence agent/heuristic rows are reused as hints
and re-resolved, and an independent agent re-confirmation corroborates and
promotes them. Make the mapping agent emit an evidence source and reject
any attribution that lacks concrete evidence, and let it declare a
first-party verdict. Skip the speculative agent for PRE_EXISTING-source
patterns, whose low signal invites invented vendors.
Add proboctl "ctp mark-first-party" and an --attribution list filter to
audit and remediate existing wrong links, and a cursor rule documenting
migration naming so the timestamp is taken from date -u, not invented.
Signed-off-by: Émile Ré <emile@probo.com>
The enrichment agent now reads full cookie-database and vendor pages
through the browser. The facts it gathers (setter, purpose, data
stored) are not protectable, but a source page's written description
is original expression. Instruct the agent to state the facts in its
own words and never copy a sentence verbatim, so stored descriptions
stay clear of third-party prose.
Signed-off-by: Émile Ré <emile@probo.com>
The tracker-mapping and common-pattern enrichment agents only had
web search, which returns title/url/snippet, so they could never
open a cookie-database or cookie-policy page to read which vendor
actually sets a tracker. This mis-attributed setters whose snippet
is misleading (e.g. _li_* read as LinkedIn rather than LiveIntent).
Wire the read-only headless-browser toolset into both agents, gated
on a configured Chrome endpoint, mirroring the common-third-party
enrichment worker: agent construction moves into the run path so each
run can carry a per-run browser that is closed when the run returns.
The prompts now direct the agent to open a promising result and read
the named setter from the full page text. Both agents stay unchanged
when no Chrome endpoint is configured.
Signed-off-by: Émile Ré <emile@probo.com>
Cookie-database and consent-directory sites (Cookifi, Cookiepedia,
cookiedatabase.org, CookieServe, ...) rank highly in web search only
because they catalog cookies, not because they set them. The mapping
agent could attribute a tracker to the directory operator itself
instead of the vendor the page names.
Instruct the agent to treat such results as reference directories and
extract the named vendor, never the operator, while keeping a CMP's own
product cookie attributable (OptanonConsent -> OneTrust, CookieConsent
-> Cookiebot). Add a conservative code backstop that discards
attributions to pure aggregators, scoped to exclude CMP vendors so
legitimate own-cookie attributions survive.
Signed-off-by: Émile Ré <emile@probo.com>
Patterns like "ethereum-https://letaido.com" embed the scanned site's
own origin, appended either by a browser extension (e.g. a wallet
injecting window.ethereum) or by an owner-set tracker. The mapping
agent read that embedded domain and attributed the site to itself as a
third party ("Letaido"), but a site owner is never a third party of its
own site.
Pass the scanned site's eTLD+1 to the agent as a <scanned_site> signal
and instruct it to ignore a domain embedded in the pattern that matches
the site, returning low confidence when that is the only cue. Add a
deterministic backstop that discards any agent attribution whose vendor
name resolves to the site's own domain, so the pattern falls through to
the unmatched fallback regardless of whether the model honored the
prompt. Deterministic catalog, naming, and sibling signals run before
the agent and are untouched, so an independent strong vendor signal
still attributes normally.
Signed-off-by: Émile Ré <emile@probo.com>
The tracker-mapping and common-pattern enrichment agents ran with
4 and 3 max turns. Those budgets are smaller than the tool-call
rounds their prompts authorize (two/one DB searches plus up to
three web searches) plus the forced structured-output synthesis
turn, so the run loop tripped MaxTurnsExceededError before it
could emit JSON.
Raise the budgets to 10 across every layer that drives them: the
bootstrap env-var defaults, the Helm chart values and production
example, and the in-code fallback. Update the bootstrap test
defaults to match.
Revert the prompt instructions that told the model to reserve a
turn for its final answer; the model cannot reliably count its
remaining turns, so the fix belongs in the turn budget rather
than the prompt text.
Signed-off-by: Émile Ré <emile@probo.com>
The identification agent attributed probo_distinct_id to Mixpanel
purely on the shared distinct_id token, and the enrichment agent
returned no description for the glob ph_phc_*_posthog because it
searched the literal "*" string and found nothing.
