Ai ReflectionEngine Added

2026-05-29 18:15:42 +02:00 · 2026-05-29 18:15:42 +02:00 · e7be1c94fd
commit e7be1c94fd
parent 4760dbb3a3
5 changed files with 754 additions and 16 deletions
--- a/README.me
+++ b/README.me
@ -1 +1,314 @@
-AISE AI Code Editor
+# AI Code Assistant
+
+Ein modularer KI-gestützter Coding-Assistent mit Tool-Ausführung, Datei-Kontextverwaltung und optionaler selbstreflektierender Agentenlogik.
+
+---
+
+# Übersicht
+
+Dieses Projekt ist ein erweiterbarer AI-Code-Assistent, der entwickelt wurde, um:
+
+* Code zu verstehen und zu erklären
+* Dateien zu bearbeiten und zu verwalten
+* Code auszuführen und Fehler zu analysieren
+* Dynamisch Tools zu verwenden
+* Antworten iterativ zu verbessern
+
+Die Anwendung kombiniert einen agentischen Workflow mit einer Reflexions-Engine, um qualitativ hochwertige Antworten und Code-Verbesserungen zu erzeugen.
+
+---
+
+# Projektstruktur
+
+```text
+main.py
+
+src/
+│
+├── backend/
+│   ├── chat_manager.py
+│   ├── reflection_engine.py
+│   ├── execution_engine.py
+│   ├── file_manager.py
+│   ├── llm_client.py
+│   ├── system_prompter.py
+│   ├── tools.py
+│   └── ...weitere Backend-Module
+│
+└── ui/
+    └── ...Frontend- und UI-Dateien
+```
+
+---
+
+# Architektur
+
+## `main.py`
+
+Der Einstiegspunkt der Anwendung.
+
+Verantwortlich für:
+
+* Initialisierung der Anwendung
+* Starten der Streamlit-Oberfläche
+* Verbinden aller Backend-Komponenten
+* Laden der UI
+
+---
+
+# Backend (`src/backend`)
+
+Enthält die gesamte Kernlogik der Anwendung.
+
+---
+
+## `chat_manager.py`
+
+Zentrale Steuerungseinheit des Assistenten.
+
+Aufgaben:
+
+* Verwaltung der Konversation
+* Agentischer Tool-Loop
+* Datei-Kontextverwaltung
+* Integration der Reflection Engine
+* Tool-Ausführung
+* Kommunikation zwischen UI und Backend
+
+Der `ChatManager` bildet das Herzstück der Anwendung.
+
+---
+
+## `reflection_engine.py`
+
+Implementiert den selbstreflektierenden Agenten-Workflow.
+
+Funktionen:
+
+* Bewertung von Antworten
+* Analyse von Schwächen
+* Iterative Verbesserung von Antworten
+* Qualitätsbewertung über Score-System
+* Automatisches Stoppen bei erreichter Qualität
+* Unterstützung von Tool-Aufrufen während der Verbesserung
+
+---
+
+## `execution_engine.py`
+
+Verantwortlich für:
+
+* Ausführen von Dateien
+* Ausführen von Code-Snippets
+* Erfassen von `stdout` und `stderr`
+* Rückgabe von Debug-Logs
+
+---
+
+## `file_manager.py`
+
+Verwaltet:
+
+* Lesen von Dateien
+* Schreiben von Dateien
+* Aktive Datei-Kontexte
+
+---
+
+## `tools.py`
+
+Definiert verfügbare Tools und das Tool-Registry-System.
+
+Beispiele:
+
+* `read_file`
+* `edit_file`
+* `execute_file`
+* `execute_code`
+* `search`
+
+---
+
+## `llm_client.py`
+
+Kommunikation mit dem verwendeten LLM/API.
+
+---
+
+## `system_prompter.py`
+
+Erstellt dynamische Systemprompts inklusive:
+
+* Datei-Kontext
+* Tool-Beschreibungen
+* Laufzeitinformationen
+
+---
+
+# UI (`src/ui`)
