code-review-ai-cli 1.3.2__tar.gz → 1.3.4__tar.gz

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-review-ai-cli
-Version: 1.3.2
+Version: 1.3.4
 Summary: Automated AI-powered code review CLI for Azure DevOps / TFS Pull Requests
 License: MIT
 Keywords: code-review,ai,azure-devops,tfs,pull-request,llm
@@ -351,9 +351,8 @@ flowchart TD
   R --> S[Limit files with max_diff_files]
   S --> T[Build changed-files summary]
   T --> U[Truncate each file with max_diff_lines]
-  U --> V[Run AI general review]
-  V --> W[Run AI structured comment generation]
-  W --> X[Preview review and suggested comments]
+  U --> V[Run single AI call: narrative review + structured comments]
+  V --> X[Preview review and suggested comments]
   X --> Y{Dry-run enabled?}
   Y -->|Yes| Z[Stop after preview]
   Y -->|No| AA{Auto-post enabled?}
@@ -392,6 +391,7 @@ Options:
 - `--format {terminal,markdown,json}`
 - `--output`, `-o`
 - `--no-color`
+- `--debug-dump`
 - `--model`, `-m`
 - `--provider`, `-p`
 - `--config`
@@ -428,3 +428,11 @@ tfs:
 ```
 
 Avoid `verify_ssl: false` except for temporary troubleshooting.
+
+### Debug Dump
+
+```yaml
+debug:
+  dump: true
+  dump_file: logs/llm_prompt_debug.log
+```

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/README.md

@@ -314,9 +314,8 @@ flowchart TD
   R --> S[Limit files with max_diff_files]
   S --> T[Build changed-files summary]
   T --> U[Truncate each file with max_diff_lines]
-  U --> V[Run AI general review]
-  V --> W[Run AI structured comment generation]
-  W --> X[Preview review and suggested comments]
+  U --> V[Run single AI call: narrative review + structured comments]
+  V --> X[Preview review and suggested comments]
   X --> Y{Dry-run enabled?}
   Y -->|Yes| Z[Stop after preview]
   Y -->|No| AA{Auto-post enabled?}
@@ -355,6 +354,7 @@ Options:
 - `--format {terminal,markdown,json}`
 - `--output`, `-o`
 - `--no-color`
+- `--debug-dump`
 - `--model`, `-m`
 - `--provider`, `-p`
 - `--config`
@@ -390,4 +390,12 @@ tfs:
   ca_bundle: C:/certs/corporate-root-ca.pem
 ```
 
-Avoid `verify_ssl: false` except for temporary troubleshooting.
+Avoid `verify_ssl: false` except for temporary troubleshooting.
+
+### Debug Dump
+
+```yaml
+debug:
+  dump: true
+  dump_file: logs/llm_prompt_debug.log
+```

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/code_review_ai_cli.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-review-ai-cli
-Version: 1.3.2
+Version: 1.3.4
 Summary: Automated AI-powered code review CLI for Azure DevOps / TFS Pull Requests
 License: MIT
 Keywords: code-review,ai,azure-devops,tfs,pull-request,llm
@@ -351,9 +351,8 @@ flowchart TD
   R --> S[Limit files with max_diff_files]
   S --> T[Build changed-files summary]
   T --> U[Truncate each file with max_diff_lines]
-  U --> V[Run AI general review]
-  V --> W[Run AI structured comment generation]
-  W --> X[Preview review and suggested comments]
+  U --> V[Run single AI call: narrative review + structured comments]
+  V --> X[Preview review and suggested comments]
   X --> Y{Dry-run enabled?}
   Y -->|Yes| Z[Stop after preview]
   Y -->|No| AA{Auto-post enabled?}
@@ -392,6 +391,7 @@ Options:
 - `--format {terminal,markdown,json}`
 - `--output`, `-o`
 - `--no-color`
+- `--debug-dump`
 - `--model`, `-m`
 - `--provider`, `-p`
 - `--config`
@@ -428,3 +428,11 @@ tfs:
 ```
 
 Avoid `verify_ssl: false` except for temporary troubleshooting.
+
+### Debug Dump
+
+```yaml
+debug:
+  dump: true
+  dump_file: logs/llm_prompt_debug.log
+```

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "code-review-ai-cli"
-version = "1.3.2"
+version = "1.3.4"
 description = "Automated AI-powered code review CLI for Azure DevOps / TFS Pull Requests"
 readme = "README.md"
 requires-python = ">=3.10"

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/src/ai_review.py

@@ -45,6 +45,7 @@ import os
 import sys
 import time
 import threading
+from datetime import datetime
 
 
 def _configure_console_streams() -> None:
@@ -254,6 +255,10 @@ def _add_global_options(parser: argparse.ArgumentParser) -> None:
         "--no-color", action="store_true",
         help="Disable terminal colors"
     )
+    group_output.add_argument(
+        "--debug-dump", action="store_true",
+        help="Dump the diff and full LLM prompt/context sent to the LLM into a debug log file"
+    )
 
     group_config = parser.add_argument_group("Configuration")
     group_config.add_argument(
@@ -321,6 +326,8 @@ def run_review(args: argparse.Namespace) -> int:
         config.dry_run = True
     if getattr(args, "auto_post", False):
         config.auto_post_comments = True
+    if getattr(args, "debug_dump", False):
+        config.debug_dump = True
 
     # --- Validate configuration ---
     issues = config.validate()
@@ -410,6 +417,12 @@ def run_pr_review_workflow(args: argparse.Namespace, config: ReviewConfig,
     # Show PR details
     print(formatter.format_pr_details(pr_details))
 
+    if config.debug_dump and not config.debug_dump_file:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        config.debug_dump_file = os.path.join("logs", f"pr_{pr_id}_{timestamp}_debug.log")
+        os.makedirs(os.path.dirname(config.debug_dump_file), exist_ok=True)
+        print(formatter.format_info(f"🐛 Debug dump enabled: {config.debug_dump_file}"))
+
     # --- Get PR diff ---
     print(formatter.format_progress("Getting Pull Request diff"))
 
@@ -492,17 +505,7 @@ def run_pr_review_workflow(args: argparse.Namespace, config: ReviewConfig,
 
     try:
         llm = LLMClient(config)
-
-        # Get general review as text
-        review_text = llm.review(
-            diff=diff_truncated,
-            files_summary=files_summary,
-            context=getattr(args, "context", ""),
-            review_scope=config.review_scope,
-        )
-
-        # Get structured comments to post
-        structured_comments = llm.review_pr_structured(
+        review_text, structured_comments = llm.review_pr(
             diff=diff_truncated,
             files_summary=files_summary,
             context=getattr(args, "context", ""),

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/src/config.py

@@ -103,6 +103,10 @@ class ReviewConfig:
     output_file: str = ""                   # Path to save output
     color_output: bool = True               # Terminal colors
 
+    # --- Debug ----------------------------------------------------------
+    debug_dump: bool = False                # Dump diff and full LLM prompt to a log file
+    debug_dump_file: str = ""               # Custom path (empty = logs/pr_<id>_debug.log)
+
     def get_effective_model(self) -> str:
         """Returns the effective model (configured or provider default)."""
