@morphllm/morphsdk 0.2.57 → 0.2.59

package/dist/tools/warp_grep/harness.cjs

@@ -30,23 +30,127 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 // tools/warp_grep/harness.ts
 var harness_exports = {};
 __export(harness_exports, {
+  AGENT_CONFIG: () => AGENT_CONFIG,
   DEFAULT_EXCLUDES: () => DEFAULT_EXCLUDES,
+  LLMResponseParser: () => LLMResponseParser,
   LocalRipgrepProvider: () => LocalRipgrepProvider,
-  MAX_TURNS: () => MAX_TURNS,
   SYSTEM_PROMPT: () => SYSTEM_PROMPT,
-  TIMEOUT_MS: () => TIMEOUT_MS,
-  formatAnalyseTree: () => formatAnalyseTree,
-  formatToolResult: () => formatToolResult,
+  buildInitialState: () => buildInitialState,
+  calculateContextBudget: () => calculateContextBudget,
+  formatListDirectoryTree: () => formatListDirectoryTree,
+  formatToolResult: () => formatAgentToolOutput,
   formatTurnMessage: () => formatTurnMessage,
+  getSystemPrompt: () => getSystemPrompt,
+  normalizeFinishFiles: () => normalizeFinishFiles,
   parseToolCalls: () => parseToolCalls,
-  resolveFinishFiles: () => resolveFinishFiles
+  readFinishFiles: () => readFinishFiles,
+  resolveFinishFiles: () => resolveFinishFiles,
+  toolGrep: () => toolGrep,
+  toolListDirectory: () => toolListDirectory,
+  toolRead: () => toolRead
 });
 module.exports = __toCommonJS(harness_exports);
 
