agent-roast 0.12.1 → 0.13.0

package/README.md

@@ -40,7 +40,8 @@ npx agent-roast          # review what you have not committed yet (on a clean tr
 npx agent-roast hook install          # Claude Code, Codex and Cursor
 ```
 
-For Claude Code there is also a plugin: `/plugin install agent-roast --marketplace Nari-Ab/agent-roast`.
+For Claude Code there is also a plugin: `/plugin install agent-roast --marketplace Nari-Ab/agent-roast`. Any agent that speaks
+MCP can also call agent-roast as a read-only tool, `agent-roast mcp` ([setup](docs/integrations.md#mcp-server)).
 
 When the agent is about to finish, agent-roast reviews everything it changed. If a check got weaker, the agent is **sent back
 once**, with the file, the line and the before and after, and told to fix the code instead. If it tries again, you get a
@@ -110,7 +111,7 @@ old debt with a baseline: [docs/reference.md](docs/reference.md#configuration).
 Measured, with the denominators and the caveats in [docs/benchmarks.md](docs/benchmarks.md). **Weakened checks are rare, mostly
 deliberate, and not specific to AI. agent-roast is a tripwire that stays silent until a check gets weaker.**
 
-- **64 of 64** hand-labeled cheating edits are caught and **44 of 44** harmless look-alikes (reformats, moves, raised
+- **71 of 71** hand-labeled cheating edits are caught and **51 of 51** harmless look-alikes (reformats, moves, raised
   thresholds, environment-gated skips) stay silent. The cases were written with the detector, so they guard against regressions;
   they do not prove accuracy. Three gaps are documented and still open.
 - **12,645 commits** that touch tests or CI/config in 166 public repositories: blocking rules fired on 0.3% of AI-signed commits
@@ -120,6 +121,10 @@ deliberate, and not specific to AI. agent-roast is a tripwire that stays silent
 - **About one finding in five was a false alarm**, and most of the rest were deliberate, explained changes a reviewer should
   see rather than mistakes. The labels come from one reader (an AI assistant), so treat them as a first estimate.
 - `replay` ran on 179 public repositories that use coding agents: no commit would have been blocked in 149 of them.
+- The rules for **eight more languages** (Go, Rust, Java, Kotlin, C#, Ruby, PHP, Swift) ran on the newest 300 commits of 39
+  repositories and every finding was read: 26 of 99 were skips that depend on the environment, now fixed; of the 73 left, 56 are
+  tests that really are switched off and 17 are deliberate test data. One reader, and the fixes were made on the same
+  repositories, so expect a few new false alarms elsewhere.
 
 ## How it compares
 

package/dist/{chunk-AP5Y4CH6.js → chunk-W6JP5YWF.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 // src/index.ts
-import path11 from "path";
+import path12 from "path";
 
 // src/git.ts
 import { spawn, spawnSync } from "child_process";
@@ -1191,7 +1191,8 @@ var GO_GUARD = {
 var RUBY_GUARD = {
   modifier: /^\s*(?:skip|pending)\b.*\b(?:if|unless)\b/,
   conditional: /^\s*(?:if|unless|else|elsif|case|when|begin|rescue|ensure|while|until)\b/,
-  test: /^\s*(?:it|specify|example|scenario|test|before|after|around)\b|^\s*def\s+test_/,
+  // Hooks (before, after, around) are not listed: a hook declared under `elsif RUBY_VERSION ...` gates its skip like an example does.
+  test: /^\s*(?:it|specify|example|scenario|test)\b|^\s*def\s+test_/,
   helper: /^\s*def\s+\w+/
 };
 var PHP_GUARD = {
@@ -1201,9 +1202,10 @@ var PHP_GUARD = {
   helper: /^\s*(?:(?:public|protected|private|static)\s+)*function\s+\w+\s*\(/
 };
 var SWIFT_GUARD = {
+  preprocessor: { open: /^\s*#if\b/, close: /^\s*#endif\b/, other: /^\s*#else(?:if)?\b/ },
   conditional: /^\s*(?:\}\s*)?(?:if|else|guard|switch|case|for|while|do|catch)\b/,
-  test: /^\s*(?:@\w+\s+)*(?:(?:private|fileprivate|static)\s+)*func\s+test\w*\s*\(/,
-  helper: /^\s*(?:@\w+\s+)*(?:(?:private|fileprivate|static)\s+)*func\s+\w+\s*\(/
+  test: /^\s*(?:@\w+\s+)*(?:(?:public|open|internal|private|fileprivate|static|final|override|class|nonisolated)\s+)*func\s+test\w*\s*\(/,
+  helper: /^\s*(?:@\w+\s+)*(?:(?:public|open|internal|private|fileprivate|static|final|override|class|nonisolated)\s+)*func\s+\w+\s*\(/
 };
 var LANGUAGE_LINE_RULES = [
   { files: /_test\.go$/i, rules: [
@@ -1226,7 +1228,9 @@ var LANGUAGE_LINE_RULES = [
     { type: "trivial-assert", regex: /\bassertThat\s*\(\s*true\s*\)\s*\.\s*isTrue\s*\(/, name: "trivial assert (assertThat(true).isTrue())" }
   ] },
   { files: /\.cs$/i, rules: [
-    { type: "test-skip", regex: /\[\s*Ignore\b|\b(?:Fact|Theory)\s*\(\s*(?:[^)]*,\s*)?Skip\s*=/, name: "ignored .NET test ([Ignore] / Skip =)" },
+    // `[Ignore]` is checked separately: it also names a mapping attribute (AutoMapper) and must sit next to a test attribute.
+    // Skip = with SkipUnless, SkipWhen or SkipType depends on the environment.