         if self.model:
@@ -222,6 +226,9 @@ class ReviewConfig:
             "output_format": ("output", "format"),
             "output_file": ("output", "file"),
             "color_output": ("output", "color"),
+            # Debug
+            "debug_dump": ("debug", "dump"),
+            "debug_dump_file": ("debug", "dump_file"),
         }
 
         for attr, keys in mapping.items():

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/src/llm_client.py

@@ -29,7 +29,7 @@ class LLMError(Exception):
 SYSTEM_PROMPTS = {
     "quick": {
         "pt": (
-            "És um code reviewer experiente em .Net C#, TypeScript e SQL. Analisa o diff de código fornecido "
+            "És um code reviewer senior experiente. Analisa o diff de código fornecido "
             "e dá um review CONCISO e direto. Foca-te nos problemas mais críticos:\n"
             "- Bugs e erros lógicos\n"
             "- Problemas de segurança\n"
@@ -37,7 +37,7 @@ SYSTEM_PROMPTS = {
             "Formato: Lista de bullet points com o ficheiro e linha quando possível. "
         ),
         "en": (
-            "You are an experienced .Net C#, TypeScript and SQL code reviewer. Analyze the provided code diff "
+            "You are an experienced Senior Code Reviewer. Analyze the provided code diff "
             "and give a CONCISE review. Focus on critical issues:\n"
             "- Bugs and logic errors\n"
             "- Security issues\n"
@@ -47,7 +47,7 @@ SYSTEM_PROMPTS = {
     },
     "detailed": {
         "pt": (
-            "És um code reviewer experiente em .Net C#, TypeScript e SQL. Analisa o diff de código "
+            "És um code reviewer senior experiente. Analisa o diff de código "
             "fornecido e retorna apenas comentários inline.\n\n"
             "Formato de output — para cada problema encontrado, escreve exatamente:\n"
             "- Linha <número_linha>: <descrição do problema> \n\n"
@@ -59,7 +59,7 @@ SYSTEM_PROMPTS = {
             "- Sê específico, objetivo e conciso. Responde em português."
         ),
         "en": (
-            "You are an expert .Net C#, TypeScript e SQL code reviewer. Analyze the provided code and return only inline comments.\n\n"
+            "You are an experienced Senior Code Reviewer. Analyze the provided code and return only inline comments.\n\n"
             "Output format — for each issue found, output exactly:\n"
             "- Line <line_number>: <issue description>\n\n"
             "Rules:\n"
@@ -72,7 +72,7 @@ SYSTEM_PROMPTS = {
     },
     "security": {
         "pt": (
-            "És um especialista em segurança de aplicações .Net C#, TypeScript e SQL. Analisa o diff "
+            "És um especialista em segurança de aplicações. Analisa o diff "
             "de código fornecido com foco EXCLUSIVO em segurança.\n\n"
             "Procura por:\n"
             "- SQL Injection\n"
@@ -90,7 +90,7 @@ SYSTEM_PROMPTS = {
             "Responde em português."