 // tools/warp_grep/agent/parser.ts
-var VALID_COMMANDS = ["analyse", "grep", "read", "finish"];
+var VALID_COMMANDS = ["list_directory", "grep", "read", "finish"];
+function isValidCommand(name) {
+  return VALID_COMMANDS.includes(name);
+}
+function getXmlElementText(xml, tagName) {
+  const regex = new RegExp(`<${tagName}>([\\s\\S]*?)</${tagName}>`, "i");
+  const match = xml.match(regex);
+  return match ? match[1].trim() : null;
+}
+function parseNestedXmlTools(text) {
+  const tools = [];
+  const toolRegex = /<([a-z_][a-z0-9_]*)>([\s\S]*?)<\/\1>/gi;
+  let match;
+  while ((match = toolRegex.exec(text)) !== null) {
+    const rawToolName = match[1].toLowerCase();
+    const content = match[2];
+    if (!isValidCommand(rawToolName)) continue;
+    const toolName = rawToolName;
+    if (toolName === "list_directory") {
+      const path4 = getXmlElementText(content, "path");
+      const pattern = getXmlElementText(content, "pattern");
+      if (path4) {
+        tools.push({ name: "list_directory", arguments: { path: path4, pattern } });
+      }
+    } else if (toolName === "grep") {
+      const pattern = getXmlElementText(content, "pattern");
+      const subDir = getXmlElementText(content, "sub_dir");
+      const glob = getXmlElementText(content, "glob");
+      if (pattern) {
+        tools.push({
+          name: "grep",
+          arguments: {
+            pattern,
+            path: subDir || ".",
+            ...glob && { glob }
+          }
+        });
+      }
+    } else if (toolName === "read") {
+      const path4 = getXmlElementText(content, "path");
+      const linesStr = getXmlElementText(content, "lines");
+      if (path4) {
+        const args = { path: path4 };
+        if (linesStr) {
+          const ranges = [];
+          for (const rangeStr of linesStr.split(",")) {
+            const trimmed = rangeStr.trim();
+            if (!trimmed) continue;
+            const [s, e] = trimmed.split("-").map((v) => parseInt(v.trim(), 10));
+            if (Number.isFinite(s) && Number.isFinite(e)) {
+              ranges.push([s, e]);
+            } else if (Number.isFinite(s)) {
+              ranges.push([s, s]);
+            }
+          }
+          if (ranges.length === 1) {
+            args.start = ranges[0][0];
+            args.end = ranges[0][1];
+          } else if (ranges.length > 1) {
+            args.lines = ranges;
+          }
+        }
+        tools.push({ name: "read", arguments: args });
+      }
+    } else if (toolName === "finish") {
+      const fileRegex = /<file>([\s\S]*?)<\/file>/gi;
+      const files = [];
+      let fileMatch;
+      while ((fileMatch = fileRegex.exec(content)) !== null) {
+        const fileContent = fileMatch[1];
+        const filePath = getXmlElementText(fileContent, "path");
+        const linesStr = getXmlElementText(fileContent, "lines");
+        if (filePath && linesStr) {
+          const ranges = [];
+          for (const rangeStr of linesStr.split(",")) {
+            if (rangeStr.trim() === "*") {
+              ranges.push([1, 999999]);
+            } else {
+              const [s, e] = rangeStr.split("-").map((v) => parseInt(v.trim(), 10));
+              if (Number.isFinite(s) && Number.isFinite(e)) {
+                ranges.push([s, e]);
+              }
+            }
+          }
+          if (ranges.length > 0) {
+            files.push({ path: filePath, lines: ranges });
+          }
+        }
+      }
+      if (files.length > 0) {
+        tools.push({ name: "finish", arguments: { files } });
+      }
+    }
+  }
+  return tools;
+}
 function preprocessText(text) {
   let processed = text.replace(/<think>[\s\S]*?<\/think>/gi, "");
+  const nestedTools = parseNestedXmlTools(processed);
   const openingTagRegex = /<tool_call>|<tool>/gi;
   const closingTagRegex = /<\/tool_call>|<\/tool>/gi;
   const openingMatches = processed.match(openingTagRegex) || [];
@@ -83,7 +187,7 @@ function preprocessText(text) {
       }
     }
   }
-  return toolCallLines;
+  return { lines: toolCallLines, nestedTools };
 }
 var LLMResponseParser = class {
   finishSpecSplitRe = /,(?=[^,\s]+:)/;
@@ -91,8 +195,8 @@ var LLMResponseParser = class {
     if (typeof text !== "string") {
       throw new TypeError("Command text must be a string.");
     }
-    const lines = preprocessText(text);
-    const commands = [];
+    const { lines, nestedTools } = preprocessText(text);
+    const commands = [...nestedTools];
     let finishAccumulator = null;
     lines.forEach((line) => {
       if (!line || line.startsWith("#")) return;
@@ -100,8 +204,8 @@ var LLMResponseParser = class {
       if (parts.length === 0) return;
       const cmd = parts[0];
       switch (cmd) {
-        case "analyse":
-          this.handleAnalyse(parts, line, commands);
+        case "list_directory":
+          this.handleListDirectory(parts, line, commands);
           break;
         case "grep":
           this.handleGrep(parts, line, commands);
@@ -119,8 +223,8 @@ var LLMResponseParser = class {
     if (finishAccumulator) {
       const map = finishAccumulator;
       const entries = [...map.entries()];
-      const filesPayload = entries.map(([path3, ranges]) => ({
-        path: path3,
+      const filesPayload = entries.map(([path4, ranges]) => ({
+        path: path4,
         lines: [...ranges].sort((a, b) => a[0] - b[0])
       }));
       commands.push({ name: "finish", arguments: { files: filesPayload } });
@@ -152,18 +256,17 @@ var LLMResponseParser = class {
   skip(message) {
     return { name: "_skip", arguments: { message } };
   }
-  handleAnalyse(parts, rawLine, commands) {
+  handleListDirectory(parts, rawLine, commands) {
     if (parts.length < 2) {
       commands.push(this.skip(
-        `[SKIPPED] Your command "${rawLine}" is missing a path. Correct format: analyse <path> [pattern]. Example: analyse src/`
+        `[SKIPPED] Your command "${rawLine}" is missing a path. Correct format: list_directory <path> [pattern]. Example: list_directory src/`
       ));
       return;
     }
-    const path3 = parts[1];
+    const path4 = parts[1];
     const pattern = parts[2]?.replace(/^"|"$/g, "") ?? null;
-    commands.push({ name: "analyse", arguments: { path: path3, pattern } });
+    commands.push({ name: "list_directory", arguments: { path: path4, pattern } });
   }
-  // no glob tool in MCP
   handleGrep(parts, rawLine, commands) {
     if (parts.length < 3) {
       commands.push(this.skip(
@@ -234,187 +337,189 @@ var LLMResponseParser = class {
   }
 };
 
-// tools/warp_grep/agent/formatter.ts
-var ToolOutputFormatter = class {
-  format(toolName, args, output, options = {}) {
-    const name = (toolName ?? "").trim();
-    if (!name) {
-      return "";
-    }
-    const payload = output?.toString?.()?.trim?.() ?? "";
-    const isError = Boolean(options.isError);
-    const safeArgs = args ?? {};
-    if (!payload && !isError) {
-      return "";
-    }
-    switch (name) {
-      case "read":
-        return this.formatRead(safeArgs, payload, isError);
-      case "analyse":
-        return this.formatAnalyse(safeArgs, payload, isError);
-      case "grep":
-        return this.formatGrep(safeArgs, payload, isError);
-      default:
-        return payload ? `<tool_output>
-${payload}
-</tool_output>` : "";
-    }
-  }
-  formatRead(args, payload, isError) {
-    if (isError) {
-      return payload;
-    }
-    const path3 = this.asString(args.path) || "...";
-    return `<file path="${path3}">
-${payload}
-</file>`;
-  }
-  formatAnalyse(args, payload, isError) {
-    const path3 = this.asString(args.path) || ".";
-    if (isError) {
-      return `<analyse_results path="${path3}" status="error">
-${payload}
-</analyse_results>`;
-    }
-    return `<analyse_results path="${path3}">
-${payload}
-</analyse_results>`;
-  }
-  formatGrep(args, payload, isError) {
-    const pattern = this.asString(args.pattern);
-    const path3 = this.asString(args.path);
-    const attributes = [];
-    if (pattern !== void 0) {
-      attributes.push(`pattern="${pattern}"`);
-    }
-    if (path3 !== void 0) {
-      attributes.push(`path="${path3}"`);
-    }
-    if (isError) {
-      attributes.push('status="error"');
-    }
-    const attrText = attributes.length ? ` ${attributes.join(" ")}` : "";
-    return `<grep_output${attrText}>
-${payload}
-</grep_output>`;
-  }
-  asString(value) {
-    if (value === null || value === void 0) {
-      return void 0;
-    }
-    return String(value);
-  }
-};
-var sharedFormatter = new ToolOutputFormatter();
-function formatAgentToolOutput(toolName, args, output, options = {}) {
-  return sharedFormatter.format(toolName, args, output, options);
-}
-
 // tools/warp_grep/agent/prompt.ts
-var SYSTEM_PROMPT = `You are a code search agent. Your task is to find all relevant code for a given query.
+var SYSTEM_PROMPT = `You are a code search agent. Your task is to find all relevant code for a given search_string.
 