+
+Enthält sämtliche Frontend- und Benutzeroberflächen-Komponenten.
+
+Mögliche Funktionen:
+
+* Chat-Oberfläche
+* Dateiansicht
+* Tool-Aktivitätsanzeige
+* Reflection-Status
+* Ausgabe von Logs und Fehlern
+
+---
+
+# Reflection-System
+
+Der Assistent kann Antworten automatisch analysieren und verbessern.
+
+Ablauf:
+
+1. Initiale Antwort erzeugen
+2. Antwort bewerten
+3. Schwächen identifizieren
+4. Antwort verbessern
+5. Wiederholen bis Qualitätsziel erreicht ist
+
+---
+
+## Reflection Workflow
+
+```text
+Benutzeranfrage
+        ↓
+Erste Antwort
+        ↓
+Kritik & Analyse
+        ↓
+Qualitätsbewertung
+        ↓
+Verbesserung
+        ↓
+Optimierte Antwort
+```
+
+---
+
+## Reflection-Konfiguration
+
+Im `ChatManager`:
+
+```python
+self.reflection_enabled = True
+self.reflection_max_rounds = 3
+self.reflection_threshold = 8.5
+```
+
+---
+
+## Reflection aktivieren/deaktivieren
+
+```python
+chat_manager.set_reflection(
+    enabled=True,
+    max_rounds=3,
+    threshold=8.5
+)
+```
+
+---
+
+# Features
+
+* Tool-basierter KI-Agent
+* Iterative Antwortverbesserung
+* Selbstreflektierende Analyse
+* Datei-Kontextintegration
+* Code-Ausführung
+* Modulare Architektur
+* Streamlit-UI
+* Dynamische Tool-Registry
+* Debug- und Fehleranalyse
+* JSON-basierte Agent-Kommunikation
+
+---
+
+# Installation
+
+## 1. Repository klonen
+
+```bash
+git clone <repository-url>
+cd <projektname>
+```
+
+---
+
+## 2. Virtuelle Umgebung erstellen
+
+### Windows
+
+```bash
+python -m venv venv
+venv\Scripts\activate
+```
+
+### Linux / macOS
+
+```bash
+python3 -m venv venv
+source venv/bin/activate
+```
+
+---
+
+## 3. Abhängigkeiten installieren
+
+```bash
+pip install -r requirements.txt
+```
+
+---
+
+# Anwendung starten
+
+Die Anwendung wird über Streamlit gestartet:
+
+```bash
+streamlit run main.py
+```
+
+---
+
+# Beispielanfragen
+
+## Code erklären
+
+```text
+Was macht diese Funktion?
+```
+
+---
+
+## Datei bearbeiten
+
+```text
+Füge Kommentare zu dieser Datei hinzu
+```
+
+---
+
+## Datei ausführen
+
+```text
+Führe die aktuelle Datei aus
+```
+
+---
+
+## Fehler analysieren
+
+```text
+Warum tritt dieser Fehler auf?
+```
+
+---
+
+
--- a/requirements.txt
+++ b/requirements.txt
@ -27,6 +27,7 @@ MarkupSafe==3.0.3
 matplotlib-inline==0.2.2
 narwhals==2.21.2
 numpy==2.4.6
+openai
 packaging==26.2
 pandas==3.0.3
 parso==0.8.7
@ -57,6 +58,8 @@ soupsieve==2.8
 stack-data==0.6.3
 starlette==1.0.0
 streamlit==1.57.0
+streamlit-ace
+streamlit_antd_components
 streamlit-tree-select==0.0.5
 tenacity==9.1.4
 toml==0.10.2
--- a/src/backend/chat_manager.py
+++ b/src/backend/chat_manager.py
@ -5,6 +5,7 @@ from .system_prompter import SystemPrompter
 from .file_manager import FileManager
 from .tools import TOOL_REGISTRY, build_tool_description
 from .execution_engine import ExecutionEngine
+from .reflection_engine import ReflectionEngine
 import backend.tools as tools_module


