@menteeai/menteeswe 0.1.12 → 0.1.13

package/dist/cli.js

@@ -194,6 +194,154 @@ var init_exec = __esm({
   }
 });
 
+// src/tools/testing.ts
+import fs4 from "fs";
+import path4 from "path";
+function readPackageJson(cwd) {
+  try {
+    return JSON.parse(fs4.readFileSync(path4.join(cwd, "package.json"), "utf8"));
+  } catch {
+    return null;
+  }
+}
+function readTextIfExists(cwd, file) {
+  try {
+    return fs4.readFileSync(path4.join(cwd, file), "utf8");
+  } catch {
+    return null;
+  }
+}
+function fileExists(cwd, file) {
+  try {
+    return fs4.existsSync(path4.join(cwd, file));
+  } catch {
+    return false;
+  }
+}
+function detectTestCommand(cwd) {
+  const pkg = readPackageJson(cwd);
+  const testScript = pkg && typeof pkg.scripts?.test === "string" ? pkg.scripts.test : null;
+  if (testScript && !/error|no test specified/i.test(testScript)) {
+    return "npm test";
+  }
+  if (fileExists(cwd, "pytest.ini") || fileExists(cwd, "tests")) return "pytest -q";
+  const pyproject = readTextIfExists(cwd, "pyproject.toml");
+  if (pyproject && /\[tool\.pytest/.test(pyproject)) return "pytest -q";
+  if (fileExists(cwd, "go.mod")) return "go test ./...";
+  if (fileExists(cwd, "Cargo.toml")) return "cargo test";
+  return null;
+}
+function detectLintCommand(cwd) {
+  if (fileExists(cwd, "eslint.config.js") || fileExists(cwd, "eslint.config.mjs") || fileExists(cwd, ".eslintrc.json") || fileExists(cwd, ".eslintrc.js")) {
+    return "npx --no-install eslint .";
+  }
+  if (fileExists(cwd, "biome.json")) return "npx --no-install biome check .";
+  if (fileExists(cwd, ".ruff.toml") || fileExists(cwd, "ruff.toml")) return "ruff check .";
+  const pyproject = readTextIfExists(cwd, "pyproject.toml");
+  if (pyproject && /\[tool\.ruff/.test(pyproject)) return "ruff check .";
+  return null;
+}
+function detectTypecheckCommand(cwd) {
+  if (fileExists(cwd, "tsconfig.json")) return "npx --no-install tsc --noEmit";
+  if (fileExists(cwd, "mypy.ini")) return "python -m mypy .";
+  const pyproject = readTextIfExists(cwd, "pyproject.toml");
+  if (pyproject && /\[tool\.mypy/.test(pyproject)) return "python -m mypy .";
+  if (fileExists(cwd, "go.mod")) return "go vet ./...";
+  if (fileExists(cwd, "Cargo.toml")) return "cargo check";
+  return null;
+}
+function makeRunnerTool(config) {
+  return {
+    name: config.name,
+    description: config.description,
+    risk: "safe",
+    timeoutMs: DEFAULT_TIMEOUT_MS2,
+    parameters: {
+      type: "object",
+      properties: {
+        command: {
+          type: "string",
+          description: "Exact command to run. If omitted, the project type is auto-detected."