+    { type: "test-skip", regex: /\b(?:Fact|Theory)\s*\(\s*(?:[^)]*,\s*)?Skip\s*=(?![^\]]*\bSkip(?:Unless|When|Type)\b)/, name: "ignored .NET test (Skip =)" },
     { type: "trivial-assert", regex: /\bAssert\s*\.\s*(?:True|IsTrue)\s*\(\s*true\s*\)/, name: "trivial assert (Assert.True(true))" },
     { type: "trivial-assert", regex: /\bAssert\s*\.\s*(?:False|IsFalse)\s*\(\s*false\s*\)/, name: "trivial assert (Assert.False(false))" },
     { type: "trivial-assert", regex: /\bAssert\s*\.\s*(?:AreEqual|Equal)\s*\(\s*(\d+|true|false|null)\s*,\s*\1\s*\)/, name: "trivial assert (Assert.Equal(1, 1))" }
@@ -1236,7 +1240,9 @@ var LANGUAGE_LINE_RULES = [
     { type: "test-focus", regex: /^\s*(?:fit|fdescribe|fcontext|fspecify|fexample|fscenario)\b/, name: "focused RSpec example (fit / fdescribe)" },
     { type: "test-focus", regex: /\b(?:it|describe|context|specify|scenario)\b[^\n]*[,(]\s*(?::focus\b|focus:\s*true\b)/, name: "focused RSpec example (focus: true)" },
     { type: "test-skip", regex: /^\s*(?:xit|xdescribe|xcontext|xspecify|xexample|xscenario|xfeature)\b/, name: "disabled RSpec example (xit / xdescribe)" },
-    { type: "test-skip", regex: /\b(?:it|describe|context|specify|scenario)\b[^\n]*[,(]\s*(?::skip\b|skip:\s*(?:true\b|["']))/, name: "disabled RSpec example (skip: true)" },
+    // `:skip => !feature?` or `skip: RUBY_VERSION < "3"` gates the example on its environment; only a bare flag, true or a reason
+    // disables it. The line is matched with its strings blanked, so a reason shows as nothing between `skip:` and `do`, `{`, `,` or `)`.
+    { type: "test-skip", regex: /\b(?:it|describe|context|specify|scenario)\b[^\n]*[,(]\s*(?::skip\b(?!\s*=>)|(?::skip\s*=>|skip:)\s*(?:true\b|(?=do\b|\{|,|\)|$)))/, name: "disabled RSpec example (skip: true)" },
     { type: "test-skip", regex: /^\s*skip(?:\s|\(|$)/, name: "skipped Ruby test (skip)", guard: RUBY_GUARD },
     { type: "test-skip", regex: /^\s*pending(?:\s+["'(]|\s*$)/, name: "pending Ruby example (pending)", guard: RUBY_GUARD },
     { type: "trivial-assert", regex: /\bexpect\s*\(\s*(true|false|1|0|nil)\s*\)\s*\.\s*to\s+(?:eq|eql|be|equal)\b\s*\(?\s*\1\b/, name: "trivial expectation (expect(true).to eq(true))" },
@@ -1251,29 +1257,53 @@ var LANGUAGE_LINE_RULES = [
   ] },
   { files: /\.swift$/i, rules: [
     { type: "test-skip", regex: /\bthrow\s+XCTSkip\s*\(/, name: "skipped XCTest (XCTSkip)", guard: SWIFT_GUARD },
-    { type: "test-skip", regex: /@(?:Test|Suite)\s*\([^)]*\.disabled\b/, name: "disabled Swift Testing test (.disabled)" },
+    // `.disabled(if: condition)` depends on the environment; `.disabled()` and `.disabled("reason")` switch the test off.
+    { type: "test-skip", regex: /@(?:Test|Suite)\s*\([^)]*\.disabled\s*\((?!\s*if\b)/, name: "disabled Swift Testing test (.disabled)" },
     { type: "trivial-assert", regex: /\bXCTAssertTrue\s*\(\s*true\s*\)/, name: "trivial assert (XCTAssertTrue(true))" },
     { type: "trivial-assert", regex: /\bXCTAssertFalse\s*\(\s*false\s*\)/, name: "trivial assert (XCTAssertFalse(false))" },
     { type: "trivial-assert", regex: /\bXCTAssertEqual\s*\(\s*(\d+|true|false|nil)\s*,\s*\1\s*\)/, name: "trivial assert (XCTAssertEqual(1, 1))" },
     { type: "trivial-assert", regex: /#expect\s*\(\s*true\s*\)/, name: "trivial expectation (#expect(true))" }
   ] }
 ];
+function inTestAttributeList(lines, index) {
+  const marker = /\[\s*(?:Test|TestCase|TestCaseSource|TestMethod|DataTestMethod|TestFixture|TestClass|Fact|Theory)\b/;
+  if (marker.test(lines[index].code)) return true;
+  for (const step of [-1, 1]) {
+    for (let i = index + step, seen = 0; i >= 0 && i < lines.length && seen < 6; i += step, seen++) {
+      if (lines[i].number !== lines[i - step].number + step || !/^\s*\[/.test(lines[i].code)) break;
+      if (marker.test(lines[i].code)) return true;
+    }
+  }
+  return false;
+}
 function isGuardedBlockSkip(lines, profile) {
   const indentOf = (text) => text.length - text.trimStart().length;
   const mentionsCi = (raw) => /\bCI\b|\bGITHUB_ACTIONS\b/.test(raw);
   const own = lines[lines.length - 1];
   if (profile.modifier?.test(own.code) || profile.conditional.test(own.code)) return !mentionsCi(own.raw);
+  const compiledConditionally = (from) => {
+    const directives = profile.preprocessor;
+    if (!directives) return false;
+    let depth = 0;
+    for (let i2 = from; i2 < lines.length - 1; i2++) {
+      if (directives.open.test(lines[i2].code)) depth++;
+      else if (directives.close.test(lines[i2].code)) depth = Math.max(0, depth - 1);
+    }
+    return depth > 0;
+  };
   let limit = indentOf(own.code);
-  for (let i = lines.length - 2; i >= 0 && lines.length - 1 - i <= 200; i--) {
-    if (lines[i].number !== lines[i + 1].number - 1) return false;
+  let i = lines.length - 2;
+  for (; i >= 0 && lines.length - 1 - i <= 200; i--) {
+    if (lines[i].number !== lines[i + 1].number - 1) return compiledConditionally(i + 1);
     const { code, raw } = lines[i];
     if (!code.trim() || indentOf(code) >= limit) continue;
+    if (profile.preprocessor && (profile.preprocessor.open.test(code) || profile.preprocessor.close.test(code) || profile.preprocessor.other.test(code))) continue;
     if (profile.conditional.test(code)) return !mentionsCi(raw);
-    if (profile.test.test(code)) return false;
+    if (profile.test.test(code)) return compiledConditionally(i + 1);
     if (profile.helper.test(code)) return true;
     limit = indentOf(code);
   }
-  return false;
+  return compiledConditionally(i + 1);
 }
 var JS_TEST_API = /(?<![\w$.])(?:(?:it|test|describe|suite)(?:\s*\.\s*(?:concurrent|serial|parallel|describe|failing|sequential))*\s*\.\s*(skip|todo|only)|((?:xit|xtest|xdescribe|fit|fdescribe)))\b((?:\s*\.\s*(?:each|failing|concurrent|serial|parallel|sequential))*)\s*(?=\(|\u0001)/g;
