io.github.cdeust/automatised-pipeline

---

Every AI coding assistant hits the same wall: you ask it to change handle_tool_call, and it either hallucinates a function that was renamed last week, edits something in the wrong community of the codebase, or silently breaks a call chain three modules away. Agents operate on strings; codebases have structure. The gap is where bugs live.

This server gives it the structure. It parses your repository into a real graph of symbols and their relationships, then answers structural questions directly — with source links, and with its own limitations stated alongside the answer.

What you get

• Real answers about your code. Who calls this function, what breaks if I change it, how does execution reach here, what belongs together — resolved across files, not guessed from text matches.
• Your agent stops re-reading files. One graph query replaces the open-five-files-and-scroll loop. On our offline evaluation that is 14.26× fewer payload tokens and 5.20× fewer tool calls than a Grep/Glob/Read baseline (protocol, caveats and what it does not prove ↓).
• 11 languages. Rust, Python, TypeScript, Java, Kotlin, Swift, Objective-C, C, C++, Go get full extraction; Ruby gets a shallow pass (node kinds, no deep extraction).
• Read-only by design. It supplies evidence for the next stage — a fix, a PRD, a review. It never edits your code.
• It says what it cannot see. Every answer carries its analysis limitations, so you can tell a real "no callers" from an unresolved import.

Sovereign intelligence, eco-responsible by intent

Sovereign is what it is today. Native Rust and tree-sitter parse your code on your machine — no model is called to read it, and nothing is uploaded. The graph is a local database file you own.

Eco-responsible is what we're aiming at. The dominant energy cost in an AI-assisted workflow is not this binary's CPU; it is inference spent re-reading files to answer a question one query could have settled. Reducing that demand is the lever we work on, and we measure it — while publishing no energy or CO₂ figure, because this repository measures no joules and a token proxy is not a watt-hour. What we measure, and what we refuse to claim ↓

One pipeline stage = one MCP tool. 10 stages. 26 tools. 1700+ tests. Zero clippy warnings, enforced in CI.

---

What an agent can ask it

For inferred Rust receiver calls, pass lsp: true to analyze_codebase and install rust-analyzer. The response's lsp_status.state distinguishes disabled, completed, completed_unresolved, and failed; failures retain their error and analysis continues on the available graph, which may contain partial LSP results. lsp_resolve retains the pass's counts; resolve.phase = "static" identifies the separate static-resolution receipt. Completion does not mean every call was resolved — completed_unresolved (issue #282) is the explicit signal for "the pass ran and resolved nothing" (at least one site was attempted), as opposed to completed, which also covers "there was nothing to resolve." A target that sits under a parent Cargo workspace which does not list it as a member — or any other condition the language server itself reports as health: "error" — fails the phase outright as lsp_workspace_load_failed before a single resolution request is sent, naming the cargo-level fix (add the package to the parent's workspace.members, or analyze the workspace root). lsp_status.server_health and lsp_resolve's own server_health field carry the server's last-reported health, message, and readiness signal even on success.

Analysis persists its coverage report for query_graph(graph="missed") and returns the same summary. Rust processes use explicit #[test] and #[kani::proof] attributes, with separate test and proof entry kinds. These are source declarations, not evidence of execution or successful proof. (Rust testing attributes, Kani proof attributes.) For a Rust codebase, the coverage report also carries outside_build_targets (issue #284): .rs files the walker indexed — their declarations, including #[kani::proof] harnesses, are in the graph — that sit outside every compiled Cargo target per cargo metadata --no-deps (a proof harness under kani/, a fuzz/ directory excluded from the workspace, …). Calls out of these files cannot be resolved by the language server, whose crate graph never contains them. The bucket is populated only when a root Cargo.toml is readable and cargo metadata succeeds; anything else (no manifest, cargo missing, or a workspace that fails to load) leaves it empty rather than guessing. When lsp: true runs against such a graph, lsp_resolve's outside_targets_count reports how many unresolved call sites the pass skipped for exactly this reason — no textDocument/definition request is ever issued for them, so they never count as failed — and get_impact's unresolved_callsites_outside_targets names the same attribution (plus the file it came from) for any target those sites call. Graphs created before entry metadata was stored require a full reindex: analyze_codebase rebuilds them, and index_codebase automatically falls back to a full index when its incremental compatibility check detects the old schema.

analyze_codebase(path: "/path/to/project", output_dir: "/tmp/run")
  → index + resolve + cluster + build search index in one call
  → 430 nodes, 400 edges, 216 communities, 35 processes on our own codebase

search_codebase(graph_path, query: "process incoming tool requests")
  → hybrid ranked results: BM25 lexical + sparse TF-IDF semantic + RRF fusion
  → returns: handle_tool_call (score 0.021), dispatch_request (0.020), ...

get_context(graph_path, qualified_name: "src/main.rs::handle_tool_call")
  → 360° view: community membership, process participation,
    incoming calls, outgoing calls, types used, types that use it
  → did-you-mean suggestions when the symbol isn't found exactly

get_impact(graph_path, qualified_name)
  → candidate impact: callers, communities and processes in the available graph
  → evidence for choosing what to inspect and recheck after a change

detect_changes(graph_path, diff_text OR base_ref+head_ref)
  → git diff → affected symbols → impacted communities → touched processes
  → risk score for the change

validate_prd_against_graph(prd_path, graph_path)
  → does the PRD reference real symbols? (symbol hallucination check)
  → does "scoped to X" match the actual community count?
  → does "doesn't affect main" hold against the call graph?

