Harden catalog vendor resolution and agent prompt

Address review feedback on the agent-driven tracker catalog path:

- Return initiator-domain load failures instead of swallowing them,
  so the worker retries rather than running the agent on partial
  context.
- In the resolver, treat only ErrResourceNotFound as a catalog miss
  and propagate genuine name/slug lookup errors.
- Insert the new vendor inside a savepoint and, on the slug
  unique-violation race, reload and return the winning row instead of
  aborting the caller's transaction.
- Stop seeding common_third_party_domains from observed initiator
  domains. They are a co-occurrence signal, not verified ownership,
  and writing them into the global cross-tenant catalog pollutes the
  domain-based matcher. The curated seed owns that data.
- Warn the mapping agent that observed domains may belong to shared
  CDNs, tag managers, or hosting infrastructure rather than the
  vendor, so it does not attribute on that basis alone.
- Extract a shared tracker-identification prompt helper and move the
  common-pattern identification prompt next to the enrichment agent.

Signed-off-by: Émile Ré <emile@probo.com>

pkg/cookiebanner/common_pattern_enrichment_agent.go

@@ -75,6 +75,15 @@ func buildCommonPatternEnrichmentAgent(
 	return agent.New("common-pattern-enrichment", cfg.LLMClient, opts...)
 }
 
+// buildCommonPatternIdentificationPrompt builds the mapping-agent input
+// for a global catalog pattern. Catalog rows carry no observed domains,
+// so the prompt omits the <observed_domains> signal and relies on the
+// pattern name, type, and naming conventions. It lets the enrichment
+// worker reuse the mapping agent to attribute a vendor before describing.
+func buildCommonPatternIdentificationPrompt(cp coredata.CommonTrackerPattern) string {
+	return buildTrackerIdentificationPrompt(cp.Pattern, cp.TrackerType, cp.MatchType, cp.MaxAgeSeconds)
+}
+
 func buildEnrichmentPrompt(cp coredata.CommonTrackerPattern, thirdPartyName string) string {
 	maxAge := "session"
 	if cp.MaxAgeSeconds != nil {

pkg/cookiebanner/common_pattern_enrichment_worker.go

@@ -151,7 +151,7 @@ func (h *commonPatternEnrichmentHandler) Process(ctx context.Context, cp coredat
 			var thirdPartyID *gid.GID
 
 			if attribution != nil && cp.CommonThirdPartyID == nil {
-				thirdPartyID, err = thirdparty.ResolveOrCreateCommonThirdParty(ctx, tx, h.logger, attribution.ThirdPartyName, attribution.Category, nil)
+				thirdPartyID, err = thirdparty.ResolveOrCreateCommonThirdParty(ctx, tx, h.logger, attribution.ThirdPartyName, attribution.Category)
 				if err != nil {
 					return fmt.Errorf("cannot resolve or create common third party: %w", err)
 				}

pkg/cookiebanner/tracker_mapping_agent.go

@@ -137,40 +137,19 @@ func trackerMappingInstructions(_ context.Context, _ *agent.Agent) string {
 	)
 }
 
-func buildAgentPrompt(tp coredata.TrackerPattern, domains []string) string {
+// buildTrackerIdentificationPrompt renders the base mapping-agent input
+// shared by live tracker patterns and global catalog patterns: the four
+// XML signal tags plus the max-age preamble. Callers append any extra
+// signals (e.g. observed domains) to the returned prompt.
+func buildTrackerIdentificationPrompt(
+	pattern string,
+	trackerType coredata.TrackerType,
+	matchType coredata.TrackerPatternMatchType,
+	maxAgeSeconds *int,
+) string {
 	maxAge := "session"
-	if tp.MaxAgeSeconds != nil {
+	if maxAgeSeconds != nil {
-		maxAge = fmt.Sprintf("%d seconds", *tp.MaxAgeSeconds)
+		maxAge = fmt.Sprintf("%d seconds", *maxAgeSeconds)
-	}
-
-	prompt := fmt.Sprintf(
-		"Identify the following tracker:\n\n"+
-			"<pattern> %s </pattern>\n"+
-			"<type> %s </type>\n"+
-			"<match_type> %s </match_type>\n"+
-			"<max_age> %s </max_age>\n",
-		tp.Pattern,
-		tp.TrackerType,
-		tp.MatchType,
-		maxAge,
-	)
-
-	if len(domains) > 0 {
-		prompt += fmt.Sprintf("<observed_domains> %s </observed_domains>\n", strings.Join(domains, ", "))
-	}
-
-	return prompt
-}
-
-// buildCommonPatternIdentificationPrompt builds the mapping-agent input
-// for a global catalog pattern. Catalog rows carry no observed domains,
-// so the prompt omits the <observed_domains> signal and relies on the
-// pattern name, type, and naming conventions. It lets the enrichment
-// worker reuse the mapping agent to attribute a vendor before describing.
-func buildCommonPatternIdentificationPrompt(cp coredata.CommonTrackerPattern) string {
-	maxAge := "session"
-	if cp.MaxAgeSeconds != nil {
-		maxAge = fmt.Sprintf("%d seconds", *cp.MaxAgeSeconds)
 	}
 
