Raise default agent token budget for reasoning models

The tracker mapping, common-pattern enrichment, and third-party
disambiguation agents default to a small max-tokens budget on the
premise that their final output is tiny structured JSON. On
reasoning models such as the GPT-5 family, reasoning tokens count
against max_tokens, so a small budget is consumed by reasoning and
the JSON is truncated, surfacing as "unexpected end of JSON input".

Raise the defaults to 4096 (1024 -> 4096 for tracker mapping, 512
-> 4096 for disambiguation) to leave headroom for the reasoning
phase. Update the bootstrap builder default, its test, and the
production values example to match.

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

contrib/helm/charts/probo/values-production.yaml.example

@@ -174,7 +174,7 @@ probo:
   #   provider: "openai"
   #   modelName: "gpt-4o-mini"
   #   temperature: "0.1"
-  #   maxTokens: "1024"
+  #   maxTokens: "4096"
 
   # Tracker mapping worker tuning (optional; seconds for interval/agentTimeout).
   # Keep concurrency modest to stay under OpenAI/Firecrawl limits and the DB pool.

pkg/bootstrap/builder.go

@@ -209,11 +209,12 @@ func (b *Builder) Build() (*probodconfig.FullConfig, error) {
 				TrackerMapping: probodconfig.LLMAgentConfig{
 					Provider:  b.getEnvOrDefault("AGENT_TRACKER_MAPPING_PROVIDER", ""),
 					ModelName: b.getEnvOrDefault("AGENT_TRACKER_MAPPING_MODEL_NAME", ""),
-					// The tracker agents emit tiny structured JSON, so
-					// they default to a smaller token budget than the
-					// shared default rather than inheriting it.
+					// The tracker agents emit tiny structured JSON, but
+					// the budget must leave headroom for reasoning
+					// models whose reasoning tokens count against
+					// max_tokens; too small a budget truncates the JSON.
 					Temperature: b.getEnvFloatPtr("AGENT_TRACKER_MAPPING_TEMPERATURE"),
-					MaxTokens:   new(b.getEnvIntOrDefault("AGENT_TRACKER_MAPPING_MAX_TOKENS", 1024)),
+					MaxTokens:   new(b.getEnvIntOrDefault("AGENT_TRACKER_MAPPING_MAX_TOKENS", 4096)),
 				},
 				Tools: probodconfig.AgentToolsConfig{
 					FirecrawlAPIKey: b.getEnv("FIRECRAWL_API_KEY"),

pkg/bootstrap/builder_test.go

@@ -218,7 +218,7 @@ func TestBuilder_Build_Defaults(t *testing.T) {
 	assert.Empty(t, cfg.Probod.Agents.TrackerMapping.Provider)
 	assert.Empty(t, cfg.Probod.Agents.TrackerMapping.ModelName)
 	assert.Nil(t, cfg.Probod.Agents.TrackerMapping.Temperature)
-	assert.Equal(t, new(1024), cfg.Probod.Agents.TrackerMapping.MaxTokens)
+	assert.Equal(t, new(4096), cfg.Probod.Agents.TrackerMapping.MaxTokens)
 
 	// Tracker worker tuning — defaults
 	assert.Equal(t, 10, cfg.Probod.TrackerMappingWorker.Interval)

pkg/cookiebanner/tracker_mapping_agent.go

@@ -40,10 +40,14 @@ const (
 	defaultMappingMaxTurns    = 4
 	defaultEnrichmentMaxTurns = 3
 
-	// defaultAgentMaxTokens caps the structured output of the mapping
-	// and enrichment agents when the agent config carries no max-tokens
-	// budget. Both outputs are tiny structured JSON.
-	defaultAgentMaxTokens = 1024
+	// defaultAgentMaxTokens caps the output of the mapping and
+	// enrichment agents when the agent config carries no max-tokens
+	// budget. Both final outputs are tiny structured JSON, but the
+	// budget must leave ample headroom for reasoning models (e.g. the
+	// GPT-5 family): their reasoning tokens count against max_tokens,
+	// so too small a budget gets consumed by reasoning and truncates
+	// the JSON, surfacing as "unexpected end of JSON input".
+	defaultAgentMaxTokens = 4096
 
 	agentThirdPartyConfidenceThreshold = 0.6
 	// agentSourceConfidence is the fixed confidence stored on catalog

pkg/thirdparty/disambiguation_agent.go

@@ -44,10 +44,14 @@ const (
 	// provider, not a real budget.
 	defaultDisambiguationTimeout = 45 * time.Second
 
-	// defaultDisambiguationMaxTokens caps the agent's structured
-	// output when the config carries no max-tokens budget. The output
-	// is a single id plus a one-sentence rationale.
-	defaultDisambiguationMaxTokens = 512
+	// defaultDisambiguationMaxTokens caps the agent's output when the
+	// config carries no max-tokens budget. The final output is tiny (a
+	// single id plus a one-sentence rationale), but the budget must
+	// leave ample headroom for reasoning models (e.g. the GPT-5
+	// family): their reasoning tokens count against max_tokens, so too
+	// small a budget gets consumed by reasoning and truncates the JSON,
+	// surfacing as "unexpected end of JSON input".
+	defaultDisambiguationMaxTokens = 4096
 )
 
 // DisambiguationConfig configures the third-party disambiguation