 function isConditionalSkip(source, raw, open2) {
@@ -1470,6 +1500,13 @@ function auditFileDiffs(files, commitHash = "HEAD", commitDate = (/* @__PURE__ *
         }
       }
     }
+    if (/\.cs$/i.test(file.path)) {
+      maskedLines.forEach((line, index) => {
+        if (/\[\s*Ignore\b/.test(line.code) && inTestAttributeList(maskedLines, index)) {
+          infractions.push({ type: "test-skip", file: file.path, lineNumber: line.number, lineSnippet: line.raw.trim(), commitHash, commitDate, reason: "ignored .NET test ([Ignore])" });
+        }
+      });
+    }
     let source = "";
     let code = "";
     let raw = "";
@@ -3674,10 +3711,10 @@ var plural = (count, word) => `${count} ${word}${count === 1 ? "" : "s"}`;
 function clipEvidence(text, limit = 200) {
   if (text.length <= limit) return text;
   const arrow = text.indexOf(" \u2192 ");
-  const clip2 = (part, size) => part.length > size ? `${part.slice(0, size - 1)}\u2026` : part;
-  if (arrow < 0) return clip2(text, limit);
+  const clip3 = (part, size) => part.length > size ? `${part.slice(0, size - 1)}\u2026` : part;
+  if (arrow < 0) return clip3(text, limit);
   const side = Math.floor((limit - 3) / 2);
-  return `${clip2(text.slice(0, arrow), side)} \u2192 ${clip2(text.slice(arrow + 3), side)}`;
+  return `${clip3(text.slice(0, arrow), side)} \u2192 ${clip3(text.slice(arrow + 3), side)}`;
 }
 async function replayHistory(options = {}) {
   assertNotAborted(options.signal);
@@ -3811,11 +3848,11 @@ function describe(finding2, index) {
   const evidence = clip(finding2.lineSnippet || finding2.reason, 220);
   return [`${index}. ${where} [${finding2.type}]`, `   ${evidence}`];
 }
-function formatAgentFeedback(result) {
+function formatAgentFeedback(result, subject = "the uncommitted changes") {
   const ranked = mostImportantFirst(result.infractions, result.failedRules);
   const blocking = ranked.filter((finding2) => result.failedRules.includes(finding2.type));
   const notes = ranked.filter((finding2) => !result.failedRules.includes(finding2.type) && isCheckRule(finding2.type));
-  const lines = ["agent-roast: the uncommitted changes make the checks easier to pass instead of making the code correct.", ""];
+  const lines = [`agent-roast: ${subject} make the checks easier to pass instead of making the code correct.`, ""];
   blocking.slice(0, LISTED).forEach((finding2, index) => lines.push(...describe(finding2, index + 1)));
   if (blocking.length > LISTED) lines.push(`   \u2026 and ${blocking.length - LISTED} more.`);
   if (notes.length) {
@@ -4079,10 +4116,278 @@ function installInstructions(options = {}) {
   });
 }
 
+// src/mcp.ts
+import * as fs9 from "fs/promises";
+import path10 from "path";
+import readline from "readline";
+var MODERN_VERSIONS = ["2026-07-28"];
+var LEGACY_VERSIONS = ["2025-11-25", "2025-06-18", "2025-03-26", "2024-11-05"];
+var VERSION_KEY = "io.modelcontextprotocol/protocolVersion";
+var TOOL_TIMEOUT_MS = 25e3;
+var CACHE_MS = 36e5;
+var MAX_TEXT = 6e3;
+var LISTED_NOTES = 3;
+var INSTRUCTIONS = "Call review_worktree before you tell the user a task is done. It compares your uncommitted changes with HEAD and reports any test, assertion, coverage threshold, lint rule or CI step that was made weaker instead of fixing the code. If it reports findings, fix the code and leave the checks as they were; if a change to a check is really intended, ask the user first.";
+var READ_ONLY = { readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false };
+var PATH_SCHEMA = { type: "string", description: "A directory inside the project. Defaults to the project the server was started in." };
+var MCP_TOOLS = [
+  {
+    name: "review_worktree",
+    title: "Review uncommitted changes",
+    description: "Compare everything not committed yet (staged, unstaged and untracked files) with HEAD and report changes that make tests, assertions, coverage thresholds, lint rules or CI easier to pass instead of fixing the code. Read-only and offline: nothing is run, written or sent.",
+    inputSchema: { type: "object", properties: { path: PATH_SCHEMA }, additionalProperties: false },
+    annotations: { title: "Review uncommitted changes", ...READ_ONLY }
+  },
+  {
+    name: "review_changes",
+    title: "Review a range of commits",
+    description: "Compare two committed revisions, for example a branch against main, and report changes that weaken tests, assertions, coverage thresholds, lint rules or CI. The rules come from the base revision, so the change under review cannot switch them off. Read-only and offline.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        base: { type: "string", description: "The revision to compare against, for example main or origin/main." },
+        head: { type: "string", description: "The revision under review. Defaults to HEAD." },
+        path: PATH_SCHEMA
+      },
+      required: ["base"],
+      additionalProperties: false
+    },
+    annotations: { title: "Review a range of commits", ...READ_ONLY }
+  },
+  {
+    name: "stop_hook",
+    title: "Stop hook adapter",
+    description: "Runs the same check as `agent-roast hook run` for plugin hooks of type mcp_tool and answers with the hook's JSON: it sends an agent back once when its changes weaken checks. Meant to be called by a hook, not by hand.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        cwd: PATH_SCHEMA,
+        stop_hook_active: { type: "string", description: '"true" when the agent is already continuing because of a Stop hook.' },
+        loop_count: { type: "string", description: "How many times the agent was already sent back, for agents that count." }
+      },
+      additionalProperties: true
+    },
+    annotations: { title: "Stop hook adapter", ...READ_ONLY }
+  }
+];
+var ToolInputError = class extends Error {
+};
