v1.1.0

Breaking changes

• Verify report schema is now 2. Consumers must read the separate outcomes: pass, findings failure, incomplete analysis, and non-certifying baseline maintenance.
• --update-baseline now exits 1: it is maintenance, not a passing verification.
• --fail-on applies only to dead code that is explicitly opted in; dead code is advisory by default.
• MCP baseline writes require --allow-baseline-update (baseline-write authority is separate from execution).

Changed

• Verify schema 2 separates pass, findings failure, incomplete analysis, and non-certifying baseline maintenance. Required failures and invalid Git bases no longer pass.
• Independent per-check enforcement; dead code is advisory by default. CLI/MCP library mode and policy agree; baseline-write authority is separate from execution.
• Pinned Git baseline mode, atomic maintenance, explicit baseline read errors, preserved unobserved provider entries, complete totals and coverage caveats.
• Provider exit/output failure handling and an offline gate regression corpus.
• Init checks actual Git worktree status rather than any ancestor named .git.
• See migration details before upgrading gate consumers.

Limits (honest)

• No native Windows or macOS validation of 1.1.0; only Linux was verified for this release (see README for what was tested earlier on Windows).
• Trusted verification mode has documented limits: see docs/VERIFY.md. Tier-0 analysis is name-based; a pass covers only the selected checks.
• The attached binary is linux-x64, built locally with cargo build --release --locked (not CI-built). Verify with sha256sum -c SHA256SUMS.txt.

Plumbgraph v1.0.3

Fixes for false positives and install problems found by running 1.0.2 on a real Unity project (Assets/UYF).

Fixed

• Unity: message methods (LateUpdate, FixedUpdate, OnDisable, ...) on a partial class are entry points when any part declares a Unity base (MonoBehaviour, ScriptableObject, NetworkBehaviour, EditorWindow, ...), even when the base is declared in another file. Parts are merged per type within one Assets/ root; the other parts of the type are then live too. Indirect bases (Hero : CharacterBase : MonoBehaviour) are resolved through classes in the index. dead-code --verbose prints unity-message@partial-merged ... for every suppressed message.
• Unity: a message-named method on a class whose Unity base cannot be decided (base type not in the index) under Assets/ is now reported as MEDIUM at most, with the reason unity-message-name-but-host-type-unresolved, instead of being silently live (partial classes) or HIGH. A class whose bases all resolve to non-Unity classes (DerivedState : PlainBase) is no longer treated as a component, so its Update(float dt) stays reportable.
• Unity: ScriptableWizard is an editor base; OnWizardCreate, OnWizardUpdate, OnWizardOtherButton are messages.

Added

• dead-code prints two notes when C# files under Assets/ are indexed (Unity pack active; do not mass-delete HIGH findings) and --help links the Unity pack limitations. docs/PACKS.md documents the Unity gaps (scenes/prefabs/UnityEvents, string-based invocation, animation events, Addressables/Resources, scip-dotnet, multiple asmdefs).
• plumb mcp startup line on stderr includes the version and pid.
• plumb doctor warns (Windows) when it cannot overwrite its own executable, e.g. because an MCP host still runs plumb mcp.
• Docs: stop MCP hosts before cargo install on Windows (or use --root); plumb mcp --root should point at the workspace you are working on.

Changed

• Index schema 4 / extractor 9 (new unity column on symbols): existing indexes are rebuilt once.
• DeadCodeResult (JSON) has an optional notes array.

Already correct in 1.0.2 (regression tests added only)

• A plain class with Update(float dt) and no base class is not an engine callback; ordinary methods that merely contain "Update" are not entry points; same-file MonoBehaviour.Update is suppressed; a private unreferenced helper on a MonoBehaviour is reported; [MenuItem], [InitializeOnLoad], [InitializeOnLoadMethod], [RuntimeInitializeOnLoadMethod], Editor and EditorWindow handling; MCP initialize already carried serverInfo.version (now asserted).

Plumbgraph v1.0.2

• Docs only: README now matches the released version, the tested Windows scope, supported weakening languages and known limits.

Plumbgraph v1.0.1

Fixes for problems found by using v1.0.0 on real projects (Windows, Unity). Not released; the crate version is still 1.0.0.

