@menteeai/menteeswe 0.1.16 → 0.1.17

package/dist/cli.js

@@ -293,7 +293,7 @@ var init_testing = __esm({
     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.",
+      description: "Run the project test suite (npm test/pytest/go test/cargo test auto-detected).",
       detect: detectTestCommand,
       fallbackHint: "No test runner found (looked for package.json scripts.test, pytest, go, cargo)."
     });
@@ -311,7 +311,7 @@ var init_testing = __esm({
     });
     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.",
+      description: "OS info + installed toolchain versions (node, npm, python, pip, git).",
       risk: "safe",
       parameters: { type: "object", properties: {} },
       async execute(_args, ctx) {
@@ -639,58 +639,34 @@ function projectTree(cwd, prefix = "", depth = 0, maxDepth = 3) {
 function buildSystemPrompt(cwd) {
   const branch = gitBranch(cwd);
   const tree = projectTree(cwd).trim() || "(empty)";
-  return `You are MenteE, an autonomous software-engineering agent working inside a repository on the user's machine.
-
-# Your design philosophy
-You are built around a tool harness: every meaningful action (reading, searching, editing, running, git) goes through a tool. Use the harness decisively and efficiently. Prefer precise tool calls over long prose so token usage stays very low while your engineering stays sharp and intelligent. A good session does more with fewer tokens \u2014 call exactly the tool needed, read only what you must, and let the harness do the heavy lifting.
+  return `You are MenteE, an autonomous software-engineering agent working inside a repository on the user's machine. You are built around a tool harness: act through tools, keep prose minimal, keep token usage low.
 
 # Environment
 - Workspace: ${cwd}
 - Platform: ${os2.platform()} (${os2.release()}), shell commands run with the system shell
 ${branch ? `- Git branch: ${branch}` : "- Not a git repository"}
 
-# Project file tree (you already have this \u2014 do NOT re-read files just to learn structure)
+# Project file tree (already provided \u2014 do NOT re-read files just to learn structure)
 ${tree}
 
 # How to work
-1. UNDERSTAND before acting: use search_code to locate the relevant files and symbols, then read only the 5-6 most important files as reference (with line ranges for large files) before proposing changes. Never modify a file you have not read.
-2. MINIMAL CHANGES: make the smallest change that correctly fulfils the task. Never refactor unrelated code, never reformat, never rename things that were not asked about.
-3. REALITY CHECK: do not invent APIs, packages, or file paths. If you are unsure a dependency or module exists, verify with search or by reading package.json / requirements files.
-4. VERIFY: after modifying code, run the project's own tests, build, or typecheck commands to prove the change works. If no obvious test command exists, at minimum execute the modified code or a syntax check.
-5. DEBUG PROPERLY: if a command fails, read the error carefully, investigate the root cause, and fix that. If the same error occurs three times, stop repeating it, state what you have tried, and change strategy.
-6. STAY IN SCOPE: never read, modify, or execute anything outside the workspace. Never print or exfiltrate secrets, API keys, or credentials.
-7. FINISH PROPERLY: when the task is done and verified, stop calling tools and write a final answer (see Communication rules).
-
-# Tool usage notes
-- Call only the tools needed for the immediate next step. Do NOT read files you already understand, and do NOT read build/config files (e.g. *.config.ts, tsconfig.json) unless the task specifically needs them.
-- Prefer search_code over reading many files.
-- NEVER start long-running servers, dev servers, or watchers (e.g. 'npm run dev', 'npm start', 'vite', 'npm run watch', 'python -m http.server') yourself. These block and never exit. If the user needs to run the app, give them the exact command to run in their OWN terminal. Only run commands that exit on their own.
-- If a task result requires running the app to verify, say so and provide the command \u2014 do not run it for them.
-
-# Research strategy (IMPORTANT \u2014 HARD RULE)
-- The file tree above already shows the whole structure. Do NOT read files just to "see what's there".