+var ok = (id, result) => ({ jsonrpc: "2.0", id, result });
+var fail = (id, code, message, data) => ({ jsonrpc: "2.0", id, error: data === void 0 ? { code, message } : { code, message, data } });
+var clip2 = (text) => text.length > MAX_TEXT ? `${text.slice(0, MAX_TEXT - 1)}\u2026` : text;
+function optionalText(value, name) {
+  if (value === void 0 || value === null || value === "") return void 0;
+  if (typeof value !== "string" || value.length > 300 || value.includes("\0")) throw new ToolInputError(`${name} must be a revision name of at most 300 characters.`);
+  return value;
+}
+function summarize(result, subject) {
+  const blocking = result.infractions.filter((finding2) => result.failedRules.includes(finding2.type));
+  const lines = [];
+  if (blocking.length) lines.push(formatAgentFeedback(result, subject));
+  else {
+    lines.push(result.filesChanged ? `agent-roast: no weakened checks in ${subject} (${result.filesInspected} of ${result.filesChanged} changed files compared).` : `agent-roast: nothing changed in ${subject}, so there is nothing to review.`);
+    const notes = result.infractions.filter((finding2) => isCheckRule(finding2.type));
+    if (notes.length) {
+      lines.push("Worth a second look (these do not block):");
+      for (const note of notes.slice(0, LISTED_NOTES)) {
+        const where = `${terminalText(note.file ?? "unknown location")}${note.lineNumber ? `:${note.lineNumber}` : ""}`;
+        lines.push(`- ${where} [${note.type}] ${terminalText(note.lineSnippet || note.reason).replace(/\s+/g, " ").trim().slice(0, 120)}`);
+      }
+      if (notes.length > LISTED_NOTES) lines.push(`- \u2026 and ${notes.length - LISTED_NOTES} more.`);
+    }
+  }
+  if (!result.complete) {
+    const first = result.skippedFiles[0];
+    lines.push(`agent-roast could not compare ${result.skippedFiles.length} changed file${result.skippedFiles.length === 1 ? "" : "s"}${first ? ` (for example ${terminalText(first.file)}: ${terminalText(first.reason).slice(0, 120)})` : ""}, so this result may be incomplete.`);
+  }
+  return lines.join("\n");
+}
+function createMcpServer(options) {
+  const env = options.env ?? process.env;
+  const timeoutMs = options.timeoutMs ?? TOOL_TIMEOUT_MS;
+  const serverInfo = { name: "agent-roast", title: "Agent Roast", version: options.version };
+  const inflight = /* @__PURE__ */ new Map();
+  const cancelled = /* @__PURE__ */ new Set();
+  let allowed;
+  const startDirectory = () => path10.resolve(options.cwd ?? (env.CLAUDE_PROJECT_DIR || env.CURSOR_PROJECT_DIR || process.cwd()));
+  const roots = () => allowed ??= Promise.all(
+    (options.roots?.length ? options.roots : [startDirectory()]).map((root) => fs9.realpath(path10.resolve(root)).catch(() => path10.resolve(root)))
+  );
+  async function gitCommonDirectory(directory) {
+    try {
+      const output = gitOutputPath((await readGit(directory, ["rev-parse", "--git-common-dir"], void 0, 8192)).toString());
+      return await fs9.realpath(path10.resolve(directory, output));
+    } catch {
+      return null;
+    }
+  }
+  async function permitted(real, list) {
+    for (const root of list) {
+      const relative2 = path10.relative(root, real);
+      if (relative2 === "" || relative2 !== ".." && !relative2.startsWith(".." + path10.sep) && !path10.isAbsolute(relative2)) return true;
+    }
+    const common = await gitCommonDirectory(real);
+    if (!common) return false;
+    for (const root of list) if (common === await gitCommonDirectory(root)) return true;
+    return false;
+  }
+  async function project(requested) {
+    if (requested !== void 0 && requested !== null && requested !== "" && (typeof requested !== "string" || requested.includes("\0"))) throw new ToolInputError("path must be a directory path.");
+    const list = await roots();
+    if (!options.roots?.length && list.every((root) => path10.parse(root).root === root)) {
+      throw new ToolInputError("The server was started outside a project (its working directory is the filesystem root). Add --root <directory> to the server's command line.");
+    }
+    const target = typeof requested === "string" && requested ? path10.resolve(list[0], requested) : list[0];
+    let real;
+    try {
+      real = await fs9.realpath(target);
+    } catch {
+      throw new ToolInputError(`${terminalText(typeof requested === "string" && requested ? requested : target)} does not exist.`);
+    }
+    if (!await permitted(real, list)) throw new ToolInputError("That path is outside the project this server was started in. Add --root <directory> to the server's command line to allow another location.");
+    return real;
+  }
+  async function callTool(name, args, signal) {
+    switch (name) {
+      case "review_worktree":
+        return summarize(await reviewWorktreeChanges({ cwd: await project(args.path), signal }), "the uncommitted changes");
+      case "review_changes": {
+        const base = optionalText(args.base, "base");
+        if (!base) throw new ToolInputError("base is required, for example main or origin/main.");
+        const head = optionalText(args.head, "head");
+        const cwd = await project(args.path);
+        return summarize(await reviewChanges({ cwd, base, head, signal }), `the changes between ${terminalText(base)} and ${terminalText(head ?? "HEAD")}`);
+      }
+      case "stop_hook": {
+        const cwd = await project(args.cwd);
+        const input = { hook_event_name: "Stop", cwd, stop_hook_active: args.stop_hook_active === true || args.stop_hook_active === "true" };
+        const loops = Number(args.loop_count);
+        if (args.loop_count !== void 0 && args.loop_count !== "" && Number.isFinite(loops)) input.loop_count = loops;
+        const outcome2 = await runHook(JSON.stringify(input), { agent: "claude", cwd, env, timeoutMs });
+        if (outcome2.stderr) process.stderr.write(outcome2.stderr);
