# The project brain: a navigable evidence map

Use this when the user asks how work connects, wants to explore impact, or needs a graph-informed handoff. The map is an index and navigation aid; authoritative Zen records remain Markdown. A graph cannot prove that work was executed, that a check passed or that a model knows the entire project.

## Assemble a useful, bounded view

Start with the selected workspace, question and access scope. Use `zen_state` to find record metadata and relationships, then `zen_record_get` or `zen_research_search` for only the relevant records. A workspace filter narrows results; it is not a security boundary for an agent credential that can reach registered workspaces. Never expose never-copy records, raw credentials or inaccessible sources. Mark local-only and stale material appropriately. Do not use private material to produce a public graph.

Present meaningful nodes such as an idea, decision, specification, task, document, artifact selection, feedback, memory, session, milestone, check, source file or actually installed skill. Show the relationship and an inspectable source/reference for each connection. Where available, show revision, timestamp, owner/agent origin and evidence. Distinguish an explicit recorded link from an extracted code relationship, a proposed or inferred relationship, and an unresolved link. A missing link is unknown, not evidence that no relationship exists; do not invent a skill-to-task link merely because both exist.

Begin with a readable local neighborhood and let the user search or expand it. Provide a list or structured text alternative to the graph, with concise explanations and links to underlying records. For a change, surface affected tasks, decisions and evidence; identify stale/unverified material and which conclusion needs review. Show exactly which sources support an answer. An inferred edge is a lead to investigate, not a decision or factual claim.

## Optional Graphify enrichment

The referenced project is [Graphify-Labs/graphify](https://github.com/Graphify-Labs/graphify). Its documented deterministic code parsing can run locally; semantic extraction of documents or media may use a selected model or configured external service. The Graphify index, configuration, source selection and generated graph files can contain private project data. Verify the installed version, runtime, source scope, provider and actual data flow before invoking it. No account or remote service is required merely to use Zen's records. Do not install a package or send source content to a model/service without the applicable user authorization.

If an appropriate local Graphify integration exists, select the intended workspace, source revision and permitted folders, then use the discovered installed capabilities to build or query a bounded graph. Preserve Graphify's actual evidence/provenance tags, exact source references, parse errors and index revision. Join imported nodes to Zen by stable record ID or an explicit verified mapping; do not infer an identity match solely from a title. Treat graph data as a rebuildable projection, never a second writer or source of truth. A changed source, missing index or unavailable adapter is stale, unavailable or unknown—not an empty definitive map.

If Graphify is absent, say so and use the existing Zen record links or a clearly labelled manual sketch. Do not invent a Graphify command result, a Zen graph MCP tool or automatic indexing. When discussing source or licensing, inspect the installed upstream version and applicable files; the current project publishes Apache-2.0 and an additional MIT license file. Do not embed third-party source without its required notices.