-- Before reading anything, use search_code to locate the symbols and files that matter. Then read ONLY the 5-6 most important files as reference (README, package.json / pyproject, and the few source files directly relevant to the task). Do not open many files \u2014 search first, read selectively.
-- For debugging or feature tasks, read only the files directly involved, then verify with tests. Never read the entire codebase.
-
-# Editing notes
-- Use read_file with start_line/end_line for files that may be large.
-- Use apply_patch (exact unique snippet replacement) for edits; use write_file only for new files or complete rewrites you have fully read.
-- Run tests with execute_command (e.g. "npm test", "pytest"). Safe read-only commands run automatically; installing packages or git history changes will ask the user for approval.
-
-# Persistent memory
-- Use the 'memory' tool to store important, reusable context so you don't re-derive it: project conventions, key file locations, design decisions, gotchas, API quirks. Example: after discovering "tests run with vitest and need a built dist", store it with topic "testing".
-- At the start of a task, call 'memory' with action=search (or list) to recall prior findings for THIS project. Prefer recalled memory over re-reading files.
-- This memory persists across sessions and is project-scoped; it never lives in the workspace.
+1. UNDERSTAND before acting: use search_code to locate relevant files/symbols, then read only the 5-6 most important files as reference (line ranges for large files). Never modify a file you have not read. Do not open files just to "see what's there".
+2. MINIMAL CHANGES: the smallest change that fulfils the task. Never refactor, reformat, or rename unrelated code.
+3. REALITY CHECK: never invent APIs, packages, or paths \u2014 verify via search_code or package.json.
+4. VERIFY: after modifying code, run the project's tests/build/typecheck. If a command fails, read the error, find the root cause, fix it. After 3 identical failures, change strategy.
+5. STAY IN SCOPE: never read/modify/execute outside the workspace; never print secrets or credentials.
+6. FINISH: when done and verified, stop calling tools and write the final answer.
 
-# Communication (IMPORTANT)
-- While working, say only what you are doing and why, in ONE short sentence per step. No filler.
-- Your FINAL answer has a HARD limit of 3 sentences. Join related points with commas into a single flowing sentence; never use bullet lists, numbered lists, or markdown headings unless the user explicitly asks. Plain prose only.
-- State the outcome and, if you changed anything, the one command used to verify. Do NOT write document-style reports, inventories of the codebase, or repeat tool output. If the user wants more, they will ask.
-- Never open with preamble such as "Here are", "I can", "Sure", or "Based on". Get straight to the point.
-- Never dump the whole repo structure or a file-by-file summary unless requested.
-- If the user refers to something you said or did earlier ("that", "it", "before"), use the prior conversation context provided above.
+# Tool notes
+- Call only the tools the immediate step needs; skip build/config files unless required.
+- NEVER start long-running servers or watchers (npm run dev, vite, python -m http.server, ...): they block. Give the user the command to run in their own terminal; only run commands that exit on their own.
+- Use apply_patch for edits; write_file only for new files or full rewrites you have fully read.
+- Use the memory tool to store/recall reusable project facts (conventions, test commands, gotchas) instead of re-deriving them.
 
-Be concise in your visible text between tool calls: one short sentence about what you are doing and why is enough. Low token usage is a core goal \u2014 be terse everywhere.`;
+# Communication
+- Between tools: ONE short sentence on what and why. Low token usage is a core goal \u2014 be terse everywhere.
+- FINAL answer: hard limit of 3 sentences, plain prose, no bullets/headings unless asked. State the outcome and the command used to verify. No preamble ("Sure", "Here are"), no repo inventories, no repeated tool output.