         ),
         "en": (
-            "You are an application security .Net C#, TypeScript and SQL specialist. Analyze the "
+            "You are an application security. Analyze the "
             "provided code diff with EXCLUSIVE focus on security.\n\n"
             "Look for:\n"
             "- SQL Injection\n"
@@ -109,44 +109,69 @@ SYSTEM_PROMPTS = {
     },
 }
 
-# Special prompt for PR review with structured comments
+# Special prompt for PR review with a combined narrative summary + structured comments
 PR_COMMENT_PROMPT = {
     "pt": (
-        "Analisa o diff de código de um Pull Request e retorna os teus comentários em formato JSON estruturado.\n\n"
-        "Para CADA problema encontrado, retorna um objeto JSON com:\n"
-        '- "file": caminho do ficheiro (ex: "src/auth.py")\n'
-        '- "line": número da linha no diff (inteiro, ou 0 se geral)\n'
-        '- "type": tipo de issue ("bug", "security", "performance", "style", "suggestion", "praise")\n'
-        '- "comment": descrição direta do problema em português, sem saudações e sem emojis\n'
-        '- "suggestion": sugestão de correção (opcional, string vazia se não aplicável)\n'
-        '- "reference": fonte ou referência para o problema (URL de documentação, padrão ou princípio). Importante: incluir SEMPRE uma referência relevante.\n\n'
+        "Analisa o diff de código de um Pull Request e retorna a tua resposta num único objeto JSON "
+        "com dois campos:\n"
+        '- "summary": um resumo narrativo do review, em texto (string). Se não encontrares problemas, '
+        'escreve algo como "Nenhum problema encontrado."\n'
+        '- "comments": um array de objetos, um por cada problema encontrado, com:\n'
+        '  - "file": caminho do ficheiro (ex: "src/auth.py")\n'
+        '  - "line": número da linha no diff (inteiro, ou 0 se geral)\n'
+        '  - "type": tipo de issue ("bug", "security", "performance", "style", "suggestion", "praise")\n'
+        '  - "comment": descrição direta do problema em português, sem saudações e sem emojis\n'
+        '  - "suggestion": sugestão de correção (opcional, string vazia se não aplicável)\n'
+        '  - "reference": fonte ou referência para o problema (URL de documentação, padrão ou princípio). '
+        'Importante: incluir SEMPRE uma referência relevante.\n\n'
         "No campo 'comment', escreve de forma objetiva e curta. "
         "Não uses introduções como 'Olá' ou 'Como code reviewer sénior'.\n"
         "No campo 'reference', inclui uma fonte confiável, padrão ou link para documentação relevante.\n\n"
-        "Responde APENAS com um JSON array válido. Exemplo:\n"
-        '[\n'
-        '  {\n'
-        '    "file": "src/auth.py",\n'
-        '    "line": 42,\n'
-        '    "type": "security",\n'
-        '    "comment": "Password armazenada em texto simples sem hashing",\n'
-        '    "suggestion": "Usar bcrypt ou argon2 para hash de passwords",\n'
-        '    "reference": "OWASP - Password Storage Cheat Sheet"\n'
-        '  }\n'
-        ']\n\n'
+        "Responde APENAS com um JSON object válido. Exemplo:\n"
+        '{\n'
+        '  "summary": "Resumo narrativo curto do review...",\n'
+        '  "comments": [\n'
+        '    {\n'
+        '      "file": "src/auth.py",\n'
+        '      "line": 42,\n'
+        '      "type": "security",\n'
+        '      "comment": "Password armazenada em texto simples sem hashing",\n'
+        '      "suggestion": "Usar bcrypt ou argon2 para hash de passwords",\n'
+        '      "reference": "OWASP - Password Storage Cheat Sheet"\n'
+        '    }\n'
+        '  ]\n'
+        '}\n\n'
     ),
     "en": (
-        "Analyze the Pull Request code diff and return your comments in structured JSON format.\n\n"
-        "For EACH issue found, return a JSON object with:\n"
-        '- "file": file path (e.g., "src/auth.py")\n'
-        '- "line": line number in diff (integer, or 0 if general)\n'
-        '- "type": issue type ("bug", "security", "performance", "style", "suggestion", "praise")\n'
-        '- "comment": direct description of the issue, with no greetings and no emojis\n'
-        '- "suggestion": fix suggestion (optional, empty string if not applicable)\n'
-        '- "reference": source or reference for the issue (e.g., "OWASP Top 10", "PEP 8", documentation URL, standard or principle). Important: ALWAYS include a relevant reference.\n\n'
+        "Analyze the Pull Request code diff and return your response as a single JSON object with two "
+        "fields:\n"
+        '- "summary": a narrative text summary of the review (string). If no issues are found, write '
+        'something like "No issues found."\n'
+        '- "comments": an array of objects, one per issue found, with:\n'
+        '  - "file": file path (e.g., "src/auth.py")\n'
+        '  - "line": line number in diff (integer, or 0 if general)\n'
+        '  - "type": issue type ("bug", "security", "performance", "style", "suggestion", "praise")\n'
+        '  - "comment": direct description of the issue, with no greetings and no emojis\n'
+        '  - "suggestion": fix suggestion (optional, empty string if not applicable)\n'
+        '  - "reference": source or reference for the issue (e.g., "OWASP Top 10", "PEP 8", documentation '
+        'URL, standard or principle). Important: ALWAYS include a relevant reference.\n\n'
         "In 'comment', use a short and objective tone. "
         "Do not include intros like 'Hello' or 'As a senior reviewer'.\n"
         "In 'reference', include a trusted source, standard or link to relevant documentation.\n\n"
+        "Respond ONLY with a valid JSON object. Example:\n"
+        '{\n'
+        '  "summary": "Short narrative summary of the review...",\n'
+        '  "comments": [\n'
+        '    {\n'
+        '      "file": "src/auth.py",\n'
+        '      "line": 42,\n'
+        '      "type": "security",\n'
+        '      "comment": "Password stored in plain text without hashing",\n'
+        '      "suggestion": "Use bcrypt or argon2 to hash passwords",\n'
+        '      "reference": "OWASP - Password Storage Cheat Sheet"\n'
+        '    }\n'
+        '  ]\n'
+        '}\n\n'
     ),
 }
 
@@ -187,10 +212,11 @@ def get_pr_comment_prompt(language: str) -> str:
     If the language key is not present, ``"pt"`` (Portuguese) is used as the
     default.
 
-    The returned prompt instructs the LLM to produce a valid JSON array where
-    each element represents a single review comment with the fields
-    ``file``, ``line``, ``type``, ``comment``, ``suggestion``,
-    and ``reference``.
+    The returned prompt instructs the LLM to produce a single valid JSON
+    object with a ``"summary"`` field (narrative review text) and a
+    ``"comments"`` field (array of review comment objects, each with the
+    fields ``file``, ``line``, ``type``, ``comment``, ``suggestion``,
+    and ``reference``).
 
     Args:
         language: Response language code — ``"en"`` for English,
@@ -210,8 +236,8 @@ def get_scope_guidance(review_scope: str, language: str, structured: bool = Fals
     * Only added lines (``+``) appear in the diff section.
     * A ``### FULL_FILE_CONTEXT_START: <path> ###`` /
       ``### FULL_FILE_CONTEXT_END ###`` block is embedded in the payload for
-      each changed file, containing the complete new-version file content as
-      **read-only** background.
+      each changed file, containing the complete new-version (after changes)
+      file content as **read-only** background.