-<workflow>
+### workflow
 You have exactly 4 turns. The 4th turn MUST be a \`finish\` call. Each turn allows up to 8 parallel tool calls.
 
-- Turn 1: Map the territory OR dive deep (based on query specificity)
+- Turn 1: Map the territory OR dive deep (based on search_string specificity)
 - Turn 2-3: Refine based on findings
 - Turn 4: MUST call \`finish\` with all relevant code locations
 - You MAY call \`finish\` early if confident\u2014but never before at least 1 search turn.
+- The user strongly prefers if you can call the finish tool early, but you must be correct
 
-Remember, if the task feels easy to you, it is strongly desirable to call \`finish\` early using fewer turns, but quality over speed.
-</workflow>
+Remember, if the task feels easy to you, it is strongly desirable to call 'finish' early using fewer turns, but quality over speed
 
-<tools>
-### \`analyse <path> [pattern]\`
-Directory tree or file search. Shows structure of a path, optionally filtered by regex pattern.
-- \`path\`: Required. Directory or file path (use \`.\` for repo root)
-- \`pattern\`: Optional regex to filter results
+### tools
+Tool calls use nested XML elements:
+\`\`\`xml
+<tool_name>
+  <parameter>value</parameter>
+</tool_name>
+\`\`\`
+
+### \`list_directory\`
+Directory tree view. Shows structure of a path, optionally filtered by regex pattern.
+
+Elements:
+- \`<path>\` (required): Directory path to list (use \`.\` for repo root)
+- \`<pattern>\` (optional): Regex to filter results
 
 Examples:
 \`\`\`
-analyse .
-analyse src/api
-analyse . ".*\\.ts$"
-analyse src "test.*"
+<list_directory>
+  <path>src/services</path>
+</list_directory>
+
+<list_directory>
+  <path>lib/utils</path>
+  <pattern>.*\\.(ts|js)$</pattern>
+</list_directory>
 \`\`\`
 
-### \`read <path>[:start-end]\`
-Read file contents. Line range is 1-based, inclusive.
+### \`read\`
+Read file contents. Supports multiple line ranges.
 - Returns numbered lines for easy reference
-- Omit range to read entire file
+- ALWAYS include import statements (usually lines 1-20). Better to over-include than miss context.
+
+Elements:
+- \`<path>\` (required): File path to read
+- \`<lines>\` (optional): Line ranges like "1-50,75-80,100-120" (omit to read entire file)
 
 Examples:
 \`\`\`
-read src/main.py
-read src/db/conn.py:10-50
-read package.json:1-20
+<read>
+  <path>src/main.py</path>
+</read>
+
+<read>
+  <path>src/auth.py</path>
+  <lines>1-20,45-80,150-200</lines>
+</read>
 \`\`\`
 
-### \`grep '<pattern>' <path>\`
-Ripgrep search. Finds pattern matches across files.
-- \`'<pattern>'\`: Required. Regex pattern wrapped in single quotes
-- \`<path>\`: Required. Directory or file to search (use \`.\` for repo root)
+### \`grep\`
+Search for pattern matches across files. Returns matches with 1 line of context above and below.
+- Match lines use \`:\` separator \u2192 \`filepath:linenum:content\`
+- Context lines use \`-\` separator \u2192 \`filepath-linenum-content\`
+
+Elements:
+- \`<pattern>\` (required): Search pattern (regex). Use \`(a|b)\` for OR patterns.
+- \`<sub_dir>\` (optional): Subdirectory to search in (defaults to \`.\`)
+- \`<glob>\` (optional): File pattern filter like \`*.py\` or \`*.{ts,tsx}\`
 
 Examples:
 \`\`\`
-grep 'class.*Service' src/
-grep 'def authenticate' .
-grep 'import.*from' src/components/
-grep 'TODO' .
+<grep>
+  <pattern>(authenticate|authorize|login)</pattern>
+  <sub_dir>src/auth/</sub_dir>
+</grep>
+
+<grep>
+  <pattern>class.*(Service|Controller)</pattern>
+  <glob>*.{ts,js}</glob>
+</grep>
+
+<grep>
+  <pattern>(DB_HOST|DATABASE_URL|connection)</pattern>
+  <glob>*.{py,yaml,env}</glob>
+  <sub_dir>lib/</sub_dir>
+</grep>
 \`\`\`
 