+- If the user says "that", "it", or "earlier", they mean the prior conversation above.`;
 }
 var SKIP_DIRS;
 var init_prompts = __esm({
@@ -753,7 +729,7 @@ function appendTurn(cwd, task, response) {
 function formatConversationContext(cwd, limit = 10) {
   const turns = loadConversation(cwd, limit);
   if (turns.length === 0) return "";
-  const MAX_PRIOR_CHARS = 5e3;
+  const MAX_PRIOR_CHARS = 3e3;
   const header = "# Prior conversation in this project (most recent shown)";
   const blocks = [];
   let total = header.length;
@@ -936,6 +912,8 @@ ${systemExtra}` : "");
     }
   };
   const messages = [{ role: "user", content: task }];
+  const readMsgIndex = /* @__PURE__ */ new Map();
+  const normalizeRel = (p) => p.replace(/\\/g, "/");
   const setPhase = (p) => {
     state.phase = p;
     bus.emit("phase", p);
@@ -1137,6 +1115,26 @@ Investigate the root cause, fix it, then finish with a final answer.`
           new_string: result.data.new_string
         });
       }
+      const readHeader = result.success && tool.name === "read_file" ? result.output.split("\n")[0] ?? "" : "";
+      const readRel = readHeader ? normalizeRel(readHeader.replace(/\s+\(\d+ lines\)\s*$/, "").trim()) : "";
+      if (readRel) readMsgIndex.set(readRel, messages.length);
+      if ((tool.name === "apply_patch" || tool.name === "write_file") && result.success) {
+        const patchedRel = normalizeRel(
+          typeof result.data?.path === "string" && result.data.path ? result.data.path : typeof args.path === "string" ? args.path : ""
+        );
+        const idx = patchedRel ? readMsgIndex.get(patchedRel) : void 0;
+        if (idx !== void 0) {
+          const old = messages[idx];
+          if (old && old.role === "tool") {
+            messages[idx] = {
+              role: "tool",
+              tool_call_id: old.tool_call_id ?? "elided",
+              content: `\u27E8${patchedRel} was read earlier; its content is elided because it has since been modified. Re-read only if needed.\u27E9`
+            };
+            readMsgIndex.delete(patchedRel);
+          }
+        }
+      }
       messages.push({
         role: "tool",
         tool_call_id: call.id,
@@ -1147,9 +1145,9 @@ Investigate the root cause, fix it, then finish with a final answer.`
       finalText = "Task cancelled.";
       break;
     }
+    trimOldToolResults(messages);
     if (estimateContextChars(messages, system) > MAX_CONTEXT_CHARS) {
-      bus.emit("warning", "Context is large; trimming oldest tool outputs.");
-      trimOldToolResults(messages);
+      bus.emit("warning", "Context is still very large after trimming oldest tool outputs.");
     }
   }
   if (!sawFinish && !finalText) {
@@ -1179,11 +1177,11 @@ var init_loop = __esm({
     init_state();
     init_base();
     init_testing();
-    MAX_CONTEXT_CHARS = 7e4;
+    MAX_CONTEXT_CHARS = 4e4;
     RATE_LIMIT_MAX_ATTEMPTS = 8;
     MAX_VERIFY = 3;
     EDIT_TOOLS = /* @__PURE__ */ new Set(["apply_patch", "write_file", "move_path"]);
-    KEEP_RECENT_TOOL = 4;
+    KEEP_RECENT_TOOL = 3;
   }
 });
 
@@ -1191,7 +1189,7 @@ var init_loop = __esm({
 var version;
 var init_package = __esm({
   "package.json"() {
-    version = "0.1.16";
+    version = "0.1.17";
   }
 });
 
@@ -2379,19 +2377,13 @@ function relativeToWorkspace(cwd, target) {
 }
 var listFiles = {
   name: "list_files",
-  description: "List files and directories under a path inside the workspace. Respects common ignore rules (node_modules, .git, dist, ...). Use this to orient yourself in the repository.",
+  description: "List files/dirs under a path (ignores node_modules etc.). The system prompt already contains a file tree.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      path: {
-        type: "string",
-        description: "Relative path inside the workspace. Defaults to the workspace root."