+        return outcome2.stdout.trim() || "{}";
+      }
+      default:
+        throw new ToolInputError(`Unknown tool: ${terminalText(name).slice(0, 100)}`);
+    }
+  }
+  async function toolResult(id, name, rawArguments) {
+    const controller = new AbortController();
+    inflight.set(id, controller);
+    const timer = setTimeout(() => controller.abort(), timeoutMs);
+    try {
+      const args = rawArguments ?? {};
+      if (typeof args !== "object" || Array.isArray(args)) throw new ToolInputError("arguments must be an object.");
+      return { text: clip2(await callTool(name, args, controller.signal)), isError: false };
+    } catch (error) {
+      if (cancelled.has(id)) return null;
+      if (controller.signal.aborted) {
+        const limit = timeoutMs >= 1e3 ? `${Math.round(timeoutMs / 1e3)} seconds` : `${timeoutMs} ms`;
+        return { text: `agent-roast: the check took longer than ${limit} and was stopped. Nothing was changed.`, isError: true };
+      }
+      const message = terminalText(error instanceof Error ? error.message : String(error));
+      return { text: clip2(error instanceof ToolInputError ? message : `agent-roast: ${message}`), isError: true };
+    } finally {
+      clearTimeout(timer);
+      inflight.delete(id);
+      cancelled.delete(id);
+    }
+  }
+  async function dispatch(id, method, params) {
+    const meta = params._meta && typeof params._meta === "object" ? params._meta : {};
+    const requested = meta[VERSION_KEY];
+    const modern = typeof requested === "string";
+    if (modern && !MODERN_VERSIONS.includes(requested)) {
+      return fail(id, -32022, "Unsupported protocol version", { supported: [...MODERN_VERSIONS, ...LEGACY_VERSIONS], requested });
+    }
+    const capabilities = { tools: { listChanged: false } };
+    switch (method) {
+      case "initialize": {
+        const asked = params.protocolVersion;
+        const protocolVersion = typeof asked === "string" && LEGACY_VERSIONS.includes(asked) ? asked : LEGACY_VERSIONS[0];
+        return ok(id, { protocolVersion, capabilities, serverInfo, instructions: INSTRUCTIONS });
+      }
+      case "ping":
+        return ok(id, {});
+      case "server/discover":
+        return ok(id, { resultType: "complete", supportedVersions: [...MODERN_VERSIONS, ...LEGACY_VERSIONS], capabilities, serverInfo, instructions: INSTRUCTIONS, ttlMs: CACHE_MS, cacheScope: "public" });
+      case "tools/list":
+        return ok(id, modern ? { resultType: "complete", tools: MCP_TOOLS, ttlMs: CACHE_MS, cacheScope: "public" } : { tools: MCP_TOOLS });
+      case "tools/call": {
+        const name = params.name;
+        if (typeof name !== "string" || !MCP_TOOLS.some((tool) => tool.name === name)) {
+          return fail(id, -32602, `Unknown tool: ${typeof name === "string" ? terminalText(name).slice(0, 100) : "(no name)"}`);
+        }
+        const outcome2 = await toolResult(id, name, params.arguments);
+        if (!outcome2) return void 0;
+        const content = [{ type: "text", text: outcome2.text }];
+        return ok(id, modern ? { resultType: "complete", content, isError: outcome2.isError } : { content, isError: outcome2.isError });
+      }
+      default:
+        return fail(id, -32601, `Method not found: ${terminalText(method).slice(0, 100)}`);
+    }
+  }
+  async function handleOne(message) {
+    if (!message || typeof message !== "object" || Array.isArray(message) || message.jsonrpc !== "2.0") return fail(null, -32600, "Invalid Request");
+    const request = message;
+    if (typeof request.method !== "string") return void 0;
+    const params = request.params && typeof request.params === "object" && !Array.isArray(request.params) ? request.params : {};
+    if (typeof request.id !== "string" && typeof request.id !== "number") {
+      const target = params.requestId;
+      if (request.method === "notifications/cancelled" && (typeof target === "string" || typeof target === "number") && inflight.has(target)) {
+        cancelled.add(target);
+        inflight.get(target)?.abort();
+      }
+      return void 0;
+    }
+    return dispatch(request.id, request.method, params);
+  }
+  return {
+    async handle(message) {
+      if (!Array.isArray(message)) return handleOne(message);
+      if (!message.length) return fail(null, -32600, "Invalid Request");
+      const responses = (await Promise.all(message.map(handleOne))).filter((response) => response !== void 0);
+      return responses.length ? responses : void 0;
+    }
+  };
+}
+async function serveMcp(options) {
+  const server = createMcpServer(options);
+  const output = options.output ?? process.stdout;
+  output.on("error", () => {
+  });
+  const write = (message) => {
+    output.write(JSON.stringify(message) + "\n");
+  };
+  const lines = readline.createInterface({ input: options.input ?? process.stdin, crlfDelay: Infinity });
+  const running = /* @__PURE__ */ new Set();
+  lines.on("line", (line) => {
+    if (!line.trim()) return;
+    let message;
+    try {
+      message = JSON.parse(line);
+    } catch {
+      write(fail(null, -32700, "Parse error"));
+      return;
+    }
+    const task = server.handle(message).then((response) => {
+      if (response !== void 0) write(response);
+    }).catch((error) => {
+      process.stderr.write(`agent-roast mcp: ${terminalText(error instanceof Error ? error.message : String(error))}
+`);
+    }).finally(() => {
+      running.delete(task);
+    });
+    running.add(task);
+  });
+  await new Promise((resolve2) => lines.once("close", () => resolve2()));
+  await Promise.allSettled([...running]);
+}
+
 // src/demo.ts
-import * as fs9 from "fs";
+import * as fs10 from "fs";
 import os2 from "os";
-import path10 from "path";
+import path11 from "path";