-### \`finish <file1:ranges> [file2:ranges ...]\`
-Submit final answer with all relevant code locations.
-- Include generous line ranges\u2014don't be stingy with context
-- Ranges are comma-separated: \`file.py:10-30,50-60\`
-- ALWAYS include import statements at the top of files (usually lines 1-20)
-- If code spans multiple files, include ALL of them
-- Small files can be returned in full
+### \`finish\`
+Submit final answer with all relevant code locations. Uses nested \`<file>\` elements.
+
+File elements:
+- \`<path>\` (required): File path
+- \`<lines>\` (optional): Line ranges like "1-50,75-80" (\`*\` for entire file)
+
+ALWAYS include import statements (usually lines 1-20). Better to over-include than miss context.
 
 Examples:
 \`\`\`
-finish src/auth.py:1-15,25-50,75-80 src/models/user.py:1-10,20-45
-finish src/index.ts:1-100
+<finish>
+  <file>
+    <path>src/auth.py</path>
+    <lines>1-15,25-50,75-80</lines>
+  </file>
+  <file>
+    <path>src/models/user.py</path>
+    <lines>*</lines>
+  </file>
+</finish>
 \`\`\`
 </tools>
 
 <strategy>
-**Before your first tool call, classify the query:**
+**Before your first tool call, classify the search_string:**
 
-| Query Type | Turn 1 Strategy | Early Finish? |
-|------------|-----------------|---------------|
-| **Specific** (function name, error string, unique identifier) | 8 parallel greps on likely paths | Often by turn 2 |
-| **Conceptual** (how does X work, where is Y handled) | analyse + 2-3 broad greps | Rarely early |
-| **Exploratory** (find all tests, list API endpoints) | analyse at multiple depths | Usually needs 3 turns |
+| Search_string Type | Round 1 Strategy | Early Finish? |
+|------------|------------------|---------------|
+| **Specific** (function name, error string, unique identifier) | 8 parallel greps on likely paths | Often by round 2 |
+| **Conceptual** (how does X work, where is Y handled) | list_directory + 2-3 broad greps | Rarely early |
+| **Exploratory** (find all tests, list API endpoints) | list_directory at multiple depths | Usually needs 3 rounds |
 
 **Parallel call patterns:**
 - **Shotgun grep**: Same pattern, 8 different directories\u2014fast coverage
 - **Variant grep**: 8 pattern variations (synonyms, naming conventions)\u2014catches inconsistent codebases
-- **Funnel**: 1 analyse + 7 greps\u2014orient and search simultaneously
+- **Funnel**: 1 list_directory + 7 greps\u2014orient and search simultaneously
 - **Deep read**: 8 reads on files you already identified\u2014gather full context fast
+
+**Tool call expectations:**
+- Low quality tool calls are ones that give back sparse information. This either means they are not well thought out and are not educated guesses OR, they are too broad and give back too many results.
+- High quality tool calls strike a balance between complexity in the tool call to exclude results we know we don't want, and how wide the search space is so that we don't miss anything. It is ok to start off with wider search spaces, but is imperative that you use your intuition from there on out and seek high quality tool calls only.
+- You are not starting blind, you have some information about root level repo structure going in, so use that to prevent making trivial repo wide queries.
+- The grep tool shows you which file path and line numbers the pattern was found in, use this information smartly when trying to read the file.
 </strategy>
 
 <output_format>
 EVERY response MUST follow this exact format:
 
 1. First, wrap your reasoning in \`<think>...</think>\` tags containing:
-   - Query classification (specific/conceptual/exploratory)
-   - Confidence estimate (can I finish in 1-2 turns?)
-   - This turn's parallel strategy
+   - Search_string classification (specific/conceptual/exploratory)
+   - Confidence estimate (can I finish in 1-2 rounds?)
+   - This round's parallel strategy
    - What signals would let me finish early?
 
-2. Then, output tool calls wrapped in \`<tool_call>...</tool_call>\` tags, one per line.
+2. Then, output up to 8 tool calls using nested XML elements.
 
 Example:
 \`\`\`
 <think>
-This is a specific query about authentication. I'll grep for auth-related patterns.
-High confidence I can finish in 2 turns if I find the auth module.
+This is a specific search_string about authentication. I'll grep for auth-related patterns.
+High confidence I can finish in 2 rounds if I find the auth module. I have already been shown the repo's structure at root
 Strategy: Shotgun grep across likely directories.
 </think>
-<tool_call>grep 'authenticate' src/</tool_call>
-<tool_call>grep 'login' src/</tool_call>
-<tool_call>analyse src/auth</tool_call>
+<grep>
+  <pattern>(authenticate|login|session)</pattern>
+  <sub_dir>src/auth/</sub_dir>
+</grep>
+<grep>
+  <pattern>(middleware|interceptor)</pattern>
+  <glob>*.{ts,js}</glob>
+</grep>
+<list_directory>
+  <path>src/auth</path>
+</list_directory>
+\`\`\`
+
+Finishing example:
+\`\`\`
+<think>
+I think I have a rough idea, but this is my last turn so I must call the finish tool regardless.
+</think>
+<finish>
+  <file>
+    <path>src/auth/login.py</path>
+    <lines>1-50</lines>
+  </file>
+  <file>
+    <path>src/middleware/session.py</path>
+    <lines>10-80</lines>
+  </file>
+</finish>
 \`\`\`
 
