AGENTS.md

AGENTS.md — codegraph-rs

A deterministic tree-sitter + SQLite/FTS5 code knowledge graph: it parses a codebase, extracts symbols and their relationships, persists them to a per-project SQLite database (with an FTS5 search index), and exposes the result through a CLI and an MCP (Model Context Protocol) stdio server. No AI / vector / LLM anywhere in the binary — output is byte-stable.

Hard invariants (never break)

• Golden .schema byte-stability — verified by crates/codegraph-bench/tests/equivalence.rs against the fixed golden artifacts under reference/golden/. Fixtures: the existing upstream corpus plus reference/golden/godot/ (corpus crates/codegraph-bench/fixtures/godot/; guards F1 autoload-call edges + F2 signal-handler edges + UID-form autoloads — a sidecar-UID script autoload (*.gd.uid) and a header-UID scene autoload (.tscn uid=) — byte-for-byte) and reference/golden/ruby/ (corpus crates/codegraph-bench/fixtures/ruby/; guards #1110 Ruby receiver.method extraction — instance/class-method Calls, Const.new Instantiates, bare include Implements — byte-for-byte), reference/golden/python/ (corpus crates/codegraph-bench/fixtures/python/; guards six bare class-as-value References edges — return, assignment RHS, registry pair, call argument, list literal, and one cross-file unique match — plus tuple-return and bare-method negatives, byte-for-byte), reference/golden/kotlin/ (corpus crates/codegraph-bench/fixtures/kotlin/; guards #1495 Kotlin callable signatures for explicit, inferred, generic, multiline nullable, and extension returns, plus a primary-constructor negative — byte-for-byte), and reference/golden/cpp/ (corpus crates/codegraph-bench/fixtures/cpp/; guards #1043 C++ class/struct inheritance incl. templated-base stripping byte-for-byte, and retroactively the earlier C++ extraction work). The Python cross-file edge is resolved by Gate 3a's unique-name fallback; import Gate 3b is deliberately unreachable for real Python nodes because the extractor does not mark them exported. Regen recipe: docs/equivalence.md "Godot fixture" / "Ruby fixture" / "Python fixture" / "Kotlin fixture" / "C++ fixture" sections.
• node-id formula: {kind}:{sha256("{filePath}:{kind}:{name}:{line}").hex[:32]}; file nodes are the literal file:{relpath}; lines are 1-based; paths relative with /.
• No AI / vector / LLM crates — enforced by scripts/guardrail.sh (CI gate): no surrealdb / rig / qdrant / lancedb / candle / onnx / ort.
• Deterministic extraction + resolution; sync output must equal index --force byte-for-byte.

Workspace layout (10 crates)

codegraph-core (types/config/logger) · codegraph-store (SQLite+FTS5) · codegraph-extract (tree-sitter walker + embedded + custom extractors; incl. C++ base_class_clause → Extends inheritance extraction with templated-base stripping, #1043) · codegraph-graph (traversal + FTS search) · codegraph-resolve (import + name matcher + FrameworkResolver; concrete GodotResolver impl — autoload-call + signal-handler resolution) · codegraph-mcp (stdio JSON-RPC; codegraph_explore runs a change-surface rescue, #1064, that surfaces a callable's buried parameter/return-type files into the explored subgraph; every source-emitting path goes through a request-memoized freshness probe — size + ms-mtime, then sha256 on stat mismatch — so a drifted file is served whole under FILE_MODE_MAX_LINES or omitted, never sliced at stored line numbers, and the response carries the staleness banner only when it actually cites one) · codegraph-cli (single binary, owns logger; also hosts the install/uninstall agent-config installer in src/installer/) · codegraph-daemon · codegraph-watch · codegraph-bench (benchmark harness + golden oracle).

The published crate is codegraph-rs (the codegraph-cli package); the installed binary is codegraph. The library crates publish as codegraph-{core,store,extract,graph,resolve,mcp,daemon,watch}. codegraph-bench is publish = false.