check_security_gates(graph_path, changed_symbols)
  → auth-critical community touch · unsafe symbol · public API change ·
    unresolved imports · test coverage gap

verify_semantic_diff(before_graph_path, after_graph_path)
  → what nodes/edges appeared, what disappeared, what dangles,
    new cycles via Tarjan SCC, regression score with verdict

These tools establish different kinds of evidence. Stage 2's verified receipt means schema checks, clarification completeness and caller acknowledgement passed; its transcript digest binds the recorded bytes, not the truth of the finding. gates_passed means no critical flag was emitted by the available security checks. Inspect report.assessment_complete as well: it is false for an empty symbol list, skipped checks, or changed symbols that could not be resolved. Review those items and warnings even when gates_passed is true. The unresolved-import gate reports unresolved imports in a changed symbol's file; a single graph snapshot cannot establish when they were introduced. A semantic-diff clean verdict requires a structural regression score below the configured threshold and no positive unresolved-import delta. Any increase in unresolved imports produces at least concerning, even below that threshold. A clean result does not establish behavioral equivalence. Tests, compiler checks or formal proofs must establish that separate property.

Impact and process results depend on the relationships the graph captured. Process traversal stops at depth 20; it is graph reachability, not an observed runtime trace or an exhaustive account of effects. Preserve coverage and resolution qualifiers, and confirm important absence claims against source even when the coverage report contains no flagged files.

---

Getting started

Prerequisites

• Rust 1.95.0 — pinned by rust-toolchain.toml, so rustup installs and selects it for you; the same compiler builds CI and the releases
• CMake (LadybugDB builds its C++ core from source — ~5 minutes first build, cached after)

Clone + build

git clone https://github.com/cdeust/ai-architect-mcp-codebase.git
cd ai-architect-mcp-codebase
cargo build --release
# First build: ~5 minutes (compiles LadybugDB C++ core)
# Subsequent builds: <1 second incremental

Register the MCP server

The repo ships a .mcp.json that Claude Code picks up automatically when you open the directory:

{
  "mcpServers": {
    "ai-architect": {
      "command": "cargo",
      "args": ["run", "--quiet", "--release", "--manifest-path", "Cargo.toml", "--", "--profile", "core"]
    }
  }
}

Or register globally (recommended agent setup — the core profile):

claude mcp add ai-architect -- /absolute/path/to/target/release/ai-architect-mcp-codebase --profile core

Tool profiles

The server registers one of two tool sets, chosen once at startup:

Profile	Tools	Who it's for
core	8 — health_check · analyze_codebase · search_codebase · get_context · get_symbol · get_impact · query_graph · detect_changes	Recommended for agents. The read-only code-intelligence surface: analyze once, then search, inspect symbols, and measure blast radius.
full	all 26	The ai-architect pipeline orchestrator — adds the internal finding → PRD stages (1/2/4/6/8/9) and the manual graph passes (index_codebase, resolve_graph, cluster_graph, lsp_resolve, get_processes, index_history).

Select with the --profile flag or the AP_PROFILE environment variable (the flag wins):

ai-architect-mcp-codebase --profile core   # agent-facing 8
AP_PROFILE=core ai-architect-mcp-codebase  # same, via env
ai-architect-mcp-codebase                  # default: full (all 26)

The default stays full until the next major version — shrinking the default tool surface is a breaking change. New agent installations should opt into core: analyze_codebase already runs index + resolve + cluster in one call, so the 18 hidden tools are pipeline plumbing an agent never needs, and hiding them keeps the tool prompt small.

First run

# Run the binary directly to verify the handshake
./target/release/ai-architect-mcp-codebase

# Or exercise it via stdio JSON-RPC:
printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"health_check","arguments":{}}}' \
  | ./target/release/ai-architect-mcp-codebase

Use with other MCP hosts

The server is a self-contained stdio binary — any MCP host can launch it. Install once:

cargo install ai-architect-mcp-codebase   # installs the `ai-architect-mcp-codebase` binary into ~/.cargo/bin

Install into your agent host (auto-config)

One command detects your installed hosts and writes the right MCP config for each — never clobbering the rest of the file:

ai-architect-mcp-codebase install

It configures the top six hosts it detects: Claude Code (~/.claude.json), Codex CLI (~/.codex/config.toml), Gemini CLI (~/.gemini/settings.json), Cursor (~/.cursor/mcp.json), VS Code (Code/User/mcp.json), and Zed (~/.config/zed/settings.json).

