Files
archie/docs/prds/code-graph-mcp.md
T
sam 2eb900dabd docs(prds): design the code graph as an MCP surface
Supersedes symbol-reference-index.md, which scoped only the data layer and
left the interface unstated. Deleted rather than archived: it was written
today, never implemented, and keeping both would leave two documents claiming
authority over one capability.

Three things the narrower doc did not settle.

The interface is MCP. archied consumes MCP and serves none, so exposing the
graph that way reaches Claude Code, pi, codex and Archie's own agents through
one surface with no coupling to the daemon. Diagrams are deliberately not a
tool: a model handed an accurate subgraph writes its own Mermaid, and the
valuable half was never the rendering.

Every answer declares what it could not see. RTA over-approximates interface
dispatch and reflection is invisible, so the *extract yaegi tables make
symbols look reachable regardless. An agent trusting a confident wrong graph
is worse off than one that greps.

A code graph does not power a documentation chatbot. It answers structural
questions about Go symbols; answering "how do I configure a second identity"
means retrieving prose from markdown. I had previously implied one substrate
served both. It does not, and building it that way would serve neither well.
2026-09-18 13:42:33 +10:00

5.3 KiB

Code graph over MCP (tools/codegraph)

Status: proposed. Supersedes symbol-reference-index.md, which scoped only the data layer.

The problem

Establishing what a change touches currently costs an agent eight to fifteen greps, and the answers are wrong in specific ways. grep -rn "Register(" cannot separate plugin.Registry.Register from captureintake.Receiver.Register from forgerpc.Server.Register, all of which exist here. It also matches the sibling checkouts under .worktrees/, which are near-complete copies of the module, so every count is silently inflated.

That cost lands hardest in investigation, triage and the TDD loop, where the first question is always "what does this touch".

Decision: exact analysis, not semantic retrieval

Tools like Greptile parse with tree-sitter, summarise each node, embed the summaries, and retrieve semantically. That design exists because they must serve any language. This repository is one Go module, and Go's type checker is ground truth:

Question Source of truth
Who references this symbol? go/packages + go/types, resolved by identity
What calls what? callgraph RTA over SSA
What imports what? go list -deps

No vector database, no embedding cost, no edges invented by a summary that drifted from the code. For a single-language repository this is both cheaper and strictly more accurate. Copy the interface those tools expose, not their architecture.

A code graph does not power a documentation chatbot. It answers structural questions about Go symbols. Answering "how do I configure a second identity" means retrieving prose from markdown, which needs chunking, embedding and similarity search over a different corpus entirely. That remains a separate capability with its own design. Conflating the two would build one system that does neither well.

Interface: an MCP server

archied already consumes MCP as a client (internal/app/agentworker starts providers) and serves none. Exposing the graph over MCP reaches Claude Code, pi, codex and Archie's own agents through one surface, with no coupling to the daemon: it is configured like any other entry in [[mcp_servers]].

Four tools, each returning JSON with file:line citations:

  • references(symbol) — every reference site, classified prod / test / generated / dynamic
  • callers(symbol, depth) — inbound call edges
  • callees(symbol, depth) — outbound call edges
  • blast_radius(file | symbol) — the transitive neighbourhood a change reaches, which is the single call meant to replace the grep sequence

Diagrams are deliberately not a tool. A model given an accurate subgraph writes its own Mermaid; a diagram endpoint would only fix the rendering of something the model already has.

Decisions

Every answer declares what it could not see. RTA over-approximates interface dispatch, and reflection is invisible: every reflect.ValueOf entry in wfextract, gateextract, pluginextract, secretextract and logextract makes a symbol look reachable whether or not anything calls it. Each response carries a dynamic_edges field naming those boundaries. An agent trusting a confident wrong graph is worse off than one that greps, so the graph states its own limits rather than implying completeness.

Warm in memory, keyed on HEAD, never persisted to disk. deadcode does whole-program RTA in under seven seconds; this does less work. The server rebuilds when HEAD changes and refuses its cache outright when the tree is dirty. Nothing is written to disk, because a persisted index becomes a rendered view that outlives its evidence. That is the failure that made STATUS.md report 90 of 91 findings re-verified while most statuses were 77 commits stale, and the same one that made reachaudit serve a previous tree's analysis from /tmp until its cache was keyed on HEAD.

tools/codegraph, not internal/. It is developer tooling, not daemon runtime, so it belongs beside docsgen, contractaudit and reachaudit in the tools module, which task test:tools already covers. Start as a single Go file; split only when it earns it.

No new dependency. golang.org/x/tools is already in go.sum at v0.49.0.

How it is verified

Red first. Table-driven tests against small fixture modules built in t.TempDir(), asserting the behaviour that distinguishes this from grep:

  1. Two methods spelled Register on different types resolve to two separate records, and a reference to one does not increment the other.
  2. A symbol referenced only from _test.go reports zero prod references.
  3. A symbol referenced only from a reflect.ValueOf table reports zero prod references and one dynamic, and names the table in dynamic_edges.
  4. A generated file's references land in generated, not prod.

Mutation-verify each: reintroduce the conflated behaviour and watch the assertion fail.

Acceptance is a live cross-check rather than a unit test. Run it against this tree and confirm it reports zero production references for worktree.Manager.Resume, webui.NewDangerousService and secret.DefaultRegistry. All three were confirmed dead by hand, and deadcode either stayed silent on them or, in the case of a package var, cannot represent them at all. Agreeing with deadcode is not the bar. Catching what deadcode structurally cannot is.