 No commentary outside \`<think>\`. No explanations after tool calls.
@@ -427,15 +532,116 @@ When calling \`finish\`:
 - Include any type definitions, interfaces, or constants used
 - Better to over-include than leave the user missing context
 - If unsure about boundaries, include more rather than less
-</finishing_requirements>
+</finishing_requirements>`;
+function getSystemPrompt() {
+  return SYSTEM_PROMPT;
+}
+
+// tools/warp_grep/agent/formatter.ts
+var ToolOutputFormatter = class {
+  format(toolName, args, output, options = {}) {
+    const name = (toolName ?? "").trim();
+    if (!name) {
+      return "";
+    }
+    const payload = output?.toString?.()?.trim?.() ?? "";
+    const isError = Boolean(options.isError);
+    const safeArgs = args ?? {};
+    if (!payload && !isError) {
+      return "";
+    }
+    switch (name) {
+      case "read":
+        return this.formatRead(safeArgs, payload, isError);
+      case "list_directory":
+        return this.formatListDirectory(safeArgs, payload, isError);
+      case "grep":
+        return this.formatGrep(safeArgs, payload, isError);
+      default:
+        return payload ? `<tool_output>
+${payload}
+</tool_output>` : "";
+    }
+  }
+  formatRead(args, payload, isError) {
+    if (isError) {
+      return payload;
+    }
+    const path4 = this.asString(args.path) || "...";
+    const start = args.start;
+    const end = args.end;
+    const linesArray = args.lines;
+    const attributes = [`path="${path4}"`];
+    if (linesArray && linesArray.length > 0) {
+      const rangeStr = linesArray.map(([s, e]) => `${s}-${e}`).join(",");
+      attributes.push(`lines="${rangeStr}"`);
+    } else if (start !== void 0 && end !== void 0) {
+      attributes.push(`lines="${start}-${end}"`);
+    }
+    return `<read ${attributes.join(" ")}>
+${payload}
+</read>`;
+  }
+  formatListDirectory(args, payload, isError) {
+    const path4 = this.asString(args.path) || ".";
+    const pattern = this.asString(args.pattern);
+    const attributes = [`path="${path4}"`];
+    if (pattern) {
+      attributes.push(`pattern="${pattern}"`);
+    }
+    if (isError) {
+      attributes.push('status="error"');
+    }
+    return `<list_directory ${attributes.join(" ")}>
+${payload}
+</list_directory>`;
+  }
+  formatGrep(args, payload, isError) {
+    const pattern = this.asString(args.pattern);
+    const subDir = this.asString(args.path);
+    const glob = this.asString(args.glob);
+    const attributes = [];
+    if (pattern !== void 0) {
+      attributes.push(`pattern="${pattern}"`);
+    }
+    if (subDir !== void 0) {
+      attributes.push(`sub_dir="${subDir}"`);
+    }
+    if (glob !== void 0) {
+      attributes.push(`glob="${glob}"`);
+    }
+    if (isError) {
+      attributes.push('status="error"');
+    }
+    const attrText = attributes.length ? ` ${attributes.join(" ")}` : "";
+    return `<grep${attrText}>
+${payload}
+</grep>`;
+  }
+  asString(value) {
+    if (value === null || value === void 0) {
+      return void 0;
+    }
+    return String(value);
+  }
+};
+var sharedFormatter = new ToolOutputFormatter();
+function formatAgentToolOutput(toolName, args, output, options = {}) {
+  return sharedFormatter.format(toolName, args, output, options);
+}
 
-Begin your exploration now to find code relevant to the query.`;
+// tools/warp_grep/agent/helpers.ts
+var import_path = __toESM(require("path"), 1);
 
 // tools/warp_grep/agent/config.ts
 var AGENT_CONFIG = {
-  // Give the model freedom; failsafe cap to prevent infinite loops
-  MAX_ROUNDS: 10,
-  TIMEOUT_MS: 3e4
+  MAX_TURNS: 4,
+  TIMEOUT_MS: 3e4,
+  MAX_CONTEXT_CHARS: 54e4,
+  MAX_OUTPUT_LINES: 200,
+  MAX_READ_LINES: 800,
+  MAX_LIST_DEPTH: 3,
+  LIST_TIMEOUT_MS: 2e3
 };
 var BUILTIN_EXCLUDES = [
   // Version control
@@ -518,7 +724,124 @@ var BUILTIN_EXCLUDES = [
 ];
 var DEFAULT_EXCLUDES = (process.env.MORPH_WARP_GREP_EXCLUDE || "").split(",").map((s) => s.trim()).filter(Boolean).concat(BUILTIN_EXCLUDES);
 