• Never clobbers. The existing config is parsed; only our ai-architect entry is added or updated; every other server survives. A file it cannot safely parse is never overwritten — it prints the exact entry to paste by hand.
• Zed JSONC. Zed's settings.json allows comments, which strict JSON editing would destroy, so install refuses to edit it in place and prints the snippet + instructions instead (your comments stay byte-for-byte).
• Codex TOML is edited comment- and format-preserving (via toml_edit).
• Flags: --dry-run (print planned changes, write nothing), --only <host> / --skip <host> (filter; --only forces a host even if undetected), --with-hooks (also register the Grep/Glob PreToolUse hook, see below). Re-running is idempotent (a second run reports "no change").
• Uninstall: ai-architect-mcp-codebase uninstall removes exactly our entries (and the hook), leaving everything else intact.

ai-architect-mcp-codebase install --dry-run                 # preview
ai-architect-mcp-codebase install --only cursor --only zed  # just these
ai-architect-mcp-codebase install --with-hooks              # + the grep→graph hook
ai-architect-mcp-codebase uninstall                         # remove our entries

Binary → first query. Measured on this machine (2026-07): install completes in ~1.3 s (dominated by process/DB startup; the config write itself is sub-second); analyze_codebase on this repo's own src/ (114 files → 16.5k nodes, 16.3k edges — index + resolve + cluster) takes ~12 s wall; the first search_codebase returns instantly. So once the binary exists, install → analyze → first graph query is ~15 s — well under the 2-minute target. The one-time cargo build --release (~5 min, compiling the LadybugDB C++ core) is a separate, before-the-clock step.

Fail-open grep→graph hook

ai-architect-mcp-codebase install --with-hooks registers a Claude Code PreToolUse hook (matcher Grep|Glob) that runs ai-architect-mcp-codebase hook-augment. Before a Grep/Glob in a project that has an ai-architect graph, it injects a one-line suggestion to consider search_codebase/query_graph first. Cardinal rule: it never blocks the tool call — no graph, an unparseable payload, or any error → it prints nothing and exits 0. Hook registration is opt-in (the --with-hooks flag), never default.

Or configure a host by hand

The CLI commands below assume ~/.cargo/bin is on your PATH. GUI hosts (Cursor, Windsurf, VS Code) may not inherit your shell PATH — in the JSON configs, replace ai-architect-mcp-codebase with the output of which ai-architect-mcp-codebase. Use the core profile (8 read-only tools) for agent hosts.

Gemini CLI

gemini mcp add -e AP_PROFILE=core ai-architect ai-architect-mcp-codebase

Or install as an extension (this repo ships a gemini-extension.json):

gemini extensions install https://github.com/cdeust/ai-architect-mcp-codebase

The extension also exposes three host-native workflows from skills/: understand-codebase, impact-analysis, and validate-change-plan. They use only the eight tools in the core profile and explicitly surface index coverage gaps before accepting negative graph results.

Claude Code plugin (primary interface)

claude plugin marketplace add cdeust/ai-architect-mcp-codebase
claude plugin install ai-architect-mcp-codebase@ai-architect-mcp-codebase-marketplace

Fresh marketplace installs require GitHub CLI 2.68 or newer. The bootstrap verifies the release's attached Sigstore bundle against the fixed cdeust/ai-architect-mcp-codebase/.github/workflows/release.yml signer before installing any executable; it never accepts a manifest-provided trust anchor. The bundle avoids a Rekor transparency-log lookup, but gh can still need the network to refresh Sigstore's TUF trust root on a cold cache. This protects the official package and makes a minimal-diff fork that changes only metadata fail closed; it cannot make arbitrary code from a hostile fork trustworthy, because such a fork can also replace the bootstrap itself. Verify that the marketplace slug is exactly cdeust/ai-architect-mcp-codebase.

Developer escape hatch: running a local dev build in place of the release

bin/ensure-binary.sh pins the installed binary to a verified release digest (see Security) — that pin rejects any binary it did not download and verify itself, including one you legitimately rebuilt from source. Set AI_ARCHITECT_SOURCE_CHECKOUT=1 to opt out of the pin for a local dev build. The bootstrap accepts two shapes under this flag, both requiring the explicit opt-in — it is never inferred from metadata:

• Plain source checkout — $CLAUDE_PLUGIN_ROOT itself contains .git (you registered a clone directly as the plugin root).
• Live-mount montage — the installed binary at target/release/ai-architect-mcp-codebase is a symlink whose fully resolved target lies outside $CLAUDE_PLUGIN_ROOT and sits inside its own .git-bearing checkout (e.g. a marketplace cache whose binary was replaced with a symlink into a separate dev clone, so you can iterate without reinstalling the plugin after every rebuild). Added in #208 — a plain .git-at-root check cannot see this shape, because a marketplace cache has no .git of its own.

What the flag skips, precisely: only the release-binary digest verification (sha256sum against the cached/pinned digest) and, for a fresh install, the download + Sigstore provenance check — for that one launch. It does not skip the Cargo.toml / plugin.json presence checks (still fatal if either file is missing), and for a plain source checkout it still runs the freshness rebuild (cargo build --release when src/ is newer than the binary). For the montage shape specifically, nothing rebuilds the binary — the bootstrap trusts the already-built binary the symlink resolves to, as-is.