-      },
-      max_entries: {
-        type: "number",
-        description: "Maximum number of entries to return. Defaults to 500."
-      }
+      path: { type: "string", description: "Relative path (default root)." },
+      max_entries: { type: "number", description: "Max entries (default 500)." }
     }
   },
   async execute(args, ctx) {
@@ -2450,13 +2442,13 @@ var listFiles = {
 };
 var movePath = {
   name: "move_path",
-  description: "Move or rename a file or directory within the workspace. Requires user approval. Parent directories of the destination are created automatically.",
+  description: "Move or rename a file/directory within the workspace. Requires user approval.",
   risk: "restricted",
   parameters: {
     type: "object",
     properties: {
-      from: { type: "string", description: "Source path relative to the workspace root." },
-      to: { type: "string", description: "Destination path relative to the workspace root." }
+      from: { type: "string", description: "Source path (relative to root)." },
+      to: { type: "string", description: "Destination path (relative to root)." }
     },
     required: ["from", "to"]
   },
@@ -2479,14 +2471,14 @@ var movePath = {
 };
 var readFile = {
   name: "read_file",
-  description: "Read a file from the workspace, optionally restricted to a line range. Output is prefixed with line numbers. Prefer reading ranges of large files instead of whole files.",
+  description: "Read a file (line-numbered), optionally a line range. Prefer ranges for large files.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      path: { type: "string", description: "File path relative to the workspace root." },
-      start_line: { type: "number", description: "First line to read (1-based). Defaults to 1." },
-      end_line: { type: "number", description: "Last line to read (inclusive). Defaults to start_line + 2000." }
+      path: { type: "string", description: "Path relative to workspace root." },
+      start_line: { type: "number", description: "First line (1-based)." },
+      end_line: { type: "number", description: "Last line inclusive (default start+400)." }
     },
     required: ["path"]
   },
@@ -2500,7 +2492,7 @@ var readFile = {
     }
     const lines = content.split(/\r?\n/);
     const start = Math.max(1, firstNumber(args, "start_line") ?? 1);
-    const defaultEnd = start + 2e3 - 1;
+    const defaultEnd = start + 400 - 1;
     const end = Math.min(lines.length, firstNumber(args, "end_line") ?? defaultEnd);
     if (start > lines.length) {
       return {
@@ -2509,7 +2501,7 @@ var readFile = {
       };
     }
     const slice = lines.slice(start - 1, end).map((line, i) => `${start + i}: ${line}`);
-    const { text, truncated } = truncateOutput(slice.join("\n"), 5e4);
+    const { text, truncated } = truncateOutput(slice.join("\n"), 12e3);
     const note = end < lines.length ? `
 ...[${lines.length - end} more lines. Use start_line=${end + 1} to continue reading.]` : "";
     return {
@@ -2522,13 +2514,13 @@ ${text}${note}`,
 };
 var writeFile = {
   name: "write_file",
-  description: "Create a new file or fully overwrite an existing file with the given content inside the workspace. For modifying existing files prefer apply_patch.",
+  description: "Create or fully overwrite a file. Prefer apply_patch for edits to existing files.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      path: { type: "string", description: "File path relative to the workspace root." },
-      content: { type: "string", description: "The full file content to write." }
+      path: { type: "string", description: "Path relative to workspace root." },
+      content: { type: "string", description: "Full file content." }
     },
     required: ["path", "content"]
   },
@@ -2545,14 +2537,14 @@ var writeFile = {
 };
 var applyPatch = {
   name: "apply_patch",
-  description: "Apply a minimal edit to an existing file by replacing an exact unique snippet (old_string) with new_string. The snippet must appear exactly once in the file. This is the preferred way to modify files: it makes the smallest possible change.",
+  description: "Replace an exact unique snippet in a file: old_string (must occur exactly once) becomes new_string. Preferred edit method.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      path: { type: "string", description: "File path relative to the workspace root." },
-      old_string: { type: "string", description: "The exact text to replace. Must occur exactly once in the file." },
-      new_string: { type: "string", description: "The replacement text." }
+      path: { type: "string", description: "Path relative to workspace root." },
