ai-architect-mcp-codebase

The support policy is stated at https://github.com/cdeust/automatised-pipeline/blob/main/SECURITY.md — the latest minor release on main receives security patches. The upgrade path is a reinstall (cargo install ai-architect-mcp, a new release tarball, or the plugin pin moving), and there is no user-authored persisted state to migrate: graph artifacts are derived data, regenerated by re-indexing. Breaking changes are recorded per release in https://github.com/cdeust/automatised-pipeline/blob/main/CHANGELOG.md

GitHub Issues, with structured templates for bug reports, enhancement requests and audit findings: https://github.com/cdeust/automatised-pipeline/issues

No vulnerability has been reported and resolved in the last 12 months, so there is no reporter to credit. The standing commitment to credit reporters in the release notes for the patched version, unless they request anonymity, is at https://github.com/cdeust/automatised-pipeline/blob/main/SECURITY.md#recognition

https://github.com/cdeust/automatised-pipeline/blob/main/SECURITY.md#response-sla documents the response process: a per-severity first-response and patch SLA, and a five-step coordinated-disclosure timeline from private report to public advisory, including the default 30-day disclosure window and reporter credit.

The primary language is Rust; the standard is rustfmt plus the project-specific rules named at https://github.com/cdeust/automatised-pipeline/blob/main/CONTRIBUTING.md#coding-standards-excerpt, which cites the full zetetic coding standards it derives from (no warnings, no unjustified unwrap/expect, no unsafe without a safety comment, sourced constants, file and function size caps, Result over panics at API boundaries).

Automatically enforced by two FLOSS tools as required status checks on every push and pull request: cargo fmt --all --check for formatting and cargo clippy --workspace --all-targets -- -D warnings for the lint rules (https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/ci.yml). Exceptions must be a single #[allow(...)] with an inline justification, per CONTRIBUTING.md.

Native binaries are produced, and the standard variables are honoured rather than replaced. Cargo passes RUSTFLAGS through to rustc, and the C/C++ dependencies (the tree-sitter runtime, zstd, the LadybugDB core) build through the cc crate, which reads CC, CFLAGS, CXX and CXXFLAGS. Nothing in https://github.com/cdeust/automatised-pipeline/blob/main/.cargo/config.toml or the workflows overrides them — the release job only adds ZSTD_SYS_USE_PKG_CONFIG, which extends the build environment rather than replacing a user-provided value.

Debugging information is controlled by the Cargo profile and honoured by cargo build and cargo install, which is how a developer or packager builds this project; nothing in the build system discards it. Disclosure: the release workflow runs a best-effort strip on the packaged tarball binary for download size (https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/release.yml), so the prebuilt artifact is stripped. A user who needs symbols builds from source with the toolchain pinned in rust-toolchain.toml and gets an identical binary — see build_repeatable.

One Cargo workspace with a single, accurate dependency graph across all six members (https://github.com/cdeust/automatised-pipeline/blob/main/Cargo.toml). Cargo resolves cross-member dependencies itself; there is no recursive make and no per-subdirectory independent build.

Measured, not assumed. The compiler is pinned in https://github.com/cdeust/automatised-pipeline/blob/main/rust-toolchain.toml (1.95.0) and the dependency graph in Cargo.lock. On 2026-07-27 (macOS aarch64, rustc 1.95.0), cargo build --release --bin automatised-pipeline produced sha256 5b3dcc46f4dc2fcbd0969fa1b27ff75ce576b75ce2e833a9aee8977246df8650; the target directory was then deleted and the build repeated from scratch, producing the identical digest. Stated limit: the target path is embedded, so a build into a different CARGO_TARGET_DIR differs (389394362d5e...); bit-identity across different paths would additionally require --remap-path-prefix, which is not configured today.

Three commonly-used conventions, all documented at https://github.com/cdeust/automatised-pipeline#getting-started: the language package manager (cargo install ai-architect-mcp, uninstall with cargo uninstall), a prebuilt platform tarball from the GitHub release, and the .mcpb bundle / plugin marketplace entry that MCP hosts install and remove through their own UI.

