@probelabs/probe 0.6.0-rc301 → 0.6.0-rc303

package/cjs/index.cjs

@@ -50610,6 +50610,10 @@ var init_parser4 = __esm({
 });
 
 // node_modules/@probelabs/maid/out/diagrams/sequence/semantics.js
+function isEscapedEntitySemicolon(image, semicolonIdx) {
+  const uptoSemicolon = image.slice(0, semicolonIdx + 1);
+  return /(?:#\d+|&#\d+|&[A-Za-z][A-Za-z0-9]+);$/.test(uptoSemicolon);
+}
 function analyzeSequence(_cst, _tokens) {
   const ctx = { tokens: _tokens };
   const v = new SequenceSemanticsVisitor(ctx);
@@ -50710,6 +50714,35 @@ function analyzeSequence(_cst, _tokens) {
     if (arrowIdx > 0) {
       const from = grabActorRef(arr, 0);
       const to = grabActorRef(arr, arrowIdx + 1);
+      const colonIdx = arr.findIndex((tk, idx) => idx > arrowIdx && tk.tokenType === Colon3);
+      if (colonIdx !== -1) {
+        let semicolonColumn = null;
+        for (let i = colonIdx + 1; i < arr.length && semicolonColumn == null; i++) {
+          const tk = arr[i];
+          const img = tk.image || "";
+          if (!img.includes(";"))
+            continue;
+          for (let j = 0; j < img.length; j++) {
+            if (img[j] !== ";")
+              continue;
+            if (isEscapedEntitySemicolon(img, j))
+              continue;
+            semicolonColumn = (tk.startColumn ?? 1) + j;
+            break;
+          }
+        }
+        if (semicolonColumn != null) {
+          errs.push({
+            line: ln,
+            column: semicolonColumn,
+            severity: "error",
+            code: "SE-MSG-SEMICOLON-UNESCAPED",
+            message: "Semicolons in sequence message text must be escaped as '#59;'.",
+            hint: "Replace ';' with '#59;' in the message text.",
+            length: 1
+          });
+        }
+      }
       if (from || to) {
         const plusTok = arr.find((tk) => tk.tokenType === Plus);
         const minusTok = arr.find((tk) => tk.tokenType === Minus);
@@ -52284,6 +52317,20 @@ function computeFixes(text, errors, level = "safe") {
     out = out.split(SENT_Q).join("&quot;");
     return out;
   }
+  function escapeUnescapedSemicolons(textPart) {
+    let out = "";
+    for (let i = 0; i < textPart.length; i++) {
+      const ch = textPart[i];
+      if (ch !== ";") {
+        out += ch;
+        continue;
+      }
+      const upto = textPart.slice(0, i + 1);
+      const isEntity = /(?:#\d+|&#\d+|&[A-Za-z][A-Za-z0-9]+);$/.test(upto);
+      out += isEntity ? ";" : "#59;";
+    }
+    return out;
+  }
   for (const e of errors) {
     const key = `${e.code}@${e.line}:${e.column}:${e.length ?? 1}`;
     if (seen.has(key))
@@ -53191,6 +53238,35 @@ function computeFixes(text, errors, level = "safe") {
       }
       continue;
     }
+    if (is("SE-MSG-SEMICOLON-UNESCAPED", e)) {
+      const lineText = lineTextAt(text, e.line);
+      const arrows = ["<<-->>", "<<->>", "-->>", "->>", "-->", "->", "--x", "-x", "--)", "-)"];
+      let ai = -1;
+      let alen = 0;
+      for (const a of arrows) {
+        const idx = lineText.indexOf(a);
+        if (idx !== -1 && (ai === -1 || idx < ai)) {
+          ai = idx;
+          alen = a.length;
+        }
+      }
+      if (ai !== -1) {
+        const colonIdx = lineText.indexOf(":", ai + alen);
+        if (colonIdx !== -1) {
+          const head2 = lineText.slice(0, colonIdx + 1);
+          const tail = lineText.slice(colonIdx + 1);
+          const fixedTail = escapeUnescapedSemicolons(tail);
+          if (fixedTail !== tail) {
+            edits.push({
+              start: { line: e.line, column: 1 },
+              end: { line: e.line, column: lineText.length + 1 },
+              newText: head2 + fixedTail
+            });
+          }
+        }
+      }
+      continue;
+    }
     if (is("SE-NOTE-MALFORMED", e)) {
       const lineText = lineTextAt(text, e.line);
       const mLR = /^(\s*)Note\s+(left|right)\s+of\s+(.+?)\s+(.+)$/.exec(lineText);
@@ -93984,10 +94060,9 @@ function parseSimpleCommand(command) {
     // Command substitution $()
     /`/,
     // Command substitution ``
-    />/,
-    // Redirection >
-    /</,
-    // Redirection <
+    // Note: > and < (redirection) are intentionally NOT in this list.
+    // They are not command separators — they redirect I/O on a single command.