Godot framework resolver (codegraph-resolve)

The GodotResolver is the first concrete FrameworkResolver impl. It fires on GDScript files and synthesizes edges that tree-sitter alone cannot produce. Three behaviors are active:

• F1 — autoload-call→func edges: a call Autoload.method() in a .gd file emits a Calls edge to the UNIQUE same-named func in the autoload's bound target script (binding read from project.godot [autoload] section, Name="*res://path.gd" form only). Determinism rule: edge built ONLY when exactly one matching func exists in that script; 0 or ≥2 matches → no edge. Files: crates/codegraph-resolve/src/frameworks/godot.rs, crates/codegraph-resolve/src/frameworks/godot_script.rs.

• F2 — signal handler extraction: connect_handler now extracts handlers from .connect(_h.bind(x)) (head segment before .bind() and Callable(self,"h")/Callable(this,"h") forms, in addition to bare .connect(_h). Other receivers, variable handlers, or non-literal method names stay dynamic sentinels (unresolved). File: crates/codegraph-resolve/src/frameworks/godot_script.rs.

• F3 — impact/affected↔audit unification: codegraph impact (file-node targets) and codegraph affected now also consume path-keyed unresolved_refs restricted to Godot ReferenceSubkinds (script_attach, ext_resource, scene_instance, group_member, signal_method, autoload), so their output agrees with codegraph audit --impact. Query-side only; zero extraction change. New function: dependent_file_paths_unresolved in crates/codegraph-store/src/queries.rs; CLI wired in crates/codegraph-cli/src/main.rs. codegraph affected additionally emits an affectedFiles key — the sorted+deduped union of every traversed dependent plus the test-file set (affectedFiles ⊇ affectedTests) — so it LISTS the complete affected set instead of only counting it via totalDependentsTraversed. Additive: changedFiles, affectedTests, totalDependentsTraversed are byte-for-byte unchanged.

Full Godot static-analysis scope, static-vs-runtime boundary, and honesty signals: docs/godot.md.

MCP protocol surface (codegraph-mcp, rmcp 3.0.1)

codegraph-mcp builds on rmcp 3.0.1 (crates/codegraph-mcp/Cargo.toml, dep and dev-dep; features server, client, transport-io, transport-streamable-http-server). The 2.1 → 3.0.1 upgrade cost three lines in rmcp_handler.rs: call_tool returns the #[non_exhaustive] CallToolResponse enum instead of CallToolResult (we only ever build Complete, via .into()), with_stateful_mode became with_legacy_session_mode, and a stale get_info() comment was corrected. The 3.x breaking changes that bit third-party code were all CLIENT-side (InitializeResult → ServerPeerInfo, optional server_info, OAuth resolve_metadata()), none of which touch a server ServerHandler. Peer::list_roots is still #[deprecated] (SEP-2577) with still no replacement, so the #[allow(deprecated)] at rmcp_handler.rs:334 stays.

We serve five protocol revisions, and that is the SDK default rather than a choice. ProtocolVersion is struct ProtocolVersion(Cow<'static, str>) plus associated constants (NOT an enum); rmcp's negotiate_protocol_version echoes back any client-requested version present in KNOWN_VERSIONS. So get_info()'s .with_protocol_version(V_2024_11_05) is only the FALLBACK for an unknown client version — it is not a ceiling, and we do not pin or force 2024-11-05:

client requests	negotiated	resultType in results	HTTP session
2024-11-05	2024-11-05	absent	no header (our config)
2025-03-26	2025-03-26	absent	no header (our config)
2025-06-18	2025-06-18	absent	no header (our config)
2025-11-25	2025-11-25	absent	no header (our config)
2026-07-28	2026-07-28	"complete"	no header (per spec)

