@probelabs/probe 0.6.0-rc292 → 0.6.0-rc293

package/build/agent/shared/prompts.js

@@ -90,9 +90,9 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - Do not add code comments unless the logic is genuinely complex and non-obvious.
 
 # Before Implementation
-- Focus on high-level design patterns and system organization
-- Identify architectural patterns and component relationships
-- Evaluate system structure and suggest architectural improvements
+- Read tests first — find existing test files for the module you're changing. They reveal expected behavior, edge cases, and the project's testing patterns.
+- Read neighboring files — understand naming conventions, error handling patterns, import style, and existing utilities before creating new ones.
+- Trace the call chain — follow how the code you're changing is called and what depends on it. Check interfaces, types, and consumers.
 - Focus on backward compatibility
 - Consider scalability, maintainability, and extensibility in your analysis
 
@@ -117,6 +117,20 @@ Before building or testing, determine the project's toolchain:
 - Read README for build/test instructions if the above are unclear
 - Common patterns: \`make build\`/\`make test\`, \`npm run build\`/\`npm test\`, \`cargo build\`/\`cargo test\`, \`go build ./...\`/\`go test ./...\`, \`python -m pytest\`
 
+# File Editing Rules
+You have access to the \`edit\`, \`create\`, and \`multi_edit\` tools for modifying files. You MUST use these tools for ALL code changes. They are purpose-built, atomic, and safe.
+
+DO NOT use sed, awk, echo/cat redirection, or heredocs to modify source code. These commands cause real damage in practice: truncated lines, duplicate code blocks, broken syntax. Every bad edit wastes iterations on fix-up commits.
+
+Use the right tool:
+1. To MODIFY existing code → \`edit\` tool (old_string → new_string, or start_line/end_line)
+2. To CREATE a new file → \`create\` tool
+3. To CHANGE multiple files at once → \`multi_edit\` tool
+4. To READ code → \`extract\` or \`search\` tools
+5. If \`edit\` fails with "file has not been read yet" → use \`extract\` with the EXACT same file path you will pass to \`edit\`. Relative vs absolute path mismatch causes this error. Use the same path format consistently. If it still fails, use bash \`cat\` to read the file, then use \`create\` to write the entire modified file. Do NOT fall back to sed.
+
+Bash is fine for: formatters (gofmt, prettier, black), build/test/lint commands, git operations, and read-only file inspection (cat, head, tail). sed/awk should ONLY be used for trivial non-code tasks (e.g., config file tweaks) where the replacement is a simple literal string swap.
+
 # During Implementation
 - Always create a new branch before making changes to the codebase.
 - Fix problems at the root cause, not with surface-level patches. Prefer general solutions over special cases.
@@ -143,6 +157,22 @@ Before committing or creating a PR, run through this checklist:
 
 Do NOT skip verification. Do NOT proceed to PR creation with a broken build or failing tests.
 
+# Output Integrity
+Your final output MUST accurately reflect what ACTUALLY happened. Do NOT fabricate, hallucinate, or report aspirational results.
+
+- Only report PR URLs you actually created or updated with \`gh pr create\` or \`git push\`. If you checked out an existing PR but did NOT push changes to it, do NOT claim you updated it.
+- Describe what you ACTUALLY DID, not what you planned or intended to do. If you ran out of iterations, say so. If tests failed, say so.
+- Only list files you actually modified AND committed.
+- If you could not complete the task — ran out of iterations, tests failed, build broken, push rejected — report the real reason honestly.
+
+NEVER claim success when:
+- You did not run \`git push\` successfully
+- Tests failed and you did not fix them
+- You hit the iteration limit before completing the work
+- You only analyzed/investigated but did not implement changes
+
+A false success report is WORSE than an honest failure — it misleads the user into thinking work is done when it is not.
+
 # GitHub Integration
 - Use the \`gh\` CLI for all GitHub operations: issues, pull requests, checks, releases.
 - To view issues or PRs: \`gh issue view <number>\`, \`gh pr view <number>\`.

package/build/tools/fileTracker.js

@@ -12,9 +12,22 @@
  */
 
 import { createHash } from 'crypto';
-import { resolve, isAbsolute } from 'path';
+import { resolve, isAbsolute, normalize } from 'path';
 import { findSymbol } from './symbolEdit.js';
 
+/**
+ * Normalize a file path for consistent storage and lookup.
+ * Resolves '.', '..', double slashes, and ensures absolute paths are canonical.
+ * Does NOT resolve symlinks (that would be expensive and might fail for non-existent files).
+ * @param {string} filePath - Path to normalize
+ * @returns {string} Normalized path
+ */
+function normalizePath(filePath) {
+  if (!filePath) return filePath;
+  // resolve() handles '.', '..', double slashes, and makes the path absolute
+  return resolve(filePath);
+}
+
 /**
  * Compute a SHA-256 content hash for a code block.
  * Normalizes trailing whitespace per line for robustness against editor formatting.
@@ -106,10 +119,11 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   markFileSeen(resolvedPath) {
-    this._seenFiles.add(resolvedPath);
-    this._textEditCounts.set(resolvedPath, 0);
+    const normalized = normalizePath(resolvedPath);
+    this._seenFiles.add(normalized);
+    this._textEditCounts.set(normalized, 0);
     if (this.debug) {
-      console.error(`[FileTracker] Marked as seen: ${resolvedPath}`);
+      console.error(`[FileTracker] Marked as seen: ${normalized}`);
     }
   }
 
@@ -119,7 +133,7 @@ export class FileTracker {
    * @returns {boolean}
    */
   isFileSeen(resolvedPath) {
-    return this._seenFiles.has(resolvedPath);
+    return this._seenFiles.has(normalizePath(resolvedPath));
   }
 
   /**
@@ -132,7 +146,7 @@ export class FileTracker {
    * @param {string} [source='extract'] - How the content was obtained
    */
   trackSymbolContent(resolvedPath, symbolName, code, startLine, endLine, source = 'extract') {
-    const key = `${resolvedPath}#${symbolName}`;
+    const key = `${normalizePath(resolvedPath)}#${symbolName}`;
     const contentHash = computeContentHash(code);
     this._contentRecords.set(key, {
       contentHash,
@@ -154,7 +168,7 @@ export class FileTracker {
    * @returns {Object|null} The stored record or null
    */
   getSymbolRecord(resolvedPath, symbolName) {
-    return this._contentRecords.get(`${resolvedPath}#${symbolName}`) || null;
+    return this._contentRecords.get(`${normalizePath(resolvedPath)}#${symbolName}`) || null;
   }
 
   /**
@@ -165,7 +179,7 @@ export class FileTracker {
    * @returns {{ok: boolean, reason?: string, message?: string}}
    */
   checkSymbolContent(resolvedPath, symbolName, currentCode) {
-    const key = `${resolvedPath}#${symbolName}`;
+    const key = `${normalizePath(resolvedPath)}#${symbolName}`;
     const record = this._contentRecords.get(key);
 
     if (!record) {
@@ -253,7 +267,7 @@ export class FileTracker {
    * @returns {{ok: boolean, reason?: string, message?: string}}
    */
   checkBeforeEdit(resolvedPath) {
-    if (!this._seenFiles.has(resolvedPath)) {
+    if (!this._seenFiles.has(normalizePath(resolvedPath))) {
       return {
         ok: false,
         reason: 'untracked',
@@ -269,8 +283,9 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   async trackFileAfterWrite(resolvedPath) {
-    this._seenFiles.add(resolvedPath);
-    this.invalidateFileRecords(resolvedPath);
+    const normalized = normalizePath(resolvedPath);
+    this._seenFiles.add(normalized);
+    this.invalidateFileRecords(normalized);
   }
 