Installation goes through cargo, which honours the standard conventions for install location: --root and CARGO_INSTALL_ROOT select the destination, CARGO_HOME relocates the default. No bespoke prefix handling is introduced.

git clone then cargo build and cargo test installs every dependency, the test fixtures and the test environment in one commonly-used command each — cargo resolves the whole graph from the committed Cargo.lock. Documented at https://github.com/cdeust/automatised-pipeline/blob/main/CONTRIBUTING.md#dev-setup

Declared in a computer-processable form and version-locked: https://github.com/cdeust/automatised-pipeline/blob/main/Cargo.toml with the rationale for each direct crate, and Cargo.lock pinning the full transitive graph. A CycloneDX SBOM of that graph is generated at release time by https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/release.yml

Two independent mechanisms. cargo audit and cargo deny run daily against the RUSTSEC database and block on any advisory that is not explicitly accepted with a written reason (https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/supply-chain-audit.yml, https://github.com/cdeust/automatised-pipeline/blob/main/deny.toml). Dependabot opens weekly grouped update PRs for both cargo and github-actions (https://github.com/cdeust/automatised-pipeline/blob/main/.github/dependabot.yml). The one advisory currently in the tree is analysed as unexploitable here and blocked upstream, with the evidence written into deny.toml. GitHub's own Dependabot alerts and security updates were enabled on 2026-07-28, so a newly published advisory raises an alert and an upgrade PR without waiting for the next scheduled audit run. Verify: gh api repos/cdeust/automatised-pipeline --jq '.security_and_analysis.dependabot_security_updates.status' returns enabled, and gh api repos/cdeust/automatised-pipeline/vulnerability-alerts returns HTTP 204.

No vendored, forked or convenience-copied component. Every external component is a normal versioned crate resolved by cargo from crates.io (https://github.com/cdeust/automatised-pipeline/blob/main/deny.toml denies unknown registries and unknown git sources), so updating one is a version bump in Cargo.toml or an automated Dependabot PR.

The stack is current: the pinned toolchain is Rust 1.95.0, and the load-bearing dependencies are on their current majors (tantivy 0.26.1, lbug 0.18.3, tree-sitter 0.25 with ABI-15 grammars). No deprecated or obsolete API is relied on. The single frozen transitive pin (cxx 1.0.138, held by lbug's exact requirement) is documented with the upstream evidence in https://github.com/cdeust/automatised-pipeline/blob/main/deny.toml rather than left as unexplained drift.

Every push to main and every pull request runs the full suite plus the structural graph-accuracy gate, and both are required status checks that report pass or fail on the PR: https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/ci.yml. The per-stage tests/*_integration.rs suites drive the pipeline end to end against fixture data, and several spawn a real process and exercise the stdio wire protocol.

Well above 50%: 34 test files in https://github.com/cdeust/automatised-pipeline/tree/main/tests carry the issue number of the defect they regress-test in their header, covering the bugs fixed over the last six months (for example #16 Cypher injection, #57 coverage honesty, #92 type-usage edges, #148 fuzz OOM, #143 disk-full diagnostics). The rule that a bug fix ships with a test that fails on the pre-fix code is written at https://github.com/cdeust/automatised-pipeline/blob/main/CONTRIBUTING.md#testing-policy-mandatory

Enforced, not merely measured: the cargo llvm-cov (80% line floor) job in https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/ci.yml runs cargo llvm-cov --workspace on every pull request and every push to main and fails the build below --fail-under-lines 80. It is a required status check on main, so a change that drops coverage under the threshold cannot merge. Latest measurement from that job: 87.73% line coverage on the Linux CI runner, the figure the badge answers to (the same workspace measures 86.83% on macOS aarch64 — see the note on platform variance in the #160 entry) — up from 81.56% after issue #169 covered the five MCP handler modules that had 0% (verification_ops, verification_core, prd_handlers, symbol_handlers) or 43.83% (search_context_handlers). The README badge is machine-checked against the same job by https://github.com/cdeust/automatised-pipeline/blob/main/scripts/check_doc_claims.py, which fails the build if the advertised figure overstates the measured one — so this answer cannot silently become false the way it could when the number came from a hand-run command.

The formal written policy is at https://github.com/cdeust/automatised-pipeline/blob/main/CONTRIBUTING.md#testing-policy-mandatory: new functionality MUST ship with tests in the same pull request, every bug fix MUST ship a regression test that fails on the pre-fix code, a defence that only a reviewer enforces MUST be replaced by a mechanical guard, and coverage MUST NOT regress below 80%. The consequence is stated — a PR that adds behaviour and no test is not ready for review.