     * The review must focus **exclusively** on the changed lines (``+``); the
       full-file section exists only to prevent the model from hallucinating
       about the surrounding code.
@@ -250,7 +276,7 @@ def get_scope_guidance(review_scope: str, language: str, structured: bool = Fals
         if language == "en":
             return (
                 "Review scope: diff_only. The diff contains only added lines (+) — context and deletions were removed. "
-                "A full file content section (between ### FULL_FILE_CONTEXT_START and ### FULL_FILE_CONTEXT_END markers) "
+                "The complete new-version file content, after the changes (between ### FULL_FILE_CONTEXT_START and ### FULL_FILE_CONTEXT_END markers) "
                 "is provided for each file as read-only context. "
                 "Use it to understand the surrounding code, but focus your review EXCLUSIVELY on the changed lines (marked + in the diff). "
                 "Do NOT report issues in unchanged lines unless they directly affect the correctness of the changes. "
@@ -259,8 +285,7 @@ def get_scope_guidance(review_scope: str, language: str, structured: bool = Fals
             )
         return (
             "Ambito de review: diff_only. O diff contém apenas linhas adicionadas (+) — contexto e eliminações foram removidos. "
-            "Uma secção com o conteúdo completo do ficheiro (entre os marcadores ### FULL_FILE_CONTEXT_START e ### FULL_FILE_CONTEXT_END) "
-            "é fornecida como contexto de leitura. "
+            "É fornecido o conteúdo completo do ficheiro na versão nova, depois das alterações (entre os marcadores ### FULL_FILE_CONTEXT_START e ### FULL_FILE_CONTEXT_END) como contexto de leitura. "
             "Usa-a para compreender o código envolvente, mas foca o teu review EXCLUSIVAMENTE nas linhas alteradas (marcadas com + no diff). "
             "NÃO reportes problemas em linhas não alteradas, exceto se afetarem diretamente a correção das alterações. "
             "Para cada problema, DEVE ser fornecido file e line válidos (>0) para comentário inline. "
@@ -270,14 +295,14 @@ def get_scope_guidance(review_scope: str, language: str, structured: bool = Fals
     if language == "en":
         return (
             "Review scope: diff_only. The diff contains only added lines (+). "
-            "A full file content section (between ### FULL_FILE_CONTEXT_START and ### FULL_FILE_CONTEXT_END markers) "
+            "The new-version (post-change) full file content (between ### FULL_FILE_CONTEXT_START and ### FULL_FILE_CONTEXT_END markers) "
             "is provided for each file as read-only context. "
             "Use it to understand the surrounding code, but focus your review EXCLUSIVELY on the changed lines. "
             "Do NOT report issues in unchanged lines unless they directly affect the correctness of the changes."
         )
     return (
         "Ambito de review: diff_only. O diff contém apenas linhas adicionadas (+). "
-        "Uma secção com o conteúdo completo do ficheiro (entre os marcadores ### FULL_FILE_CONTEXT_START e ### FULL_FILE_CONTEXT_END) "
+        "Uma secção com o conteúdo completo do ficheiro na versão nova, após as alterações (entre os marcadores ### FULL_FILE_CONTEXT_START e ### FULL_FILE_CONTEXT_END) "
         "é fornecida como contexto de leitura. "
         "Usa-a para compreender o código envolvente, mas foca o teu review EXCLUSIVAMENTE nas linhas alteradas. "
         "NÃO reportes problemas em linhas não alteradas, exceto se afetarem diretamente a correção das alterações."
@@ -316,6 +341,34 @@ class LLMClient:
     def __init__(self, config: ReviewConfig):
         self.config = config
 
+    def _dump_prompt_debug(self, system_prompt: str, user_message: str) -> None:
+        """
+        Appends the full prompt/context sent to the LLM to a debug log file.
+
+        Only active when ``config.debug_dump`` is enabled. The log may
+        contain source code and PR content, so it must never be committed
+        to version control.
+        """
+        if not getattr(self.config, "debug_dump", False):
+            return
+
+        log_path = self.config.debug_dump_file or os.path.join("logs", "llm_prompt_debug.log")
+        log_dir = os.path.dirname(log_path)
+        if log_dir:
+            os.makedirs(log_dir, exist_ok=True)
+        try:
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(f"\n{'=' * 80}\n")
+                f.write(f"[FULL CONTEXT SENT TO LLM] {datetime.datetime.now().isoformat()}\n")
+                f.write(f"Provider: {self.config.llm_provider} | Model: {self.config.get_effective_model()}\n")
+                f.write(f"{'=' * 80}\n")
+                f.write("--- SYSTEM PROMPT ---\n")
+                f.write(system_prompt + "\n")
+                f.write("--- USER MESSAGE ---\n")
+                f.write(user_message + "\n")
+        except OSError:
+            pass
+
     def _load_custom_prompt_text(self) -> str:
         """Loads extra instructions from a configurable Markdown file."""
         path = (self.config.custom_prompt_file or "").strip()
@@ -332,73 +385,30 @@ class LLMClient:
         except Exception:
             return ""
 
-    def review(self, diff: str, files_summary: list[dict],
-               context: str = "", review_scope: str = "diff_only") -> str:
-        """
-        Sends the diff to the LLM and returns the review as text.
+    def review_pr(self, diff: str, files_summary: list[dict],
+                  context: str = "", review_scope: str = "diff_only") -> tuple[str, list[dict]]:
         """
-        base_prompt = get_system_prompt(
-            self.config.verbosity,
-            self.config.review_language,
-        )
-        custom_prompt = self._load_custom_prompt_text()
+        Sends the diff to the LLM in a single call and returns both the
+        narrative review text and the structured PR comments, parsed from one
+        combined JSON response.
 
-        scope_guidance = get_scope_guidance(
-            review_scope=review_scope,
-            language=self.config.review_language,
-            structured=False,
-        )
+        This replaces the previous two-call approach (separate narrative and
+        structured-comments requests), halving the diff/context payload sent
+        to the LLM per PR review.