@ -19,12 +20,35 @@ class ChatManager:
        self.active_file_content = None
        self.on_file_written = None
        self.on_tool_used = None
+
+        # ---- Reflection settings ----
+        self.reflection_enabled = True
+        self.reflection_max_rounds = 3
+        self.reflection_threshold = 8.5
+
        tools_module.on_tool_used = self._tool_used_callback

+    # ------------------------------------------------------------------ #
+    #  Callbacks                                                            #
+    # ------------------------------------------------------------------ #
+
    def _tool_used_callback(self, tool_name: str, summary: str):
        if self.on_tool_used:
            self.on_tool_used(tool_name, summary)

+    def _reflection_progress_callback(self, round_num: int, step):
+        if self.on_tool_used:
+            icon = "✅" if step.stop_reason in ("threshold_reached",) else "🔁"
+            self.on_tool_used(
+                "reflect",
+                f"{icon} Reflexion Runde {round_num} | Score {step.score:.1f}/10"
+                + (f" | {step.stop_reason}" if step.stop_reason not in ("continued", "") else ""),
+            )
+
+    # ------------------------------------------------------------------ #
+    #  File context                                                         #
+    # ------------------------------------------------------------------ #
+
    def set_active_file(self, file_path: str, content: str = None):
        self.active_file_path = file_path
        self.active_file_content = content if content else None
@ -77,7 +101,55 @@ To give a final answer (NO tool needed):
 - Never use execute_code just to print an explanation — answer directly instead
 """

+    # ------------------------------------------------------------------ #
+    #  Core generation loop                                                 #
+    # ------------------------------------------------------------------ #
+
    def generate_response(self, prompt: str, execution_results: dict = None) -> str:
+        """
+        1. Agentic loop → initial answer (hat Tool-Zugriff)
+        2. Wenn Reflection aktiviert: ReflectionEngine verbessert die Antwort.
+           Die Engine bekommt jetzt agent_loop + debug_logs übergeben,
+           damit Refinement-Runden ebenfalls Tools aufrufen können.
+        """
+        initial_answer = self._run_agent_loop(prompt, execution_results)
+
+        if not self.reflection_enabled:
+            return initial_answer
+
+        # Sehr kurze oder Fehler-Antworten nicht reflektieren
+        if len(initial_answer.strip()) < 30 or initial_answer.strip().upper().startswith("ERROR"):
+            return initial_answer
+
+        engine = ReflectionEngine(
+            client=self.client,
+            max_rounds=self.reflection_max_rounds,
+            threshold=self.reflection_threshold,
+            model=MODEL,
+            on_progress=self._reflection_progress_callback,
+            # NEU: agent_loop-Callback damit Refinement Tools nutzen kann
+            agent_loop=self._run_agent_loop,
+        )
+
+        context = self.active_file_content or ""
+        result = engine.reflect(
+            user_question=prompt,
+            initial_answer=initial_answer,
+            context=context,
+            # NEU: Debug-Logs werden durchgereicht
+            debug_logs=execution_results,
+        )
+
+        if self.on_tool_used and result.improved:
+            self.on_tool_used("reflect", f"✨ {result.summary}")
+
+        return result.final_answer
+
+    # ------------------------------------------------------------------ #
+    #  Agentic tool-use loop                                               #
+    # ------------------------------------------------------------------ #
+
+    def _run_agent_loop(self, prompt: str, execution_results: dict = None) -> str:
        system_prompt = self._build_system_prompt()
        user_message = f"User request: {prompt}"

@ -90,6 +162,10 @@ To give a final answer (NO tool needed):
            user_message += "\n[Execution Feedback — errors found]:\n"
            for log in execution_results.get("logs", []):
                user_message += f"  - [{log['level']}] {log['message']}\n"
+            if execution_results.get("stderr", "").strip():
+                user_message += f"[stderr]:\n{execution_results['stderr'].strip()}\n"
+            if execution_results.get("stdout", "").strip():
+                user_message += f"[stdout]:\n{execution_results['stdout'].strip()}\n"

        messages = [
            {"role": "system", "content": system_prompt},
@ -160,9 +236,18 @@ To give a final answer (NO tool needed):
                    "content": "Your response was not valid JSON. Reply with a single JSON object only."
                })

-        # Hard fallback — plain chat without JSON constraint
+        # Hard fallback
        return chat(self.client, messages, model=MODEL, temperature=0.7)