The requirement is documented for contributors at https://github.com/cdeust/automatised-pipeline/blob/main/CONTRIBUTING.md#testing-policy-mandatory

Maximally strict and enforced, not merely enabled: cargo clippy --workspace --all-targets -- -D warnings plus cargo fmt --all --check run as a required status check on every push and pull request (https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/ci.yml, the clippy job). Allow-attributes require an inline justification per https://github.com/cdeust/automatised-pipeline/blob/main/CONTRIBUTING.md#coding-standards-excerpt

https://github.com/cdeust/automatised-pipeline/blob/main/docs/ASSURANCE-CASE.md Claim 4 maps each principle to code: fail-safe defaults and allowlisting (validate_safe_id accepts only [A-Za-z0-9._-] and rejects empty, over-long, leading-dot and .. ids; require_absolute rejects relative paths and any parent-dir component), complete mediation (validation sits at the single boundary every tool call crosses, so a new tool cannot forget it), economy of mechanism (one escaping function, one size-config resolver, one response budget, each with its own tests), and least privilege (every workflow declares contents: read and elevates only where it must, which is why Scorecard's Token-Permissions check scores 10).

No security mechanism depends on cryptography, so there is no default algorithm choice whose weakening would matter. See the crypto_working justification for the two non-security hash uses.

No security mechanism depends on a cryptographic algorithm, so there is no algorithm choice for a user to switch. The two hashes in the tree are used for a local integrity digest and for content addressing.

The software processes no authentication credentials and no private cryptographic keys. It authenticates nothing and connects to nothing.

The software performs no network communication. It speaks MCP over stdio and opens no socket; indexing is entirely local. Verified rather than assumed: cargo tree -e normal resolves 172 crates and none of them is reqwest, hyper, ureq, curl, tokio, async-std, rustls or openssl (checked 2026-07-27), and the exclusion of an HTTP client is a stated design decision in https://github.com/cdeust/automatised-pipeline#dependencies

The software does not use TLS; it performs no network communication at all. See crypto_used_network.

The software does not use TLS, so there is no certificate verification to perform. See crypto_used_network.

The software does not use TLS and sends no HTTP headers, private or otherwise. See crypto_used_network.

Met as of v0.8.3, measured 2026-07-29. Every release asset carries a provenance attestation AND publishes the Sigstore bundle beside it: gh attestation verify <asset> --repo cdeust/automatised-pipeline succeeds for all three platform tarballs, the .mcpb and the CycloneDX SBOM, and the release page carries a sibling <asset>.sigstore.json for each. Publishing the bundle is what makes the signature visible at all to an offline consumer and to Scorecard, whose Signed-Releases check scores on the presence of such an asset rather than on the attestation API (ossf/scorecard docs/checks.md). Release run 30402899199. The windows-x86_64 asset is absent from v0.8.3 — its build failed on an MSVC link error, tracked as issue #176 — so this criterion is answered for the assets that shipped, not for a platform that did not.

All input from potentially untrusted sources is checked against an allowlist at the boundary it crosses. Tool arguments: validate_safe_id accepts only [A-Za-z0-9._-]+ with a length cap and rejects empty, leading-dot and .. ids; require_absolute rejects relative paths and any parent-dir component; each tool rejects a missing required field with a named error rather than a default. Repository content: symbol names and paths derived from indexed files pass through the single graph_store::cypher_str escaping choke point, whose reintroduction guard is https://github.com/cdeust/automatised-pipeline/blob/main/tests/no_naive_cypher_escape.rs, and parser input is size- and time-bounded. Responses are budgeted to a derived 100,000-character cap (https://github.com/cdeust/automatised-pipeline/blob/main/src/response_budget.rs). The full argument is Claims 1, 3 and 4 of https://github.com/cdeust/automatised-pipeline/blob/main/docs/ASSURANCE-CASE.md

https://github.com/cdeust/automatised-pipeline/blob/main/docs/ASSURANCE-CASE.md — the threat model (what the tool reads, why its input is untrusted, what privileges it runs with), a table of the six trust boundaries and where each is mediated, five claims with the evidence and the stated limit of each, an explicit argument that secure design principles are applied, and a table mapping each common implementation weakness class to what counters it and what residual risk remains. It closes with what the case does not claim.

CodeQL runs with queries: security-and-quality, which includes the common-vulnerability query packs, over the rust and actions languages: https://github.com/cdeust/automatised-pipeline/blob/main/.github/workflows/codeql.yml

The fuzz targets run under AddressSanitizer: cargo-fuzz's --sanitizer option defaults to address (rust-fuzz/cargo-fuzz, src/options.rs) and the workflow does not override it. The fuzz profile also keeps debug-assertions and overflow-checks on (https://github.com/cdeust/automatised-pipeline/blob/main/fuzz/Cargo.toml), so a violated invariant becomes a crash the fuzzer can find. This matters here because the parser runtime underneath is C.