resultType (SEP-2322) is absent for pre-2026 peers because upstream #1038 strips it for legacy peers, and 2026-07-28 is stateless per SEP-2567 — no Mcp-Session-Id — unconditionally, regardless of with_legacy_session_mode, which now governs legacy versions only. We pass with_legacy_session_mode(false) (rmcp_handler.rs:737), so the legacy versions issue no Mcp-Session-Id either: THIS server never sends one at any revision. The two absences differ in cause, though — legacy is our configuration and would come back if that flag flipped to true, while 2026-07-28 is mandated and cannot be turned back on. SEP-2243 standard headers are validated in both directions: matching MCP-Protocol-Version / Mcp-Method / Mcp-Name reach the tool, while a missing protocol header or a mismatched Mcp-Method / Mcp-Name is rejected with HTTP 400 + JSON-RPC code -32020. Coverage lives in tests/rmcp_l3.rs (version echo, resultType presence/absence) and tests/rmcp_http.rs (no session header, SEP-2243 both branches).

Not implemented, deliberately: Tasks (SEP-2663), MRTR / elicitation-in-tool, and subscriptions. We advertise only the tools capability and always return CallToolResponse::Complete; we never construct Task or InputRequired. accepted_subscription_filter keeps its None default, while legacy subscribe / unsubscribe return method-not-found. The request surface is not literally tools-only, however. Beyond the mandatory initialize handshake — itself an inherited default (handler/server.rs:314-326), which is why negotiation runs automatically over our get_info() and why the V_2024_11_05 there is only a fallback — these inherited defaults also answer Ok. discover returns real content (supported_protocol_versions() plus get_info()); complete returns an empty CompleteResult; list_prompts / list_resources / list_resource_templates (handler/server.rs:362-383) return EMPTY results, so prompts/list, resources/list, and resources/templates/list answer with nothing rather than method-not-found; and ping succeeds on the legacy revisions only (method-not-found at 2026-07-28, handler/server.rs:112-118). The set_level, get_prompt, read_resource, and task-lifecycle defaults return method-not-found. These are SDK defaults, not our implementations, and they add no advertised capabilities.

MCP golden fixtures are STRUCTURAL, not byte-stable. tests/support/parity.rs:244-300 and tests/golden_mcp.rs compare only NAMED fields — initialize: protocolVersion, capabilities, serverInfo.name, serverInfo.version, instructions; tools/list: names + order + inputSchema + annotations; tool result: content[0].type, isError, and the sorted set of non-empty text lines; error: code + message — never whole-object equality. That is why an added top-level resultType could not drift the 15 existing fixtures, and equally why it needed its own explicit test. (The EXTRACTION goldens under reference/golden/ are a different contract and genuinely ARE byte-stable — see "Hard invariants" above.)

HTTP MCP server: background mode + addr-keyed registry

serve --mcp (stdio) uses the PER-PROJECT daemon (.codegraph/daemon.pid + socket; the whole rendezvous is derived from IndexPaths::current_root). serve --http (streamable-HTTP) is different: HTTP servers are keyed by BIND ADDR — a global server (no --path) spans many projects — so they use a GLOBAL, addr-keyed registry, NOT the per-project root. The registry lives in codegraph-daemon/src/http_registry.rs: one <addr-sanitized>.json file per running server (HttpServerInfo { pid, addr, mode, project, started_at, version, log_file }) under $XDG_STATE_HOME/codegraph/http (else ~/.local/state/codegraph/http; %LOCALAPPDATA%\codegraph\http on Windows; CODEGRAPH_HTTP_REGISTRY_DIR overrides). Entries are pruned when their pid is dead (self-heal, gated on is_process_alive).

serve --http stays FOREGROUND by default; serve --http --detach runs in the BACKGROUND via spawn::spawn_detached_http (generalized from spawn_detached_daemon over the shared detach() primitive; the child carries CODEGRAPH_HTTP_DETACH_INTERNAL=1 so it runs the foreground serve path and does NOT re-detach). On startup serve --http prunes dead entries, ERRORS on a live same-addr conflict (listing the running instance), and notes any other live servers when the addr is free. The codegraph http {list, status, stop} subcommand group inspects and terminates registered servers (stop uses process::terminate_pid — SIGTERM on unix / TerminateProcess on Windows). None of this touches extraction/golden equivalence.