Threat model. This is an explicit, user-set opt-in, never something packaged metadata can trigger. An attacker who can already write to your plugin cache — replacing the installed binary with a symlink to force this path — can just as easily replace bin/ensure-binary.sh or bin/launch-plugin.sh themselves, so the digest pin was never a defense against that attacker; it defends the default path (flag unset) where the bootstrap is the thing standing between a marketplace download and your shell. The default path is unchanged by this hatch and remains a hard fatal on any digest mismatch. Every accepted bypass is announced on stderr even in quiet mode:

ai-architect-mcp-codebase: bootstrap verification skipped (source-checkout mode)
ai-architect-mcp-codebase: live-mounted dev symlink: <plugin-cache>/target/release/ai-architect-mcp-codebase -> <resolved dev path> (source checkout at <resolved .git root>)

Diagnosing the failure mode without the flag. If a marketplace-cache binary is replaced by a montage symlink and AI_ARCHITECT_SOURCE_CHECKOUT is not set, the plugin dies silently from Claude Code's point of view — you only see MCP error -32000: Connection closed. The real cause is on stderr, which Claude Code does not surface for a failed MCP launch; run the launcher by hand with CLAUDE_PLUGIN_ROOT set to the plugin cache directory to see it:

CLAUDE_PLUGIN_ROOT=/path/to/plugin/cache bin/launch-plugin.sh
# ai-architect-mcp-codebase: FATAL: cached binary digest mismatch; reinstall the plugin

Operational gotcha: export AI_ARCHITECT_SOURCE_CHECKOUT=1 in ~/.zshrc alone is not enough. ~/.zshrc is read only by interactive shells; the Claude Code plugin launcher and its hooks run in non-interactive ones and never see it. Put the export in ~/.zshenv (or your shell's equivalent non-interactive startup file) instead.

If the former Automatised Pipeline plugin is installed, remove it before installing the canonical package:

claude plugin uninstall automatised-pipeline@automatised-pipeline-marketplace
claude plugin marketplace remove automatised-pipeline-marketplace

Claude MCP allowlists and permissions must also replace every prefix listed in revoked_claude_tool_prefixes in the contract with mcp__plugin_ai-architect-mcp-codebase_ai-architect__<tool>. The final ai-architect segment is intentionally stable: it is the MCP server key, not the plugin's distribution name. The machine-readable source of truth is mcp-contract.json; consumer repositories validate their allowlists against its derived claude_tool_prefix instead of maintaining an independent spelling.

Contract schema 1 requires distribution, claude_plugin, claude_marketplace, mcp_server, claude_tool_prefix, and revoked_claude_tool_prefixes. Consumers must pin the raw contract URL to the full commit SHA (tags can be moved), validate that the prefix equals mcp__plugin_<claude_plugin>_<mcp_server>__, and remove revoked prefixes from allowlists rather than retaining them as aliases. Consumer PRs record the full producer commit in their contract URL; the v0.11.1 release must not be assumed available until its verified-release workflow completes. The same contract is included in the crate, MCPB, and signed release assets.

OpenAI Codex CLI (also picked up by the ChatGPT desktop app and Codex IDE extension — they share ~/.codex/config.toml)

codex mcp add ai-architect -- ai-architect-mcp-codebase --profile core

Or in ~/.codex/config.toml:

[mcp_servers.ai-architect]
command = "ai-architect-mcp-codebase"
args = ["--profile", "core"]

Or install the packaged Codex plugin and its three matching skills from this repository's marketplace:

cargo install ai-architect-mcp-codebase
codex plugin marketplace add cdeust/ai-architect-mcp-codebase
codex plugin add ai-architect-mcp-codebase@ai-architect-mcp-codebase

The Codex package lives under plugins/ai-architect-mcp-codebase/, with its own .mcp.json fixed to --profile core. This isolation is intentional: the root .mcp.json remains the existing Claude project configuration and keeps the server's backward-compatible full default.

For Gemini CLI, uninstall the former extension identity before reinstalling from the renamed repository:

gemini extensions uninstall ai-architect
gemini extensions install https://github.com/cdeust/ai-architect-mcp-codebase

Cursor — .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "ai-architect": {
      "command": "ai-architect-mcp-codebase",
      "args": ["--profile", "core"]
    }
  }
}

Windsurf — ~/.codeium/windsurf/mcp_config.json: same mcpServers block as Cursor.

VS Code — .vscode/mcp.json:

{
  "servers": {
    "ai-architect": {
      "type": "stdio",
      "command": "ai-architect-mcp-codebase",
      "args": ["--profile", "core"]
    }
  }
}

OpenAI Agents SDK (Python)

from agents.mcp import MCPServerStdio

async with MCPServerStdio(
    name="ai-architect",
    params={"command": "ai-architect-mcp-codebase", "args": ["--profile", "core"]},
) as server:
    agent = Agent(name="Assistant", mcp_servers=[server])

---

The pipeline

Every stage is a tool. Stages build on each other but are independently callable. The pipeline is serial in logical order but MCP calls are stateless — you can re-run stages 3a-3d on a fresh codebase without re-running stages 1-2.

#	Tool(s)	What it does
0	health_check	Handshake + protocol + tool count
1	extract_finding, refine_finding	Deterministic finding extraction + orchestrator-aware prompt refinement
2	start_verification, append_clarification, finalize_verification, abort_verification	Human-gated clarification loop with SHA-256 transcript digest, atomic single-file session state
3a	index_codebase, query_graph, get_symbol	tree-sitter AST → LadybugDB graph (16 node labels, 36+ relationship tables); user-configurable exclude_dirs and graceful skip of unreadable directories (issue #249)
3b	resolve_graph, lsp_resolve	Import/call/impl resolution with confidence scoring + optional LSP deep resolution (rust-analyzer / pyright / typescript-language-server)
3c	cluster_graph, get_processes, get_impact	Leiden-class community detection (Louvain + C2 repair) + BFS execution-flow tracing from entry points
3d	search_codebase, get_context, analyze_codebase, detect_changes	Hybrid BM25 + sparse TF-IDF + RRF search · 360° symbol view · all-in-one analysis · git-diff impact
4	prepare_prd_input	Bundle verified finding + graph intel → artifact for ai-architect-mcp-spec
6	validate_prd_against_graph	Symbol hallucination · community consistency · process-impact contradiction
8	check_security_gates	Auth-critical community · unsafe symbol · public-API change · unresolved-import presence · test-coverage gap
9	verify_semantic_diff	Before/after graph diff with Tarjan SCC cycle detection and regression scoring

Stages 5 (PRD generation), 7 (implementation), 10 (benchmark), 11 (deployment), 12 (PR) belong to other systems in the pipeline: ai-architect-mcp-spec, the coding agent, CI, and gh. This project is the read-only intelligence half.