+    # ------------------------------------------------------------------ #
+    #  Session management                                                   #
+    # ------------------------------------------------------------------ #
+
+    def set_reflection(self, enabled: bool, max_rounds: int = 3, threshold: float = 8.5):
+        self.reflection_enabled = enabled
+        self.reflection_max_rounds = max(1, min(max_rounds, 5))
+        self.reflection_threshold = max(1.0, min(threshold, 10.0))
+
    def get_history(self) -> list:
        return self.conversation_history

--- a/src/backend/reflection_engine.py
+++ b/src/backend/reflection_engine.py
@ -0,0 +1,280 @@
+import json
+from dataclasses import dataclass, field
+from typing import Optional, Callable
+from .llm_client import chat_json, MODEL
+
+
+# ------------------------------------------------------------------ #
+#  Datastructures                                                      #
+# ------------------------------------------------------------------ #
+
+@dataclass
+class ReflectionStep:
+    round: int
+    critique: str
+    score: float
+    improved_answer: str
+    stop_reason: str = ""   # "threshold_reached" | "no_improvement" | "max_rounds" | "continued"
+
+
+@dataclass
+class ReflectionResult:
+    original_answer: str
+    final_answer: str
+    rounds_taken: int
+    steps: list[ReflectionStep] = field(default_factory=list)
+    improved: bool = False
+
+    @property
+    def summary(self) -> str:
+        if not self.steps:
+            return "Keine Reflexion durchgeführt."
+        last = self.steps[-1]
+        return (
+            f"🔁 {self.rounds_taken} Reflexionsrunde(n) | "
+            f"Score: {last.score:.1f}/10 | "
+            f"Stop: {last.stop_reason}"
+        )
+
+
+# ------------------------------------------------------------------ #
+#  Engine                                                               #
+# ------------------------------------------------------------------ #
+
+class ReflectionEngine:
+    CRITIC_SYSTEM = """You are a strict code-assistant quality reviewer.
+You evaluate an AI assistant's answer to a coding/debugging question.
+
+You receive:
+- The user question
+- The full file content (if available)
+- Debug logs / execution output (if available)
+- The assistant's answer
+
+Use ALL of this to spot errors: wrong line numbers, fixes that don't match the
+actual code, ignored error messages, missing edge cases, etc.
+
+Respond ONLY with a JSON object:
+{
+  "score": <float 0-10>,
+  "critique": "<concise bullet list of weaknesses, max 5 points>",
+  "needs_improvement": <true|false>
+}
+
+Scoring guide:
+  9-10 : Nearly perfect – correct, clear, complete, matches actual code
+  7-8  : Good but minor issues (verbosity, one missed case, slight imprecision)
+  5-6  : Mediocre – partly correct, ignores evidence from logs or code
+  3-4  : Poor – significant errors or contradicts the actual file/logs
+  0-2  : Wrong or harmful
+
+Set needs_improvement=false when score >= 8.5."""
+
+    # Refinement-System-Prompt für den Fallback (kein agent_loop verfügbar)
+    REFINE_FALLBACK_SYSTEM = """You are a code assistant improving your own previous answer.
+You receive: the original user question, file content, debug logs, your previous answer,
+and a critique of it.
+
+Produce a strictly better answer that addresses ALL critique points and is consistent
+with the actual file content and debug logs shown.
+Do NOT mention the reflection process or that this is a revised answer.
+
+Respond ONLY with a JSON object:
+{
+  "answer": "<your improved answer in Markdown>"
+}"""
+
+    def __init__(
+        self,
+        client,
+        max_rounds: int = 3,
+        threshold: float = 8.5,
+        model: str = MODEL,
+        on_progress: Optional[Callable] = None,
+        agent_loop: Optional[Callable] = None,
+    ):
+        self.client = client
+        self.max_rounds = max_rounds
+        self.threshold = threshold
+        self.model = model
+        self.on_progress = on_progress
+        # agent_loop(prompt: str, execution_results: dict | None) -> str
+        self.agent_loop = agent_loop
+
+    # -------------------------------------------------------------- #
+
+    def reflect(
+        self,
+        user_question: str,
+        initial_answer: str,
+        context: str = "",
+        debug_logs: Optional[dict] = None,
+    ) -> ReflectionResult:
+
+        result = ReflectionResult(
+            original_answer=initial_answer,
+            final_answer=initial_answer,
+            rounds_taken=0,
+        )
+
+        current_answer = initial_answer
+        prev_score: Optional[float] = None
+
+        for round_num in range(1, self.max_rounds + 1):
+
+            # ---- 1. Critique ----
+            critique_data = self._critique(
+                user_question, current_answer, context, debug_logs
+            )
+            score = critique_data.get("score", 5.0)
+            critique_text = critique_data.get("critique", "")
+            needs_improvement = critique_data.get("needs_improvement", True)
+
+            # ---- 2. Stop conditions ----
+            stop_reason = ""
+            if score >= self.threshold or not needs_improvement:
+                stop_reason = "threshold_reached"
+            elif prev_score is not None and score <= prev_score:
+                stop_reason = "no_improvement"
+
+            if stop_reason:
+                step = ReflectionStep(
+                    round=round_num,
+                    critique=critique_text,
+                    score=score,
+                    improved_answer=current_answer,
+                    stop_reason=stop_reason,
+                )
+                result.steps.append(step)
+                result.rounds_taken = round_num
+                if self.on_progress:
+                    self.on_progress(round_num, step)
+                break
+
+            # ---- 3. Refine ----
+            improved = self._refine(
+                user_question, current_answer, critique_text, context, debug_logs
+            )
+
+            step = ReflectionStep(
+                round=round_num,
+                critique=critique_text,
+                score=score,
+                improved_answer=improved,
+                stop_reason="continued",
+            )
+            result.steps.append(step)
+            result.rounds_taken = round_num
+
+            if self.on_progress:
+                self.on_progress(round_num, step)
+
+            prev_score = score
+            current_answer = improved
+
+        else:
+            if result.steps:
+                result.steps[-1].stop_reason = "max_rounds"
+
+        result.final_answer = current_answer
+        result.improved = current_answer != initial_answer
+        return result
+
+    # -------------------------------------------------------------- #
+    #  Internal helpers                                                   #
+    # -------------------------------------------------------------- #
+
+    def _build_debug_block(self, debug_logs: Optional[dict]) -> str:
+        """Wandelt execution_results in einen lesbaren Text-Block um."""
+        if not debug_logs:
+            return ""
+        parts = []
+        if debug_logs.get("stdout", "").strip():
+            parts.append(f"stdout:\n{debug_logs['stdout'].strip()}")
+        if debug_logs.get("stderr", "").strip():
+            parts.append(f"stderr:\n{debug_logs['stderr'].strip()}")
+        for log in debug_logs.get("logs", []):
+            parts.append(f"[{log.get('level','?')}] {log.get('message','')}")
+        return "\n\nDebug / execution output:\n" + "\n".join(parts) if parts else ""
+
+    def _critique(
+        self,
+        question: str,
+        answer: str,
+        context: str,
+        debug_logs: Optional[dict],
+    ) -> dict:
+        ctx_block = f"\n\nFull file content:\n{context[:3000]}" if context else ""
+        dbg_block = self._build_debug_block(debug_logs)
+
+        messages = [
+            {"role": "system", "content": self.CRITIC_SYSTEM},
+            {
+                "role": "user",
+                "content": (
+                    f"User question:\n{question}"
+                    f"{ctx_block}"
+                    f"{dbg_block}\n\n"
+                    f"Assistant answer to evaluate:\n{answer}"
+                ),
+            },
+        ]
+        try:
+            raw = chat_json(self.client, messages, model=self.model, temperature=0.3)
+            return json.loads(raw)
+        except Exception:
+            return {"score": 5.0, "critique": "Evaluation failed.", "needs_improvement": True}
+
+    def _refine(
+        self,
+        question: str,
+        prev_answer: str,
+        critique: str,
+        context: str,
+        debug_logs: Optional[dict],
+    ) -> str:
+
+        ctx_block = f"\n\nFull file content:\n{context[:3000]}" if context else ""
+        dbg_block = self._build_debug_block(debug_logs)
+
+        if self.agent_loop is not None:
+
+            refine_prompt = (
+                f"[REFLECTION REFINEMENT]\n"
+                f"The following critique was raised about your last answer:\n"
+                f"{critique}\n\n"
+                f"Original user question: {question}"
+                f"{ctx_block}"
+                f"{dbg_block}\n\n"
+                f"Your previous answer:\n{prev_answer}\n\n"
+                f"Please fix all issues from the critique. "
+                f"You may use tools (read_file, edit_file, search, execute_code) "
+                f"if needed to produce a correct and complete answer. "
+                f"Do NOT mention this refinement process in your answer."
+            )
+            try:
+                return self.agent_loop(refine_prompt, None)
+            except Exception:
+                pass  # Fallback to chat_json
+
+        # ---- Fallback ohne Tool-Zugriff ----
+        messages = [
+            {"role": "system", "content": self.REFINE_FALLBACK_SYSTEM},
+            {
+                "role": "user",
+                "content": (
+                    f"Original user question:\n{question}"
+                    f"{ctx_block}"
+                    f"{dbg_block}\n\n"
+                    f"Your previous answer:\n{prev_answer}\n\n"
+                    f"Critique:\n{critique}\n\n"
+                    "Now write your improved answer."
+                ),
+            },
+        ]
+        try:
+            raw = chat_json(self.client, messages, model=self.model, temperature=0.4)
+            data = json.loads(raw)
+            return data.get("answer", prev_answer)
+        except Exception:
+            return prev_answer
--- a/src/main.py
+++ b/src/main.py
@ -22,7 +22,8 @@ st.markdown("""
    body.chat-open .main .block-container { margin-right: 400px; }
    iframe[height="1"] { display: none !important; }
    [data-testid="stSidebar"] { z-index: 99 !important; }
-    /* Tool-use badge styling */
+
+    /* Tool-use & reflection badge styling */
    .tool-badge {
        display: inline-block;
        background: #1e3a5f;
@ -34,6 +35,18 @@ st.markdown("""
        margin: 2px 0 6px 0;
        font-family: monospace;
    }
+    .reflect-badge {
+        display: inline-block;
+        background: #1e3a1e;
+        color: #7ef77e;
+        border: 1px solid #2d9e2d;
+        border-radius: 6px;
+        padding: 3px 10px;
+        font-size: 0.82em;
+        margin: 2px 0 6px 0;
+        font-family: monospace;
+    }
+
    /* Pin chat input to bottom of its column */
    [data-testid="stChatInput"] {
        position: sticky !important;
@ -45,13 +58,20 @@ st.markdown("""
 </style>
 """, unsafe_allow_html=True)