   /**
@@ -279,10 +294,11 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   recordTextEdit(resolvedPath) {
-    const count = (this._textEditCounts.get(resolvedPath) || 0) + 1;
-    this._textEditCounts.set(resolvedPath, count);
+    const normalized = normalizePath(resolvedPath);
+    const count = (this._textEditCounts.get(normalized) || 0) + 1;
+    this._textEditCounts.set(normalized, count);
     if (this.debug) {
-      console.error(`[FileTracker] Text edit #${count} for ${resolvedPath}`);
+      console.error(`[FileTracker] Text edit #${count} for ${normalized}`);
     }
   }
 
@@ -292,7 +308,7 @@ export class FileTracker {
    * @returns {{ok: boolean, editCount?: number, message?: string}}
    */
   checkTextEditStaleness(resolvedPath) {
-    const count = this._textEditCounts.get(resolvedPath) || 0;
+    const count = this._textEditCounts.get(normalizePath(resolvedPath)) || 0;
     if (count >= this.maxConsecutiveTextEdits) {
       return {
         ok: false,
@@ -323,7 +339,7 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   invalidateFileRecords(resolvedPath) {
-    const prefix = resolvedPath + '#';
+    const prefix = normalizePath(resolvedPath) + '#';
     for (const key of this._contentRecords.keys()) {
       if (key.startsWith(prefix)) {
         this._contentRecords.delete(key);
@@ -340,7 +356,7 @@ export class FileTracker {
    * @returns {boolean}
    */
   isTracked(resolvedPath) {
-    return this.isFileSeen(resolvedPath);
+    return this.isFileSeen(normalizePath(resolvedPath));
   }
 
   /**

package/build/tools/vercel.js

@@ -385,7 +385,7 @@ export const searchTool = (options = {}) => {
 			? searchDelegateDescription
 			: searchDescription,
 		inputSchema: searchSchema,
-		execute: async ({ query: searchQuery, path, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage }) => {
+		execute: async ({ query: searchQuery, path, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage, workingDirectory }) => {
 			// Auto-quote mixed-case and underscore terms to prevent unwanted stemming/splitting
 			// Skip when exact=true since that already preserves the literal string
 			if (!exact && searchQuery) {
@@ -399,15 +399,18 @@ export const searchTool = (options = {}) => {
 			// Use parameter maxTokens if provided, otherwise use the default
 			const effectiveMaxTokens = paramMaxTokens || maxTokens;
 
+			// Use workingDirectory (injected by _buildNativeTools at runtime) > cwd from config > fallback
+			const effectiveSearchCwd = workingDirectory || options.cwd || '.';
+
 			// Parse and resolve paths (supports comma-separated and relative paths)
 			let searchPaths;
 			if (path) {
-				searchPaths = parseAndResolvePaths(path, options.cwd);
+				searchPaths = parseAndResolvePaths(path, effectiveSearchCwd);
 			}
 
 			// Default to cwd or '.' if no paths provided
 			if (!searchPaths || searchPaths.length === 0) {
-				searchPaths = [options.cwd || '.'];
+				searchPaths = [effectiveSearchCwd];
 			}
 
 			// Join paths with space for CLI (probe search supports multiple paths)
@@ -416,7 +419,7 @@ export const searchTool = (options = {}) => {
 			const searchOptions = {
 				query: searchQuery,
 				path: searchPath,
-				cwd: options.cwd, // Working directory for resolving relative paths
+				cwd: effectiveSearchCwd, // Working directory for resolving relative paths
 				allowTests: allow_tests ?? true,
 				exact,
 				json: false,
@@ -473,7 +476,7 @@ export const searchTool = (options = {}) => {
 					const result = maybeAnnotate(await runRawSearch());
 					// Track files found in search results for staleness detection
 					if (options.fileTracker && typeof result === 'string') {
-						options.fileTracker.trackFilesFromOutput(result, options.cwd || '.').catch(() => {});
+						options.fileTracker.trackFilesFromOutput(result, effectiveSearchCwd).catch(() => {});
 					}
 					return result;
 				} catch (error) {
@@ -532,7 +535,7 @@ export const searchTool = (options = {}) => {
 					}
 					const fallbackResult = maybeAnnotate(await runRawSearch());
 					if (options.fileTracker && typeof fallbackResult === 'string') {
-						options.fileTracker.trackFilesFromOutput(fallbackResult, options.cwd || '.').catch(() => {});
+						options.fileTracker.trackFilesFromOutput(fallbackResult, effectiveSearchCwd).catch(() => {});
 					}
 					return fallbackResult;
 				}
@@ -614,7 +617,7 @@ export const searchTool = (options = {}) => {
 				try {
 					const fallbackResult2 = maybeAnnotate(await runRawSearch());
 					if (options.fileTracker && typeof fallbackResult2 === 'string') {
-						options.fileTracker.trackFilesFromOutput(fallbackResult2, options.cwd || '.').catch(() => {});
+						options.fileTracker.trackFilesFromOutput(fallbackResult2, effectiveSearchCwd).catch(() => {});
 					}
 					return fallbackResult2;
 				} catch (fallbackError) {
@@ -693,10 +696,10 @@ export const extractTool = (options = {}) => {
 		name: 'extract',
 		description: extractDescription,
 		inputSchema: extractSchema,
-		execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format }) => {
+		execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format, workingDirectory }) => {
 			try {
-				// Use the cwd from config for working directory
-				const effectiveCwd = options.cwd || '.';
+				// Use workingDirectory (injected by _buildNativeTools at runtime) > cwd from config > fallback
+				const effectiveCwd = workingDirectory || options.cwd || '.';
 
 				if (debug) {
 					if (targets) {

package/cjs/agent/ProbeAgent.cjs

@@ -27566,7 +27566,7 @@ var init_vercel = __esm({
         name: "search",
         description: searchDelegate ? searchDelegateDescription : searchDescription,
         inputSchema: searchSchema,
-        execute: async ({ query: searchQuery, path: path9, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage }) => {
+        execute: async ({ query: searchQuery, path: path9, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage, workingDirectory }) => {
           if (!exact && searchQuery) {
             const originalQuery = searchQuery;
             searchQuery = autoQuoteSearchTerms(searchQuery);
@@ -27575,18 +27575,19 @@ var init_vercel = __esm({
             }
           }
           const effectiveMaxTokens = paramMaxTokens || maxTokens;
+          const effectiveSearchCwd = workingDirectory || options.cwd || ".";
           let searchPaths;
           if (path9) {
-            searchPaths = parseAndResolvePaths(path9, options.cwd);
+            searchPaths = parseAndResolvePaths(path9, effectiveSearchCwd);
           }
           if (!searchPaths || searchPaths.length === 0) {
-            searchPaths = [options.cwd || "."];
+            searchPaths = [effectiveSearchCwd];
           }
           const searchPath = searchPaths.join(" ");
           const searchOptions = {
             query: searchQuery,
             path: searchPath,
-            cwd: options.cwd,
+            cwd: effectiveSearchCwd,
             // Working directory for resolving relative paths
             allowTests: allow_tests ?? true,
             exact,
@@ -27637,7 +27638,7 @@ var init_vercel = __esm({
             try {
               const result = maybeAnnotate(await runRawSearch());
               if (options.fileTracker && typeof result === "string") {
-                options.fileTracker.trackFilesFromOutput(result, options.cwd || ".").catch(() => {
+                options.fileTracker.trackFilesFromOutput(result, effectiveSearchCwd).catch(() => {
                 });
               }
               return result;
@@ -27690,7 +27691,7 @@ var init_vercel = __esm({
               }
               const fallbackResult = maybeAnnotate(await runRawSearch());
               if (options.fileTracker && typeof fallbackResult === "string") {
-                options.fileTracker.trackFilesFromOutput(fallbackResult, options.cwd || ".").catch(() => {
+                options.fileTracker.trackFilesFromOutput(fallbackResult, effectiveSearchCwd).catch(() => {
                 });
               }
               return fallbackResult;
@@ -27753,7 +27754,7 @@ var init_vercel = __esm({
             try {
               const fallbackResult2 = maybeAnnotate(await runRawSearch());
               if (options.fileTracker && typeof fallbackResult2 === "string") {
-                options.fileTracker.trackFilesFromOutput(fallbackResult2, options.cwd || ".").catch(() => {
+                options.fileTracker.trackFilesFromOutput(fallbackResult2, effectiveSearchCwd).catch(() => {
                 });
               }
               return fallbackResult2;
@@ -27807,9 +27808,9 @@ var init_vercel = __esm({
         name: "extract",
         description: extractDescription,
         inputSchema: extractSchema,
-        execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format }) => {
+        execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format, workingDirectory }) => {
           try {
-            const effectiveCwd = options.cwd || ".";
+            const effectiveCwd = workingDirectory || options.cwd || ".";
             if (debug) {
               if (targets) {
                 console.error(`Executing extract with targets: "${targets}", cwd: "${effectiveCwd}", context lines: ${context_lines || 10}`);
@@ -48638,6 +48639,10 @@ var init_file_lister = __esm({
 });
 
 // src/tools/fileTracker.js
+function normalizePath(filePath) {
+  if (!filePath) return filePath;
+  return (0, import_path10.resolve)(filePath);
+}
 function computeContentHash(content) {
   const normalized = (content || "").split("\n").map((l) => l.trimEnd()).join("\n");
   return (0, import_crypto3.createHash)("sha256").update(normalized).digest("hex").slice(0, 16);
@@ -48700,10 +48705,11 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       markFileSeen(resolvedPath) {
-        this._seenFiles.add(resolvedPath);
-        this._textEditCounts.set(resolvedPath, 0);
+        const normalized = normalizePath(resolvedPath);
+        this._seenFiles.add(normalized);
+        this._textEditCounts.set(normalized, 0);
         if (this.debug) {
-          console.error(`[FileTracker] Marked as seen: ${resolvedPath}`);
+          console.error(`[FileTracker] Marked as seen: ${normalized}`);
         }
       }
       /**
@@ -48712,7 +48718,7 @@ var init_fileTracker = __esm({
        * @returns {boolean}
        */
       isFileSeen(resolvedPath) {
-        return this._seenFiles.has(resolvedPath);
+        return this._seenFiles.has(normalizePath(resolvedPath));
       }
       /**
        * Store a content hash for a symbol in a file.
@@ -48724,7 +48730,7 @@ var init_fileTracker = __esm({
        * @param {string} [source='extract'] - How the content was obtained
        */
       trackSymbolContent(resolvedPath, symbolName, code, startLine, endLine, source = "extract") {
-        const key = `${resolvedPath}#${symbolName}`;
+        const key = `${normalizePath(resolvedPath)}#${symbolName}`;
         const contentHash = computeContentHash(code);
         this._contentRecords.set(key, {
           contentHash,
@@ -48745,7 +48751,7 @@ var init_fileTracker = __esm({
        * @returns {Object|null} The stored record or null
        */
       getSymbolRecord(resolvedPath, symbolName) {
-        return this._contentRecords.get(`${resolvedPath}#${symbolName}`) || null;
+        return this._contentRecords.get(`${normalizePath(resolvedPath)}#${symbolName}`) || null;
       }
       /**
        * Check if a symbol's current content matches what was stored.
@@ -48755,7 +48761,7 @@ var init_fileTracker = __esm({
        * @returns {{ok: boolean, reason?: string, message?: string}}
        */
       checkSymbolContent(resolvedPath, symbolName, currentCode) {
-        const key = `${resolvedPath}#${symbolName}`;
+        const key = `${normalizePath(resolvedPath)}#${symbolName}`;
         const record2 = this._contentRecords.get(key);
         if (!record2) {
           return { ok: true };
@@ -48832,7 +48838,7 @@ var init_fileTracker = __esm({
        * @returns {{ok: boolean, reason?: string, message?: string}}
        */
       checkBeforeEdit(resolvedPath) {
-        if (!this._seenFiles.has(resolvedPath)) {
+        if (!this._seenFiles.has(normalizePath(resolvedPath))) {
           return {
             ok: false,
             reason: "untracked",
@@ -48847,8 +48853,9 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       async trackFileAfterWrite(resolvedPath) {
-        this._seenFiles.add(resolvedPath);
-        this.invalidateFileRecords(resolvedPath);
+        const normalized = normalizePath(resolvedPath);
+        this._seenFiles.add(normalized);
+        this.invalidateFileRecords(normalized);
       }
       /**
        * Record a text-mode edit (old_string/new_string) to a file.
@@ -48856,10 +48863,11 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       recordTextEdit(resolvedPath) {
-        const count = (this._textEditCounts.get(resolvedPath) || 0) + 1;
-        this._textEditCounts.set(resolvedPath, count);
+        const normalized = normalizePath(resolvedPath);
+        const count = (this._textEditCounts.get(normalized) || 0) + 1;
+        this._textEditCounts.set(normalized, count);
         if (this.debug) {
-          console.error(`[FileTracker] Text edit #${count} for ${resolvedPath}`);
+          console.error(`[FileTracker] Text edit #${count} for ${normalized}`);
         }
       }
       /**
@@ -48868,7 +48876,7 @@ var init_fileTracker = __esm({
        * @returns {{ok: boolean, editCount?: number, message?: string}}
        */
       checkTextEditStaleness(resolvedPath) {
-        const count = this._textEditCounts.get(resolvedPath) || 0;
+        const count = this._textEditCounts.get(normalizePath(resolvedPath)) || 0;
         if (count >= this.maxConsecutiveTextEdits) {
           return {
             ok: false,
@@ -48897,7 +48905,7 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       invalidateFileRecords(resolvedPath) {
-        const prefix = resolvedPath + "#";
+        const prefix = normalizePath(resolvedPath) + "#";
         for (const key of this._contentRecords.keys()) {
           if (key.startsWith(prefix)) {
             this._contentRecords.delete(key);
@@ -48913,7 +48921,7 @@ var init_fileTracker = __esm({
        * @returns {boolean}
        */
       isTracked(resolvedPath) {
-        return this.isFileSeen(resolvedPath);
+        return this.isFileSeen(normalizePath(resolvedPath));
       }
       /**
        * Clear all tracking state.
@@ -69026,7 +69034,7 @@ var init_graph_builder = __esm({
       applyLinkStyles() {
         if (!this.pendingLinkStyles.length || !this.edges.length)
           return;
-        const normalize3 = (s) => {
+        const normalize4 = (s) => {
           const out = {};
           for (const [kRaw, vRaw] of Object.entries(s)) {
             const k = kRaw.trim().toLowerCase();
@@ -69047,7 +69055,7 @@ var init_graph_builder = __esm({
           return out;
         };
         for (const cmd of this.pendingLinkStyles) {
-          const style = normalize3(cmd.props);
+          const style = normalize4(cmd.props);
           for (const idx of cmd.indices) {
             if (idx >= 0 && idx < this.edges.length) {
               const e = this.edges[idx];
@@ -76314,7 +76322,7 @@ var require_layout = __commonJS({
     "use strict";
     var _ = require_lodash2();
     var acyclic = require_acyclic();
-    var normalize3 = require_normalize();
+    var normalize4 = require_normalize();
     var rank = require_rank();
     var normalizeRanks = require_util().normalizeRanks;
     var parentDummyChains = require_parent_dummy_chains();
@@ -76376,7 +76384,7 @@ var require_layout = __commonJS({
         removeEdgeLabelProxies(g);
       });
       time3("    normalize.run", function() {
-        normalize3.run(g);
+        normalize4.run(g);
       });
       time3("    parentDummyChains", function() {
         parentDummyChains(g);
@@ -76403,7 +76411,7 @@ var require_layout = __commonJS({
         removeBorderNodes(g);
       });
       time3("    normalize.undo", function() {
-        normalize3.undo(g);
+        normalize4.undo(g);
       });
       time3("    fixupEdgeLabelCoords", function() {
         fixupEdgeLabelCoords(g);
@@ -82774,8 +82782,8 @@ var require_resolve = __commonJS({
       }
       return count;
     }
-    function getFullPath(resolver, id = "", normalize3) {
-      if (normalize3 !== false)
+    function getFullPath(resolver, id = "", normalize4) {
+      if (normalize4 !== false)
         id = normalizeId(id);
       const p = resolver.parse(id);
       return _getFullPath(resolver, p);
@@ -84115,7 +84123,7 @@ var require_fast_uri = __commonJS({
     "use strict";
     var { normalizeIPv6, removeDotSegments, recomposeAuthority, normalizeComponentEncoding, isIPv4, nonSimpleDomain } = require_utils2();
     var { SCHEMES, getSchemeHandler } = require_schemes();
-    function normalize3(uri, options) {
+    function normalize4(uri, options) {
       if (typeof uri === "string") {
         uri = /** @type {T} */
         serialize(parse11(uri, options), options);
@@ -84351,7 +84359,7 @@ var require_fast_uri = __commonJS({
     }
     var fastUri = {
       SCHEMES,
-      normalize: normalize3,
+      normalize: normalize4,
       resolve: resolve9,
       resolveComponent,
       equal,
@@ -88558,9 +88566,9 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - Do not add code comments unless the logic is genuinely complex and non-obvious.
 
 # Before Implementation
-- Focus on high-level design patterns and system organization
-- Identify architectural patterns and component relationships
-- Evaluate system structure and suggest architectural improvements
+- Read tests first \u2014 find existing test files for the module you're changing. They reveal expected behavior, edge cases, and the project's testing patterns.
+- Read neighboring files \u2014 understand naming conventions, error handling patterns, import style, and existing utilities before creating new ones.
+- Trace the call chain \u2014 follow how the code you're changing is called and what depends on it. Check interfaces, types, and consumers.
 - Focus on backward compatibility
 - Consider scalability, maintainability, and extensibility in your analysis
 
@@ -88585,6 +88593,20 @@ Before building or testing, determine the project's toolchain:
 - Read README for build/test instructions if the above are unclear
 - Common patterns: \`make build\`/\`make test\`, \`npm run build\`/\`npm test\`, \`cargo build\`/\`cargo test\`, \`go build ./...\`/\`go test ./...\`, \`python -m pytest\`
 
+# File Editing Rules
+You have access to the \`edit\`, \`create\`, and \`multi_edit\` tools for modifying files. You MUST use these tools for ALL code changes. They are purpose-built, atomic, and safe.
+
+DO NOT use sed, awk, echo/cat redirection, or heredocs to modify source code. These commands cause real damage in practice: truncated lines, duplicate code blocks, broken syntax. Every bad edit wastes iterations on fix-up commits.
+
+Use the right tool:
+1. To MODIFY existing code \u2192 \`edit\` tool (old_string \u2192 new_string, or start_line/end_line)
+2. To CREATE a new file \u2192 \`create\` tool
+3. To CHANGE multiple files at once \u2192 \`multi_edit\` tool
+4. To READ code \u2192 \`extract\` or \`search\` tools
+5. If \`edit\` fails with "file has not been read yet" \u2192 use \`extract\` with the EXACT same file path you will pass to \`edit\`. Relative vs absolute path mismatch causes this error. Use the same path format consistently. If it still fails, use bash \`cat\` to read the file, then use \`create\` to write the entire modified file. Do NOT fall back to sed.
+
+Bash is fine for: formatters (gofmt, prettier, black), build/test/lint commands, git operations, and read-only file inspection (cat, head, tail). sed/awk should ONLY be used for trivial non-code tasks (e.g., config file tweaks) where the replacement is a simple literal string swap.
+
 # During Implementation
 - Always create a new branch before making changes to the codebase.
 - Fix problems at the root cause, not with surface-level patches. Prefer general solutions over special cases.
@@ -88611,6 +88633,22 @@ Before committing or creating a PR, run through this checklist:
 
 Do NOT skip verification. Do NOT proceed to PR creation with a broken build or failing tests.
 
+# Output Integrity
+Your final output MUST accurately reflect what ACTUALLY happened. Do NOT fabricate, hallucinate, or report aspirational results.
+
+- Only report PR URLs you actually created or updated with \`gh pr create\` or \`git push\`. If you checked out an existing PR but did NOT push changes to it, do NOT claim you updated it.
+- Describe what you ACTUALLY DID, not what you planned or intended to do. If you ran out of iterations, say so. If tests failed, say so.
+- Only list files you actually modified AND committed.
+- If you could not complete the task \u2014 ran out of iterations, tests failed, build broken, push rejected \u2014 report the real reason honestly.
+
+NEVER claim success when:
+- You did not run \`git push\` successfully
+- Tests failed and you did not fix them
+- You hit the iteration limit before completing the work
+- You only analyzed/investigated but did not implement changes
+
+A false success report is WORSE than an honest failure \u2014 it misleads the user into thinking work is done when it is not.
+
 # GitHub Integration
 - Use the \`gh\` CLI for all GitHub operations: issues, pull requests, checks, releases.
 - To view issues or PRs: \`gh issue view <number>\`, \`gh pr view <number>\`.

package/cjs/index.cjs

@@ -27260,6 +27260,10 @@ var init_symbolEdit = __esm({
 });
 
 // src/tools/fileTracker.js
+function normalizePath(filePath) {
+  if (!filePath) return filePath;
+  return (0, import_path6.resolve)(filePath);
+}
 function computeContentHash(content) {
   const normalized = (content || "").split("\n").map((l) => l.trimEnd()).join("\n");
   return (0, import_crypto2.createHash)("sha256").update(normalized).digest("hex").slice(0, 16);
@@ -27322,10 +27326,11 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       markFileSeen(resolvedPath) {
-        this._seenFiles.add(resolvedPath);
-        this._textEditCounts.set(resolvedPath, 0);
+        const normalized = normalizePath(resolvedPath);
+        this._seenFiles.add(normalized);
+        this._textEditCounts.set(normalized, 0);
         if (this.debug) {
-          console.error(`[FileTracker] Marked as seen: ${resolvedPath}`);
+          console.error(`[FileTracker] Marked as seen: ${normalized}`);
         }
       }
       /**
@@ -27334,7 +27339,7 @@ var init_fileTracker = __esm({
        * @returns {boolean}
        */
       isFileSeen(resolvedPath) {
-        return this._seenFiles.has(resolvedPath);
+        return this._seenFiles.has(normalizePath(resolvedPath));
       }
       /**
        * Store a content hash for a symbol in a file.
@@ -27346,7 +27351,7 @@ var init_fileTracker = __esm({
        * @param {string} [source='extract'] - How the content was obtained
        */
       trackSymbolContent(resolvedPath, symbolName, code, startLine, endLine, source = "extract") {
-        const key = `${resolvedPath}#${symbolName}`;
+        const key = `${normalizePath(resolvedPath)}#${symbolName}`;
         const contentHash = computeContentHash(code);
         this._contentRecords.set(key, {
           contentHash,
@@ -27367,7 +27372,7 @@ var init_fileTracker = __esm({
        * @returns {Object|null} The stored record or null
        */
       getSymbolRecord(resolvedPath, symbolName) {
-        return this._contentRecords.get(`${resolvedPath}#${symbolName}`) || null;
+        return this._contentRecords.get(`${normalizePath(resolvedPath)}#${symbolName}`) || null;
       }
       /**
        * Check if a symbol's current content matches what was stored.
@@ -27377,7 +27382,7 @@ var init_fileTracker = __esm({
        * @returns {{ok: boolean, reason?: string, message?: string}}
        */
       checkSymbolContent(resolvedPath, symbolName, currentCode) {
-        const key = `${resolvedPath}#${symbolName}`;
+        const key = `${normalizePath(resolvedPath)}#${symbolName}`;
         const record2 = this._contentRecords.get(key);
         if (!record2) {
           return { ok: true };
@@ -27454,7 +27459,7 @@ var init_fileTracker = __esm({
        * @returns {{ok: boolean, reason?: string, message?: string}}
        */
       checkBeforeEdit(resolvedPath) {
-        if (!this._seenFiles.has(resolvedPath)) {
+        if (!this._seenFiles.has(normalizePath(resolvedPath))) {
           return {
             ok: false,
             reason: "untracked",
@@ -27469,8 +27474,9 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       async trackFileAfterWrite(resolvedPath) {
-        this._seenFiles.add(resolvedPath);
-        this.invalidateFileRecords(resolvedPath);
+        const normalized = normalizePath(resolvedPath);
+        this._seenFiles.add(normalized);
+        this.invalidateFileRecords(normalized);
       }
       /**
        * Record a text-mode edit (old_string/new_string) to a file.
@@ -27478,10 +27484,11 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       recordTextEdit(resolvedPath) {
-        const count = (this._textEditCounts.get(resolvedPath) || 0) + 1;
-        this._textEditCounts.set(resolvedPath, count);
+        const normalized = normalizePath(resolvedPath);
+        const count = (this._textEditCounts.get(normalized) || 0) + 1;
+        this._textEditCounts.set(normalized, count);
         if (this.debug) {
-          console.error(`[FileTracker] Text edit #${count} for ${resolvedPath}`);
+          console.error(`[FileTracker] Text edit #${count} for ${normalized}`);
         }
       }
       /**
@@ -27490,7 +27497,7 @@ var init_fileTracker = __esm({
        * @returns {{ok: boolean, editCount?: number, message?: string}}
        */
       checkTextEditStaleness(resolvedPath) {
-        const count = this._textEditCounts.get(resolvedPath) || 0;
+        const count = this._textEditCounts.get(normalizePath(resolvedPath)) || 0;
         if (count >= this.maxConsecutiveTextEdits) {
           return {
             ok: false,
@@ -27519,7 +27526,7 @@ var init_fileTracker = __esm({
        * @param {string} resolvedPath - Absolute path to the file
        */
       invalidateFileRecords(resolvedPath) {
-        const prefix = resolvedPath + "#";
+        const prefix = normalizePath(resolvedPath) + "#";
         for (const key of this._contentRecords.keys()) {
           if (key.startsWith(prefix)) {
             this._contentRecords.delete(key);
@@ -27535,7 +27542,7 @@ var init_fileTracker = __esm({
        * @returns {boolean}
        */
       isTracked(resolvedPath) {
-        return this.isFileSeen(resolvedPath);
+        return this.isFileSeen(normalizePath(resolvedPath));
       }
       /**
        * Clear all tracking state.
@@ -31588,7 +31595,7 @@ var init_esm3 = __esm({
 });
 
 // node_modules/path-scurry/dist/esm/index.js
-var import_node_path, import_node_url, import_fs4, actualFS, import_promises, realpathSync2, defaultFS, fsFromOption, uncDriveRegexp, uncToDrive, eitherSep, UNKNOWN, IFIFO, IFCHR, IFDIR, IFBLK, IFREG, IFLNK, IFSOCK, IFMT, IFMT_UNKNOWN, READDIR_CALLED, LSTAT_CALLED, ENOTDIR, ENOENT, ENOREADLINK, ENOREALPATH, ENOCHILD, TYPEMASK, entToType, normalizeCache, normalize, normalizeNocaseCache, normalizeNocase, ResolveCache, ChildrenCache, setAsCwd, PathBase, PathWin32, PathPosix, PathScurryBase, PathScurryWin32, PathScurryPosix, PathScurryDarwin, Path, PathScurry;
+var import_node_path, import_node_url, import_fs4, actualFS, import_promises, realpathSync2, defaultFS, fsFromOption, uncDriveRegexp, uncToDrive, eitherSep, UNKNOWN, IFIFO, IFCHR, IFDIR, IFBLK, IFREG, IFLNK, IFSOCK, IFMT, IFMT_UNKNOWN, READDIR_CALLED, LSTAT_CALLED, ENOTDIR, ENOENT, ENOREADLINK, ENOREALPATH, ENOCHILD, TYPEMASK, entToType, normalizeCache, normalize2, normalizeNocaseCache, normalizeNocase, ResolveCache, ChildrenCache, setAsCwd, PathBase, PathWin32, PathPosix, PathScurryBase, PathScurryWin32, PathScurryPosix, PathScurryDarwin, Path, PathScurry;
 var init_esm4 = __esm({
   "node_modules/path-scurry/dist/esm/index.js"() {
     init_esm2();
@@ -31643,7 +31650,7 @@ var init_esm4 = __esm({
     TYPEMASK = 1023;
     entToType = (s) => s.isFile() ? IFREG : s.isDirectory() ? IFDIR : s.isSymbolicLink() ? IFLNK : s.isCharacterDevice() ? IFCHR : s.isBlockDevice() ? IFBLK : s.isSocket() ? IFSOCK : s.isFIFO() ? IFIFO : UNKNOWN;
     normalizeCache = /* @__PURE__ */ new Map();
-    normalize = (s) => {
+    normalize2 = (s) => {
       const c = normalizeCache.get(s);
       if (c)
         return c;
@@ -31656,7 +31663,7 @@ var init_esm4 = __esm({
       const c = normalizeNocaseCache.get(s);
       if (c)
         return c;
-      const n = normalize(s.toLowerCase());
+      const n = normalize2(s.toLowerCase());
       normalizeNocaseCache.set(s, n);
       return n;
     };
@@ -31823,7 +31830,7 @@ var init_esm4 = __esm({
        */
       constructor(name15, type = UNKNOWN, root2, roots, nocase, children, opts) {
         this.name = name15;
-        this.#matchName = nocase ? normalizeNocase(name15) : normalize(name15);
+        this.#matchName = nocase ? normalizeNocase(name15) : normalize2(name15);
         this.#type = type & TYPEMASK;
         this.nocase = nocase;
         this.roots = roots;
@@ -31916,7 +31923,7 @@ var init_esm4 = __esm({
           return this.parent || this;
         }
         const children = this.children();
-        const name15 = this.nocase ? normalizeNocase(pathPart) : normalize(pathPart);
+        const name15 = this.nocase ? normalizeNocase(pathPart) : normalize2(pathPart);
         for (const p of children) {
           if (p.#matchName === name15) {
             return p;
@@ -32161,7 +32168,7 @@ var init_esm4 = __esm({
        * directly.
        */
       isNamed(n) {
-        return !this.nocase ? this.#matchName === normalize(n) : this.#matchName === normalizeNocase(n);
+        return !this.nocase ? this.#matchName === normalize2(n) : this.#matchName === normalizeNocase(n);
       }
       /**
        * Return the Path object corresponding to the target of a symbolic link.
@@ -32300,7 +32307,7 @@ var init_esm4 = __esm({
       #readdirMaybePromoteChild(e, c) {
         for (let p = c.provisional; p < c.length; p++) {
           const pchild = c[p];
-          const name15 = this.nocase ? normalizeNocase(e.name) : normalize(e.name);
+          const name15 = this.nocase ? normalizeNocase(e.name) : normalize2(e.name);
           if (name15 !== pchild.#matchName) {
             continue;
           }
@@ -53617,7 +53624,7 @@ var init_graph_builder = __esm({
       applyLinkStyles() {
         if (!this.pendingLinkStyles.length || !this.edges.length)
           return;
-        const normalize3 = (s) => {
+        const normalize4 = (s) => {
           const out = {};
           for (const [kRaw, vRaw] of Object.entries(s)) {
             const k = kRaw.trim().toLowerCase();
@@ -53638,7 +53645,7 @@ var init_graph_builder = __esm({
           return out;
         };
         for (const cmd of this.pendingLinkStyles) {
-          const style = normalize3(cmd.props);
+          const style = normalize4(cmd.props);
           for (const idx of cmd.indices) {
             if (idx >= 0 && idx < this.edges.length) {
               const e = this.edges[idx];
@@ -60905,7 +60912,7 @@ var require_layout = __commonJS({
     "use strict";
     var _ = require_lodash2();
     var acyclic = require_acyclic();
-    var normalize3 = require_normalize();
+    var normalize4 = require_normalize();
     var rank = require_rank();
     var normalizeRanks = require_util().normalizeRanks;
     var parentDummyChains = require_parent_dummy_chains();
@@ -60967,7 +60974,7 @@ var require_layout = __commonJS({
         removeEdgeLabelProxies(g);
       });
       time3("    normalize.run", function() {
-        normalize3.run(g);
+        normalize4.run(g);
       });
       time3("    parentDummyChains", function() {
         parentDummyChains(g);
@@ -60994,7 +61001,7 @@ var require_layout = __commonJS({
         removeBorderNodes(g);
       });
       time3("    normalize.undo", function() {
-        normalize3.undo(g);
+        normalize4.undo(g);
       });
       time3("    fixupEdgeLabelCoords", function() {
         fixupEdgeLabelCoords(g);
@@ -67365,8 +67372,8 @@ var require_resolve = __commonJS({
       }
       return count;
     }
-    function getFullPath(resolver, id = "", normalize3) {
-      if (normalize3 !== false)
+    function getFullPath(resolver, id = "", normalize4) {
+      if (normalize4 !== false)
         id = normalizeId(id);
       const p = resolver.parse(id);
       return _getFullPath(resolver, p);
@@ -68706,7 +68713,7 @@ var require_fast_uri = __commonJS({
     "use strict";
     var { normalizeIPv6, removeDotSegments, recomposeAuthority, normalizeComponentEncoding, isIPv4, nonSimpleDomain } = require_utils();
     var { SCHEMES, getSchemeHandler } = require_schemes();
-    function normalize3(uri, options) {
+    function normalize4(uri, options) {
       if (typeof uri === "string") {
         uri = /** @type {T} */
         serialize(parse11(uri, options), options);
@@ -68942,7 +68949,7 @@ var require_fast_uri = __commonJS({
     }
     var fastUri = {
       SCHEMES,
-      normalize: normalize3,
+      normalize: normalize4,
       resolve: resolve9,
       resolveComponent,
       equal,
@@ -73149,9 +73156,9 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - Do not add code comments unless the logic is genuinely complex and non-obvious.
 
 # Before Implementation
-- Focus on high-level design patterns and system organization
-- Identify architectural patterns and component relationships
-- Evaluate system structure and suggest architectural improvements
+- Read tests first \u2014 find existing test files for the module you're changing. They reveal expected behavior, edge cases, and the project's testing patterns.
+- Read neighboring files \u2014 understand naming conventions, error handling patterns, import style, and existing utilities before creating new ones.
+- Trace the call chain \u2014 follow how the code you're changing is called and what depends on it. Check interfaces, types, and consumers.
 - Focus on backward compatibility
 - Consider scalability, maintainability, and extensibility in your analysis
 
@@ -73176,6 +73183,20 @@ Before building or testing, determine the project's toolchain:
 - Read README for build/test instructions if the above are unclear
 - Common patterns: \`make build\`/\`make test\`, \`npm run build\`/\`npm test\`, \`cargo build\`/\`cargo test\`, \`go build ./...\`/\`go test ./...\`, \`python -m pytest\`
 
+# File Editing Rules
+You have access to the \`edit\`, \`create\`, and \`multi_edit\` tools for modifying files. You MUST use these tools for ALL code changes. They are purpose-built, atomic, and safe.
+
+DO NOT use sed, awk, echo/cat redirection, or heredocs to modify source code. These commands cause real damage in practice: truncated lines, duplicate code blocks, broken syntax. Every bad edit wastes iterations on fix-up commits.
+
+Use the right tool:
+1. To MODIFY existing code \u2192 \`edit\` tool (old_string \u2192 new_string, or start_line/end_line)
+2. To CREATE a new file \u2192 \`create\` tool
+3. To CHANGE multiple files at once \u2192 \`multi_edit\` tool
+4. To READ code \u2192 \`extract\` or \`search\` tools
+5. If \`edit\` fails with "file has not been read yet" \u2192 use \`extract\` with the EXACT same file path you will pass to \`edit\`. Relative vs absolute path mismatch causes this error. Use the same path format consistently. If it still fails, use bash \`cat\` to read the file, then use \`create\` to write the entire modified file. Do NOT fall back to sed.
+
+Bash is fine for: formatters (gofmt, prettier, black), build/test/lint commands, git operations, and read-only file inspection (cat, head, tail). sed/awk should ONLY be used for trivial non-code tasks (e.g., config file tweaks) where the replacement is a simple literal string swap.
+
 # During Implementation
 - Always create a new branch before making changes to the codebase.
 - Fix problems at the root cause, not with surface-level patches. Prefer general solutions over special cases.
@@ -73202,6 +73223,22 @@ Before committing or creating a PR, run through this checklist:
 
 Do NOT skip verification. Do NOT proceed to PR creation with a broken build or failing tests.
 
+# Output Integrity
+Your final output MUST accurately reflect what ACTUALLY happened. Do NOT fabricate, hallucinate, or report aspirational results.
+
+- Only report PR URLs you actually created or updated with \`gh pr create\` or \`git push\`. If you checked out an existing PR but did NOT push changes to it, do NOT claim you updated it.
+- Describe what you ACTUALLY DID, not what you planned or intended to do. If you ran out of iterations, say so. If tests failed, say so.
+- Only list files you actually modified AND committed.
+- If you could not complete the task \u2014 ran out of iterations, tests failed, build broken, push rejected \u2014 report the real reason honestly.
+
+NEVER claim success when:
+- You did not run \`git push\` successfully
+- Tests failed and you did not fix them
+- You hit the iteration limit before completing the work
+- You only analyzed/investigated but did not implement changes
+
+A false success report is WORSE than an honest failure \u2014 it misleads the user into thinking work is done when it is not.
+
 # GitHub Integration
 - Use the \`gh\` CLI for all GitHub operations: issues, pull requests, checks, releases.
 - To view issues or PRs: \`gh issue view <number>\`, \`gh pr view <number>\`.
@@ -101404,7 +101441,7 @@ var init_vercel = __esm({
         name: "search",
         description: searchDelegate ? searchDelegateDescription : searchDescription,
         inputSchema: searchSchema,
-        execute: async ({ query: searchQuery, path: path9, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage }) => {
+        execute: async ({ query: searchQuery, path: path9, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage, workingDirectory }) => {
           if (!exact && searchQuery) {
             const originalQuery = searchQuery;
             searchQuery = autoQuoteSearchTerms(searchQuery);
@@ -101413,18 +101450,19 @@ var init_vercel = __esm({
             }
           }
           const effectiveMaxTokens = paramMaxTokens || maxTokens;
+          const effectiveSearchCwd = workingDirectory || options.cwd || ".";
           let searchPaths;
           if (path9) {
-            searchPaths = parseAndResolvePaths(path9, options.cwd);
+            searchPaths = parseAndResolvePaths(path9, effectiveSearchCwd);
           }
           if (!searchPaths || searchPaths.length === 0) {
-            searchPaths = [options.cwd || "."];
+            searchPaths = [effectiveSearchCwd];
           }
           const searchPath = searchPaths.join(" ");
           const searchOptions = {
             query: searchQuery,
             path: searchPath,
-            cwd: options.cwd,
+            cwd: effectiveSearchCwd,
             // Working directory for resolving relative paths
             allowTests: allow_tests ?? true,
             exact,
@@ -101475,7 +101513,7 @@ var init_vercel = __esm({
             try {
               const result = maybeAnnotate(await runRawSearch());
               if (options.fileTracker && typeof result === "string") {
-                options.fileTracker.trackFilesFromOutput(result, options.cwd || ".").catch(() => {
+                options.fileTracker.trackFilesFromOutput(result, effectiveSearchCwd).catch(() => {
                 });
               }
               return result;
@@ -101528,7 +101566,7 @@ var init_vercel = __esm({
               }
               const fallbackResult = maybeAnnotate(await runRawSearch());
               if (options.fileTracker && typeof fallbackResult === "string") {
-                options.fileTracker.trackFilesFromOutput(fallbackResult, options.cwd || ".").catch(() => {
+                options.fileTracker.trackFilesFromOutput(fallbackResult, effectiveSearchCwd).catch(() => {
                 });
               }
               return fallbackResult;
@@ -101591,7 +101629,7 @@ var init_vercel = __esm({
             try {
               const fallbackResult2 = maybeAnnotate(await runRawSearch());
               if (options.fileTracker && typeof fallbackResult2 === "string") {
-                options.fileTracker.trackFilesFromOutput(fallbackResult2, options.cwd || ".").catch(() => {
+                options.fileTracker.trackFilesFromOutput(fallbackResult2, effectiveSearchCwd).catch(() => {
                 });
               }
               return fallbackResult2;
@@ -101645,9 +101683,9 @@ var init_vercel = __esm({
         name: "extract",
         description: extractDescription,
         inputSchema: extractSchema,
-        execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format }) => {
+        execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format, workingDirectory }) => {
           try {
-            const effectiveCwd = options.cwd || ".";
+            const effectiveCwd = workingDirectory || options.cwd || ".";
             if (debug) {
               if (targets) {
                 console.error(`Executing extract with targets: "${targets}", cwd: "${effectiveCwd}", context lines: ${context_lines || 10}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@probelabs/probe",
-  "version": "0.6.0-rc292",
+  "version": "0.6.0-rc293",
   "description": "Node.js wrapper for the probe code search tool",
   "main": "src/index.js",
   "module": "src/index.js",

package/src/agent/shared/prompts.js

@@ -90,9 +90,9 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - Do not add code comments unless the logic is genuinely complex and non-obvious.
 
 # Before Implementation
-- Focus on high-level design patterns and system organization
-- Identify architectural patterns and component relationships
-- Evaluate system structure and suggest architectural improvements
+- Read tests first — find existing test files for the module you're changing. They reveal expected behavior, edge cases, and the project's testing patterns.
+- Read neighboring files — understand naming conventions, error handling patterns, import style, and existing utilities before creating new ones.
+- Trace the call chain — follow how the code you're changing is called and what depends on it. Check interfaces, types, and consumers.
 - Focus on backward compatibility
 - Consider scalability, maintainability, and extensibility in your analysis
 
@@ -117,6 +117,20 @@ Before building or testing, determine the project's toolchain:
 - Read README for build/test instructions if the above are unclear
 - Common patterns: \`make build\`/\`make test\`, \`npm run build\`/\`npm test\`, \`cargo build\`/\`cargo test\`, \`go build ./...\`/\`go test ./...\`, \`python -m pytest\`
 
+# File Editing Rules
+You have access to the \`edit\`, \`create\`, and \`multi_edit\` tools for modifying files. You MUST use these tools for ALL code changes. They are purpose-built, atomic, and safe.
+
+DO NOT use sed, awk, echo/cat redirection, or heredocs to modify source code. These commands cause real damage in practice: truncated lines, duplicate code blocks, broken syntax. Every bad edit wastes iterations on fix-up commits.
+
+Use the right tool:
+1. To MODIFY existing code → \`edit\` tool (old_string → new_string, or start_line/end_line)
+2. To CREATE a new file → \`create\` tool
+3. To CHANGE multiple files at once → \`multi_edit\` tool
+4. To READ code → \`extract\` or \`search\` tools
+5. If \`edit\` fails with "file has not been read yet" → use \`extract\` with the EXACT same file path you will pass to \`edit\`. Relative vs absolute path mismatch causes this error. Use the same path format consistently. If it still fails, use bash \`cat\` to read the file, then use \`create\` to write the entire modified file. Do NOT fall back to sed.
+
+Bash is fine for: formatters (gofmt, prettier, black), build/test/lint commands, git operations, and read-only file inspection (cat, head, tail). sed/awk should ONLY be used for trivial non-code tasks (e.g., config file tweaks) where the replacement is a simple literal string swap.
+
 # During Implementation
 - Always create a new branch before making changes to the codebase.
 - Fix problems at the root cause, not with surface-level patches. Prefer general solutions over special cases.
@@ -143,6 +157,22 @@ Before committing or creating a PR, run through this checklist:
 
 Do NOT skip verification. Do NOT proceed to PR creation with a broken build or failing tests.
 
+# Output Integrity
+Your final output MUST accurately reflect what ACTUALLY happened. Do NOT fabricate, hallucinate, or report aspirational results.
+
+- Only report PR URLs you actually created or updated with \`gh pr create\` or \`git push\`. If you checked out an existing PR but did NOT push changes to it, do NOT claim you updated it.
+- Describe what you ACTUALLY DID, not what you planned or intended to do. If you ran out of iterations, say so. If tests failed, say so.
+- Only list files you actually modified AND committed.
+- If you could not complete the task — ran out of iterations, tests failed, build broken, push rejected — report the real reason honestly.
+
+NEVER claim success when:
+- You did not run \`git push\` successfully
+- Tests failed and you did not fix them
+- You hit the iteration limit before completing the work
+- You only analyzed/investigated but did not implement changes
+
+A false success report is WORSE than an honest failure — it misleads the user into thinking work is done when it is not.
+
 # GitHub Integration
 - Use the \`gh\` CLI for all GitHub operations: issues, pull requests, checks, releases.
 - To view issues or PRs: \`gh issue view <number>\`, \`gh pr view <number>\`.

package/src/tools/fileTracker.js

@@ -12,9 +12,22 @@
  */
 
 import { createHash } from 'crypto';
-import { resolve, isAbsolute } from 'path';
+import { resolve, isAbsolute, normalize } from 'path';
 import { findSymbol } from './symbolEdit.js';
 
+/**
+ * Normalize a file path for consistent storage and lookup.
+ * Resolves '.', '..', double slashes, and ensures absolute paths are canonical.
+ * Does NOT resolve symlinks (that would be expensive and might fail for non-existent files).
+ * @param {string} filePath - Path to normalize
+ * @returns {string} Normalized path
+ */
+function normalizePath(filePath) {
+  if (!filePath) return filePath;
+  // resolve() handles '.', '..', double slashes, and makes the path absolute
+  return resolve(filePath);
+}
+
 /**
  * Compute a SHA-256 content hash for a code block.
  * Normalizes trailing whitespace per line for robustness against editor formatting.
@@ -106,10 +119,11 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   markFileSeen(resolvedPath) {
-    this._seenFiles.add(resolvedPath);
-    this._textEditCounts.set(resolvedPath, 0);
+    const normalized = normalizePath(resolvedPath);
+    this._seenFiles.add(normalized);
+    this._textEditCounts.set(normalized, 0);
     if (this.debug) {
-      console.error(`[FileTracker] Marked as seen: ${resolvedPath}`);
+      console.error(`[FileTracker] Marked as seen: ${normalized}`);
     }
   }
 
@@ -119,7 +133,7 @@ export class FileTracker {
    * @returns {boolean}
    */
   isFileSeen(resolvedPath) {
-    return this._seenFiles.has(resolvedPath);
+    return this._seenFiles.has(normalizePath(resolvedPath));
   }
 