-// tools/warp_grep/tools/finish.ts
+// tools/warp_grep/agent/helpers.ts
+function formatTurnMessage(turnsUsed, maxTurns) {
+  const turnsRemaining = maxTurns - turnsUsed;
+  if (turnsRemaining === 1) {
+    return `
+You have used ${turnsUsed} turns, you only have 1 turn remaining. You have run out of turns to explore the code base and MUST call the finish tool now`;
+  }
+  return `
+You have used ${turnsUsed} turn${turnsUsed === 1 ? "" : "s"} and have ${turnsRemaining} remaining`;
+}
+function calculateContextBudget(messages) {
+  const totalChars = messages.reduce((sum, m) => sum + m.content.length, 0);
+  const maxChars = AGENT_CONFIG.MAX_CONTEXT_CHARS;
+  const percent = Math.round(totalChars / maxChars * 100);
+  const usedK = Math.round(totalChars / 1e3);
+  const maxK = Math.round(maxChars / 1e3);
+  return `<context_budget>${percent}% (${usedK}K/${maxK}K chars)</context_budget>`;
+}
+async function buildInitialState(repoRoot, query, provider) {
+  try {
+    const entries = await provider.listDirectory({
+      path: ".",
+      maxResults: AGENT_CONFIG.MAX_OUTPUT_LINES,
+      maxDepth: 2
+    });
+    const treeLines = entries.map((e) => {
+      const indent = "  ".repeat(e.depth);
+      const name = e.type === "dir" ? `${e.name}/` : e.name;
+      return `${indent}${name}`;
+    });
+    const repoName = import_path.default.basename(repoRoot);
+    const treeOutput = treeLines.length > 0 ? `${repoName}/
+${treeLines.join("\n")}` : `${repoName}/`;
+    return `<repo_structure>
+${treeOutput}
+</repo_structure>
+
+<search_string>
+${query}
+</search_string>`;
+  } catch {
+    const repoName = import_path.default.basename(repoRoot);
+    return `<repo_structure>
+${repoName}/
+</repo_structure>
+
+<search_string>
+${query}
+</search_string>`;
+  }
+}
+function formatListDirectoryTree(entries) {
+  if (!entries.length) return "empty";
+  return entries.map((e) => {
+    const indent = "  ".repeat(e.depth);
+    const name = e.type === "dir" ? `${e.name}/` : e.name;
+    return `${indent}${name}`;
+  }).join("\n");
+}
+
+// tools/warp_grep/agent/tools/grep.ts
+async function toolGrep(provider, args) {
+  const res = await provider.grep(args);
+  if (res.error) {
+    return { output: res.error };
+  }
+  if (!res.lines.length) {
+    return { output: "no matches" };
+  }
+  return { output: res.lines.join("\n") };
+}
+
+// tools/warp_grep/agent/tools/read.ts
+async function toolRead(provider, args) {
+  if (args.lines && args.lines.length > 0) {
+    const chunks = [];
+    for (const [start, end] of args.lines) {
+      const res2 = await provider.read({ path: args.path, start, end });
+      if (res2.error) return res2.error;
+      chunks.push(res2.lines.join("\n"));
+    }
+    if (chunks.every((c) => c === "")) return "(empty file)";
+    return chunks.join("\n...\n");
+  }
+  const res = await provider.read({ path: args.path, start: args.start, end: args.end });
+  if (res.error) {
+    return res.error;
+  }
+  if (!res.lines.length) return "(empty file)";
+  return res.lines.join("\n");
+}
+
+// tools/warp_grep/agent/tools/list_directory.ts
+async function toolListDirectory(provider, args) {
+  const list = await provider.listDirectory({
+    path: args.path,
+    pattern: args.pattern ?? null,
+    maxResults: args.maxResults ?? AGENT_CONFIG.MAX_OUTPUT_LINES,
+    maxDepth: args.maxDepth ?? AGENT_CONFIG.MAX_LIST_DEPTH
+  });
+  if (!list.length) return "empty";
+  if (list.length >= AGENT_CONFIG.MAX_OUTPUT_LINES) {
+    return "query not specific enough, tool called tried to return too much context and failed";
+  }
+  return list.map((e) => {
+    const indent = "  ".repeat(e.depth);
+    const name = e.type === "dir" ? `${e.name}/` : e.name;
+    return `${indent}${name}`;
+  }).join("\n");
+}
+
+// tools/warp_grep/agent/tools/finish.ts
+function normalizeFinishFiles(files) {
+  return files.map((f) => {
+    const merged = mergeRanges(f.lines);
+    return { path: f.path, lines: merged };
+  });
+}
 async function readFinishFiles(repoRoot, files, reader) {
   const out = [];
   for (const f of files) {
@@ -553,7 +876,7 @@ function mergeRanges(ranges) {
 
 // tools/warp_grep/providers/local.ts
 var import_promises2 = __toESM(require("fs/promises"), 1);
-var import_path2 = __toESM(require("path"), 1);
+var import_path3 = __toESM(require("path"), 1);
 
 // tools/warp_grep/utils/ripgrep.ts
 var import_child_process = require("child_process");
@@ -618,21 +941,21 @@ async function runRipgrep(args, opts) {
 