+      old_string: { type: "string", description: "Exact text to replace, unique in file." },
+      new_string: { type: "string", description: "Replacement text." }
     },
     required: ["path", "old_string", "new_string"]
   },
@@ -2831,15 +2823,15 @@ function ripgrepAvailable() {
 }
 var searchCode = {
   name: "search_code",
-  description: "Search file contents across the workspace for a pattern (regex supported). Faster and more context-efficient than reading many files. Returns matching file paths, line numbers, and lines.",
+  description: "Regex-search file contents workspace-wide; returns path:line:match. Cheaper than reading files.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      pattern: { type: "string", description: "Regex or literal text to search for." },
-      path: { type: "string", description: "Subdirectory to search. Defaults to the workspace root." },
-      glob: { type: "string", description: 'Optional file glob filter, e.g. "*.ts" or "src/**".' },
-      max_results: { type: "number", description: "Maximum matches to return. Defaults to 50." }
+      pattern: { type: "string", description: "Regex or literal text." },
+      path: { type: "string", description: "Subdirectory (default root)." },
+      glob: { type: "string", description: 'Glob filter, e.g. "*.ts".' },
+      max_results: { type: "number", description: "Max matches (default 50)." }
     },
     required: ["pattern"]
   },
@@ -2919,14 +2911,14 @@ ${text}`,
 };
 var searchFiles = {
   name: "search_files",
-  description: "Find files by name pattern (supports * and ? wildcards) inside the workspace. Use when you know roughly what a file is called.",
+  description: "Find files by name with * and ? wildcards.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      pattern: { type: "string", description: 'File name pattern with wildcards, e.g. "*.test.ts" or "config*". Matches against the full relative path too.' },
-      path: { type: "string", description: "Subdirectory to search. Defaults to the workspace root." },
-      max_results: { type: "number", description: "Maximum results. Defaults to 100." }
+      pattern: { type: "string", description: 'Name pattern with wildcards, e.g. "*.test.ts".' },
+      path: { type: "string", description: "Subdirectory (default root)." },
+      max_results: { type: "number", description: "Max results (default 100)." }
     },
     required: ["pattern"]
   },
@@ -3140,7 +3132,7 @@ var DEFAULT_TIMEOUT_MS = 12e4;
 var MAX_OUTPUT_CHARS = 1e5;
 var executeCommand = {
   name: "execute_command",
-  description: "Execute a shell command in the workspace. Safe read-only commands (ls, git status, npm test, ...) run automatically. Install/git-write commands require user approval. Dangerous commands (rm -rf, sudo, ...) are blocked.",
+  description: "Run a shell command in the workspace. Read-only commands auto-run; installs/git-writes need approval; dangerous ones are blocked.",
   risk: "restricted",
   timeoutMs: DEFAULT_TIMEOUT_MS,
   dynamicRisk: (args) => {
@@ -3150,8 +3142,8 @@ var executeCommand = {
   parameters: {
     type: "object",
     properties: {
-      command: { type: "string", description: "The shell command to execute." },
-      timeout_ms: { type: "number", description: `Timeout in milliseconds. Defaults to ${DEFAULT_TIMEOUT_MS}, max 600000.` }
+      command: { type: "string", description: "The command." },
+      timeout_ms: { type: "number", description: `Timeout ms (default ${DEFAULT_TIMEOUT_MS}).` }
     },
     required: ["command"]
   },
@@ -3200,19 +3192,16 @@ function save(cwd, entries) {
 }
 var memoryTool = {
   name: "memory",
-  description: 'Persist and recall key information across iterations AND across sessions (project-scoped, stored under ~/.mentee/memory, never in the workspace). Use action=add to store important findings, decisions, conventions, or context you discovered (e.g. "project uses TypeScript 5 strict mode", "auth lives in src/auth.ts"). Use action=search (query) to recall before re-deriving something you may have learned earlier. Use list to see everything, forget to remove by id or topic.',
+  description: "Persist notes for this project across sessions. Actions: add (content, topic?), search (query), list, forget (id or topic).",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      action: { type: "string", enum: ["add", "search", "list", "forget"], description: "Operation to perform." },
-      content: { type: "string", description: "The note text (for action=add)." },
-      topic: {
-        type: "string",
-        description: "Optional short tag/category for the note (action=add) or, for action=forget, the topic to clear."