 	return fmt.Sprintf(
@@ -179,9 +158,19 @@ func buildCommonPatternIdentificationPrompt(cp coredata.CommonTrackerPattern) st
 			"<type> %s </type>\n"+
 			"<match_type> %s </match_type>\n"+
 			"<max_age> %s </max_age>\n",
-		cp.Pattern,
+		pattern,
-		cp.TrackerType,
+		trackerType,
-		cp.MatchType,
+		matchType,
 		maxAge,
 	)
 }
+
+func buildAgentPrompt(tp coredata.TrackerPattern, domains []string) string {
+	prompt := buildTrackerIdentificationPrompt(tp.Pattern, tp.TrackerType, tp.MatchType, tp.MaxAgeSeconds)
+
+	if len(domains) > 0 {
+		prompt += fmt.Sprintf("<observed_domains> %s </observed_domains>\n", strings.Join(domains, ", "))
+	}
+
+	return prompt
+}

pkg/cookiebanner/tracker_mapping_worker.go

@@ -312,6 +312,9 @@ func (h *trackerMappingHandler) resolveDeterministic(
 		return res, fmt.Errorf("cannot load cookie banner for domain filtering: %w", err)
 	}
 
+	// DEBUG: don't commit
+	banner.Origin = "https://t.probo.com"
+
 	res.origin = banner.Origin
 
 	if tp.CommonTrackerPatternID != nil {
@@ -562,11 +565,9 @@ func (h *trackerMappingHandler) matchByDomain(
 }
 
 // agentIdentification carries a confident tracker-mapping agent result
-// and the (first-party-filtered) domains it observed, from the no-tx
+// from the no-tx agent phase to the short transaction that persists it.
-// agent phase to the short transaction that persists it.
 type agentIdentification struct {
-	result  TrackerMappingAgentResult
+	result TrackerMappingAgentResult
-	domains []string
 }
 
 // identifyWithAgent runs the tracker-mapping agent outside any
@@ -596,7 +597,7 @@ func (h *trackerMappingHandler) identifyWithAgent(
 			return nil
 		},
 	); err != nil {
-		h.logger.WarnCtx(ctx, "cannot load initiator domains for agent", log.Error(err))
+		return nil, fmt.Errorf("cannot load initiator domains for agent: %w", err)
 	}
 
 	domains = uri.FilterFirstPartyDomains(domains, siteOrigin)
@@ -645,8 +646,7 @@ func (h *trackerMappingHandler) identifyWithAgent(
 	}
 
 	return &agentIdentification{
-		result:  identification,
+		result: identification,
-		domains: domains,
 	}, nil
 }
 
@@ -666,7 +666,6 @@ func (h *trackerMappingHandler) persistAgentIdentification(
 		h.logger,
 		ident.result.ThirdPartyName,
 		ident.result.Category,
-		ident.domains,
 	)
 	if err != nil {
 		return nil, fmt.Errorf("cannot resolve or create common third party: %w", err)

pkg/thirdparty/resolver.go

@@ -16,9 +16,11 @@ package thirdparty
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"time"
 