Tighten the identification prompt so attribution requires a perfect
pattern match or a meaningful prefix that belongs to the vendor; a
generic token behind a different prefix is not a match. Teach the
enrichment prompt to strip wildcard and variable parts before
searching, and to treat a vendor name embedded in the key as
corroboration so clearly-named trackers still get a description.
Signed-off-by: Émile Ré <emile@probo.com>
The enrichment worker no longer invents a description when a tracker's
purpose cannot be substantiated; it records an empty description and
marks the row enriched so the stale-recovery loop does not retry it.
Vendor identification is the mapping pipeline's job, so the enricher
reuses the existing tracker-mapping agent to attribute a third party
for an unlinked common pattern before describing it. A confident
catalog match seeds the enrichment prompt and links the pattern, but
the enricher never creates or overrides an attribution.
When a blank, unlinked catalog row later gains a third party through
the mapping pipeline's upsert, enrichment is re-armed so the now-known
vendor gets a second, better-informed description attempt.
Signed-off-by: Émile Ré <emile@probo.com>
The agent returned a single confidence that conflated two unrelated
judgments: whether an artifact is a meaningful web tracker and which
vendor set it. The prompt's tracker-worthiness skepticism drove the
number down for extension state like __darkreader__wasEnabledForHost,
pushing it below the gate and dropping the attribution entirely, so a
clearly-named vendor never reached the catalog.
Rename the agent field to ThirdPartyConfidence and scope it to the
attribution alone. The identify gate now checks that a vendor is named
with sufficient confidence; on success the catalog row is stored at a
fixed agent confidence like the other heuristic signals, and on failure
the unmatched fallback still records the pattern with no third party.
The stored pattern confidence was only used for ordering and as agent
context, never as a gate, so a separate LLM-provided number is dropped
rather than split out.
Signed-off-by: Émile Ré <emile@probo.com>
Tracker descriptions were only filled on the agent-identification path,
so patterns resolved by domain, sibling, or fallback stayed without one,
and empty mapping upserts could clobber a researched description on the
shared catalog row.
Move description ownership to a dedicated, global common-pattern
enrichment worker. New catalog rows are queued on insert; the worker
researches a compliance-grade description with web search, records it on
the common pattern, and fans it out to every linked tracker pattern. The
mapping worker no longer generates descriptions and only propagates an
already-enriched one at link time.
Rename TrackerMappingConfig to TrackerAgentsConfig since the mapping and
enrichment agents now share it.
Signed-off-by: Émile Ré <emile@probo.com>
Tracker scripts loaded through a first-party reverse proxy (e.g.
t.probo.com proxying PostHog) share the scanned site's eTLD+1 and
were incorrectly matched against the site owner's own
CommonThirdParty entry in matchByDomain. This caused trackers like
ph_phc_* to be attributed to the site owner instead of PostHog.
Load the CookieBanner origin in Process and pass it to both
matchByDomain and identifyWithAgent. Both now filter out initiator
domains whose eTLD+1 matches the site before querying the catalog
or feeding domains to the LLM agent. The prompt is also updated to
warn about proxy domains.
Signed-off-by: Émile Ré <emile@probo.com>
Generate the category list from coredata.ThirdPartyCategories() at
runtime instead of hardcoding it in the prompt text. Type the
TrackerIdentification.Category field as coredata.ThirdPartyCategory so
JSON unmarshaling validates values automatically.
Also documents the .txt.tmpl template file naming convention.
Signed-off-by: Émile Ré <emile@probo.com>
Firecrawl provides higher quality search results than SearXNG.
When configured (firecrawl-endpoint + firecrawl-api-key), the
tracker-mapping agent and search toolset prefer it over the
SearXNG backend. Also improves the tracker identification prompt
with multi-strategy search queries that leverage domain signals
and adapt to tracker type.
Signed-off-by: Émile Ré <emile@probo.com>
When both pattern matching and domain matching fail to identify a
tracker, an opt-in LLM agent can now attempt identification using
internal database searches and optional web search. The agent returns
structured output (third party name, category, description, confidence)
and the worker auto-creates CommonThirdParty records when needed.
The feature is gated behind the `llm.tracker-mapping.provider` config
field; when unset the worker behaves exactly as before.
Signed-off-by: Émile Ré <emile@probo.com>