---

26 MCP Tools

Every tool takes structured JSON arguments via the MCP protocol and returns a structured JSON response. No LLM is called from inside any tool — intelligence is the agent's job; the tool's job is safe, fast data movement with invariants.

Stage 0:  health_check
Stage 1:  extract_finding · refine_finding
Stage 2:  start_verification · append_clarification · finalize_verification · abort_verification
Stage 3:  ingest_traces
Stage 3a: index_codebase · index_status · query_graph · get_symbol
Stage 3b: resolve_graph · lsp_resolve
Stage 3c: cluster_graph · get_processes · get_impact
Stage 3d: search_codebase · get_context · analyze_codebase · detect_changes
Stage 3e: index_history
Stage 4:  prepare_prd_input
Stage 6:  validate_prd_against_graph
Stage 8:  check_security_gates
Stage 9:  verify_semantic_diff

Each tool has a JSON Schema enforced at the wire, reason codes on error (no cryptic protocol errors), and a receipt-style response with timing and counts.

Agent installs rarely need all 26 — the core profile (see Tool profiles) registers just the 8 code-intelligence tools.

Team-shared graph artifact (optional)

index_codebase can commit a compressed snapshot of the graph so teammates who clone the repo never have to cold-index it.

• index_codebase with "export_artifact": true writes, after a successful index, a tar → zstd snapshot to <path>/.ai-architect-mcp-codebase/graph.zst plus a graph.meta.json sidecar (schema version, git sha, tool version, node/edge counts). It also appends a .gitattributes entry (.ai-architect-mcp-codebase/graph.zst binary merge=ours) so the committed binary never produces merge conflicts across branches. Commit both files. A repo indexed before the project rename (issue #195) carries this snapshot under the old .automatised-pipeline/ directory name; the first touch of the artifact (export, bootstrap, or even a hook-augment Grep/Glob check) migrates it to the current name in place — a one-shot rename, not a permanent dual-path read.

• index_codebase with "bootstrap": true — when there is no local graph at <output_dir>/graph but a committed artifact is present — decompresses the snapshot instead of cold-indexing. Staleness is checked first by comparing the artifact's git sha with the repo's current HEAD:

  • shas equal → import as-is (nothing to fill), response source='artifact_bootstrap', graph_state='fresh';
  • shas differ → by default the snapshot is imported, then incrementally filled up to the working tree (only the artifact→HEAD diff is re-parsed), response source='artifact_bootstrap_fill', graph_state='filled_to_working_tree', carrying fill_method and {changed, added, deleted, renamed, unchanged} counts;
  • "accept_stale": true → import the stale snapshot anyway and skip the fill, and the response carries a stale_artifact {artifact_sha, head_sha, commits_behind} report so a stale graph is never mistaken for a fresh one.

  A fill that fails (no git diff and no bundled manifest) falls back to a full index, as does an import failure — both explicit (logged to stderr, never a silent partial graph) and reported via a bootstrap_skipped note.

Excluding directories from the walk (issue #249)

Both index_codebase and analyze_codebase accept "exclude_dirs" (default []) — directory names or paths to prune from the walk in addition to the built-in build/dependency skip list (node_modules, .venv, vendor, target, …). This is for directories that must never be read (a secrets folder, a credentials mount), not a performance prune:

• An entry without a path separator (e.g. "secrets") is a bare name, matched anywhere in the tree — like the built-in list.
• An entry with a path separator (e.g. "config/secrets") is a path relative to path, matched as exactly one subtree. No glob support.
• Exclusion wins over every dependency_scope tier, including full — it is checked before, and independently of, dependency-directory descent.
• Pruned directories are never silently dropped: each appears in the coverage sidecar as skipped with reason user_excluded, and the response's coverage.skipped.user_excluded_count carries the exact count.
• Changing exclude_dirs on an existing graph requires "full": true — like dependency_scope, the incremental-index manifest does not capture it.

Independent of exclude_dirs, a directory the OS refuses to read (EACCES/PermissionDenied) no longer aborts the whole index: it is recorded in the coverage sidecar with reason unreadable and the walk continues past it, so one locked-down subdirectory can no longer discard an otherwise-successful index.

All three flags default to false, so existing behavior and the core/core8 profiles are unchanged. The artifact is entirely optional: without it, index_codebase cold-indexes exactly as before.

Post-import incremental fill (re-index only the artifact_commit..HEAD diff instead of a full re-index) is tracked in #62 — it needs a changed-files-only indexer, which AP does not yet have.