 // tools/warp_grep/utils/paths.ts
 var import_fs = __toESM(require("fs"), 1);
-var import_path = __toESM(require("path"), 1);
+var import_path2 = __toESM(require("path"), 1);
 function resolveUnderRepo(repoRoot, targetPath) {
-  const absRoot = import_path.default.resolve(repoRoot);
-  const resolved = import_path.default.resolve(absRoot, targetPath);
+  const absRoot = import_path2.default.resolve(repoRoot);
+  const resolved = import_path2.default.resolve(absRoot, targetPath);
   ensureWithinRepo(absRoot, resolved);
   return resolved;
 }
 function ensureWithinRepo(repoRoot, absTarget) {
-  const rel = import_path.default.relative(import_path.default.resolve(repoRoot), import_path.default.resolve(absTarget));
-  if (rel.startsWith("..") || import_path.default.isAbsolute(rel)) {
+  const rel = import_path2.default.relative(import_path2.default.resolve(repoRoot), import_path2.default.resolve(absTarget));
+  if (rel.startsWith("..") || import_path2.default.isAbsolute(rel)) {
     throw new Error(`Path outside repository root: ${absTarget}`);
   }
 }
 function toRepoRelative(repoRoot, absPath) {
-  return import_path.default.relative(import_path.default.resolve(repoRoot), import_path.default.resolve(absPath));
+  return import_path2.default.relative(import_path2.default.resolve(repoRoot), import_path2.default.resolve(absPath));
 }
 function isSymlink(p) {
   try {
@@ -675,10 +998,18 @@ var LocalRipgrepProvider = class {
     this.excludes = excludes;
   }
   async grep(params) {
-    const abs = resolveUnderRepo(this.repoRoot, params.path);
+    let abs;
+    try {
+      abs = resolveUnderRepo(this.repoRoot, params.path);
+    } catch (err) {
+      return {
+        lines: [],
+        error: `[PATH ERROR] ${err instanceof Error ? err.message : String(err)}`
+      };
+    }
     const stat = await import_promises2.default.stat(abs).catch(() => null);
     if (!stat) return { lines: [] };
-    const targetArg = abs === import_path2.default.resolve(this.repoRoot) ? "." : toRepoRelative(this.repoRoot, abs);
+    const targetArg = abs === import_path3.default.resolve(this.repoRoot) ? "." : toRepoRelative(this.repoRoot, abs);
     const args = [
       "--no-config",
       "--no-heading",
@@ -687,6 +1018,9 @@ var LocalRipgrepProvider = class {
       "--color=never",
       "--trim",
       "--max-columns=400",
+      "-C",
+      "1",
+      ...params.glob ? ["--glob", params.glob] : [],
       ...this.excludes.flatMap((e) => ["-g", `!${e}`]),
       params.pattern,
       targetArg || "."
@@ -711,29 +1045,24 @@ Details: ${res.stderr}` : ""}`
       };
     }
     const lines = (res.stdout || "").trim().split(/\r?\n/).filter((l) => l.length > 0);
-    return { lines };
-  }
-  async glob(params) {
-    const abs = resolveUnderRepo(this.repoRoot, params.path);
-    const targetArg = abs === import_path2.default.resolve(this.repoRoot) ? "." : toRepoRelative(this.repoRoot, abs);
-    const args = [
-      "--no-config",
-      "--files",
-      "-g",
-      params.pattern,
-      ...this.excludes.flatMap((e) => ["-g", `!${e}`]),
-      targetArg || "."
-    ];
-    const res = await runRipgrep(args, { cwd: this.repoRoot });
-    if (res.exitCode === -1) {
-      console.warn(`[warp_grep] ripgrep not available for glob: ${res.stderr || "execution failed"}`);
-      return { files: [] };
+    if (lines.length > AGENT_CONFIG.MAX_OUTPUT_LINES) {
+      return {
+        lines: [],
+        error: "query not specific enough, tool tried to return too much context and failed"
+      };
     }
-    const files = (res.stdout || "").trim().split(/\r?\n/).filter((l) => l.length > 0);
-    return { files };
+    return { lines };
   }
   async read(params) {
-    const abs = resolveUnderRepo(this.repoRoot, params.path);
+    let abs;
+    try {
+      abs = resolveUnderRepo(this.repoRoot, params.path);
+    } catch (err) {
+      return {
+        lines: [],
+        error: `[PATH ERROR] ${err instanceof Error ? err.message : String(err)}`
+      };
+    }
     const stat = await import_promises2.default.stat(abs).catch(() => null);
     if (!stat || !stat.isFile()) {
       return {
@@ -753,7 +1082,15 @@ Details: ${res.stderr}` : ""}`
         error: `[UNREADABLE FILE] You tried to read "${params.path}" but this file is either too large or not a text file, so it cannot be read. Try a different file.`
       };
     }
-    const lines = await readAllLines(abs);
+    let lines;
+    try {
+      lines = await readAllLines(abs);
+    } catch (err) {