+        },
+        timeout_ms: { type: "number", description: `Timeout in ms. Defaults to ${DEFAULT_TIMEOUT_MS2}.` }
+      }
+    },
+    async execute(args, ctx) {
+      let command = firstString(args, "command");
+      if (!command) {
+        command = config.detect(ctx.cwd) ?? void 0;
+        if (!command) {
+          return {
+            success: false,
+            output: `Could not auto-detect a command for this project. ${config.fallbackHint} Use the command argument to run one explicitly.`
+          };
+        }
+      }
+      const timeoutMs = Math.min(6e5, firstNumber(args, "timeout_ms") ?? DEFAULT_TIMEOUT_MS2);
+      const outcome = await runShellCommand(command, ctx.cwd, timeoutMs);
+      const formatted = formatOutcome(outcome);
+      return { success: formatted.success, output: formatted.output, truncated: formatted.truncated };
+    }
+  };
+}
+var DEFAULT_TIMEOUT_MS2, runTests, runLinter, runTypecheck, inspectEnv;
+var init_testing = __esm({
+  "src/tools/testing.ts"() {
+    "use strict";
+    init_base();
+    init_exec();
+    DEFAULT_TIMEOUT_MS2 = 18e4;
+    runTests = makeRunnerTool({
+      name: "run_tests",
+      description: "Run the project test suite. Auto-detects npm test / pytest / go test / cargo test from the project files, or run an exact command. Use this to verify changes.",
+      detect: detectTestCommand,
+      fallbackHint: "No test runner found (looked for package.json scripts.test, pytest, go, cargo)."
+    });
+    runLinter = makeRunnerTool({
+      name: "run_linter",
+      description: "Run the project linter (eslint / biome / ruff auto-detected).",
+      detect: detectLintCommand,
+      fallbackHint: "No linter config found (looked for eslint, biome, ruff)."
+    });
+    runTypecheck = makeRunnerTool({
+      name: "run_typecheck",
+      description: "Run static type checking (tsc --noEmit / mypy / go vet / cargo check auto-detected).",
+      detect: detectTypecheckCommand,
+      fallbackHint: "No typechecker config found (looked for tsconfig.json, mypy, go, cargo)."
+    });
+    inspectEnv = {
+      name: "inspect_env",
+      description: "Inspect the environment: OS info, and which toolchains are installed with versions (node, npm, python, pip, git). Use this to adapt commands to the machine.",
+      risk: "safe",
+      parameters: { type: "object", properties: {} },
+      async execute(_args, ctx) {
+        const lines = [
+          `os: ${process.platform} ${process.version} ${process.arch}`,
+          `workspace: ${ctx.cwd}`
+        ];
+        const probes = [
+          ["node", "node --version"],
+          ["npm", "npm --version"],
+          ["npx", "npx --version"],
+          ["python", "python --version"],
+          ["pip", "pip --version"],
+          ["git", "git --version"],
+          ["docker", "docker --version"]
+        ];
+        const results = await Promise.all(
+          probes.map(async ([label, probe]) => {
+            const outcome = await runShellCommand(probe, ctx.cwd, 8e3);
+            const version2 = `${outcome.stdout} ${outcome.stderr}`.trim().split("\n")[0] ?? "";
+            return `${label}: ${outcome.exitCode === 0 ? version2 || "installed (version unknown)" : "not available"}`;
+          })
+        );
+        lines.push(...results);
+        return { success: true, output: lines.join("\n") };
+      }
+    };
+  }
+});
+
 // src/events.ts
 var EventBus;
 var init_events = __esm({
@@ -635,6 +783,7 @@ function createAgentState(task, maxIterations) {
     sameErrorCount: 0,
     totalToolCalls: 0,
     readCount: 0,
+    verifyAttempts: 0,
     usage: { inputTokens: 0, outputTokens: 0, modelRequests: 0 },
     startedAt: Date.now()
   };
@@ -724,6 +873,31 @@ function trimOldToolResults(messages) {