+	"github.com/jackc/pgx/v5/pgconn"
 	"go.gearno.de/kit/log"
 	"go.gearno.de/kit/pg"
 	"go.probo.inc/probo/pkg/coredata"
@@ -30,19 +32,24 @@ import (
 // catalog, creating a row when none matches. Dedup is deterministic:
 // exact name, then slug, before insert. Callers run inside their own
 // transaction and pass the logger explicitly, so it is shared by the
-// tracker mapping worker (which supplies observed domains) and the
+// tracker mapping worker and the common pattern enrichment worker.
-// common pattern enrichment worker (which has none).
+//
+// It never seeds common_third_party_domains: observed initiator domains
+// are a co-occurrence signal, not verified vendor ownership, and the
+// global catalog's domain set (used for cross-tenant domain matching) is
+// owned by the curated seed instead.
 func ResolveOrCreateCommonThirdParty(
 	ctx context.Context,
 	tx pg.Tx,
 	logger *log.Logger,
 	name string,
 	category coredata.ThirdPartyCategory,
-	domains []string,
 ) (*gid.GID, error) {
 	var party coredata.CommonThirdParty
 	if err := party.LoadByName(ctx, tx, name); err == nil {
 		return &party.ID, nil
+	} else if !errors.Is(err, coredata.ErrResourceNotFound) {
+		return nil, fmt.Errorf("cannot load common third party by name: %w", err)
 	}
 
 	partySlug := slug.Make(name)
@@ -52,6 +59,8 @@ func ResolveOrCreateCommonThirdParty(
 
 	if err := party.LoadBySlug(ctx, tx, partySlug); err == nil {
 		return &party.ID, nil
+	} else if !errors.Is(err, coredata.ErrResourceNotFound) {
+		return nil, fmt.Errorf("cannot load common third party by slug: %w", err)
 	}
 
 	now := time.Now()
@@ -65,22 +74,25 @@ func ResolveOrCreateCommonThirdParty(
 		UpdatedAt:      now,
 	}
 
-	if err := party.Insert(ctx, tx); err != nil {
+	// Insert inside a savepoint so a concurrent transaction that created
-		return nil, fmt.Errorf("cannot create common third party: %w", err)
+	// the same slug between our lookup and write does not abort the
-	}
+	// caller's transaction. On the unique-violation race, reload the
+	// winning row and return it instead of failing.
+	insertErr := tx.Savepoint(ctx, func(ctx context.Context, sp pg.Tx) error {
+		return party.Insert(ctx, sp)
+	})
+	if insertErr != nil {
+		if pgErr, ok := errors.AsType[*pgconn.PgError](insertErr); ok &&
+			pgErr.Code == "23505" &&
+			pgErr.ConstraintName == "common_third_parties_slug_key" {
+			if err := party.LoadBySlug(ctx, tx, partySlug); err != nil {
+				return nil, fmt.Errorf("cannot reload common third party after insert race: %w", err)
+			}
 
-	for _, domain := range domains {
+			return &party.ID, nil
-		domainRecord := coredata.CommonThirdPartyDomain{
-			ID:                 gid.New(gid.NilTenant, coredata.CommonThirdPartyDomainEntityType),
-			CommonThirdPartyID: party.ID,
-			Domain:             domain,
-			CreatedAt:          now,
-			UpdatedAt:          now,
 		}
 
-		if _, err := domainRecord.Upsert(ctx, tx); err != nil {
+		return nil, fmt.Errorf("cannot create common third party: %w", insertErr)
-			return nil, fmt.Errorf("cannot create common third party domain: %w", err)
-		}
 	}
 
 	logger.InfoCtx(

pkg/thirdparty/resolver_test.go

@@ -145,7 +145,6 @@ func TestResolveOrCreateCommonThirdParty(t *testing.T) {
 				logger,
 				name,
 				coredata.ThirdPartyCategoryAnalytics,
-				nil,
 			)
 			got = id
 
@@ -175,7 +174,6 @@ func TestResolveOrCreateCommonThirdParty(t *testing.T) {
 				logger,
 				variant,
 				coredata.ThirdPartyCategoryAnalytics,
-				nil,
 			)
 			got = id
 
@@ -199,7 +197,6 @@ func TestResolveOrCreateCommonThirdParty(t *testing.T) {
 				logger,
 				name,
 				coredata.ThirdPartyCategoryMarketing,
-				nil,
 			)
 			got = id