---

Architecture

Rust MCP server, hand-rolled stdio JSON-RPC 2.0 (no SDK — we own the wire). Clean Architecture with module boundaries.

transport (stdio, JSON-RPC framing)
      ↓
server/main.rs  (request dispatch, tool registry)
      ↓
handlers (do_* functions, one per tool)
      ↓
core modules:
    graph_store        — LadybugDB port (Cypher + UNWIND + prepared statements)
    parser/{rust,python,typescript,mod}  — tree-sitter AST extractors
    indexer            — walk + parse + persist pipeline
    resolver           — cross-file import/call/impl resolution
    lsp_{client,resolver}  — optional LSP deep resolution
    clustering         — inline Louvain + C2 repair + process tracing
    search/{bm25,vector,rrf,mod}  — hybrid search (Tantivy + sparse TF-IDF + RRF)
    prd_input          — stage 4: bundle for ai-architect-mcp-spec
    prd_validator      — stage 6: validate PRD claims against graph
    security_gates     — stage 8: auth/unsafe/API/imports/coverage checks
    semantic_diff      — stage 9: before/after graph regression scoring
    git_diff           — diff parser + symbol mapping

Dependencies

Sixteen crates. Nothing speculative; everything justified.

Crate	Purpose	License	Why
serde + serde_json	Wire serialization	MIT	JSON-RPC, artifact persistence
sha2	Stage-2 transcript digest	MIT	Tamper detection
lbug (LadybugDB)	Embedded property graph + Cypher	MIT	Native Cypher, FTS-ready, the Kùzu successor
tree-sitter	Incremental parser runtime	MIT	First-class Rust bindings
tree-sitter-rust · -python · -typescript · -java · -kotlin-ng · -swift · -objc · -c · -cpp · -go	Language grammars (10)	MIT / Apache-2.0	Semantic structure without a compiler
tantivy	Lucene-grade BM25	MIT	Real ranked text search, <10ms startup

Deliberately not included: async runtime (we're stdio-blocking), HTTP client, LLM SDK, embedding model runtime (sparse TF-IDF replaces it at zero dep cost).

Storage

Graphs are per-finding by design (Lamport's isolation invariant): each finding gets its own LadybugDB instance at <output_dir>/runs/<run_id>/findings/<finding_id>/graph/. Zero-coordination concurrency, trivial cleanup, no cross-finding state leakage. Redundant indexing for shared codebases is acknowledged and mitigated in a later optional cache layer — not shoehorned into the core.

Configuration — max_db_size

Every LadybugDB Database this crate opens reserves virtual address space up front, sized by max_db_size. lbug's own default (SystemConfig::default()) is 1 << 43 = 8 TiB per instance; with graph_cache's MAX_CACHED_GRAPHS = 8 entries live in the read-path cache at once, that is a 64 TiB worst case (issue #25). src/graph_store.rs::system_config() is the single choke point every GraphStore::open_or_create call resolves through, in this precedence order:

1. AP_LBUG_TEST_MAX_DB_SIZE — test-only, set for every cargo test process via .cargo/config.toml's [env] table (512 MiB / 2^29, issue #21). Always wins when present, so cargo test behavior is independent of the production knob below.
2. AP_LBUG_MAX_DB_SIZE — production override, unset by default. Bytes, must be a power of two and at least 8 MiB (lbug's own BufferManager::verifySizeParams floor). An invalid value is rejected with an actionable error at GraphStore::open_or_create time — never a silent fallback.
3. Default: 8 TiB (1 << 43 bytes) when neither var is set — lbug's own DEFAULT_VM_REGION_MAX_SIZE, the engine's per-database VM-region ceiling on every 64-bit desktop/server platform (lbug-0.19.1/lbug-src/src/include/common/constants.h). This is an address-space reservation, not an allocation: disk and memory grow only with the data actually written. An earlier release capped the default at 8 GiB (issue #25, sized from the measurement table below); that cap aborted any ingestion whose graph outgrew it and was repealed on 2026-08-14 — an index must complete regardless of corpus size, multi-TiB included.

Set AP_LBUG_MAX_DB_SIZE to bound the reservation in address-space-constrained environments (e.g. containers with a low RLIMIT_AS); the historical measurement table below documents typical graph sizes.

Measured graph sizes (2026-07-15, du -k on every graph file found under ~/.cache/cortex/code-graphs/*/graph, ~/.cortex/ap_graph/graph, and **/.prd-gen/graphs/*/graph), top 10 of 75:

Graph	Size
repro-cortex-viz-deps (cortex-viz + node_modules)	473 MiB
bench-c2-viz-deps (cortex-viz + deps)	472 MiB
bench-c3-viz-pubapi (cortex-viz, public API surface)	460 MiB
wt-windows-launcher-96-97-* (Cortex worktree)	147 MiB
wt-homeostatic-* (Cortex worktree)	144 MiB
wt-tools-drift-* (Cortex worktree)	143 MiB
Cortex-wt-wiki-titles-*	142 MiB
wt-findings-provenance-*	126 MiB
anthropic-partnership-Cortex	126 MiB
wt-ingest-provenance-*	124 MiB

Total across all 75 measured graphs: ~4.87 GiB. Every graph other than the top 3 (which include node_modules) is under 150 MiB — the node_modules-inclusive runs are the actual worst case driving the sizing rule above.