     message.content = `\u27E8prior tool output trimmed: ${head}\u2026 (${content.length} chars)\u27E9`;
   }
 }
+async function runVerification(tools, ctx, bus, state, cwd) {
+  const checks = [
+    { tool: tools.get("run_typecheck"), detect: detectTypecheckCommand },
+    { tool: tools.get("run_tests"), detect: detectTestCommand }
+  ];
+  let combined = "";
+  for (const { tool, detect } of checks) {
+    if (!tool) continue;
+    if (!detect(cwd)) continue;
+    bus.emit("tool_started", `verify \xB7 ${tool.name}`, { tool: tool.name });
+    const res = await tool.execute({}, ctx);
+    bus.emit("tool_completed", `verify \xB7 ${tool.name}`, {
+      tool: tool.name,
+      success: res.success,
+      output: res.output.slice(0, 2e3)
+    });
+    combined += `${tool.name}: ${res.output}
+`;
+    if (!res.success) {
+      state.errors.push(res.output.slice(0, 500));
+      return { ok: false, output: combined };
+    }
+  }
+  return { ok: true, output: combined };
+}
 async function runAgent(options) {
   const {
     task,
@@ -756,7 +930,12 @@ ${systemExtra}` : "");
     }
   };
   const messages = [{ role: "user", content: task }];
+  const setPhase = (p) => {
+    state.phase = p;
+    bus.emit("phase", p);
+  };
   bus.emit("task_started", task, { provider: provider.name, model: model ?? provider.defaultModel });
+  setPhase("EXPLORATION");
   let finalText = "";
   let sawFinish = false;
   while (state.iteration < maxIterations) {
@@ -828,6 +1007,27 @@ ${systemExtra}` : "");
     if (response.toolCalls.length === 0) {
       finalText = response.content ?? "";
       sawFinish = true;
+      if (state.modifiedFiles.length > 0 && state.verifyAttempts < MAX_VERIFY) {
+        setPhase("VALIDATION");
+        const verdict = await runVerification(tools, toolCtx, bus, state, cwd);
+        if (!verdict.ok) {
+          state.verifyAttempts += 1;
+          if (state.verifyAttempts >= MAX_VERIFY) {
+            bus.emit("warning", "Verification failed after multiple attempts; finishing with the issue reported.");
+            break;
+          }
+          setPhase("DEBUGGING");
+          messages.push({
+            role: "user",
+            content: `SYSTEM NOTE: Verification failed.
+${verdict.output}
+Investigate the root cause, fix it, then finish with a final answer.`
+          });
+          sawFinish = false;
+          continue;
+        }
+        setPhase("REVIEW");
+      }
       break;
     }
     messages.push({
@@ -859,6 +1059,7 @@ ${systemExtra}` : "");
         });
         continue;
       }
+      if (EDIT_TOOLS.has(tool.name) && state.phase === "EXPLORATION") setPhase("IMPLEMENTATION");
       bus.emit("tool_started", toolCallPreview(call), { tool: tool.name, args });
       const risk = tool.dynamicRisk ? tool.dynamicRisk(args) : tool.risk;
       let approved = true;
@@ -969,7 +1170,7 @@ ${systemExtra}` : "");
   });
   return { success, finalText, state };
 }
-var MAX_CONTEXT_CHARS, READ_BUDGET, RATE_LIMIT_MAX_ATTEMPTS, KEEP_RECENT_TOOL;
+var MAX_CONTEXT_CHARS, READ_BUDGET, RATE_LIMIT_MAX_ATTEMPTS, MAX_VERIFY, EDIT_TOOLS, KEEP_RECENT_TOOL;
 var init_loop = __esm({
   "src/agent/loop.ts"() {
     "use strict";
@@ -977,9 +1178,12 @@ var init_loop = __esm({
     init_conversation();
     init_state();
     init_base();
+    init_testing();
     MAX_CONTEXT_CHARS = 7e4;
     READ_BUDGET = 6;
     RATE_LIMIT_MAX_ATTEMPTS = 8;
+    MAX_VERIFY = 3;
+    EDIT_TOOLS = /* @__PURE__ */ new Set(["apply_patch", "write_file", "move_path"]);
     KEEP_RECENT_TOOL = 4;
   }
 });
@@ -988,7 +1192,7 @@ var init_loop = __esm({
 var version;
 var init_package = __esm({
   "package.json"() {
-    version = "0.1.12";
+    version = "0.1.13";
   }
 });
 
@@ -1354,6 +1558,7 @@ function App(props) {
   const [showDetails, setShowDetails] = useState3(false);
   const [thinkingText, setThinkingText] = useState3("");