Foreground stdio serve --mcp gets a THIRD, PID-keyed registry (codegraph-daemon/src/mcp_registry.rs): one <pid>.json (McpServerInfo { pid, project: Option<String>, transport: "stdio", started_at, version }, camelCase on the wire) under the same GLOBAL state chain as the HTTP one but with an mcp leaf ($XDG_STATE_HOME/codegraph/mcp, else ~/.local/state/codegraph/mcp, %LOCALAPPDATA%\codegraph\mcp on Windows; CODEGRAPH_MCP_REGISTRY_DIR overrides). PID keying is forced by the transport: a stdio process has no addr and no per-project rendezvous, several may serve one project, and one may serve none. Registration fires from all THREE foreground exits in cmd_serve (Direct, SpawnOrProxy, and the too-broad-root home guard) and NEVER from BeDaemon, which already owns .codegraph/daemon.pid. Reads go through RegistryRead::{Available, Unavailable} so a MISSING directory ("nobody registered yet", normal) is distinguishable from an unreadable one (an outage); read_dir's NotFound only means MISSING when fs::symlink_metadata also fails, so a dangling symlink at the registry path reads as an outage instead of an empty registry. list_entries is the RAW on-disk view (stale entries included); live_entries filters its RETURN VALUE by is_process_alive rather than trusting prune_dead's deletions to have landed, so a dead entry on an undeletable (read-only) registry is still never reported as running — dead-PID and unparseable files are pruned on read as best-effort disk self-heal only. project records the launch --path and ONLY that: it is Some only when the user actually passed --path (a bare serve --mcp stores None, not its cwd), and it is purely INFORMATIONAL — never a capability boundary, because roots::resolve_project_arg probes an absolute per-call projectPath on its own merits and consults the launch default only when no path was passed, so any live server can be asked to open any indexed project's database. codegraph mcp list [--json] renders it as LAUNCH PROJECT, and BOTH holder diagnostics — index's pre-warning and the RemoveDatabase FAILURE path — therefore report ALL live entries with no narrowing: filtering by project would drop the holder in the very case they exist for (a server launched elsewhere that a client has since pointed at this project), and the failure path additionally has only a DB path. CODEGRAPH_DIR selects one validated project-local directory name, so the database remains inside the project tree. This registry is PURE OBSERVABILITY — there is no mcp stop and no terminate_pid import, because a stale entry's PID may have been reused and this workspace has no portable instance-identity primitive (try_acquire_daemon_lock is a create_new placeholder plus a recorded PID, not an OS advisory lock); list asks a human to confirm the PID with ps -p <pid> -o command= / tasklist /FI "PID eq <pid>" before offering kill <pid> / taskkill /PID <pid> /F. CLI/daemon only; extraction and golden equivalence untouched.

Agent installer (codegraph install / uninstall)