---

The zetetic standard

Inherited from zetetic-team-subagents. Not a prompt suggestion — an enforcement rule that holds in code.

Pillar	Question
Logical	Is it consistent?
Critical	Is it true?
Rational	Is it useful?
Essential	Is it necessary?

In this codebase it concretely means:

1. Every algorithm traces to a source. Louvain → Blondel et al. 2008. Leiden C2 repair → Traag et al. 2019. RRF → Cormack, Clarke, Büttcher 2009. SCC → Tarjan 1972. BM25 via Tantivy → Robertson et al. 1994.
2. Named constants should record their source or measured rationale. RRF_K = 60 cites Cormack 2009. BULK_BATCH_SIZE = 500 cites Kùzu/LadybugDB tuning. PARSE_TIMEOUT_MICROS = 5_000_000 is justified in the block above it.
3. No invented numbers. Where a value was chosen by judgment, the comment says so ("heuristic, not paper-backed") and cites its operational justification.
4. Tool responses cite the spec that governs each error reason. unsafe finding_id (spec §5.1.4, §9.3 Q4): must match [A-Za-z0-9._-]+ — callers see which rule they violated.
5. When a capability can't be proved at spec time, the tool degrades gracefully and says so in plain language. Example: lsp_resolve on a stub binary returns lsp_probe_failed: found on PATH but didn't respond as an LSP server (stdout closed immediately; likely a stub, proxy, or non-LSP binary) — not a cryptic protocol error.

---

Security

Four CRITICAL, four HIGH, three MEDIUM findings were surfaced by a security-auditor agent pass and fixed in commit 512d683:

• Cypher injection via insert_edge → centralized cypher_str() escaping (\ first, then ')
• Git argument injection → validate_git_ref rejects --, newlines, NUL; -- separator before refs
• Arbitrary binary execution via lsp_command → strict allowlist (rust-analyzer, pyright, pyright-langserver, typescript-language-server)
• Symlink traversal → fs::symlink_metadata + MAX_DEPTH
• Resource exhaustion → MAX_FILES=100_000, MAX_FILE_BYTES=10 MB, MAX_TOTAL_BYTES=2 GB, MAX_DEPTH=64
• Tree-sitter pathological input → set_timeout_micros(5_000_000) + MAX_PARSE_BYTES=1 MB
• query_graph read-only → two layers over disjoint statement families (see below)
• graph_path filesystem safety → validate_graph_path_safe() before any remove_dir_all
• LSP rootUri → RFC 3986 percent-encoding
• Diff line overflow → DIFF_LINE_MAX = u64::MAX / 2 guard

Each fix has a test that asserts the exploit is now rejected. Run cargo test to see 1700+ tests pass including the exploit-regression suite.

How query_graph is kept read-only

Two layers, covering disjoint statement families. Neither subsumes the other.

Layer	Refuses	Mechanism
Engine (GraphStore::execute_read_only_query)	every database write and DDL — CREATE, MERGE, SET, DELETE/DETACH DELETE, DROP, ALTER, however spelled	PreparedStatement::is_read_only(): the verdict comes from the compiled plan, so a mutation written in syntax no keyword scan enumerates is still refused
Lexical (FORBIDDEN_CYPHER_KEYWORDS)	filesystem and database movement — COPY … TO, EXPORT/IMPORT DATABASE, ATTACH, DETACH, USE, LOAD FROM	whole-word, case-insensitive scan of the query with string literals, backticked identifiers and comments masked out first
Lexical (READ_ONLY_PROCEDURES)	every CALL naming anything but TABLE_INFO / SHOW_TABLES — including the CALL <setting> = <value> configuration form	per-PROCEDURE classification of the identifier after each CALL token

The lexical layer is load-bearing, not defence in depth. lbug's StatementReadWriteAnalyzer overrides visitCopyFrom but leaves six statements at the base visitor's no-op — visitCopyTo, visitExportDatabase, visitImportDatabase, visitAttachDatabase, visitDetachDatabase and visitUseDatabase (parsed_statement_visitor.h:51, 57-61 on lbug 0.19.1) — so all six are classified read-only. DETACH/USE were added to the denylist on 2026-08-25 after a mechanical re-audit against those headers; before it, both passed the lexical filter AND the engine filter. Measured 2026-08-24 against lbug 0.19.1 on both available engine gates — is_read_only() and a database opened with SystemConfig::read_only(true), which reaches the same predicate via ClientContext::validateTransaction — COPY (…) TO 'f.csv' and EXPORT DATABASE 'd' execute and write the filesystem, while both correctly refuse CREATE NODE TABLE. Pinned by engine_gate_does_not_cover_filesystem_writes and, for the whole family, engine_classifies_every_filesystem_statement_as_read_only.

CALL is classified per procedure rather than refused wholesale, so schema introspection (CALL table_info('Function') RETURN *) is reachable. That distinction is load-bearing too: the same analyzer returns readOnly = true from visitStandaloneCall, so CALL threads = 8 — a configuration write — is engine-read-only, and this lexical layer is the only barrier that exists against it. Relaxing the KEYWORD rather than allowlisting the PROCEDURE would remove that barrier entirely.

A keyword introduced by : or . is an identifier, not a clause, so queries over this schema's own Import node table work unchanged: MATCH (f:File)-[:Defines_File_Import]->(n:Import) WHERE n.is_resolved = false RETURN n.path.