+    // The base command is still checked against allow/deny lists.
     /\*\*/,
     // Glob patterns (potentially dangerous)
     /^\s*\{.*,.*\}|\{.*\.\.\.*\}/
@@ -94090,12 +94165,8 @@ function isComplexPattern(pattern) {
     // Background execution
     /\$\(/,
     // Command substitution $()
-    /`/,
+    /`/
     // Command substitution ``
-    />/,
-    // Redirection >
-    /</
-    // Redirection <
   ];
   return operatorPatterns.some((p) => p.test(pattern));
 }
@@ -94204,12 +94275,14 @@ var init_bashPermissions = __esm({
        * @param {string[]} [config.deny] - Additional deny patterns (always win)
        * @param {boolean} [config.disableDefaultAllow] - Disable default allow list
        * @param {boolean} [config.disableDefaultDeny] - Disable default deny list
+       * @param {boolean} [config.allowEdit] - Whether file editing is allowed (controls output redirection)
        * @param {boolean} [config.debug] - Enable debug logging
        * @param {Object} [config.tracer] - Optional tracer for telemetry
        */
       constructor(config2 = {}) {
         this.debug = config2.debug || false;
         this.tracer = config2.tracer || null;
+        this.allowEdit = config2.allowEdit || false;
         this.defaultAllowPatterns = config2.disableDefaultAllow ? [] : [...DEFAULT_ALLOW_PATTERNS];
         this.customAllowPatterns = config2.allow && Array.isArray(config2.allow) ? [...config2.allow] : [];
         this.allowPatterns = [...this.defaultAllowPatterns, ...this.customAllowPatterns];
@@ -94295,6 +94368,24 @@ var init_bashPermissions = __esm({
           console.log(`[BashPermissions] Checking simple command: "${command}"`);
           console.log(`[BashPermissions] Parsed: ${parsed.command} with args: [${parsed.args.join(", ")}]`);
         }
+        if (!this.allowEdit && parsed.args.some((arg) => arg === ">" || arg === ">>")) {
+          const result2 = {
+            allowed: false,
+            reason: "Output redirection (> or >>) requires edit permissions (allowEdit)",
+            command,
+            parsed
+          };
+          if (this.debug) {
+            console.log(`[BashPermissions] DENIED - output redirection without allowEdit`);
+          }
+          this.recordBashEvent("permission.denied", {
+            command,
+            parsedCommand: parsed.command,
+            reason: "output_redirection_without_allow_edit",
+            isComplex: false
+          });
+          return result2;
+        }
         if (matchesAnyPattern(parsed, this.customDenyPatterns)) {
           const matchedPatterns = this.customDenyPatterns.filter((pattern) => matchesPattern(parsed, pattern));
           if (this.debug) {
@@ -94564,6 +94655,15 @@ var init_bashPermissions = __esm({
               deniedReason = parsed.error || "Component contains nested complex constructs";
               break;
             }
+            if (!this.allowEdit && parsed.args && parsed.args.some((arg) => arg === ">" || arg === ">>")) {
+              if (this.debug) {
+                console.log(`[BashPermissions] Component "${component}" has output redirection without allowEdit`);
+              }
+              allAllowed = false;
+              deniedComponent = component;
+              deniedReason = "Output redirection (> or >>) requires edit permissions (allowEdit)";
+              break;
+            }
             if (matchesAnyPattern(parsed, this.customDenyPatterns)) {
               if (this.debug) {
                 console.log(`[BashPermissions] Component "${component}" matches custom deny pattern`);
@@ -95093,6 +95193,7 @@ var init_bash = __esm({
         debug = false,
         cwd,
         allowedFolders = [],
+        allowEdit = false,
         workspaceRoot: providedWorkspaceRoot,
         tracer = null
       } = options;
@@ -95102,6 +95203,7 @@ var init_bash = __esm({
         deny: bashConfig.deny,
         disableDefaultAllow: bashConfig.disableDefaultAllow,
         disableDefaultDeny: bashConfig.disableDefaultDeny,
+        allowEdit,
         debug,
         tracer
       });
@@ -100231,19 +100333,23 @@ ${activeToolsDesc}
 - Time budget remaining: ${Math.round(remainingBudgetMs / 6e4)} minutes
 - Max per extension: ${maxPerReqMin} minutes
 