 import { spawnSync as spawnSync4 } from "child_process";
 import pc2 from "picocolors";
 var CART = `export function total(prices: number[], discount = 0): number {
@@ -4124,7 +4429,7 @@ jobs:
 ${lenient ? "        continue-on-error: true\n" : ""}        run: npm test
 `;
 async function runDemo(options = {}) {
-  const dir = fs9.mkdtempSync(path10.join(os2.tmpdir(), "agent-roast-demo-"));
+  const dir = fs10.mkdtempSync(path11.join(os2.tmpdir(), "agent-roast-demo-"));
   try {
     const git2 = (...args) => {
       const run = spawnSync4(
@@ -4138,7 +4443,7 @@ async function runDemo(options = {}) {
           "-c",
           "commit.gpgsign=false",
           "-c",
-          `core.hooksPath=${path10.join(dir, ".no-hooks")}`,
+          `core.hooksPath=${path11.join(dir, ".no-hooks")}`,
           ...args
         ],
         { cwd: dir, env: getHermeticGitEnv(), encoding: "utf8", timeout: 3e4, windowsHide: true }
@@ -4146,9 +4451,9 @@ async function runDemo(options = {}) {
       if (run.error || run.status !== 0) throw new ConfigError(`The demo needs Git to create a temporary repository: ${run.error?.message ?? run.stderr.trim()}`);
     };
     const write = (file, text) => {
-      const destination = path10.join(dir, file);
-      fs9.mkdirSync(path10.dirname(destination), { recursive: true });
-      fs9.writeFileSync(destination, text);
+      const destination = path11.join(dir, file);
+      fs10.mkdirSync(path11.dirname(destination), { recursive: true });
+      fs10.writeFileSync(destination, text);
     };
     git2("init", "-q", "-b", "main");
     write("src/cart.ts", CART);
@@ -4182,7 +4487,7 @@ async function runDemo(options = {}) {
     ].join("\n");
     return { exitCode: result.failedRules.length ? 0 : 1, output, result };
   } finally {
-    fs9.rmSync(dir, { recursive: true, force: true });
+    fs10.rmSync(dir, { recursive: true, force: true });
   }
 }
 
@@ -4194,9 +4499,9 @@ var InvalidDetectorError = class extends Error {
   }
 };
 async function roastRepository(options = {}) {
-  const cwd = path11.resolve(options.cwd || process.cwd());
+  const cwd = path12.resolve(options.cwd || process.cwd());
   const since = options.since || "90 days ago";
-  const repoName = path11.basename(cwd);
+  const repoName = path12.basename(cwd);
   const collector = new MetricsCollector({ all: options.all, ignorePaths: options.ignorePaths, rules: options.rules, panicLoopWindowMinutes: options.panicLoopWindowMinutes });
   const gitStream = streamGitLog(cwd, since);
   await parseGitLogStream(gitStream, (commit, stats) => {
@@ -4211,7 +4516,7 @@ async function roastRepository(options = {}) {
   return { summary, output };
 }
 async function auditPullRequest(options) {
-  const cwd = path11.resolve(options.cwd || process.cwd());
+  const cwd = path12.resolve(options.cwd || process.cwd());
   const format = options.format || "terminal";
   const failOn = (options.failOn || ["test-skip"]).map((s) => s.trim()).filter(Boolean);
   for (const det of failOn) {
@@ -4379,6 +4684,9 @@ export {
   INSTRUCTIONS_END,
   INSTRUCTIONS_BLOCK,
   installInstructions,
+  MCP_TOOLS,
+  createMcpServer,
+  serveMcp,
   runDemo,
   InvalidDetectorError,
   roastRepository,

package/dist/cli.js

@@ -34,10 +34,11 @@ import {
   roastRepository,
   runDemo,
   runHook,
+  serveMcp,
   terminalText,
   uninstallHooks,
   writeBaseline
-} from "./chunk-AP5Y4CH6.js";
+} from "./chunk-W6JP5YWF.js";
 
 // src/cli.ts
 import { Command } from "commander";
@@ -590,4 +591,12 @@ hook.command("resume").description("End a pause").action(async () => {
     process.exitCode = 2;
   }
 });
+program.command("mcp").description("Run as an MCP server over stdio, so any MCP-capable coding agent can review its own changes (read-only, offline)").option("--root <directory...>", "a directory the tools may review (repeatable); defaults to the project the server was started in").action(async (opts) => {
+  try {
+    await serveMcp({ version: pkg.version, roots: opts.root });
+  } catch (error) {
+    console.error(`agent-roast mcp: ${errorMessage(error)}`);
+    process.exitCode = 1;
+  }
+});
 program.parse(process.argv);

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-import { Readable } from 'stream';
+import { Readable, Writable } from 'stream';
 
 /** A reviewed allowance for a precise pattern and number of occurrences.
  * No wildcard, inline bypass, source snippet or executable policy is needed. */
@@ -493,9 +493,9 @@ interface HookOutcome {
 }
 /**
  * What the agent reads when it is sent back: the offending changes, then the instruction to fix the code rather than the
- * checks. Short on purpose, since it is added to the agent's context.
+ * checks. Short on purpose, since it is added to the agent's context. `subject` names what was reviewed.
  */
-declare function formatAgentFeedback(result: ReviewResult): string;
+declare function formatAgentFeedback(result: ReviewResult, subject?: string): string;
 /** The time a pause lasts until, or null when the hook is active. Expired pauses are cleaned up. */
 declare function hookPausedUntil(cwd: string, now?: Date): Promise<Date | null>;
 declare function pauseHook(cwd: string, until: Date): Promise<string>;
@@ -558,6 +558,129 @@ declare function installInstructions(options?: {
     action: "created" | "updated" | "unchanged";
 }[];
 
+/**
+ * A small MCP server over stdio, so any agent that speaks MCP can ask "did my changes weaken a check?" before it says it is
+ * done. It is read-only and offline like the rest of agent-roast, and written without a dependency on purpose: the built files
+ * contain no network code, which a test checks. It answers both protocol eras: the initialize handshake of 2025-11-25 and
+ * earlier, and the per-request metadata of 2026-07-28 (modelcontextprotocol.io/specification/2026-07-28/basic/versioning).