+
+def _badge_html(event: str) -> str:
+    """Return the appropriate badge HTML for a tool event."""
+    cls = "reflect-badge" if event.startswith(("🔁", "✅", "✨")) else "tool-badge"
+    return f'<div class="{cls}">{event}</div>'
+
+
 def main():
    # --- Initialize Backend Components ---
-    file_manager    = FileManager()
-    navigation      = FileNavigation()
-    chat_manager    = ChatManager(file_manager=file_manager)
+    file_manager     = FileManager()
+    navigation       = FileNavigation()
+    chat_manager     = ChatManager(file_manager=file_manager)
    execution_engine = ExecutionEngine()
-    output_display  = OutputDisplay()
+    output_display   = OutputDisplay()

    # --- UI Components ---
    chat   = ChatInterface(chat_manager=chat_manager)
@ -69,6 +89,11 @@ def main():
    if "ai_wrote_file"         not in st.session_state: st.session_state.ai_wrote_file         = False
    if "tool_events"           not in st.session_state: st.session_state.tool_events           = []

+    # Reflection state defaults
+    if "reflect_enabled"   not in st.session_state: st.session_state.reflect_enabled   = True
+    if "reflect_rounds"    not in st.session_state: st.session_state.reflect_rounds    = 3
+    if "reflect_threshold" not in st.session_state: st.session_state.reflect_threshold = 8.5
+
    # --- Restore active file context on every rerender ---
    if st.session_state.get("active_file") and st.session_state.get("editor_content"):
        chat_manager.set_active_file(
@ -78,6 +103,13 @@ def main():
    else:
        chat_manager.set_active_file(None, "")

