A coding agent asked a typed graph of a real code base where the Stripe webhook is handled. The graph answered with five pointers and no file contents. This page replays that run step by step, exactly as it was recorded on 16 September 2026.
The recorded event keeps the four expressions, the answer and the pointers. It does not keep what each step returned, so the panel describes what the verb returns by contract and does not invent a payload. The navigator in this run is a hosted model inside the server, reached through codemap_ask; an agent can also run the same verbs itself through codemap_step, with no model in between.
The answer, in the navigator's words: “So the call chain is: StripeWebhookController → StripeWebhookHandler → SubscriptionService (+ SubscriptionEventRepository for event persistence).” And five pointers. A pointer is a name, a role, a subsystem, a size and a path. The agent opens these files and no others.
The same server, the same night, a question about something this code base does not contain.
“This question falls outside what this graph indexes. … there is no ‘recommendation engine’ subsystem in the L1 index … this graph has no pointer to offer.”
An answer that can say “not here” is one you can test. The locally trained navigator abstains on every out-of-graph question in its shipped build.
Counts from graph pack 1.1.1: a Spring Boot backend and an Angular frontend.
The five most common. The rare ones carry the most meaning: 6 TRIGGERS, 11 VALIDATES, 13 CONSTRAINS.
A hop on the graph is small. Three questions answered by hand: 1.1 KB of tool output against 26.9 KB of grep, 2.2 KB once the grep was refined.
At equal content similarity, two files joined by a typed edge change together 4 to 9 times as often as two without one.
Over 35 paired tasks, agents with the served graph used 1.6 times the tokens of agents with grep, 3.5 more turns and 88.5 seconds longer.
That night's log holds 166 requests for a prose answer and none that walk the graph step by step. The comparison that matters has not been run yet.