+ */
+type RpcId = string | number;
+type RpcResponse = {
+    jsonrpc: "2.0";
+    id: RpcId | null;
+} & ({
+    result: unknown;
+} | {
+    error: {
+        code: number;
+        message: string;
+        data?: unknown;
+    };
+});
+declare const MCP_TOOLS: readonly [{
+    readonly name: "review_worktree";
+    readonly title: "Review uncommitted changes";
+    readonly description: string;
+    readonly inputSchema: {
+        readonly type: "object";
+        readonly properties: {
+            readonly path: {
+                readonly type: "string";
+                readonly description: "A directory inside the project. Defaults to the project the server was started in.";
+            };
+        };
+        readonly additionalProperties: false;
+    };
+    readonly annotations: {
+        readonly readOnlyHint: true;
+        readonly destructiveHint: false;
+        readonly idempotentHint: true;
+        readonly openWorldHint: false;
+        readonly title: "Review uncommitted changes";
+    };
+}, {
+    readonly name: "review_changes";
+    readonly title: "Review a range of commits";
+    readonly description: string;
+    readonly inputSchema: {
+        readonly type: "object";
+        readonly properties: {
+            readonly base: {
+                readonly type: "string";
+                readonly description: "The revision to compare against, for example main or origin/main.";
+            };
+            readonly head: {
+                readonly type: "string";
+                readonly description: "The revision under review. Defaults to HEAD.";
+            };
+            readonly path: {
+                readonly type: "string";
+                readonly description: "A directory inside the project. Defaults to the project the server was started in.";
+            };
+        };
+        readonly required: readonly ["base"];
+        readonly additionalProperties: false;
+    };
+    readonly annotations: {
+        readonly readOnlyHint: true;
+        readonly destructiveHint: false;
+        readonly idempotentHint: true;
+        readonly openWorldHint: false;
+        readonly title: "Review a range of commits";
+    };
+}, {
+    readonly name: "stop_hook";
+    readonly title: "Stop hook adapter";
+    readonly description: string;
+    readonly inputSchema: {
+        readonly type: "object";
+        readonly properties: {
+            readonly cwd: {
+                readonly type: "string";
+                readonly description: "A directory inside the project. Defaults to the project the server was started in.";
+            };
+            readonly stop_hook_active: {
+                readonly type: "string";
+                readonly description: "\"true\" when the agent is already continuing because of a Stop hook.";
+            };
+            readonly loop_count: {
+                readonly type: "string";
+                readonly description: "How many times the agent was already sent back, for agents that count.";
+            };
+        };
+        readonly additionalProperties: true;
+    };
+    readonly annotations: {
+        readonly readOnlyHint: true;
+        readonly destructiveHint: false;
+        readonly idempotentHint: true;
+        readonly openWorldHint: false;
+        readonly title: "Stop hook adapter";
+    };
+}];
+interface McpServerOptions {
+    /** The version reported to clients. */
+    version: string;
+    /** Directories the tools may read. Defaults to the project the server was started in. */
+    roots?: string[];
+    /** The project directory when the server was started; defaults to CLAUDE_PROJECT_DIR, CURSOR_PROJECT_DIR, then the working directory. */
+    cwd?: string;
+    env?: NodeJS.ProcessEnv;
+    /** How long one tool call may take before it is stopped. */
+    timeoutMs?: number;
+}
+interface McpServer {
+    /** Answers one JSON-RPC message (or a batch), or nothing for a notification. */
+    handle(message: unknown): Promise<RpcResponse | RpcResponse[] | undefined>;
+}
+declare function createMcpServer(options: McpServerOptions): McpServer;
+interface ServeMcpOptions extends McpServerOptions {
+    input?: Readable;
+    output?: Writable;
+}
+/** Reads newline-delimited JSON-RPC from the input and writes responses to the output until the input closes. */
+declare function serveMcp(options: ServeMcpOptions): Promise<void>;
+
 interface DemoResult {
     exitCode: number;
     output: string;
@@ -613,4 +736,4 @@ declare function roastRepository(options?: RoastOptions): Promise<{
 }>;
 declare function auditPullRequest(options: PrAuditOptions): Promise<PrAuditResult>;
 
-export { AGENT_LABELS, type AgentRoastConfig, type AiAttributionResult, CONFIG_FILENAMES, type CheckOptions, type CheckResult, type CommitInfo, ConfigError, DEFAULT_CONFIG, DEFAULT_CONFIG_TEMPLATE, type DemoResult, type DetectorType, FIELD_SEPARATOR, type FindingBaseline, type FindingException, GITHUB_WORKFLOW_TEMPLATE, GitError, GitResolutionError, HERMETIC_GIT_FLAGS, HOOK_AGENTS, type HookAgent, type HookInstallChange, type HookInstallOptions, type HookOptions, type HookOutcome, type HookScope, INSTRUCTIONS_BLOCK, INSTRUCTIONS_END, INSTRUCTIONS_START, type Infraction, type InitOptions, type InitResult, InvalidDetectorError, MAX_CONFIG_BYTES, type MetricSummary, MetricsCollector, PRE_COMMIT_HOOK_TEMPLATE, type ParseDiffResult, type ParsedCommit, type ParsedDiffLine, type ParsedFileDiff, type PrAuditOptions, type