-Decide whether the agent should get more time. EXTEND if:
+Your DEFAULT should be to EXTEND. The agent is working on a task and deserves time to finish. Only decline if you see clear evidence of a problem.
+
+EXTEND (this should be your default) if:
 - Tools are actively running (especially delegates or complex analysis) \u2014 they need time to finish
-- The agent is making clear progress on a complex task
+- The agent is making progress, even if slowly \u2014 complex tasks take time
 - New information is being gathered that will improve the final answer
+- The agent is iterating on an approach (trying different searches, refining queries) \u2014 this is normal problem-solving, not a loop
+- There is remaining budget and the task is not yet complete
+- When in doubt, extend \u2014 it's better to give the agent a chance than to cut it off prematurely
 
-DO NOT EXTEND if:
-- The agent appears stuck in a loop (repeating the same tool calls or getting the same errors)
-- The conversation shows the agent retrying failed operations without changing approach
-- The agent has enough information to answer but keeps searching for more
-- Tool calls are returning empty or error results repeatedly
-- The agent is doing redundant work (searching for things it already found)
+DO NOT EXTEND only if you see CLEAR evidence of:
+- The agent is stuck in an obvious loop \u2014 repeating the EXACT same tool calls with the EXACT same arguments and getting the same errors back-to-back (3+ times)
+- The agent is retrying a fundamentally broken operation without changing its approach at all
+- Tool calls are consistently returning errors or empty results AND the agent is not adapting
+- The conversation clearly shows the agent has all the information it needs and is just making redundant calls
 
-A stuck agent will not recover with more time \u2014 it will just burn the budget. Better to force it to answer with what it has.
+IMPORTANT: Iterating, refining, or trying variations is NOT the same as being stuck in a loop. A loop means identical repeated calls with no variation. Be generous with time \u2014 a slightly longer response time is much better than a prematurely cut-off incomplete answer.
 