 
-        if custom_prompt:
-            system_prompt = (
-                f"{base_prompt}\n\n"
-                f"{scope_guidance}\n\n"
-                "---\n"
-                "Custom user instructions (follow with priority):\n"
-                f"{custom_prompt}"
-            )
-            merged_context = (
-                f"{context}\n\n[Custom context loaded from {self.config.custom_prompt_file}]"
-                if context else
-                f"[Custom context loaded from {self.config.custom_prompt_file}]"
-            )
-        else:
-            system_prompt = f"{base_prompt}\n\n{scope_guidance}"
-            merged_context = context
-
-        user_message = build_user_message(diff, files_summary, merged_context)
-
-        provider = self.config.llm_provider.lower()
-
-        if provider == "openai":
-            return self._call_openai(system_prompt, user_message)
-        elif provider == "azure_openai":
-            return self._call_openai(system_prompt, user_message, azure=True)
-        elif provider == "gemini":
-            return self._call_gemini(system_prompt, user_message)
-        elif provider == "claude":
-            return self._call_claude(system_prompt, user_message)
-        elif provider == "ollama":
-            return self._call_ollama(system_prompt, user_message)
-        elif provider == "copilot":
-            return self._call_copilot(system_prompt, user_message)
-        elif provider == "bedrock":
-            return self._call_bedrock(system_prompt, user_message)
-        else:
-            raise LLMError(
-                f"Unsupported provider: '{provider}'.\n"
-                "Available providers: openai, azure_openai, gemini, claude, ollama, copilot, bedrock"
-            )
+        Args:
+            diff: Unified diff (or filtered/truncated diff) to review.
+            files_summary: List of dicts with 'file', 'additions', 'deletions'.
+            context: Additional free-text context supplied by the user.
+            review_scope: "diff_only" (default) or "full_code".
 
-    def review_pr_structured(self, diff: str, files_summary: list[dict],
-                             context: str = "", review_scope: str = "diff_only") -> list[dict]:
-        """
-        Sends the diff to the LLM and returns structured PR comments.
-        
         Returns:
-            List of dicts with keys: file, line, type, comment, suggestion
+            Tuple of (review_text, structured_comments), where structured_comments
+            is a list of dicts with keys: file, line, type, comment, suggestion,
+            reference.
         """
-        base_prompt = get_pr_comment_prompt(self.config.review_language)
+        base_prompt = get_system_prompt(self.config.verbosity, self.config.review_language)
+        json_schema_prompt = get_pr_comment_prompt(self.config.review_language)
         custom_prompt = self._load_custom_prompt_text()
 
         scope_guidance = get_scope_guidance(
@@ -407,10 +417,11 @@ class LLMClient:
             structured=True,
         )
 
+        combined_base = f"{base_prompt}\n\n{json_schema_prompt}\n\n{scope_guidance}"
+
         if custom_prompt:
             system_prompt = (
-                f"{base_prompt}\n\n"
-                f"{scope_guidance}\n\n"
+                f"{combined_base}\n\n"
                 "---\n"
                 "Custom user instructions (follow with priority):\n"
                 f"{custom_prompt}"
@@ -421,10 +432,11 @@ class LLMClient:
                 f"[Custom context loaded from {self.config.custom_prompt_file}]"
             )
         else:
-            system_prompt = f"{base_prompt}\n\n{scope_guidance}"
+            system_prompt = combined_base
             merged_context = context
 
         user_message = build_user_message(diff, files_summary, merged_context)
+        self._dump_prompt_debug(system_prompt, user_message)
 
         provider = self.config.llm_provider.lower()
 
@@ -443,13 +455,23 @@ class LLMClient:
         elif provider == "bedrock":
             raw = self._call_bedrock(system_prompt, user_message)
         else:
-            raise LLMError(f"Unsupported provider: '{provider}'")
+            raise LLMError(
+                f"Unsupported provider: '{provider}'.\n"
+                "Available providers: openai, azure_openai, gemini, claude, ollama, copilot, bedrock"
+            )
+
+        return self._parse_combined_response(raw)
 
-        return self._parse_structured_comments(raw)
+    def _extract_json_block(self, raw_response: str) -> str:
+        """Strips markdown code fences and isolates the outer JSON object.
+
+        Args:
+            raw_response: Raw text returned by the LLM, possibly wrapped in a
+                markdown code fence (```json ... ```` or ``` ... ````).
 
-    def _parse_structured_comments(self, raw_response: str) -> list[dict]:
-        """Parses the LLM JSON response."""
-        # Try to extract JSON from possible markdown
+        Returns:
+            The substring most likely to contain a valid JSON object.
+        """
         text = raw_response.strip()
         if text.startswith("```"):
             # Remove markdown code blocks
@@ -464,39 +486,62 @@ class LLMClient:
                     json_lines.append(line)
             text = "\n".join(json_lines).strip()
 
-        # Try to find JSON array
-        start = text.find("[")
-        end = text.rfind("]")
+        # Try to find JSON object
+        start = text.find("{")
+        end = text.rfind("}")
         if start != -1 and end != -1:
             text = text[start:end + 1]
 
+        return text
+
+    def _parse_combined_response(self, raw_response: str) -> tuple[str, list[dict]]:
+        """Parses the combined LLM response into (summary, comments).
+
+        Falls back to treating the entire raw response as the summary (with an
+        empty comments list) when the response is not valid JSON or is not a
+        JSON object, mirroring the previous fallback behavior for malformed
+        structured responses.
+
+        Args:
+            raw_response: Raw text returned by the LLM.
+
+        Returns:
+            Tuple of (summary, comments) where comments is a list of dicts with
+            keys: file, line, type, comment, suggestion, reference.