PrAuditResult, RECORD_SEPARATOR, RULE_GUIDES, type ReplayCommit, type ReplayOptions, type ReplayResult, type ReplayStatus, type ReviewFormat, type ReviewOptions, type ReviewResult, type RoastOptions, type RuleGuide, type ShieldsEndpointResponse, type StagedReviewOptions, type SuppressedFinding, VALID_DETECTORS, applyBaseline, applyExceptions, auditFileDiffs, auditPullRequest, checkProject, compareFindings, configFromText, createBaseline, detectAiAttribution, findMergeBase, findingFingerprint, formatAgentFeedback, formatCheckMarkdown, formatCheckTerminal, formatGithubAnnotations, formatGithubStepSummary, formatPrMarkdownComment, formatPrTerminalCard, formatReplayTerminal, formatReviewMarkdown, formatReviewTerminal, formatSarif, formatTerminalCard, generateShareUrl, generateShieldsJson, generateShieldsMarkdown, generateSvgBadge, getHermeticGitEnv, getScoreColor, gitOutputPath, hookCommand, hookConfigPath, hookPausedUntil, hookStatus, initProject, installHooks, installInstructions, isCheckRule, isCodeFile, isTestFile, loadBaseline, loadConfig, matchIgnorePath, mostImportantFirst, packageVersion, parseDiff, parseDuration, parseGitLogStream, pauseHook, replayHistory, repositoryRoot, resolveCommitSha, resumeHook, reviewChanges, reviewStagedChanges, reviewWorktreeChanges, roastRepository, runDemo, runHook, streamGitDiff, streamGitLog, stripStringAndRegexLiterals, terminalText, uninstallHooks, unquoteGitPath, validateExceptions, writeBaseline };
+export { AGENT_LABELS, type AgentRoastConfig, type AiAttributionResult, CONFIG_FILENAMES, type CheckOptions, type CheckResult, type CommitInfo, ConfigError, DEFAULT_CONFIG, DEFAULT_CONFIG_TEMPLATE, type DemoResult, type DetectorType, FIELD_SEPARATOR, type FindingBaseline, type FindingException, GITHUB_WORKFLOW_TEMPLATE, GitError, GitResolutionError, HERMETIC_GIT_FLAGS, HOOK_AGENTS, type HookAgent, type HookInstallChange, type HookInstallOptions, type HookOptions, type HookOutcome, type HookScope, INSTRUCTIONS_BLOCK, INSTRUCTIONS_END, INSTRUCTIONS_START, type Infraction, type InitOptions, type InitResult, InvalidDetectorError, MAX_CONFIG_BYTES, MCP_TOOLS, type McpServer, type McpServerOptions, type MetricSummary, MetricsCollector, PRE_COMMIT_HOOK_TEMPLATE, type ParseDiffResult, type ParsedCommit, type ParsedDiffLine, type ParsedFileDiff, type PrAuditOptions, type PrAuditResult, RECORD_SEPARATOR, RULE_GUIDES, type ReplayCommit, type ReplayOptions, type ReplayResult, type ReplayStatus, type ReviewFormat, type ReviewOptions, type ReviewResult, type RoastOptions, type RuleGuide, type ServeMcpOptions, type ShieldsEndpointResponse, type StagedReviewOptions, type SuppressedFinding, VALID_DETECTORS, applyBaseline, applyExceptions, auditFileDiffs, auditPullRequest, checkProject, compareFindings, configFromText, createBaseline, createMcpServer, detectAiAttribution, findMergeBase, findingFingerprint, formatAgentFeedback, formatCheckMarkdown, formatCheckTerminal, formatGithubAnnotations, formatGithubStepSummary, formatPrMarkdownComment, formatPrTerminalCard, formatReplayTerminal, formatReviewMarkdown, formatReviewTerminal, formatSarif, formatTerminalCard, generateShareUrl, generateShieldsJson, generateShieldsMarkdown, generateSvgBadge, getHermeticGitEnv, getScoreColor, gitOutputPath, hookCommand, hookConfigPath, hookPausedUntil, hookStatus, initProject, installHooks, installInstructions, isCheckRule, isCodeFile, isTestFile, loadBaseline, loadConfig, matchIgnorePath, mostImportantFirst, packageVersion, parseDiff, parseDuration, parseGitLogStream, pauseHook, replayHistory, repositoryRoot, resolveCommitSha, resumeHook, reviewChanges, reviewStagedChanges, reviewWorktreeChanges, roastRepository, runDemo, runHook, serveMcp, streamGitDiff, streamGitLog, stripStringAndRegexLiterals, terminalText, uninstallHooks, unquoteGitPath, validateExceptions, writeBaseline };

package/dist/index.js

@@ -16,6 +16,7 @@ import {
   INSTRUCTIONS_START,
   InvalidDetectorError,
   MAX_CONFIG_BYTES,
+  MCP_TOOLS,
   MetricsCollector,
   PRE_COMMIT_HOOK_TEMPLATE,
   RECORD_SEPARATOR,
@@ -29,6 +30,7 @@ import {
   compareFindings,
   configFromText,
   createBaseline,
+  createMcpServer,
   detectAiAttribution,
   findMergeBase,
   findingFingerprint,
@@ -80,6 +82,7 @@ import {
   roastRepository,
   runDemo,
   runHook,
+  serveMcp,
   streamGitDiff,
   streamGitLog,
   stripStringAndRegexLiterals,
@@ -88,7 +91,7 @@ import {
   unquoteGitPath,
   validateExceptions,
   writeBaseline
-} from "./chunk-AP5Y4CH6.js";
+} from "./chunk-W6JP5YWF.js";
 export {
   AGENT_LABELS,
   CONFIG_FILENAMES,
@@ -106,6 +109,7 @@ export {
   INSTRUCTIONS_START,
   InvalidDetectorError,
   MAX_CONFIG_BYTES,
+  MCP_TOOLS,
   MetricsCollector,
   PRE_COMMIT_HOOK_TEMPLATE,
   RECORD_SEPARATOR,
@@ -119,6 +123,7 @@ export {
   compareFindings,
   configFromText,
   createBaseline,
+  createMcpServer,
   detectAiAttribution,
   findMergeBase,
   findingFingerprint,
@@ -170,6 +175,7 @@ export {
   roastRepository,
   runDemo,
   runHook,
+  serveMcp,
   streamGitDiff,
   streamGitLog,
   stripStringAndRegexLiterals,

package/package.json

@@ -1,7 +1,8 @@
 {
   "name": "agent-roast",
-  "version": "0.12.1",
+  "version": "0.13.0",
   "description": "Stops AI coding agents (and rushed humans) from turning a build green by weakening tests, coverage gates or CI. Hook for Claude Code, Codex and Cursor, pre-commit, GitHub Action. Offline.",
+  "mcpName": "io.github.Nari-Ab/agent-roast",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",