 Respond with ONLY valid JSON (no markdown, no explanation):
 {"extend": true, "minutes": <1-${maxPerReqMin}>, "reason": "your reason here"}
@@ -100287,38 +100393,38 @@ or
               const decision = JSON.parse(jsonStr);
               if (decision.extend && decision.minutes > 0) {
                 const requestedMs = Math.min(decision.minutes, maxPerReqMin) * 6e4;
-                const grantedMs = Math.min(requestedMs, remainingBudgetMs, negotiatedTimeoutState.maxPerRequestMs);
-                const grantedMin = Math.round(grantedMs / 6e4 * 10) / 10;
+                const grantedMs2 = Math.min(requestedMs, remainingBudgetMs, negotiatedTimeoutState.maxPerRequestMs);
+                const grantedMin2 = Math.round(grantedMs2 / 6e4 * 10) / 10;
                 negotiatedTimeoutState.extensionsUsed++;
-                negotiatedTimeoutState.totalExtraTimeMs += grantedMs;
-                negotiatedTimeoutState.extensionMessage = `\u23F0 Time limit was reached. The timeout observer granted ${grantedMin} more minute(s) (reason: ${decision.reason || "work in progress"}). Extensions remaining: ${negotiatedTimeoutState.maxRequests - negotiatedTimeoutState.extensionsUsed}. Continue your work efficiently.`;
+                negotiatedTimeoutState.totalExtraTimeMs += grantedMs2;
+                negotiatedTimeoutState.extensionMessage = `\u23F0 Time limit was reached. The timeout observer granted ${grantedMin2} more minute(s) (reason: ${decision.reason || "work in progress"}). Extensions remaining: ${negotiatedTimeoutState.maxRequests - negotiatedTimeoutState.extensionsUsed}. Continue your work efficiently.`;
                 negotiatedTimeoutState.softTimeoutId = setTimeout(() => {
                   runTimeoutObserver();
-                }, grantedMs);
+                }, grantedMs2);
                 if (this.debug) {
-                  console.log(`[DEBUG] Timeout observer: granted ${grantedMin} min (reason: ${decision.reason}). Extensions: ${negotiatedTimeoutState.extensionsUsed}/${negotiatedTimeoutState.maxRequests}`);
+                  console.log(`[DEBUG] Timeout observer: granted ${grantedMin2} min (reason: ${decision.reason}). Extensions: ${negotiatedTimeoutState.extensionsUsed}/${negotiatedTimeoutState.maxRequests}`);
                 }
                 if (this.tracer) {
                   this.tracer.addEvent("negotiated_timeout.observer_extended", {
                     decision_reason: decision.reason,
                     requested_minutes: decision.minutes,
-                    granted_ms: grantedMs,
-                    granted_min: grantedMin,
+                    granted_ms: grantedMs2,
+                    granted_min: grantedMin2,
                     extensions_used: negotiatedTimeoutState.extensionsUsed,
                     max_requests: negotiatedTimeoutState.maxRequests,
                     total_extra_time_ms: negotiatedTimeoutState.totalExtraTimeMs,
-                    budget_remaining_ms: remainingBudgetMs - grantedMs,
+                    budget_remaining_ms: remainingBudgetMs - grantedMs2,
                     active_tools: activeToolsList.map((t) => t.name),
                     active_tools_count: activeToolsList.length
                   });
                 }
                 this.events.emit("timeout.extended", {
-                  grantedMs,
+                  grantedMs: grantedMs2,
                   reason: decision.reason || "work in progress",
                   extensionsUsed: negotiatedTimeoutState.extensionsUsed,
                   extensionsRemaining: negotiatedTimeoutState.maxRequests - negotiatedTimeoutState.extensionsUsed,
                   totalExtraTimeMs: negotiatedTimeoutState.totalExtraTimeMs,
-                  budgetRemainingMs: remainingBudgetMs - grantedMs
+                  budgetRemainingMs: remainingBudgetMs - grantedMs2
                 });
               } else {
                 if (this.debug) {
@@ -100340,6 +100446,18 @@ or
                 });
                 await this._initiateGracefulStop(gracefulTimeoutState, `observer declined: ${decision.reason}`);
               }
+              return {
+                decision: decision.extend ? "extended" : "declined",
+                reason: decision.reason || "",
+                ...decision.extend ? {
+                  granted_ms: grantedMs,
+                  granted_min: grantedMin,
+                  budget_remaining_ms: remainingBudgetMs - grantedMs
+                } : {},
+                extensions_used: negotiatedTimeoutState.extensionsUsed,
+                max_requests: negotiatedTimeoutState.maxRequests,
+                total_extra_time_ms: negotiatedTimeoutState.totalExtraTimeMs
+              };
             };
             try {
               if (this.tracer) {
@@ -100348,6 +100466,23 @@ or
                   "timeout.extensions_used": negotiatedTimeoutState.extensionsUsed,
                   "timeout.active_tools_count": activeToolsList.length,
                   "timeout.remaining_budget_ms": remainingBudgetMs
+                }, (span, result) => {
+                  if (result) {
+                    span.setAttributes({
+                      "observer.decision": result.decision,
+                      "observer.reason": result.reason,
+                      "observer.extensions_used": result.extensions_used,
+                      "observer.max_requests": result.max_requests,
+                      "observer.total_extra_time_ms": result.total_extra_time_ms
+                    });
+                    if (result.decision === "extended") {
+                      span.setAttributes({
+                        "observer.granted_ms": result.granted_ms,
+                        "observer.granted_min": result.granted_min,
+                        "observer.budget_remaining_ms": result.budget_remaining_ms
+                      });
+                    }
+                  }
                 });
               } else {
                 await observerFn();
@@ -100455,7 +100590,13 @@ or
                       }
                       return {
                         toolChoice: "none",
-                        userMessage: `\u26A0\uFE0F TIME LIMIT REACHED. You are running out of time. You have ${remaining} step(s) remaining. Provide your BEST answer NOW using the information you have already gathered. Do NOT call any more tools. Summarize your findings and respond completely. If something was not completed, honestly state what was not done and provide any partial results or recommendations you can offer.`
+                        userMessage: `\u26A0\uFE0F TIME BUDGET EXHAUSTED. Your allocated time for this task has run out. You have ${remaining} step(s) remaining to provide your answer.
+
+IMPORTANT: This is a time budget constraint, NOT a system shutdown or error. The system is working perfectly \u2014 you simply used all your allocated time.
+
+Do NOT say things like "the system is shutting down" or "try again later" \u2014 the user submitted a request and is waiting for YOUR answer right now.
+
+Provide your BEST answer NOW using the information you have already gathered. Do NOT call any more tools. Summarize your findings and respond completely. If something was not completed, honestly state what was not done and provide any partial results or recommendations you can offer.`
                       };
                     }
                     if (this.debug) {
@@ -100883,7 +101024,9 @@ Respond with ONLY valid JSON \u2014 no markdown, no explanation, no text outside
                     } catch {
                     }
                   }
-                  const summaryPrompt = `Your operation was interrupted by a timeout observer because the time limit was reached. Some of your tool calls were cancelled mid-execution.
+                  const summaryPrompt = `Your allocated time budget for this task has been exhausted. Some of your tool calls were cancelled mid-execution because the timeout observer determined the time limit was reached.
+
+IMPORTANT: This is a time budget constraint, NOT a system shutdown or error. The system is working perfectly \u2014 you simply used all your allocated time. Do NOT say things like "the system is shutting down" or "try again later." The user is waiting for your answer RIGHT NOW.
 