+        """
+        text = self._extract_json_block(raw_response)
+
         try:
-            comments = json.loads(text)
-            if not isinstance(comments, list):
-                comments = [comments]
+            data = json.loads(text)
         except json.JSONDecodeError:
-            # Fallback: return as a general comment
-            return [{
-                "file": "",
-                "line": 0,
-                "type": "suggestion",
-                "comment": raw_response,
-                "suggestion": "",
-                "reference": "",
-            }]
-
-        # Validate and normalize each comment
-        validated = []
-        for c in comments:
-            validated.append({
+            return raw_response, []
+
+        if not isinstance(data, dict):
+            return raw_response, []
+
+        summary = str(data.get("summary", "")).strip()
+        raw_comments = data.get("comments", [])
+        if not isinstance(raw_comments, list):
+            raw_comments = []
+
+        comments = []
+        for c in raw_comments:
+            if not isinstance(c, dict):
+                continue
+            try:
+                line = int(c.get("line", 0))
+            except (TypeError, ValueError):
+                line = 0
+            comments.append({
                 "file": str(c.get("file", "")),
-                "line": int(c.get("line", 0)),
+                "line": line,
                 "type": str(c.get("type", "suggestion")),
                 "comment": str(c.get("comment", "")),
                 "suggestion": str(c.get("suggestion", "")),
                 "reference": str(c.get("reference", "")),
             })
-        return validated
+
+        return summary or raw_response, comments
 
     # ------------------------------------------------------------------
     # OpenAI / Azure OpenAI

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/src/prompts/config.yaml.template

@@ -81,7 +81,7 @@ review:
   # Review scope:
   #   diff_only (default): focuses only on differences introduced in the PR
   #   full_code: allows full code review of changed files
-  scope: full_code
+  scope: diff_only
 
   # Markdown file with custom rules/context for the prompt
   # The content of this file is injected into LLM instructions on each review
@@ -122,3 +122,7 @@ output:
 
   # Terminal colors
   color: true
+
+debug:
+  dump: false
+  dump_file: "logs/debug.log"

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/src/tfs_client.py

@@ -300,17 +300,17 @@ class TFSClient:
         Gets the diff of a specific Pull Request.
 
         For ``diff_only`` scope (default), builds a standard unified diff for
-        each changed file and appends the full new-version file content as a
-        clearly-marked read-only block::
+        each changed file and appends the full new-version (source branch) file
+        content as a clearly-marked read-only block::
 
             ### FULL_FILE_CONTEXT_START: /path/to/file ###
-            <full file content>
+            <full file content — after changes>
             ### FULL_FILE_CONTEXT_END ###
 
         This block is preserved by
         :py:meth:`GitUtils.filter_diff_additions_only` so that the LLM
-        receives the surrounding context without being asked to review
-        unchanged lines.
+        receives the complete resulting file as context, without being asked
+        to review unchanged lines.
 
         For ``full_code`` scope, only the new-version file content is sent
         (every line prefixed with ``+``), without a ``-`` baseline.
@@ -442,8 +442,9 @@ class TFSClient:
         from the source branch, then generates a standard unified diff with
         3 lines of context using :py:mod:`difflib`.
 
-        After the diff lines, the **full new-version file content** is
-        appended as a read-only context block bounded by sentinel markers::
+        After the diff lines, the **full new-version (source branch, post-change)
+        file content** is appended as a read-only context block bounded by
+        sentinel markers::
 
             ### FULL_FILE_CONTEXT_START: /path/to/file ###
             <complete file content — not a diff, not prefixed with +/->
@@ -451,8 +452,10 @@ class TFSClient:
 
         These markers are recognised by
         :py:meth:`GitUtils.filter_diff_additions_only`, which preserves every
-        line inside the block.  This lets the LLM understand surrounding code
-        without being asked to review unchanged lines.
+        line inside the block.  This gives the LLM the complete resulting file
+        so it can correctly understand line numbers and surrounding code for
+        the added lines, avoiding confusion from showing stale pre-change
+        content.
 
         Args:
             repository: Repository name.
@@ -465,7 +468,7 @@ class TFSClient:
 
         Returns:
             List of diff strings for this file, including the
-            ``FULL_FILE_CONTEXT`` block when new content is available.
+            ``FULL_FILE_CONTEXT`` block when old content is available.
         """
         old_lines: list[str] = []
         new_lines: list[str] = []
@@ -518,10 +521,12 @@ class TFSClient:
         # so that filter_diff_by_extensions and _split_diff_sections work correctly.
         result = [f"diff --git a{original_path} b{file_path}"] + diff
 
-        # Include full file content as read-only context for the LLM.
-        # This section is preserved by filter_diff_additions_only and allows
-        # the model to understand the surrounding code without reviewing
-        # unchanged lines.
+        # Include new-version file content as read-only context for the LLM.
+        # Using the post-change (source branch) content gives the LLM the
+        # complete resulting file so it can correctly understand line numbers
+        # and surrounding code for the added lines, avoiding confusion from
+        # showing stale pre-change content.
+        # delete is excluded automatically because new_lines is empty for deleted files.