   /**
@@ -132,7 +146,7 @@ export class FileTracker {
    * @param {string} [source='extract'] - How the content was obtained
    */
   trackSymbolContent(resolvedPath, symbolName, code, startLine, endLine, source = 'extract') {
-    const key = `${resolvedPath}#${symbolName}`;
+    const key = `${normalizePath(resolvedPath)}#${symbolName}`;
     const contentHash = computeContentHash(code);
     this._contentRecords.set(key, {
       contentHash,
@@ -154,7 +168,7 @@ export class FileTracker {
    * @returns {Object|null} The stored record or null
    */
   getSymbolRecord(resolvedPath, symbolName) {
-    return this._contentRecords.get(`${resolvedPath}#${symbolName}`) || null;
+    return this._contentRecords.get(`${normalizePath(resolvedPath)}#${symbolName}`) || null;
   }
 
   /**
@@ -165,7 +179,7 @@ export class FileTracker {
    * @returns {{ok: boolean, reason?: string, message?: string}}
    */
   checkSymbolContent(resolvedPath, symbolName, currentCode) {
-    const key = `${resolvedPath}#${symbolName}`;
+    const key = `${normalizePath(resolvedPath)}#${symbolName}`;
     const record = this._contentRecords.get(key);
 
     if (!record) {
@@ -253,7 +267,7 @@ export class FileTracker {
    * @returns {{ok: boolean, reason?: string, message?: string}}
    */
   checkBeforeEdit(resolvedPath) {
-    if (!this._seenFiles.has(resolvedPath)) {
+    if (!this._seenFiles.has(normalizePath(resolvedPath))) {
       return {
         ok: false,
         reason: 'untracked',
@@ -269,8 +283,9 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   async trackFileAfterWrite(resolvedPath) {
-    this._seenFiles.add(resolvedPath);
-    this.invalidateFileRecords(resolvedPath);
+    const normalized = normalizePath(resolvedPath);
+    this._seenFiles.add(normalized);
+    this.invalidateFileRecords(normalized);
   }
 
   /**
@@ -279,10 +294,11 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   recordTextEdit(resolvedPath) {
-    const count = (this._textEditCounts.get(resolvedPath) || 0) + 1;
-    this._textEditCounts.set(resolvedPath, count);
+    const normalized = normalizePath(resolvedPath);
+    const count = (this._textEditCounts.get(normalized) || 0) + 1;
+    this._textEditCounts.set(normalized, count);
     if (this.debug) {
-      console.error(`[FileTracker] Text edit #${count} for ${resolvedPath}`);
+      console.error(`[FileTracker] Text edit #${count} for ${normalized}`);
     }
   }
 
@@ -292,7 +308,7 @@ export class FileTracker {
    * @returns {{ok: boolean, editCount?: number, message?: string}}
    */
   checkTextEditStaleness(resolvedPath) {
-    const count = this._textEditCounts.get(resolvedPath) || 0;
+    const count = this._textEditCounts.get(normalizePath(resolvedPath)) || 0;
     if (count >= this.maxConsecutiveTextEdits) {
       return {
         ok: false,
@@ -323,7 +339,7 @@ export class FileTracker {
    * @param {string} resolvedPath - Absolute path to the file
    */
   invalidateFileRecords(resolvedPath) {
-    const prefix = resolvedPath + '#';
+    const prefix = normalizePath(resolvedPath) + '#';
     for (const key of this._contentRecords.keys()) {
       if (key.startsWith(prefix)) {
         this._contentRecords.delete(key);
@@ -340,7 +356,7 @@ export class FileTracker {
    * @returns {boolean}
    */
   isTracked(resolvedPath) {
-    return this.isFileSeen(resolvedPath);
+    return this.isFileSeen(normalizePath(resolvedPath));
   }
 
   /**

package/src/tools/vercel.js

@@ -385,7 +385,7 @@ export const searchTool = (options = {}) => {
 			? searchDelegateDescription
 			: searchDescription,
 		inputSchema: searchSchema,
-		execute: async ({ query: searchQuery, path, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage }) => {
+		execute: async ({ query: searchQuery, path, allow_tests, exact, maxTokens: paramMaxTokens, language, session, nextPage, workingDirectory }) => {
 			// Auto-quote mixed-case and underscore terms to prevent unwanted stemming/splitting
 			// Skip when exact=true since that already preserves the literal string
 			if (!exact && searchQuery) {
@@ -399,15 +399,18 @@ export const searchTool = (options = {}) => {
 			// Use parameter maxTokens if provided, otherwise use the default
 			const effectiveMaxTokens = paramMaxTokens || maxTokens;
 
+			// Use workingDirectory (injected by _buildNativeTools at runtime) > cwd from config > fallback
+			const effectiveSearchCwd = workingDirectory || options.cwd || '.';
+
 			// Parse and resolve paths (supports comma-separated and relative paths)
 			let searchPaths;
 			if (path) {
-				searchPaths = parseAndResolvePaths(path, options.cwd);
+				searchPaths = parseAndResolvePaths(path, effectiveSearchCwd);
 			}
 
 			// Default to cwd or '.' if no paths provided
 			if (!searchPaths || searchPaths.length === 0) {
-				searchPaths = [options.cwd || '.'];
+				searchPaths = [effectiveSearchCwd];
 			}
 
 			// Join paths with space for CLI (probe search supports multiple paths)
@@ -416,7 +419,7 @@ export const searchTool = (options = {}) => {
 			const searchOptions = {
 				query: searchQuery,
 				path: searchPath,
-				cwd: options.cwd, // Working directory for resolving relative paths
+				cwd: effectiveSearchCwd, // Working directory for resolving relative paths
 				allowTests: allow_tests ?? true,
 				exact,
 				json: false,
@@ -473,7 +476,7 @@ export const searchTool = (options = {}) => {
 					const result = maybeAnnotate(await runRawSearch());
 					// Track files found in search results for staleness detection
 					if (options.fileTracker && typeof result === 'string') {
-						options.fileTracker.trackFilesFromOutput(result, options.cwd || '.').catch(() => {});
+						options.fileTracker.trackFilesFromOutput(result, effectiveSearchCwd).catch(() => {});
 					}
 					return result;
 				} catch (error) {
@@ -532,7 +535,7 @@ export const searchTool = (options = {}) => {
 					}
 					const fallbackResult = maybeAnnotate(await runRawSearch());
 					if (options.fileTracker && typeof fallbackResult === 'string') {
-						options.fileTracker.trackFilesFromOutput(fallbackResult, options.cwd || '.').catch(() => {});
+						options.fileTracker.trackFilesFromOutput(fallbackResult, effectiveSearchCwd).catch(() => {});
 					}
 					return fallbackResult;
 				}
@@ -614,7 +617,7 @@ export const searchTool = (options = {}) => {
 				try {
 					const fallbackResult2 = maybeAnnotate(await runRawSearch());
 					if (options.fileTracker && typeof fallbackResult2 === 'string') {
-						options.fileTracker.trackFilesFromOutput(fallbackResult2, options.cwd || '.').catch(() => {});
+						options.fileTracker.trackFilesFromOutput(fallbackResult2, effectiveSearchCwd).catch(() => {});
 					}
 					return fallbackResult2;
 				} catch (fallbackError) {
@@ -693,10 +696,10 @@ export const extractTool = (options = {}) => {
 		name: 'extract',
 		description: extractDescription,
 		inputSchema: extractSchema,
-		execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format }) => {
+		execute: async ({ targets, input_content, line, end_line, allow_tests, context_lines, format, workingDirectory }) => {
 			try {
-				// Use the cwd from config for working directory
-				const effectiveCwd = options.cwd || '.';
+				// Use workingDirectory (injected by _buildNativeTools at runtime) > cwd from config > fallback
+				const effectiveCwd = workingDirectory || options.cwd || '.';
 
 				if (debug) {
 					if (targets) {