Fixed

• Windows: plumb doctor (and enrich, rule-engine detection in verify --run) reported every provider as missing because lookup used the bare name (cargo), never cargo.exe / tsc.cmd. Bare names are now tried with the PATHEXT extensions that std::process::Command can start (.com, .exe, .bat, .cmd); Unix lookup is unchanged.
• plumb init printed a Codex config.toml snippet with a \\?\C:\... path inside a TOML basic string, which does not parse. The verbatim prefix is stripped (\\?\UNC\ becomes \\) and the path is written as a TOML literal string, or an escaped basic string when it contains ' or control characters. Report paths no longer show \\?\ either.
• plumb map --budget N was rejected; --budget is now a visible alias of --tokens.
• plumb init and .gitignore: --dry-run shows the lines it would add; it writes .plumbgraph/* + !.plumbgraph/allow.toml (the previous .plumbgraph/ made the committed allow-list impossible to commit); any existing rule mentioning .plumbgraph (including !.plumbgraph/) is respected; it works in a subdirectory of a repository; CRLF files keep CRLF; a .gitignore that is not UTF-8 (for example UTF-16 from PowerShell 5) is no longer read as empty and overwritten.

Added

• C# / Unity dead-code entry points: MonoBehaviour/ScriptableObject/NetworkBehaviour/StateMachineBehaviour/UIBehaviour messages (Awake, Start, Update, OnTriggerEnter2D, ...), Unity attributes ([RuntimeInitializeOnLoadMethod], [MenuItem], [InitializeOnLoad], ...), and Unity callback interfaces (ISerializationCallbackReceiver, IHasCustomMenu, UI EventSystem handlers). Public methods of Unity classes stay findings with a 0.25 penalty and a UnityEvent / Animation Event fp_risk. Rules and limits: docs/PACKS.md. On Unity's open-project-1 scripts (commit 608eac9) dead-code findings went from 371 to 166 (medium 97 to 28); on Newtonsoft.Json (52fa3ae) they are identical before and after (204).
• plumb init --no-gitignore.
• README: Windows install and PATH setup, and exactly what was tested on Windows.

Changed

• Index schema 3 / extractor 8 (new framework_risk column): existing indexes are rebuilt once.

Plumbgraph v1.0.0

Plumbgraph v1.0.0

Source-only GitHub release. Not published to crates.io.

Know your code. Verify the change. Plumbgraph is a local code-intelligence and pre-submit verification tool for AI coding agents (CLI plumb, MCP server over stdio).

Highlights

• plumb verify: one verdict over dead code, hallucinated/undeclared dependencies, test weakening, diagnostics, semgrep and ast-grep, diffed against plumb-baseline.json so only new findings fail. External tools run only with --run.
• plumb map (PageRank-ranked, hard token budget) and plumb impact (transitive callers, tests to run).
• Languages: Python, JavaScript, TypeScript/TSX, Rust, Go, Java, C# (tier-0, tree-sitter); SCIP overlay for compiler-grade references where you have an index.
• MCP: 13 tools and 3 resources; tested with the official TypeScript SDK and the MCP inspector CLI.
• Provider orchestration (plumb doctor, plumb enrich), plumb init for agent config.
• Every finding carries source and confidence; deps allow-list entries require a written reason.

Since the v1.0 candidate (this release PR)

• Version 1.0.0 everywhere (plumb --version).
• Dead-code false-positive fixes found by hand review of chi and gson: Rust {NAME} format captures, Java main/C# Main, Java serialization hooks, annotation-driven framework methods; app-mode hint to use --lib for libraries. The index cache rebuilds automatically on upgrade.
• [[deps_allow]] in .plumbgraph/allow.toml (path glob, optional package, mandatory reason). plumb verify . on plumbgraph itself now passes.
• plumb verify --fail-on none reports REPORT-ONLY instead of FAIL.
• New logo, social preview, README, demo recordings generated from real runs, architecture and MCP docs.

Validated end to end (see docs/VALIDATION.md)

• Real repos: requests, express, ripgrep, cobra, gson, Newtonsoft.Json, kubernetes (13,558 of 31,421 tracked files were source files and got indexed, in 41 s, no OOM), vscode (14.6k files; 60 s), Wraith, classroom-archiver.
• Fixed from that testing: large-repo memory and speed, weakening detection for Go/Java/C#, alias-import false positives, bounded check-deps/verify output, MCP map token cost, closed-pipe crashes.
• Seeded weakening recall 12/12 on Python, JS, Rust, Go and C#; Java skips and deletions 4/4 each, loosened assertions untested.

Known limitations (read before relying on it)

• Name-based analysis without a SCIP index; dead-code precision on libraries in app mode was 1/40 on chi (use --lib); recall on real code is unmeasured. See docs/EVALUATION.md.
• scip-go/scip-java/scip-dotnet/rust-analyzer SCIP were not run; no Windows testing; MCP agent hosts not tested; no prompts capability.
• Distribution is source-only: no binaries, no crate.

Full changelog: CHANGELOG.md.