   const [currentEdit, setCurrentEdit] = useState3(null);
+  const [phase, setPhase] = useState3("");
   const finalTextRef = useRef(null);
   const pendingTool = useRef(null);
   const counter = useRef(0);
@@ -1374,6 +1579,10 @@ function App(props) {
         if (delta) setThinkingText((prev) => (prev + delta).slice(-4e3));
         return;
       }
+      if (event.type === "phase") {
+        setPhase(event.message ?? "");
+        return;
+      }
       if (event.type === "model_request" || event.type === "task_completed") {
         setThinkingText("");
       }
@@ -1767,7 +1976,11 @@ function App(props) {
             SPINNER_FRAMES[frame],
             " "
           ] }),
-          /* @__PURE__ */ jsx6(Text6, { dimColor: true, children: "agent working... (Ctrl+C to exit)" })
+          /* @__PURE__ */ jsxs6(Text6, { dimColor: true, children: [
+            "agent ",
+            PHASE_LABEL[phase] ?? "working",
+            "\u2026 (Ctrl+C to exit)"
+          ] })
         ] }),
         thinkingText ? /* @__PURE__ */ jsxs6(Box6, { children: [
           /* @__PURE__ */ jsx6(Text6, { color: "cyan", children: "\u{1F4AD} " }),
@@ -1789,7 +2002,7 @@ function App(props) {
     ] })
   ] });
 }
-var SPINNER_FRAMES, PROVIDER_NAMES, PROVIDER_DESCRIPTIONS, BANNER;
+var SPINNER_FRAMES, PHASE_LABEL, PROVIDER_NAMES, PROVIDER_DESCRIPTIONS, BANNER;
 var init_App = __esm({
   "src/tui/App.tsx"() {
     "use strict";
@@ -1808,6 +2021,16 @@ var init_App = __esm({
     init_KeyDialog();
     init_package();
     SPINNER_FRAMES = ["\u280B", "\u2819", "\u2839", "\u2838", "\u283C", "\u2834", "\u2826", "\u2827", "\u2807", "\u280F"];
+    PHASE_LABEL = {
+      INITIALIZATION: "starting",
+      EXPLORATION: "exploring",
+      PLANNING: "planning",
+      IMPLEMENTATION: "implementing",
+      VALIDATION: "verifying",
+      DEBUGGING: "debugging",
+      REVIEW: "reviewing",
+      COMPLETION: "finalizing"
+    };
     PROVIDER_NAMES = ["kimi", "glm", "zai", "zai-coding", "openrouter", "nvidia", "mock"];
     PROVIDER_DESCRIPTIONS = {
       kimi: "Moonshot \xB7 api.moonshot.ai",
@@ -2324,20 +2547,43 @@ var applyPatch = {
     if (oldString === newString) {
       return { success: false, output: "old_string and new_string are identical; nothing to change." };
     }
-    const occurrences = content.split(oldString).length - 1;
-    if (occurrences === 0) {
-      return {
-        success: false,
-        output: "old_string not found in file. Read the file again and copy the exact text, including whitespace."
-      };
+    const tryMatch = (c, o, n) => {
+      const occ = c.split(o).length - 1;
+      return occ === 1 ? c.replace(o, n) : null;
+    };
+    let updated = tryMatch(content, oldString, newString);
+    let note = "";
+    if (updated === null) {
+      const lfContent = content.replace(/\r\n/g, "\n");
+      const lfOld = oldString.replace(/\r\n/g, "\n");
+      const lfNew = newString.replace(/\r\n/g, "\n");
+      const lfUpdated = tryMatch(lfContent, lfOld, lfNew);
+      if (lfUpdated !== null) {
+        updated = lfUpdated;
+        note = " (applied after normalizing CRLF \u2192 LF line endings)";
+      }
     }
-    if (occurrences > 1) {
+    if (updated === null) {
+      const strip = (s) => s.split("\n").map((l) => l.replace(/\s+$/, "")).join("\n");
+      const sUpdated = tryMatch(strip(content), strip(oldString), strip(newString));
+      if (sUpdated !== null) {
+        updated = sUpdated;
+        note = " (applied after trimming trailing whitespace to match)";
+      }
+    }
+    if (updated === null) {
+      const occ = content.split(oldString).length - 1;
+      if (occ === 0) {
+        return {
+          success: false,
+          output: "old_string not found in file. Read the file again and copy the exact text, including whitespace."