+    # Apply current reflection settings
+    chat_manager.set_reflection(
+        enabled    = st.session_state.reflect_enabled,
+        max_rounds = st.session_state.reflect_rounds,
+        threshold  = st.session_state.reflect_threshold,
+    )
+
    # --- Wire up the AI file-write callback ---
    def on_file_written(path: str, new_content: str):
        if str(path) == str(st.session_state.get("active_file")):
@ -87,7 +119,7 @@ def main():

    chat_manager.on_file_written = on_file_written

-    # --- Wire up tool-use notification ---
+    # --- Wire up tool-use / reflection notification ---
    def on_tool_used(tool_name: str, summary: str):
        st.session_state.tool_events.append(summary)

@ -95,13 +127,41 @@ def main():

    # ---------------- SIDEBAR ----------------
    st.sidebar.markdown("---")
-
    selected_file = navigation.render_sidebar()
    if selected_file:
        st.session_state.selected_file = selected_file
    st.sidebar.markdown("---")
    current_file = st.session_state.selected_file

+    # ---- Reflection controls in sidebar ----
+    with st.sidebar.expander("🧠 Selbstreflexion", expanded=False):
+        st.session_state.reflect_enabled = st.toggle(
+            "Aktiviert",
+            value=st.session_state.reflect_enabled,
+            help="Das Modell bewertet und verbessert jede Antwort selbstständig.",
+        )
+        st.session_state.reflect_rounds = st.slider(
+            "Max. Runden",
+            min_value=1, max_value=5,
+            value=st.session_state.reflect_rounds,
+            disabled=not st.session_state.reflect_enabled,
+            help="Wie oft darf das Modell seine Antwort überarbeiten?",
+        )
+        st.session_state.reflect_threshold = st.slider(
+            "Qualitäts-Schwelle",
+            min_value=5.0, max_value=10.0, step=0.5,
+            value=st.session_state.reflect_threshold,
+            disabled=not st.session_state.reflect_enabled,
+            help="Score (0–10) ab dem die Reflexion stoppt.",
+        )
+        if st.session_state.reflect_enabled:
+            st.caption(
+                f"Modell überarbeitet bis Score ≥ {st.session_state.reflect_threshold:.1f} "
+                f"oder {st.session_state.reflect_rounds} Runden erreicht."
+            )
+        else:
+            st.caption("Antworten werden ohne Selbstkritik ausgegeben.")
+
    # ---------------- FILE SWITCH ----------------
    if current_file != st.session_state.active_file:
        st.session_state.active_file = current_file
@ -155,8 +215,6 @@ def main():
        if not st.session_state.chat_open:
            st.subheader("💬 AI Assistant")

-            # Scrollable messages area — Streamlit's container(height=) gives it
-            # a fixed viewport that scrolls, so the input below is always visible.
            messages_container = st.container(height=600, border=False)
            with messages_container:
                messages = st.session_state.chat_messages
@ -170,10 +228,9 @@ def main():
                            and message.get("tool_events")
                        ):
                            for event in message["tool_events"]:
-                                st.markdown(f'<div class="tool-badge">{event}</div>', unsafe_allow_html=True)
+                                st.markdown(_badge_html(event), unsafe_allow_html=True)
                        st.markdown(message["content"])

-            # Input lives OUTSIDE the scrollable container — always at the bottom
            if prompt := st.chat_input("Ask about your code..."):
                st.session_state.tool_events = []
                st.session_state.chat_messages.append({"role": "user", "content": prompt})
@ -183,12 +240,12 @@ def main():
                        st.markdown(prompt)

                    with st.chat_message("assistant"):
-                        with st.spinner("AI is thinking..."):
+                        with st.spinner("AI is thinking…"):
                            response = chat_manager.generate_response(prompt, st.session_state.run_result)

                        tool_events = list(st.session_state.tool_events)
                        for event in tool_events:
-                            st.markdown(f'<div class="tool-badge">{event}</div>', unsafe_allow_html=True)
+                            st.markdown(_badge_html(event), unsafe_allow_html=True)

                        st.markdown(response)

@ -208,6 +265,6 @@ def main():
        chat_manager.clear_history()
        st.rerun()

+
 if __name__ == "__main__":
    main()
-# End of main.py