 Please provide a DETAILED summary of:
 1. What you were asked to do (the original task)
@@ -100918,7 +101061,14 @@ Be thorough \u2014 this is the user's only response. Include all useful informat
                   let summaryText;
                   if (this.tracer) {
                     summaryText = await this.tracer.withSpan("negotiated_timeout.abort_summary", summaryFn, {
-                      "summary.conversation_messages": currentMessages.length
+                      "summary.conversation_messages": currentMessages.length,
+                      "observer.was_timeout": true
+                    }, (span, result) => {
+                      if (result) {
+                        span.setAttributes({
+                          "observer.summary_length": result.length
+                        });
+                      }
                     });
                   } else {
                     summaryText = await summaryFn();
@@ -102930,8 +103080,23 @@ function buildSearchDelegateTask({ searchQuery, searchPath, exact, language, all
     '- Searching "getUserData" ALREADY matches "get", "user", "data" and their variations.',
     "- NEVER repeat the same search query \u2014 you will get the same results. Changing the path does NOT change this.",
     "- NEVER search trivial variations of the same keyword (e.g., AllowedIPs then allowedIps then allowed_ips). This is wasteful \u2014 probe handles it.",
-    "- If a search returns no results, the term likely does not exist. Try a genuinely DIFFERENT keyword or concept, not a variation.",
-    "- If 2-3 searches return no results for a concept, STOP searching for it and move on. Do NOT keep retrying.",
+    "",
+    "When a search returns no results:",
+    '- If you searched a SUBFOLDER (e.g., path="gateway/"), the term might exist elsewhere.',
+    "  Try searching from the workspace root (omit the path parameter) or a different directory.",
+    "  But do NOT retry the same subfolder with different quoting \u2014 that will not help.",
+    "- If you searched the WORKSPACE ROOT and got no results, the term does not exist in this codebase.",
+    '  Changing quotes, adding "func " prefix, or switching to method syntax will NOT help.',
+    "- These are ALL the same failed search, NOT different searches:",
+    '    search("func ctxGetData") \u2192 no results',
+    '    search("ctxGetData")      \u2192 no results  \u2190 WASTED, same concept, different quoting',
+    "    search(ctxGetData)         \u2192 no results  \u2190 WASTED, same concept, no quotes",
+    '    search("ctx.GetData")      \u2192 no results  \u2190 WASTED, method syntax of same concept',
+    '  After the FIRST "no results" at a given scope, either widen the search path or try',
+    "  a fundamentally different approach: search for a broader concept, use listFiles",
+    "  to discover actual function names, or extract a known file to read real code.",
+    "- If 2 searches return no results for a concept (across different scopes), the code likely",
+    "  uses different naming than you expect \u2014 discover the real names via extract or listFiles.",
     "",
     "When to use exact=true:",
     "- Use exact=true when searching for a KNOWN symbol name (function, type, variable, struct).",
@@ -102984,6 +103149,21 @@ function buildSearchDelegateTask({ searchQuery, searchPath, exact, language, all
     '  \u2192 search "ForwardMessage" \u2192 search "ForwardMessage" \u2192 search "ForwardMessage" (WRONG: repeating the exact same query)',
     '  \u2192 search "authentication" \u2192 wait \u2192 search "session management" \u2192 wait (WRONG: these are independent, run them in parallel)',
     "",
+    "  WORST pattern \u2014 retrying a non-existent function with quote/syntax variations (this wastes 30 minutes):",
+    '  \u2192 search "func ctxGetData" \u2192 no results',
+    '  \u2192 search "ctxGetData" \u2192 no results          \u2190 WRONG: same term without "func" prefix',
+    '  \u2192 search "ctx.GetData" \u2192 no results          \u2190 WRONG: method syntax of same concept',
+    '  \u2192 search "ctx.SetData" \u2192 no results          \u2190 WRONG: Set variant of same concept',
+    "  \u2192 search ctxGetData \u2192 no results             \u2190 WRONG: unquoted version of same term",
+    "  \u2192 extract api.go \u2192 extract api.go \u2192 extract api.go (8 times!) \u2190 WRONG: re-reading same file",
+    '  FIX: After "func ctxGetData" returns no results in gateway/:',
+    "  Option A: Widen scope \u2014 search from the workspace root (omit path) in case the",
+    "    function is defined in a different package (e.g., apidef/, user/, config/).",
+    "  Option B: Discover real names \u2014 extract a file you KNOW uses context (e.g., a",
+    "    middleware file) and READ what functions it actually calls.",
+    "  Option C: Browse \u2014 use listFiles to see what files exist and extract the relevant ones.",
+    "  NEVER: retry the same concept with different quoting in the same directory.",
+    "",
     "Keyword tips:",
     "- Common programming keywords are filtered as stopwords when unquoted: function, class, return, new, struct, impl, var, let, const, etc.",
     '- Avoid searching for these alone \u2014 combine with a specific term (e.g., "middleware function" is fine, "function" alone is too generic).',
@@ -103022,7 +103202,7 @@ function buildSearchDelegateTask({ searchQuery, searchPath, exact, language, all
     "   - Type references and imports \u2192 include type definitions.",
     "   - Registered handlers/middleware \u2192 include all registered items.",
     "6. If a search returns results, use extract to verify relevance. Run multiple extracts in parallel too.",
-    "7. If a search returns NO results, the term does not exist. Do NOT retry with variations. Move on.",
+    "7. If a search returns NO results: widen the path scope if you searched a subfolder, or move on. Do NOT retry with quote/syntax variations \u2014 they search the same index.",
     "8. Once you have enough targets (typically 5-15), output your final JSON answer immediately.",
     "",
     `Query: ${searchQuery}`,
@@ -103082,7 +103262,14 @@ var init_vercel = __esm({
       const previousSearches = /* @__PURE__ */ new Map();
       const dupBlockCounts = /* @__PURE__ */ new Map();
       const paginationCounts = /* @__PURE__ */ new Map();