+        };
+      }
       return {
         success: false,
-        output: `old_string occurs ${occurrences} times. Include more surrounding lines to make it unique.`
+        output: `old_string occurs ${occ} times. Include more surrounding lines to make it unique.`
       };
     }
-    const updated = content.replace(oldString, newString);
     fs2.writeFileSync(filePath, updated, "utf8");
     const removed = oldString.split(/\r?\n/).length;
     const added = newString.split(/\r?\n/).length;
@@ -2901,147 +3147,8 @@ var executeCommand = {
   }
 };
 
-// src/tools/testing.ts
-init_base();
-init_exec();
-import fs4 from "fs";
-import path4 from "path";
-var DEFAULT_TIMEOUT_MS2 = 18e4;
-function readPackageJson(cwd) {
-  try {
-    return JSON.parse(fs4.readFileSync(path4.join(cwd, "package.json"), "utf8"));
-  } catch {
-    return null;
-  }
-}
-function readTextIfExists(cwd, file) {
-  try {
-    return fs4.readFileSync(path4.join(cwd, file), "utf8");
-  } catch {
-    return null;
-  }
-}
-function fileExists(cwd, file) {
-  try {
-    return fs4.existsSync(path4.join(cwd, file));
-  } catch {
-    return false;
-  }
-}
-function detectTestCommand(cwd) {
-  const pkg = readPackageJson(cwd);
-  const testScript = pkg && typeof pkg.scripts?.test === "string" ? pkg.scripts.test : null;
-  if (testScript && !/error|no test specified/i.test(testScript)) {
-    return "npm test";
-  }
-  if (fileExists(cwd, "pytest.ini") || fileExists(cwd, "tests")) return "pytest -q";
-  const pyproject = readTextIfExists(cwd, "pyproject.toml");
-  if (pyproject && /\[tool\.pytest/.test(pyproject)) return "pytest -q";
-  if (fileExists(cwd, "go.mod")) return "go test ./...";
-  if (fileExists(cwd, "Cargo.toml")) return "cargo test";
-  return null;
-}
-function detectLintCommand(cwd) {
-  if (fileExists(cwd, "eslint.config.js") || fileExists(cwd, "eslint.config.mjs") || fileExists(cwd, ".eslintrc.json") || fileExists(cwd, ".eslintrc.js")) {
-    return "npx --no-install eslint .";
-  }
-  if (fileExists(cwd, "biome.json")) return "npx --no-install biome check .";
-  if (fileExists(cwd, ".ruff.toml") || fileExists(cwd, "ruff.toml")) return "ruff check .";
-  const pyproject = readTextIfExists(cwd, "pyproject.toml");
-  if (pyproject && /\[tool\.ruff/.test(pyproject)) return "ruff check .";
-  return null;
-}
-function detectTypecheckCommand(cwd) {
-  if (fileExists(cwd, "tsconfig.json")) return "npx --no-install tsc --noEmit";
-  if (fileExists(cwd, "mypy.ini")) return "python -m mypy .";
-  const pyproject = readTextIfExists(cwd, "pyproject.toml");
-  if (pyproject && /\[tool\.mypy/.test(pyproject)) return "python -m mypy .";
-  if (fileExists(cwd, "go.mod")) return "go vet ./...";
-  if (fileExists(cwd, "Cargo.toml")) return "cargo check";
-  return null;
-}
-function makeRunnerTool(config) {
-  return {
-    name: config.name,
-    description: config.description,
-    risk: "safe",
-    timeoutMs: DEFAULT_TIMEOUT_MS2,
-    parameters: {
-      type: "object",
-      properties: {
-        command: {
-          type: "string",
-          description: "Exact command to run. If omitted, the project type is auto-detected."