-      },
-      query: { type: "string", description: "Search terms (for action=search). Falls back to content if omitted." },
-      id: { type: "string", description: "Entry id to remove (for action=forget)." }
+      action: { type: "string", enum: ["add", "search", "list", "forget"], description: "Operation." },
+      content: { type: "string", description: "Note text (add)." },
+      topic: { type: "string", description: "Tag (add) or topic to clear (forget)." },
+      query: { type: "string", description: "Search terms (search)." },
+      id: { type: "string", description: "Entry id (forget)." }
     },
     required: ["action"]
   },
@@ -3324,13 +3313,13 @@ function deleteCommands(cwd, target) {
 }
 var webSearch = {
   name: "web_search",
-  description: "Search the web for documentation, tutorials, or solutions to problems. Returns top results with snippets.",
+  description: "Web search for docs/solutions. Returns top results with snippets.",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
       query: { type: "string", description: "Search query." },
-      maxResults: { type: "number", description: "Maximum results to return. Defaults to 10." }
+      maxResults: { type: "number", description: "Max results (default 10)." }
     },
     required: ["query"]
   },
@@ -3346,7 +3335,7 @@ var webSearch = {
 };
 var httpFetch = {
   name: "http_fetch",
-  description: "Fetch the content of a URL. Useful for retrieving API responses, documentation pages, or raw files. Returns the raw body text (truncated if large).",
+  description: "Fetch a URL and return the body text (truncated if large).",
   risk: "safe",
   parameters: {
     type: "object",
@@ -3372,7 +3361,7 @@ var httpFetch = {
 };
 var fileInfo = {
   name: "file_info",
-  description: "Get detailed information about a file or directory inside the workspace. Includes size, type, and a profile suitable for safe manual deletion.",
+  description: "File/dir info: size, type, safe-deletion profile.",
   risk: "safe",
   parameters: {
     type: "object",
@@ -3404,7 +3393,7 @@ ${steps.join("\n")}`,
 };
 var safeDeleteSuggestion = {
   name: "safe_delete_suggestion",
-  description: "Get PowerShell and CMD commands for manual deletion of a file/directory, WITH double-checking steps and verification hints. Does NOT execute deletion; returns commands for the user to run.",
+  description: "Return (not execute) PowerShell/CMD deletion commands with double-check steps for a path.",
   risk: "safe",
   parameters: {
     type: "object",
@@ -3438,12 +3427,12 @@ ${steps.join("\n")}
 };
 var processList = {
   name: "process_list",
-  description: "List currently running processes on the system. Useful for finding processes that may be locking files or consuming resources.",
+  description: "List running processes (find what locks files or uses resources).",
   risk: "safe",
   parameters: {
     type: "object",
     properties: {
-      filter: { type: "string", description: "Optional filter substring to match process names." }
+      filter: { type: "string", description: "Filter substring for process names." }
     }
   },
   async execute(args, ctx) {
@@ -3464,7 +3453,7 @@ ${trimmed}` };
 };
 var portCheck = {
   name: "port_check",
-  description: "Check if a port is in use on the local machine. useful for debugging server startup issues or verifying ports are free.",
+  description: "Check whether a local port is in use.",
   risk: "safe",
   parameters: {
     type: "object",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@menteeai/menteeswe",
-  "version": "0.1.16",
+  "version": "0.1.17",
   "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",