+      return {
+        lines: [],
+        error: `[READ ERROR] Failed to read "${params.path}": ${err instanceof Error ? err.message : String(err)}`
+      };
+    }
     const total = lines.length;
     let s = params.start ?? 1;
     let e = Math.min(params.end ?? total, total);
@@ -766,30 +1103,46 @@ Details: ${res.stderr}` : ""}`
       const content = lines[i - 1] ?? "";
       out.push(`${i}|${content}`);
     }
+    if (out.length > AGENT_CONFIG.MAX_READ_LINES) {
+      const truncated = out.slice(0, AGENT_CONFIG.MAX_READ_LINES);
+      truncated.push(`... [truncated: showing ${AGENT_CONFIG.MAX_READ_LINES} of ${out.length} lines]`);
+      return { lines: truncated };
+    }
     return { lines: out };
   }
-  async analyse(params) {
-    const abs = resolveUnderRepo(this.repoRoot, params.path);
+  async listDirectory(params) {
+    let abs;
+    try {
+      abs = resolveUnderRepo(this.repoRoot, params.path);
+    } catch {
+      return [];
+    }
     const stat = await import_promises2.default.stat(abs).catch(() => null);
     if (!stat || !stat.isDirectory()) {
       return [];
     }
-    const maxResults = params.maxResults ?? 100;
-    const maxDepth = params.maxDepth ?? 2;
+    const maxResults = params.maxResults ?? AGENT_CONFIG.MAX_OUTPUT_LINES;
+    const maxDepth = params.maxDepth ?? AGENT_CONFIG.MAX_LIST_DEPTH;
     const regex = params.pattern ? new RegExp(params.pattern) : null;
     const results = [];
+    let timedOut = false;
+    const startTime = Date.now();
     async function walk(dir, depth) {
+      if (Date.now() - startTime > AGENT_CONFIG.LIST_TIMEOUT_MS) {
+        timedOut = true;
+        return;
+      }
       if (depth > maxDepth || results.length >= maxResults) return;
       const entries = await import_promises2.default.readdir(dir, { withFileTypes: true });
       for (const entry of entries) {
-        const full = import_path2.default.join(dir, entry.name);
+        if (timedOut || results.length >= maxResults) break;
+        const full = import_path3.default.join(dir, entry.name);
         const rel = toRepoRelative(abs, full).replace(/^[.][/\\]?/, "");
-        if (DEFAULT_EXCLUDES.some((ex) => rel.split(import_path2.default.sep).includes(ex))) continue;
+        if (DEFAULT_EXCLUDES.some((ex) => rel.split(import_path3.default.sep).includes(ex))) continue;
         if (regex && !regex.test(entry.name)) continue;
-        if (results.length >= maxResults) break;
         results.push({
           name: entry.name,
-          path: toRepoRelative(import_path2.default.resolve(""), full),
+          path: toRepoRelative(import_path3.default.resolve(""), full),
           // relative display
           type: entry.isDirectory() ? "dir" : "file",
           depth
@@ -809,29 +1162,6 @@ var parser = new LLMResponseParser();
 function parseToolCalls(text) {
   return parser.parse(text);
 }
-function formatToolResult(name, args, output, options) {
-  return formatAgentToolOutput(name, args, output, options ?? {});
-}
-function formatTurnMessage(turn, maxTurns = 4) {
-  const remaining = maxTurns - turn;
-  if (remaining === 0) {
-    return `
-
-[Turn ${turn}/${maxTurns}] This is your LAST turn. You MUST call finish now.`;
-  }
-  if (remaining === 1) {
-    return `
-
-[Turn ${turn}/${maxTurns}] You have 1 turn remaining.`;
-  }
-  return `
-
-[Turn ${turn}/${maxTurns}] You have ${remaining} turns remaining.`;
-}
-function formatAnalyseTree(entries) {
-  if (!entries.length) return "empty";
-  return entries.map((e) => `${"  ".repeat(e.depth)}- ${e.type === "dir" ? "[D]" : "[F]"} ${e.name}`).join("\n");
-}
 async function resolveFinishFiles(provider, files) {
   return readFinishFiles("", files, async (p, s, e) => {
     const r = await provider.read({ path: p, start: s, end: e });
@@ -841,19 +1171,25 @@ async function resolveFinishFiles(provider, files) {
     });
   });
 }
-var MAX_TURNS = 4;
-var TIMEOUT_MS = AGENT_CONFIG.TIMEOUT_MS;
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
+  AGENT_CONFIG,
   DEFAULT_EXCLUDES,
+  LLMResponseParser,
   LocalRipgrepProvider,
-  MAX_TURNS,
   SYSTEM_PROMPT,
-  TIMEOUT_MS,
-  formatAnalyseTree,
+  buildInitialState,
+  calculateContextBudget,
+  formatListDirectoryTree,
   formatToolResult,
   formatTurnMessage,
+  getSystemPrompt,
+  normalizeFinishFiles,
   parseToolCalls,
-  resolveFinishFiles
+  readFinishFiles,
+  resolveFinishFiles,
+  toolGrep,
+  toolListDirectory,
+  toolRead
 });
 //# sourceMappingURL=harness.cjs.map