-        },
-        timeout_ms: { type: "number", description: `Timeout in ms. Defaults to ${DEFAULT_TIMEOUT_MS2}.` }
-      }
-    },
-    async execute(args, ctx) {
-      let command = firstString(args, "command");
-      if (!command) {
-        command = config.detect(ctx.cwd) ?? void 0;
-        if (!command) {
-          return {
-            success: false,
-            output: `Could not auto-detect a command for this project. ${config.fallbackHint} Use the command argument to run one explicitly.`
-          };
-        }
-      }
-      const timeoutMs = Math.min(6e5, firstNumber(args, "timeout_ms") ?? DEFAULT_TIMEOUT_MS2);
-      const outcome = await runShellCommand(command, ctx.cwd, timeoutMs);
-      const formatted = formatOutcome(outcome);
-      return { success: formatted.success, output: formatted.output, truncated: formatted.truncated };
-    }
-  };
-}
-var runTests = makeRunnerTool({
-  name: "run_tests",
-  description: "Run the project test suite. Auto-detects npm test / pytest / go test / cargo test from the project files, or run an exact command. Use this to verify changes.",
-  detect: detectTestCommand,
-  fallbackHint: "No test runner found (looked for package.json scripts.test, pytest, go, cargo)."
-});
-var runLinter = makeRunnerTool({
-  name: "run_linter",
-  description: "Run the project linter (eslint / biome / ruff auto-detected).",
-  detect: detectLintCommand,
-  fallbackHint: "No linter config found (looked for eslint, biome, ruff)."
-});
-var runTypecheck = makeRunnerTool({
-  name: "run_typecheck",
-  description: "Run static type checking (tsc --noEmit / mypy / go vet / cargo check auto-detected).",
-  detect: detectTypecheckCommand,
-  fallbackHint: "No typechecker config found (looked for tsconfig.json, mypy, go, cargo)."
-});
-var inspectEnv = {
-  name: "inspect_env",
-  description: "Inspect the environment: OS info, and which toolchains are installed with versions (node, npm, python, pip, git). Use this to adapt commands to the machine.",
-  risk: "safe",
-  parameters: { type: "object", properties: {} },
-  async execute(_args, ctx) {
-    const lines = [
-      `os: ${process.platform} ${process.version} ${process.arch}`,
-      `workspace: ${ctx.cwd}`
-    ];
-    const probes = [
-      ["node", "node --version"],
-      ["npm", "npm --version"],
-      ["npx", "npx --version"],
-      ["python", "python --version"],
-      ["pip", "pip --version"],
-      ["git", "git --version"],
-      ["docker", "docker --version"]
-    ];
-    const results = await Promise.all(
-      probes.map(async ([label, probe]) => {
-        const outcome = await runShellCommand(probe, ctx.cwd, 8e3);
-        const version2 = `${outcome.stdout} ${outcome.stderr}`.trim().split("\n")[0] ?? "";
-        return `${label}: ${outcome.exitCode === 0 ? version2 || "installed (version unknown)" : "not available"}`;
-      })
-    );
-    lines.push(...results);
-    return { success: true, output: lines.join("\n") };
-  }
-};
+// src/tools/registry.ts
+init_testing();
 
 // src/tools/memory.ts
 init_config();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@menteeai/menteeswe",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "description": "MenteE SWE — a model-agnostic autonomous software-engineering agent CLI. Bring your own intelligence: Kimi, GLM/Z.ai, and more.",
   "type": "module",
   "license": "MIT",