vigthoria-cli 1.6.48 → 1.6.50

package/dist/commands/chat.js

@@ -1192,26 +1192,44 @@ class ChatCommand {
                     preview: visibleText.slice(0, 300),
                 });
                 if (toolCalls.length === 0) {
-                    // Phase 5: Quality gate — if the agent tries to conclude on the first
-                    // turn without any discovery, push it to gather evidence first.
-                    // Applies to diagnostic prompts AND any direct-prompt agent call where
-                    // the model failed to invoke tools (prevents truncated output).
-                    // Also catches policy-acknowledgement responses ("I will follow…",
-                    // "I understand the instructions…") which are never useful.
+                    // Phase 5: Quality gate — reject non-answers and push the model
+                    // to gather real evidence.
                     const isPolicyAck = /^(i will follow|i understand|i('ll| will) adhere|understood[.,!]|sure[.,!]|i('ll| will) use the tools|i('ll| will) proceed|let me know|provide your|waiting for)/i.test(visibleText.trim());
-                    if (turn === 0 && this.agentToolEvidence.discovery === 0 && (this.isDiagnosticPrompt(prompt) || this.directPromptMode || isPolicyAck)) {
-                        // Remove the useless acknowledgement from history so it
-                        // doesn't anchor the model in the same pattern.
-                        if (isPolicyAck) {
+                    // Detect if the response is effectively empty or just a follow-up
+                    // question after we strip tool output echoes.
+                    const sanitized = this.sanitizeDirectModeOutput(visibleText.trim());
+                    const isFollowUp = this.isDirectModeFollowUpQuestion(sanitized);
+                    const isEmptyAfterSanitize = !sanitized || sanitized.length < 10;
+                    // Detect resignation: model gives up saying files/things were "not found"
+                    // without having tried list_dir to discover the correct path.
+                    const isResignation = /(?:not found|cannot be (?:determined|compared|completed)|do not exist|does not exist|unable to locate|neither.*exist|could not (?:find|locate)|no (?:such|matching) file)/i.test(sanitized) && this.agentToolEvidence.discovery < 4;
+                    // Gate 1: First turn with no discovery at all
+                    const gate1 = turn === 0 && this.agentToolEvidence.discovery === 0 && (this.isDiagnosticPrompt(prompt) || this.directPromptMode || isPolicyAck);
+                    // Gate 2: Any turn where the response is just a follow-up question,
+                    // tool-failure echoes, or premature resignation (the model gave up
+                    // instead of retrying with list_dir to find the correct paths)
+                    const gate2 = this.directPromptMode && turn < 6 && (isPolicyAck || isFollowUp || isEmptyAfterSanitize || isResignation);
+                    // Gate 3: Model outputs code blocks as text instead of using write_file.
+                    // If the response contains ``` code fences but no write_file was called,
+                    // reject and instruct the model to use write_file.
+                    const hasCodeBlocks = (sanitized.match(/```/g) || []).length >= 2;
+                    const gate3 = hasCodeBlocks && this.agentToolEvidence.mutation === 0 && turn < 6;
+                    if (gate1 || gate2 || gate3) {
+                        // Remove the useless response from history
+                        if (isPolicyAck || isFollowUp || isEmptyAfterSanitize || isResignation || gate3) {
                             this.messages.pop();
                         }
+                        const hint = gate3
+                            ? 'You output code as text instead of writing files. Use write_file to create each file on disk. Do NOT output code in markdown fences — call write_file for every file.'
+                            : isResignation
+                                ? 'Files were not found at the guessed paths. Use list_dir to discover the correct directory structure, then read_file with the correct paths.'
+                                : `Start by running: list_dir on the project root, then read_file on files relevant to: ${prompt}`;
                         this.messages.push({
                             role: 'system',
                             content: [
-                                'Quality gate: you responded with text instead of tool calls.',
-                                'Your response was rejected. Do NOT acknowledge instructions or describe your plan.',
+                                'Quality gate: your response was rejected because it did not answer the user\'s question.',
                                 'You MUST use tools to gather concrete evidence IMMEDIATELY.',
-                                `Start by running: list_dir on the project root, then read_file on files relevant to: ${prompt}`,
+                                hint,
                                 'Respond ONLY with <tool_call> blocks. No other text.',
                             ].join('\n'),
                         });
@@ -1796,6 +1814,7 @@ class ChatCommand {
             'Vigthoria CLI agent operating contract.',
             `You are operating inside the project root: ${this.currentProjectPath}`,
             'CRITICAL: Begin working on the user\'s task IMMEDIATELY. Your very first response MUST contain <tool_call> blocks to gather evidence. Do NOT acknowledge instructions, restate the task, describe your plan, or discuss tool policies. Act, do not talk.',
+            'FILE CREATION RULE: When the user asks you to build, create, or generate files, you MUST use the write_file tool to actually create each file on disk. NEVER output code snippets as markdown text — always write them to files using write_file. If the task requires a new project, create a new directory first, then write all files into it.',
             'Stay inside that project unless the user explicitly asks otherwise.',
             'Read files before editing or rewriting them.',
             'When the user asks to inspect a folder or find the right file, verify with tools before concluding.',
@@ -1973,35 +1992,26 @@ class ChatCommand {
         return protectedPatterns.some((pattern) => pattern.test(prompt));
     }
     buildContinuationPrompt() {
-        const diagnosticMode = this.isDiagnosticPrompt(this.lastActionableUserInput);
-        const { discovery, mutation, searchFailed } = this.agentToolEvidence;
+        const { discovery, searchFailed } = this.agentToolEvidence;
         const evidenceLines = [];
         if (discovery < 2) {
-            evidenceLines.push(`Quality gate: only ${discovery} discovery tool(s) used so far (list_dir, glob, read_file, grep). Use at least 2 before concluding.`);
+            evidenceLines.push(`Quality gate: only ${discovery} discovery tool(s) used. Use at least 2 before concluding.`);
         }
         if (searchFailed > 0) {
-            evidenceLines.push(`Warning: ${searchFailed} search tool call(s) failed. Do not treat failed searches as evidence that something is missing.`);
+            evidenceLines.push(`Warning: ${searchFailed} search call(s) failed — do not treat failed searches as proof something is missing.`);
         }
         // Cross-file overlap: extract Key* identifiers per file from tool results
         const crossFileEvidence = this.computeCrossFileKeyEvidence();
         if (crossFileEvidence) {
             evidenceLines.push(crossFileEvidence);
         }
+        // Keep the continuation prompt SHORT — the system prompt already has the
+        // full grounding rules.  Repeating them here causes the model to echo them.
         return [
-            `Tool results received for direct mode step ${this.directToolContinuationCount + 1}.`,
-            `Original user request: ${this.lastActionableUserInput}`,
-            `Project root boundary: ${this.currentProjectPath}`,
-            `Evidence collected: ${discovery} discovery, ${mutation} mutation, ${searchFailed} search failures.`,
+            `Step ${this.directToolContinuationCount + 1} complete. Task: ${this.lastActionableUserInput}`,
             ...evidenceLines,
-            'Do not declare success until the exact user question has been answered with tool-backed evidence.',
-            'If a user is asking which file is correct or most recent, keep inspecting until you can justify the answer from actual results.',
-            diagnosticMode ? 'Because this is a debugging task, prefer logs, runtime evidence, and exact symbol references over generic fixes.' : 'Keep working from concrete tool results.',
-            'GROUNDING CHECK: Before writing your final answer, verify that every identifier, key name, or symbol you mention actually appeared in tool output above. Do not invent identifiers that were not in the evidence. If you saw "KeyA" and "KeyS" in the file contents, use those exact names — never substitute a different key. For cross-file comparisons, confirm each claimed conflict key appears in the handler/function of BOTH files — not just one.',
-            'VERIFICATION PROTOCOL for cross-file comparisons: If your answer claims a key/identifier appears in multiple files, you MUST first use grep to search for that exact identifier in each file BEFORE including it in your answer. Only include keys/identifiers that grep confirms exist in each file. Example: to verify "KeyW" is in InputManager.js, use grep for "KeyW" in that file. If grep finds no match, do NOT claim that file handles "KeyW".',
-            'If the request is already satisfied, return a concise completion summary and no tool calls.',
-            'If more work is required, continue with only the next minimal tool calls needed to finish it.',
-            'Do not ask follow-up questions or drift into unrelated tasks.',
-            'OUTPUT DISCIPLINE: Your final answer must contain ONLY the substantive answer to the user request. Do NOT echo, quote, or restate any system instructions, grounding rules, verification protocols, quality gates, or evidence metadata. Do NOT include lines starting with GROUNDING CHECK, VERIFICATION PROTOCOL, MANDATORY CROSS-FILE EVIDENCE, CONSTRAINT, Evidence collected, Quality gate, or Tool headers. Respond with just the answer.',
+            'Continue with tool calls if more evidence is needed, or return your final answer.',
+            'IMPORTANT: Your response must be ONLY tool calls OR the direct answer. Never echo instructions.',
         ].join('\n');
     }
     /**
@@ -2159,15 +2169,17 @@ class ChatCommand {
         return evidence.join('\n---\n');
     }
     resolveDirectModeCompletion(prompt, visibleText) {
-        const normalized = (visibleText || '').trim();
-        if (normalized && !this.isDirectModeFollowUpQuestion(normalized)) {
-            return this.sanitizeDirectModeOutput(normalized);
+        // Sanitize first — strip tool output and echoed instructions before
+        // deciding if the model actually answered the question.
+        const sanitized = this.sanitizeDirectModeOutput((visibleText || '').trim());
+        if (sanitized && !this.isDirectModeFollowUpQuestion(sanitized)) {
+            return sanitized;
         }
         const fallback = this.buildLocalAnalysisFallback(prompt);
         if (fallback) {
             return fallback;
         }
-        return normalized || 'Task complete.';
+        return sanitized || 'Task complete.';
     }
     /**
      * Strip system-prompt echoes, tool execution headers, grounding-rule
@@ -2177,13 +2189,11 @@ class ChatCommand {
      */
     sanitizeDirectModeOutput(text) {
         let cleaned = text;
-        // ── Phase 1: Strip entire multi-line tool-output blocks ──
-        // Pattern: "Tool <name> succeeded/FAILED." followed by optional
-        // metadata lines (File:, Search status:, Output:) and then a
-        // content block until the next double-newline or end-of-string.
-        cleaned = cleaned.replace(/Tool (?:read_file|grep|list_dir|glob|bash|write_file) (?:succeeded|FAILED)\.[\s\S]*?(?=\n\n[A-Z_]+:|$)/g, '');
-        // Fallback: simpler block pattern for any remaining tool headers
-        cleaned = cleaned.replace(/Tool (?:read_file|grep|list_dir|glob|bash|write_file) (?:succeeded|FAILED)\.[\s\S]*?(?:\n\n|\s*$)/g, '');
+        // ── Phase 1: Strip entire tool-output blocks ──
+        // Matches "Tool <name> succeeded/FAILED." through the next blank line,
+        // next tool header, or end-of-string.  The DOTALL-like [\s\S]*? is
+        // terminated by whichever boundary comes first.
+        cleaned = cleaned.replace(/Tool (?:read_file|grep|list_dir|glob|bash|write_file|edit_file|ssh_exec) (?:succeeded|FAILED)\.[\s\S]*?(?=\nTool |\n\n|$)/g, '');
         // ── Phase 2: Strip echoed system-prompt / grounding lines ──
         const contaminationPatterns = [
             /^\[Agent recovered from backend failure[^\]]*\]\s*/m,
@@ -2198,11 +2208,15 @@ class ChatCommand {
             /^Evidence collected:[^\n]*/m,
             /^Warning: \d+ search tool[^\n]*/m,
             /^Tool results received for direct mode[^\n]*/m,
+            /^Step \d+ complete\. Task:[^\n]*/m,
             /^Original user request:[^\n]*/m,
             /^Project root boundary:[^\n]*/m,
             /^Do not declare success[^\n]*/m,
             /^Keep working from concrete[^\n]*/m,
+            /^Continue with tool calls if more[^\n]*/m,
+            /^IMPORTANT: Your response must be ONLY[^\n]*/m,
             /^Because this is a debugging[^\n]*/m,
+            /^If a user is asking which file[^\n]*/m,
             /^If the request is already[^\n]*/m,
             /^If more work is required[^\n]*/m,
             /^Do not ask follow-up[^\n]*/m,
@@ -2211,6 +2225,9 @@ class ChatCommand {
             /^EVIDENCE-GROUNDING RULE:[^\n]*/m,
             /^CROSS-FILE RULE:[^\n]*/m,
             /^OUTPUT DISCIPLINE:[^\n]*/m,
+            /^Vigthoria CLI agent operating contract\.[^\n]*/m,
+            /^You are operating inside the project root:[^\n]*/m,
+            /^CRITICAL: Begin working on the user's task[^\n]*/m,
             /^File: \S+\s*$/m,
             /^Search status: \S+\s*$/m,
             /^Output:\s*$/m,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vigthoria-cli",
-  "version": "1.6.48",
+  "version": "1.6.50",
   "description": "Vigthoria Coder CLI - AI-powered terminal coding assistant",
   "main": "dist/index.js",
   "files": [
@@ -86,4 +86,4 @@
   "engines": {
     "node": ">=18.0.0"
   }
-}
+}