+      let consecutiveNoResults = 0;
+      const MAX_CONSECUTIVE_NO_RESULTS = 4;
+      const failedConcepts = /* @__PURE__ */ new Map();
       const MAX_PAGES_PER_QUERY = 3;
+      function normalizeQueryConcept(query2) {
+        if (!query2) return "";
+        return query2.replace(/^["']|["']$/g, "").replace(/\./g, "").replace(/[_\-\s]+/g, "").toLowerCase().trim();
+      }
       return (0, import_ai5.tool)({
         name: "search",
         description: searchDelegate ? searchDelegateDescription : searchDescription,
@@ -103151,6 +103338,41 @@ var init_vercel = __esm({
               }
               previousSearches.set(searchKey, { hadResults: false });
               paginationCounts.set(searchKey, 0);
+              const normalizedKey = `${searchPath}::${normalizeQueryConcept(searchQuery)}`;
+              if (failedConcepts.has(normalizedKey) && failedConcepts.get(normalizedKey) >= 2) {
+                const conceptCount = failedConcepts.get(normalizedKey) + 1;
+                failedConcepts.set(normalizedKey, conceptCount);
+                if (debug) {
+                  console.error(`[CONCEPT-DEDUP] Blocked variation of failed concept (${conceptCount}x): "${searchQuery}" normalized to "${normalizeQueryConcept(searchQuery)}"`);
+                }
+                const isSubfolder = path9 && path9 !== effectiveSearchCwd && path9 !== ".";
+                const scopeHint = isSubfolder ? `
+- Try searching from the workspace root (omit the path parameter) \u2014 the term may exist in a different directory` : `
+- The term does not exist in this codebase at any path`;
+                return `CONCEPT ALREADY FAILED (${conceptCount} variations tried). You already searched for "${normalizeQueryConcept(searchQuery)}" with different quoting/syntax in this path and got NO results each time. Changing quotes, adding "func" prefix, or switching to method syntax will NOT change the results.
+
+Change your strategy:${scopeHint}
+- Use extract on a file you ALREADY found to read actual code and discover real function/type names
+- Use listFiles to browse directories and find what functions actually exist
+- Search for a BROADER concept (e.g., instead of "ctxGetData", try "context" or "middleware data access")
+- If you have enough information from prior searches, provide your final answer NOW`;
+              }
+              if (consecutiveNoResults >= MAX_CONSECUTIVE_NO_RESULTS) {
+                if (debug) {
+                  console.error(`[CIRCUIT-BREAKER] ${consecutiveNoResults} consecutive no-result searches, blocking: "${searchQuery}"`);
+                }
+                const isSubfolderCB = path9 && path9 !== effectiveSearchCwd && path9 !== ".";
+                const cbScopeHint = isSubfolderCB ? `
+- You have been searching in "${path9}" \u2014 try searching from the workspace root or a different directory` : "";
+                return `CIRCUIT BREAKER: Your last ${consecutiveNoResults} searches ALL returned no results. You appear to be guessing function/type names that don't match what's actually in the code.
+
+Change your approach:${cbScopeHint}
+1. Use extract on files you already found \u2014 read the actual code to discover real function names
+2. Use listFiles to browse directories and see what files/functions actually exist
+3. If you found some results earlier, those are likely sufficient \u2014 provide your final answer
+
+Retrying search query variations will not help. Discover real names from real code instead.`;
+              }
             } else {
               const pageCount = (paginationCounts.get(searchKey) || 0) + 1;
               paginationCounts.set(searchKey, pageCount);
@@ -103164,10 +103386,24 @@ var init_vercel = __esm({
             try {
               const result = maybeAnnotate(await runRawSearch());
               if (typeof result === "string" && result.includes("No results found")) {
+                consecutiveNoResults++;
+                const normalizedKey = `${searchPath}::${normalizeQueryConcept(searchQuery)}`;
+                failedConcepts.set(normalizedKey, (failedConcepts.get(normalizedKey) || 0) + 1);
+                if (debug) {
+                  console.error(`[NO-RESULTS] consecutiveNoResults=${consecutiveNoResults}, concept "${normalizeQueryConcept(searchQuery)}" failed ${failedConcepts.get(normalizedKey)}x`);
+                }
                 if (/^[A-Z]+-\d+$/.test(searchQuery.trim()) || /^[A-Z]+-\d+$/.test(searchQuery.replace(/"/g, "").trim())) {
                   return result + "\n\n\u26A0\uFE0F Your query looks like a ticket/issue ID (e.g., JIRA-1234). Ticket IDs are rarely present in source code. Search for the technical concepts described in the ticket instead (e.g., function names, error messages, variable names).";
                 }
+                if (consecutiveNoResults >= MAX_CONSECUTIVE_NO_RESULTS - 1) {
+                  const isSubfolderWarn = path9 && path9 !== effectiveSearchCwd && path9 !== ".";
+                  const warnScopeHint = isSubfolderWarn ? ` You are searching in "${path9}" \u2014 consider searching from the workspace root or a different directory.` : "";
+                  return result + `
+
+\u26A0\uFE0F WARNING: ${consecutiveNoResults} consecutive searches returned no results.${warnScopeHint} Before your next action: use extract on a file you already found to read actual code, or use listFiles to discover what functions really exist. One more failed search will trigger the circuit breaker.`;
+                }
               } else if (typeof result === "string") {
+                consecutiveNoResults = 0;
                 const entry = previousSearches.get(searchKey);
                 if (entry) entry.hadResults = true;
               }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@probelabs/probe",
-  "version": "0.6.0-rc301",
+  "version": "0.6.0-rc303",
   "description": "Node.js wrapper for the probe code search tool",
   "main": "src/index.js",
   "module": "src/index.js",
@@ -81,7 +81,7 @@
     "@anthropic-ai/claude-agent-sdk": "^0.1.46",
     "@modelcontextprotocol/sdk": "^1.0.0",
     "@nyariv/sandboxjs": "github:probelabs/SandboxJS",
-    "@probelabs/maid": "^0.0.28",
+    "@probelabs/maid": "^0.0.29",
     "acorn": "^8.15.0",
     "acorn-walk": "^8.3.4",
     "adm-zip": "^0.5.16",