The gate does not extend that exemption to an alias (AS <keyword>), where the clause detectors do. The asymmetry is deliberate: on the gate an exemption can only ever let a keyword through, so it fails closed and a bare use/create pattern variable is refused (backtick it); on the clause detectors the expensive direction is reversed, because reading AS limit as a clause would suppress the LIMIT injection. A masked literal or backticked identifier is treated as a TOKEN, never as whitespace, so no look-back can walk across one.

query_graph executes one statement per call. A trailing ; is accepted; a ;-chained request is refused with reason multi_statement_not_supported, because the read-only classification, LIMIT injection, ORDER BY detection and the offset cursor are all properties of a single statement.

The full security argument — threat model, trust boundaries, what each claim rests on, and where it stops — is in docs/ASSURANCE-CASE.md. Reporting process and response SLA: SECURITY.md. How the project is run and what happens if the maintainer stops: GOVERNANCE.md. Where it is going: docs/ROADMAP.md. OpenSSF Best Practices answers, criterion by criterion: .bestpractices.json.

---

Scale

Re-measured 2026-07-28 on the current dependency (lbug 0.18, rustc 1.95.0, macOS 26.5.1 arm64) by re-running the dba agent's nine compile-and-run probes — cargo test --release --test lbug_bulk_investigation -- --nocapture, 199 edges per strategy. The ranking is the same one the original 0.15.3 run found; the absolute figures are not comparable across the two runs, because both the engine version and the machine changed.

Strategy	ms/edge
Raw string per edge (naive)	9.658
Prepared statement, no transaction	6.924
BEGIN TRANSACTION + prepared + COMMIT	0.328
UNWIND + typed LogicalType::Struct	0.127

The chosen path is 76× faster than the naive one on this measurement.

The bulk-insert path uses UNWIND with a typed struct schema (the engineer who wrote the first version used LogicalType::Any which fails the binder — the typed struct form works). Prepared statements are cached in a RefCell<HashMap<query, PreparedStatement>> on the GraphStore. Sparse TF-IDF replaces the dense N × V × 4B matrix — 30.5× smaller on our own codebase (108 KB vs 3.2 MB) and scales linearly with non-zero terms rather than vocab size. Clustering eliminated probe_node_label_for_process (per-node Cypher round-trip) in favor of a single in-memory HashMap<id, label> population pass.

500-file synthetic Rust fixture indexes in ~38 seconds end-to-end (parse + resolve + cluster + search index), down from the pre-audit implied "5 min – 1 hour" bracket.

---

Falsifiable evidence — graph tools vs a Grep/Glob/Read baseline

This offline retrieval evaluation compares graph queries with a fixed substring-search/full-file-read protocol on an authored 4-language corpus (Python, TypeScript, Go, Rust): 20 questions across five capability dimensions. benchmarks/eval_headtohead/PRE_REGISTRATION.md records the hypotheses and protocol. The current results below are the post-#92 run in benchmarks/eval_headtohead/results.json; earlier runs remain separately saved. See that folder's MANIFEST.md for provenance and reproduce.sh for the command. The deterministic evaluation needs no API key or external corpus; building it requires the Rust toolchain and dependencies to be available.

metric (mean ± sample stdev, n=20)	AP graph tools	Grep/Glob/Read baseline	source field
retrieval precision	1.00 ± 0.00	0.65 ± 0.33	aggregate.{graph,explorer}.precision
retrieval recall	1.00 ± 0.00	1.00 ± 0.00	aggregate.*.recall
payload token proxy	43.14 ± 17.26	550.36 ± 330.28	aggregate.*.tokens
modeled tool calls	1.00 ± 0.00	5.20 ± 1.64	aggregate.*.tool_calls
mean per-question token ratio (baseline / graph)	14.26×	—	aggregate.token_ratio_explorer_over_graph.mean
mean per-question tool-call ratio	5.20×	—	aggregate.toolcall_ratio_explorer_over_graph.mean

All four hypotheses H1–H4 are SUPPORTED in the current run under this protocol. The original run FALSIFIED H4: graph recall was 0.825 against 1.00. Its five losses (go-D3, go-D4, rs-D2, rs-D4, ts-D4) remain in raw_results.2026-07-26-pre-fix-87.json; fixes #87 and #92 closed these gaps. The corpus informed those fixes, so the current result is a regression benchmark, not an unseen generalization test.

Costs are modeled, not observed AI-client bills or tool traces. The graph leg serializes a benchmark-specific compact envelope of symbol identities; the baseline counts its substring-hit transcript plus full matching files. Both use a payload-size / 4 token proxy. Graph calls are assigned one per question; baseline calls are assigned two plus the number of matching files. Indexing, client prompts, actual MCP response envelopes and model reasoning are excluded. The 14.26× figure is a mean of per-question ratios; dividing aggregate payload volumes gives 12.76×, a different statistic. The optional answer-quality judge (AP_EVAL_JUDGE_CMD) did not run. These measurements establish file-retrieval results and protocol costs, not AI-agent success, hallucination reduction or real-world token savings.

---

Green software engineering

This server's efficiency argument is demand reduction, and it is already measured — it just was not named as such until now.

The dominant energy term in an LLM-assisted coding workflow is not this binary's CPU. It is the model inference spent re-reading files to answer a question a structural query could have answered once. Every token