codegraph install writes the codegraph MCP-server entry into each supported agent's config (Claude Code, Cursor, Codex CLI, opencode, Hermes Agent, Gemini CLI, Antigravity IDE, Kiro, Trae, Qoder, Zed); uninstall reverses it. The written command launches the binary (command: "codegraph", args: ["serve", "--mcp"]). Cursor and Trae use --path ${workspaceFolder} in their global config so one entry auto-follows each project window; Kiro and Qoder write a bare global entry (no --path) that serves tools read-only off any existing index, with the agent passing the project path per call — run codegraph init --target=<ide> inside each project to write a project-local config with an absolute --path for live watch. Kiro's mcp.json also carries a //-commented HTTP alternative alongside the active stdio entry (JSONC, idempotent, injected best-effort without corrupting existing files); it uses http://localhost:8111/mcp because Kiro allows http only for localhost (remote servers must be https). Zed's settings.json likewise carries //-commented remote-development alternatives after the active context_servers.codegraph stdio entry (both install global and init project-local): an SSH-stdio bridge and an HTTP server (http://localhost:8111/mcp, marked RECOMMENDED for remote); the shared JSONC-safe injector is inject_commented_alternative(path, parent_key, entry_key, sentinel, block) in shared.rs, used by both Kiro (mcpServers) and Zed (context_servers). Non-interactive, flag-driven (--target, --global/--local/--location, --yes, --no-permissions, --print-config); the config-writing logic (paths/keys/marker sections, idempotent upsert, uninstall removal) is CLI-only and additive — it does NOT touch extraction/golden equivalence.

Verification gates (run before every commit)

make ci          # fmt-check + clippy + test + guardrail
# or individually:
cargo test --workspace          # incl. golden oracle + sync equivalence
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all --check
bash scripts/guardrail.sh
make coverage    # workspace coverage summary (informational; `make coverage-html` for the full report)

Test coverage (tracked, informational)

• Unit-test coverage is a tracked metric via cargo llvm-cov + Codecov. Run make coverage for a summary; make coverage-html for the browsable report; make coverage-lcov writes the lcov.info CI uploads.
• Target is 95%+ (aspirational). The CI gate is informational / non-blocking — the coverage job is kept out of the CI Success gate and the Codecov status is informational: true (codecov.yml), so a below-target % never turns CI red. This honors the iron rule "local green ⇒ CI green".
• Measured 94.42% line / 94.12% region / 93.41% function — close to the 95% target but not yet there. Figure measured at 4fabc0b (rmcp 3.0.1 round) via make coverage; re-measure before quoting it, since it moves with every round. The remaining gap is now concentrated in the per-language extractors rather than the resolver or watch internals: 7 of the 10 lowest-covered files are codegraph-extract/src/lang/* — r.rs 70.59%, dart.rs 72.75%, erlang.rs 75.77%, arkts.rs 81.01%, nix.rs 82.21%, jsx.rs 82.61%, lua.rs 85.94% — alongside codegraph-watch/src/migrate.rs 74.77%, codegraph-store/src/index_state.rs 80.45%, and codegraph-store/src/connection.rs 82.57%.
• Enabling Codecov: enable the repo at codecov.io. This repo is public, so tokenless upload works (no CODECOV_TOKEN needed); a private repo would need CODECOV_TOKEN in GitHub repo Secrets.

CI, hooks & release

• Pre-push hook (.githooks/pre-push): runs fmt + clippy + test + guardrail on git push (never on commit). Enable once per clone with make hooks (sets core.hooksPath). Local green ⇒ CI green.
• CI (.github/workflows/ci.yml): Test (fmt/clippy/test/guardrail) + Security Audit (cargo-audit) + CI Success gate, on push/PR to main.
• Release (.github/workflows/release-please.yml): release-please opens a release PR; merging it cuts a v<version> tag and triggers the pipeline — 6-platform binaries (linux musl x86_64/aarch64 via cargo-zigbuild, macOS x86_64/aarch64, and Windows MSVC x86_64-pc-windows-msvc on windows-latest plus aarch64-pc-windows-msvc on windows-11-arm), git-cliff release notes, and a GitHub Release with the binaries attached. The project is distributed via GitHub Releases + cargo install --git; it is NOT published to crates.io. Version bumps are owned by release-please via .release-please-manifest.json — never bump by hand.
• Commits are English Conventional Commits. feat→minor, fix→patch. The end-to-end release runbook lives in the codegraph-release skill.
• Docs are formatted with oxfmt (make fmt); .oxfmtignore excludes golden fixtures, embedded JSON, and auto-generated files.