package/src/agent/ProbeAgent.js

@@ -3746,19 +3746,23 @@ ${activeToolsDesc}
 - Time budget remaining: ${Math.round(remainingBudgetMs / 60000)} minutes
 - Max per extension: ${maxPerReqMin} minutes
 
-Decide whether the agent should get more time. EXTEND if:
+Your DEFAULT should be to EXTEND. The agent is working on a task and deserves time to finish. Only decline if you see clear evidence of a problem.
+
+EXTEND (this should be your default) if:
 - Tools are actively running (especially delegates or complex analysis) — they need time to finish
-- The agent is making clear progress on a complex task
+- The agent is making progress, even if slowly — complex tasks take time
 - New information is being gathered that will improve the final answer
+- The agent is iterating on an approach (trying different searches, refining queries) — this is normal problem-solving, not a loop
+- There is remaining budget and the task is not yet complete
+- When in doubt, extend — it's better to give the agent a chance than to cut it off prematurely
 
-DO NOT EXTEND if:
-- The agent appears stuck in a loop (repeating the same tool calls or getting the same errors)
-- The conversation shows the agent retrying failed operations without changing approach
-- The agent has enough information to answer but keeps searching for more
-- Tool calls are returning empty or error results repeatedly
-- The agent is doing redundant work (searching for things it already found)
+DO NOT EXTEND only if you see CLEAR evidence of:
+- The agent is stuck in an obvious loop — repeating the EXACT same tool calls with the EXACT same arguments and getting the same errors back-to-back (3+ times)
+- The agent is retrying a fundamentally broken operation without changing its approach at all
+- Tool calls are consistently returning errors or empty results AND the agent is not adapting
+- The conversation clearly shows the agent has all the information it needs and is just making redundant calls
 
-A stuck agent will not recover with more time — it will just burn the budget. Better to force it to answer with what it has.
+IMPORTANT: Iterating, refining, or trying variations is NOT the same as being stuck in a loop. A loop means identical repeated calls with no variation. Be generous with time — a slightly longer response time is much better than a prematurely cut-off incomplete answer.
 
 Respond with ONLY valid JSON (no markdown, no explanation):
 {"extend": true, "minutes": <1-${maxPerReqMin}>, "reason": "your reason here"}
@@ -3885,6 +3889,20 @@ or
 
             await this._initiateGracefulStop(gracefulTimeoutState, `observer declined: ${decision.reason}`);
           }
+
+          // Return decision data for span enrichment
+          return {
+            decision: decision.extend ? 'extended' : 'declined',
+            reason: decision.reason || '',
+            ...(decision.extend ? {
+              granted_ms: grantedMs,
+              granted_min: grantedMin,
+              budget_remaining_ms: remainingBudgetMs - grantedMs,
+            } : {}),
+            extensions_used: negotiatedTimeoutState.extensionsUsed,
+            max_requests: negotiatedTimeoutState.maxRequests,
+            total_extra_time_ms: negotiatedTimeoutState.totalExtraTimeMs,
+          };
         };
 
         try {
@@ -3894,6 +3912,23 @@ or
               'timeout.extensions_used': negotiatedTimeoutState.extensionsUsed,
               'timeout.active_tools_count': activeToolsList.length,
               'timeout.remaining_budget_ms': remainingBudgetMs,
+            }, (span, result) => {
+              if (result) {
+                span.setAttributes({
+                  'observer.decision': result.decision,
+                  'observer.reason': result.reason,
+                  'observer.extensions_used': result.extensions_used,
+                  'observer.max_requests': result.max_requests,
+                  'observer.total_extra_time_ms': result.total_extra_time_ms,
+                });
+                if (result.decision === 'extended') {
+                  span.setAttributes({
+                    'observer.granted_ms': result.granted_ms,
+                    'observer.granted_min': result.granted_min,
+                    'observer.budget_remaining_ms': result.budget_remaining_ms,
+                  });
+                }
+              }
             });
           } else {
             await observerFn();
@@ -4033,7 +4068,7 @@ or
                   }
                   return {
                     toolChoice: 'none',
-                    userMessage: `⚠️ TIME LIMIT REACHED. You are running out of time. You have ${remaining} step(s) remaining. Provide your BEST answer NOW using the information you have already gathered. Do NOT call any more tools. Summarize your findings and respond completely. If something was not completed, honestly state what was not done and provide any partial results or recommendations you can offer.`
+                    userMessage: `⚠️ TIME BUDGET EXHAUSTED. Your allocated time for this task has run out. You have ${remaining} step(s) remaining to provide your answer.\n\nIMPORTANT: This is a time budget constraint, NOT a system shutdown or error. The system is working perfectly — you simply used all your allocated time.\n\nDo NOT say things like "the system is shutting down" or "try again later" — the user submitted a request and is waiting for YOUR answer right now.\n\nProvide your BEST answer NOW using the information you have already gathered. Do NOT call any more tools. Summarize your findings and respond completely. If something was not completed, honestly state what was not done and provide any partial results or recommendations you can offer.`
                   };
                 }
 
@@ -4571,8 +4606,10 @@ Double-check your response based on the criteria above. If everything looks good
                 } catch {}
               }
 
-              const summaryPrompt = `Your operation was interrupted by a timeout observer because the time limit was reached. ` +
-                `Some of your tool calls were cancelled mid-execution.\n\n` +
+              const summaryPrompt = `Your allocated time budget for this task has been exhausted. ` +
+                `Some of your tool calls were cancelled mid-execution because the timeout observer determined the time limit was reached.\n\n` +
+                `IMPORTANT: This is a time budget constraint, NOT a system shutdown or error. The system is working perfectly — you simply used all your allocated time. ` +
+                `Do NOT say things like "the system is shutting down" or "try again later." The user is waiting for your answer RIGHT NOW.\n\n` +
                 `Please provide a DETAILED summary of:\n` +
                 `1. What you were asked to do (the original task)\n` +
                 `2. What you accomplished — include ALL findings, code snippets, data, and conclusions you gathered\n` +
@@ -4615,6 +4652,13 @@ Double-check your response based on the criteria above. If everything looks good
               if (this.tracer) {
                 summaryText = await this.tracer.withSpan('negotiated_timeout.abort_summary', summaryFn, {
                   'summary.conversation_messages': currentMessages.length,
+                  'observer.was_timeout': true,
+                }, (span, result) => {
+                  if (result) {
+                    span.setAttributes({
+                      'observer.summary_length': result.length,
+                    });
+                  }
                 });
               } else {
                 summaryText = await summaryFn();