A living map of your system — and of what the AI does to it
The Blueprint is a derived graph of your codebase — modules, endpoints, external integrations, datastores — rebuilt from code on every derive. Never hand-maintained, never stale by design. Human and agent knowledge attaches to it through append-only, provenance-carrying layers. Five mechanisms make AI work legible on it.
A real board — 27 components: modules, exposed endpoints, datastores, and external integrations. Blue solid wires are build-time imports, amber dashed are runtime calls, green are exposed surfaces; the dots are per-module learning state.
Before
Ghost proposals
proposals · ghost nodes
During discovery and planning, components the plan predicts it will create are registered as ghost nodes on the board — dashed, translucent, unmistakably not-yet-real. You see the planned shape of the system before a line is written.
Confirmation is mechanical, never narrated. When a later derive finds a real node matching the predicted id, the ghost becomes solid. When the plan is archived, unconfirmed ghosts are swept. The AI cannot claim it built something — the extractors have to find it in code.
Before
Change simulation
simulation · phase 2 of 5
A plan compiles into a phase-by-phase walkthrough you can scrub. Each phase highlights the modules it touches with MODIFIED / ADDED badges directly on the board — blast radius, reviewed phase by phase, before you approve execution.
The badges are honest: they come from the plan's declared files and anchors, and the same decoration engine later shows what actually changed.
During
The story layer
story · step 3/5 — the data flow
Every plan narrates a concept arc — problem, shape, data flow, decisions and rejected alternatives, risks — in at most five steps. The story compiles against the graph with strict anchor resolution: a step cannot reference a component that does not exist or is not legitimately predicted.
It renders as an interactive walkthrough card in the chat, pinned at the moment the plan was created, with two-way board sync: stepping the story decorates the board; clicking a board node reveals its step.
After
Drift detection
drift · 2 modules moved
Flow keeps a content-hash baseline of every code file. When reality moves — a teammate's merge, an agent's edits, your own changes — the board shows exactly which modules drifted from the indexed understanding, and the learning layer flags what needs re-learning.
Baselines advance only when you acknowledge a review, never silently. Timestamps are never trusted; content is.
Always
Provenance & trust tiers
provenance · anchored records only
Annotations enter the map only by promotion of anchored execution records — a decision recorded during an actual run, carrying its run id. An agent cannot write “I refactored the auth module” onto the map as fact. Contradicting notes are refused unless they corroborate or explicitly supersede with a reason.
Knowledge provenance is tiered: code is ground truth; Flow's own store is verified knowledge; foreign docs found in the repo are unverified claims until corroborated against code — a built-in doc-rot radar instead of silent poisoning. The trust model is enforced in one code path, not in prompts.