         if new_lines:
             result.append(f"### FULL_FILE_CONTEXT_START: {file_path} ###")
             result.extend(new_lines)

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/tests/test_ai_review.py

@@ -192,10 +192,10 @@ def test_run_pr_review_workflow_dry_run_saves_output(mocker, review_config) -> N
     tfs.get_pull_request_diff.return_value = diff
 
     llm = MagicMock()
-    llm.review.return_value = "General review"
-    llm.review_pr_structured.return_value = [
-        {"file": "a.py", "line": 1, "type": "bug", "comment": "msg"}
-    ]
+    llm.review_pr.return_value = (
+        "General review",
+        [{"file": "a.py", "line": 1, "type": "bug", "comment": "msg"}],
+    )
 
     git_utils = MagicMock()
     git_utils.filter_diff_additions_only.return_value = diff
@@ -258,10 +258,8 @@ def test_run_pr_review_workflow_returns_error_when_posting_fails(mocker, review_
     tfs.post_review_comments.side_effect = FakeTFSError("boom")
 
     llm = MagicMock()
-    llm.review.return_value = "General review"
-    llm.review_pr_structured.return_value = [
-        {"file": "a.py", "line": 1, "type": "bug", "comment": "msg"}
-    ]
+    structured_comments = [{"file": "a.py", "line": 1, "type": "bug", "comment": "msg"}]
+    llm.review_pr.return_value = ("General review", structured_comments)
 
     git_utils = MagicMock()
     git_utils.filter_diff_additions_only.return_value = diff
@@ -273,7 +271,7 @@ def test_run_pr_review_workflow_returns_error_when_posting_fails(mocker, review_
     mocker.patch("src.ai_review.ProgressIndicator", return_value=progress)
     mocker.patch("src.ai_review.LLMClient", return_value=llm)
     mocker.patch("src.ai_review.GitUtils.__new__", return_value=git_utils)
-    mocker.patch("src.ai_review._select_comments_to_post", return_value=llm.review_pr_structured.return_value)
+    mocker.patch("src.ai_review._select_comments_to_post", return_value=structured_comments)
     mocker.patch("src.ai_review._save_pr_review_output")
     mocker.patch("src.tfs_client.TFSClient", return_value=tfs)
     mocker.patch("src.tfs_client.TFSError", new=FakeTFSError)

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/tests/test_llm_client.py

@@ -89,7 +89,7 @@ def make_llm_config(**changes: object) -> ReviewConfig:
 
 def test_prompt_helpers_select_expected_language_and_scope() -> None:
     """It should select prompts and scope guidance consistently."""
-    assert "code reviewer" in get_system_prompt("quick", "en")
+    assert "code reviewer" in get_system_prompt("quick", "en").lower()
     assert "JSON" in get_pr_comment_prompt("pt")
     assert "full_code" in get_scope_guidance("full_code", "en")
     assert "file e line" in get_scope_guidance("diff_only", "pt", structured=True)
@@ -249,58 +249,95 @@ def test_load_custom_prompt_text_variants(tmp_path: Path, mocker) -> None:
     assert client._load_custom_prompt_text() == ""
 
 
-def test_review_dispatches_and_merges_custom_prompt(mocker, tmp_path: Path) -> None:
+def test_review_pr_dispatches_and_merges_custom_prompt(mocker, tmp_path: Path) -> None:
     """It should route reviews to the configured provider and merge custom context."""
     prompt_file = tmp_path / "prompt.md"
     prompt_file.write_text("Always mention tests", encoding="utf-8")
     config = make_llm_config(custom_prompt_file=str(prompt_file))
     client = LLMClient(config)
-    openai = mocker.patch("src.llm_client.LLMClient._call_openai", return_value="review text")
+    openai = mocker.patch(
+        "src.llm_client.LLMClient._call_openai",
+        return_value='{"summary": "review text", "comments": []}',
+    )
 
-    result = client.review("+code", [{"file": "a.py", "additions": 1, "deletions": 0}], context="Focus on bugs")
+    result = client.review_pr("+code", [{"file": "a.py", "additions": 1, "deletions": 0}], context="Focus on bugs")
 
-    assert result == "review text"
+    assert result == ("review text", [])
     system_prompt, user_message = openai.call_args.args[:2]
     assert "Custom user instructions" in system_prompt
     assert "Custom context loaded from" in user_message
 
 
-def test_review_raises_for_unsupported_provider() -> None:
+def test_review_pr_raises_for_unsupported_provider() -> None:
     """It should reject unsupported providers before any HTTP call."""
     client = LLMClient(make_llm_config(llm_provider="unknown"))
 
     with pytest.raises(LLMError, match="Unsupported provider"):
-        client.review("+code", [])
+        client.review_pr("+code", [])
 
 
-def test_review_pr_structured_dispatches_and_parses(mocker) -> None:
-    """It should dispatch structured reviews and normalize JSON comments."""
+def test_review_pr_dispatches_and_parses_combined_response(mocker) -> None:
+    """It should dispatch a single combined review and normalize summary + JSON comments."""
     client = LLMClient(make_llm_config(llm_provider="copilot"))
     copilot = mocker.patch(
         "src.llm_client.LLMClient._call_copilot",
-        return_value='[{"file": "src/app.py", "line": 5, "type": "bug", "comment": "boom", "suggestion": "fix", "reference": "Docs"}]',
+        return_value=(
+            '{"summary": "General review", "comments": '
+            '[{"file": "src/app.py", "line": 5, "type": "bug", "comment": "boom", '
+            '"suggestion": "fix", "reference": "Docs"}]}'
+        ),
     )
 
-    comments = client.review_pr_structured("+code", [{"file": "a.py", "additions": 1, "deletions": 0}])
+    summary, comments = client.review_pr("+code", [{"file": "a.py", "additions": 1, "deletions": 0}])
 
     assert copilot.called
+    assert summary == "General review"
     assert comments == [{"file": "src/app.py", "line": 5, "type": "bug", "comment": "boom", "suggestion": "fix", "reference": "Docs"}]
 
 
-def test_parse_structured_comments_handles_markdown_single_object_and_invalid_json() -> None:
-    """It should parse code fences, single objects and invalid fallback payloads."""
+def test_parse_combined_response_handles_markdown_missing_fields_and_invalid_json() -> None:
+    """It should parse code fences, tolerate missing/invalid fields and fall back on invalid payloads."""
     client = LLMClient(make_llm_config())
-    fenced = client._parse_structured_comments(
-        "```json\n[{\"file\": \"a.py\", \"line\": 1, \"comment\": \"x\"}]\n```"
+
+    fenced_summary, fenced_comments = client._parse_combined_response(
+        "```json\n{\"summary\": \"ok\", \"comments\": [{\"file\": \"a.py\", \"line\": 1, \"comment\": \"x\"}]}\n```"
     )
-    single = client._parse_structured_comments(
-        '{"file": "a.py", "line": 2, "type": "style", "comment": "y"}'
+    assert fenced_summary == "ok"
+    assert fenced_comments[0]["file"] == "a.py"
+    assert fenced_comments[0]["line"] == 1
+
+    # Missing "comments" field defaults to an empty list.
+    no_comments_summary, no_comments = client._parse_combined_response(
+        '{"summary": "only summary"}'
     )
-    fallback = client._parse_structured_comments("not json at all")
+    assert no_comments_summary == "only summary"
+    assert no_comments == []
 
-    assert fenced[0]["file"] == "a.py"
-    assert single[0]["line"] == 2
-    assert fallback[0]["comment"] == "not json at all"
+    # Non-list "comments" field is treated as if no comments were provided.
+    bad_comments_summary, bad_comments = client._parse_combined_response(
+        '{"summary": "weird", "comments": "not a list"}'
+    )
+    assert bad_comments_summary == "weird"
+    assert bad_comments == []
+
+    # Non-dict JSON value (e.g. a bare array with no nested object) falls back
+    # to the raw response text as the summary, with no comments.
+    raw_array = "[1, 2, 3]"
+    array_summary, array_comments = client._parse_combined_response(raw_array)
+    assert array_summary == raw_array
+    assert array_comments == []
+
+    # Malformed JSON falls back to raw text as summary with no comments.
+    fallback_summary, fallback_comments = client._parse_combined_response("not json at all")
+    assert fallback_summary == "not json at all"
+    assert fallback_comments == []
+
+    # A non-numeric "line" value must not crash parsing; it defaults to 0.
+    bad_line_summary, bad_line_comments = client._parse_combined_response(
+        '{"summary": "ok", "comments": [{"file": "a.py", "line": "N/A", "comment": "z"}]}'
+    )
+    assert bad_line_summary == "ok"
+    assert bad_line_comments[0]["line"] == 0
 
 
 def test_call_openai_builds_expected_payload_and_validates_configuration(mocker) -> None:

{code_review_ai_cli-1.3.2 → code_review_ai_cli-1.3.4}/tests/test_tfs_client.py

@@ -354,7 +354,7 @@ def test_build_diff_parts_and_file_content(mocker) -> None:
 # ---------------------------------------------------------------------------
 
 def test_build_unified_diff_part_appends_full_file_context_block(mocker) -> None:
-    """_build_unified_diff_part must append FULL_FILE_CONTEXT markers with the new file content."""
+    """_build_unified_diff_part must append FULL_FILE_CONTEXT markers with the new (source branch, post-change) file content."""
     client = TFSClient(make_tfs_config())
     mocker.patch(
         "src.tfs_client.TFSClient._get_file_content",
@@ -366,15 +366,36 @@ def test_build_unified_diff_part_appends_full_file_context_block(mocker) -> None
         "refs/heads/feature", "refs/heads/main",
     )
     joined = "\n".join(result)
+    context_block = joined.split("### FULL_FILE_CONTEXT_START")[1]
 
     assert "### FULL_FILE_CONTEXT_START: /src/app.py ###" in joined
-    assert "new line one" in joined
-    assert "new line two" in joined
+    assert "new line one" in context_block
+    assert "new line two" in context_block
+    assert "old line" not in context_block
     assert "### FULL_FILE_CONTEXT_END ###" in joined
     # Context block must appear after the diff headers
     assert joined.index("### FULL_FILE_CONTEXT_START") > joined.index("diff --git")
 
 
+def test_build_unified_diff_part_context_block_present_on_add(mocker) -> None:
+    """_build_unified_diff_part must append a FULL_FILE_CONTEXT block for add, containing the new content."""
+    client = TFSClient(make_tfs_config())
+    mocker.patch(
+        "src.tfs_client.TFSClient._get_file_content",
+        return_value="new content",
+    )
+
+    result = client._build_unified_diff_part(
+        "repo-a", "/src/new_file.py", "/src/new_file.py", "add",
+        "refs/heads/feature", "refs/heads/main",
+    )
+    joined = "\n".join(result)
+
+    assert "### FULL_FILE_CONTEXT_START: /src/new_file.py ###" in joined
+    assert "new content" in joined
+    assert "### FULL_FILE_CONTEXT_END ###" in joined
+
+
 def test_build_unified_diff_part_no_context_block_on_delete(mocker) -> None:
     """_build_unified_diff_part must NOT append a FULL_FILE_CONTEXT block when change_type is delete."""
     client = TFSClient(make_tfs_config())
@@ -394,7 +415,8 @@ def test_build_unified_diff_part_no_context_block_on_delete(mocker) -> None:
 
 
 def test_build_unified_diff_part_context_block_uses_new_file_path(mocker) -> None:
-    """The FULL_FILE_CONTEXT_START marker must reference the new file path (file_path), not original_path."""
+    """The FULL_FILE_CONTEXT_START marker must reference the new file path (file_path), not original_path,
+    and the block content must be the new (renamed) file's content."""
     client = TFSClient(make_tfs_config())
     mocker.patch(
         "src.tfs_client.TFSClient._get_file_content",
@@ -406,9 +428,12 @@ def test_build_unified_diff_part_context_block_uses_new_file_path(mocker) -> Non
         "refs/heads/feature", "refs/heads/main",
     )
     joined = "\n".join(result)
+    context_block = joined.split("### FULL_FILE_CONTEXT_START")[1]
 
     assert "### FULL_FILE_CONTEXT_START: /src/renamed.py ###" in joined
     assert "### FULL_FILE_CONTEXT_START: /src/original.py ###" not in joined
+    assert "new content" in context_block
+    assert "old content" not in context_block
 
 
 def test_build_unified_diff_part_no_context_block_when_content_unavailable(mocker) -> None: