glide-mcp 0.1.1__py3-none-any.whl → 0.1.8__py3-none-any.whl

src/mcp/app.py

@@ -1,18 +1,19 @@
 from src.kite_exclusive.commit_splitter.services.voyage_service import embed_code
 from src.core.LLM.cerebras_inference import complete
-from typing import Any, Dict, List, Tuple
+from src.kite_exclusive.resolve_conflicts.breeze_inference import resolve_merge_conflict
+from src.kite_exclusive.resolve_conflicts.morph_service import apply_code_edit
+from typing import Dict, List, Tuple, Optional
 import subprocess
 import json
 import os
 import asyncio
+import re
 from dotenv import load_dotenv
 import helix
 from fastmcp import FastMCP
 load_dotenv()
 
-mcp = FastMCP[Any]("glide")
-
-HELIX_API_ENDPOINT = os.getenv("HELIX_API_ENDPOINT", "")
+mcp = FastMCP("glide")
 
 
 async def find_git_root(start_path: str = None) -> str:
@@ -25,8 +26,6 @@ async def find_git_root(start_path: str = None) -> str:
     Returns:
         Path to the git repository root, or None if not in a git repository
     """
-    # First, try to get workspace root from common environment variables
-    # MCP clients like Cursor might set these
     env_vars = [
         "MCP_WORKSPACE_ROOT",
         "CURSOR_WORKSPACE_ROOT", 
@@ -38,15 +37,12 @@ async def find_git_root(start_path: str = None) -> str:
     for env_var in env_vars:
         workspace_from_env = os.getenv(env_var)
         if workspace_from_env and os.path.isdir(workspace_from_env):
-            # Try to find git root starting from this directory
             start_path = workspace_from_env
             break
     
     if start_path is None:
         start_path = os.getcwd()
     
-    # Use git rev-parse --show-toplevel to find the git root
-    # Use asyncio.create_subprocess_exec directly to avoid circular dependency
     try:
         process = await asyncio.create_subprocess_exec(
             "git",
@@ -64,40 +60,25 @@ async def find_git_root(start_path: str = None) -> str:
             if git_root:
                 return git_root
     except (FileNotFoundError, OSError):
-        # Git not found or other OS error
         pass
     
     return None
 
 
-# Helper function to run subprocess calls asynchronously to avoid blocking stdio
 async def run_subprocess(args: List[str], **kwargs) -> subprocess.CompletedProcess:
     """Run subprocess calls asynchronously to avoid blocking stdio transport."""
-    # Use asyncio.create_subprocess_exec instead of subprocess.run to avoid blocking
     capture_output = kwargs.pop('capture_output', False)
     text = kwargs.pop('text', False)
-    check = kwargs.pop('check', False)  # Handle check parameter separately
+    check = kwargs.pop('check', False)
     
-    # CRITICAL: Set stdin to DEVNULL to prevent subprocess from inheriting 
-    # the MCP stdio stdin, which causes deadlocks
     stdin = kwargs.pop('stdin', asyncio.subprocess.DEVNULL)
-    
-    # CRITICAL: Always capture stdout/stderr to PIPE to prevent subprocess output
-    # from leaking into the MCP stdio communication channel (which breaks JSON parsing)
-    # In stdio mode, parent's stdout/stderr IS the MCP communication channel, so we must
-    # always capture subprocess output to prevent git messages from breaking JSON protocol
     stdout = asyncio.subprocess.PIPE
     stderr = asyncio.subprocess.PIPE
-    # Remove any stdout/stderr from kwargs since we're overriding them
     kwargs.pop('stdout', None)
     kwargs.pop('stderr', None)
-    
-    # Only pass valid parameters to asyncio.create_subprocess_exec
-    # Filter out any subprocess.run() specific parameters that aren't valid
-    # Explicitly remove check and other invalid params to prevent errors
-    kwargs.pop('check', None)  # Extra safety: ensure check is removed
-    kwargs.pop('timeout', None)  # timeout handled by asyncio.wait_for elsewhere
-    kwargs.pop('input', None)  # input not supported in async subprocess
+    kwargs.pop('check', None)
+    kwargs.pop('timeout', None)
+    kwargs.pop('input', None)
     
     valid_exec_kwargs = {}
     allowed_params = {'cwd', 'env', 'start_new_session', 'shell', 'preexec_fn', 
@@ -106,10 +87,6 @@ async def run_subprocess(args: List[str], **kwargs) -> subprocess.CompletedProce
     for key, value in kwargs.items():
         if key in allowed_params:
             valid_exec_kwargs[key] = value
-        # Silently ignore other parameters
-    
-    # Final safety check: ensure check is not in valid_exec_kwargs
-    assert 'check' not in valid_exec_kwargs, "check parameter should not be passed to subprocess"
     
     process = await asyncio.create_subprocess_exec(
         *args,
@@ -121,7 +98,6 @@ async def run_subprocess(args: List[str], **kwargs) -> subprocess.CompletedProce
     
     stdout_data, stderr_data = await process.communicate()
     
-    # Create a CompletedProcess-like object
     result = subprocess.CompletedProcess(
         args=args,
         returncode=process.returncode,
@@ -129,7 +105,6 @@ async def run_subprocess(args: List[str], **kwargs) -> subprocess.CompletedProce
         stderr=stderr_data.decode('utf-8') if text and stderr_data else stderr_data,
     )
     
-    # If check=True, raise CalledProcessError on non-zero return code
     if check and result.returncode != 0:
         raise subprocess.CalledProcessError(
             result.returncode, args, result.stdout, result.stderr
@@ -137,19 +112,6 @@ async def run_subprocess(args: List[str], **kwargs) -> subprocess.CompletedProce
     
     return result
 
-@mcp.tool
-async def draft_pr():
-    instructions = [
-        "step 1: grep for CONTRIBUTING.md or similar documentation in the repository. If unable to find it, look for any contributing guidelines in the repository.",
-        "step 2: if not found, follow best practices for writing a pull request.",
-        "step 3: use the edit file tool to write a new PR_DRAFT.md file for the project.",
-    ]
-    result = "draft pr instructions: \n\n"
-    for i, instruction in enumerate(instructions, 1):
-        result += f"{i}. {instruction}\n\n"
-    return result
-
-
 @mcp.tool(
     name="split_commit",
     description="Splits a large unified diff / commit into smaller semantically-grouped commits.",
@@ -163,18 +125,13 @@ async def split_commit(workspace_root: str = None):
                         If not provided, will attempt to detect from environment variables or current directory.
     """
     try:
-        # Detect the git repository root
         if workspace_root:
-            # If provided, use it directly
             detected_root = await find_git_root(workspace_root)
             if detected_root:
                 workspace_root = detected_root
             elif not os.path.isdir(workspace_root):
                 return f"error: provided workspace_root '{workspace_root}' does not exist or is not a directory."
-            # If workspace_root is provided but not a git repo, we'll still try to use it
-            # (git commands will fail with a clear error if it's not a git repo)
         else:
-            # Try to auto-detect
             workspace_root = await find_git_root()
             if not workspace_root:
                 cwd = os.getcwd()
@@ -186,8 +143,6 @@ async def split_commit(workspace_root: str = None):
                     f"  2. Provide the workspace_root parameter with the path to your git repository root."
                 )
         
-        # 1) Collect changed files and per-file unified diffs
-        # Check staged, unstaged, and untracked files
         staged_proc = await run_subprocess(
             ["git", "diff", "--cached", "--name-only"], 
             capture_output=True, 
@@ -207,9 +162,6 @@ async def split_commit(workspace_root: str = None):
             cwd=workspace_root
         )
         
-        # Check if git commands failed (might indicate not a git repo)
-        # Note: git commands can return non-zero even in valid repos (e.g., no changes)
-        # Only error if we get explicit "not a git repository" messages
         error_messages = []
         if staged_proc.returncode != 0 and staged_proc.stderr:
             error_messages.append(staged_proc.stderr)
@@ -238,7 +190,6 @@ async def split_commit(workspace_root: str = None):
 
         file_to_diff: Dict[str, str] = {}
         for path in changed_files:
-            # Try staged diff first, then unstaged
             p = await run_subprocess(
                 ["git", "diff", "--cached", "--", path], 
                 capture_output=True, 
@@ -257,69 +208,54 @@ async def split_commit(workspace_root: str = None):
                 if p.returncode == 0 and p.stdout.strip():
                     file_to_diff[path] = p.stdout
                 else:
-                    # For untracked/new files, read the entire file content as the "diff"
-                    # Paths from git are relative to workspace root, so join them
                     file_path = os.path.join(workspace_root, path) if not os.path.isabs(path) else path
                     try:
                         with open(file_path, "r", encoding="utf-8") as f:
                             content = f.read()
-                            # Format as a new file addition diff
                             file_to_diff[path] = (
                                 f"diff --git a/{path} b/{path}\nnew file mode 100644\n--- /dev/null\n+++ b/{path}\n@@ -0,0 +1,{len(content.splitlines())} @@\n+{chr(10).join('+'+line for line in content.splitlines())}"
                             )
                     except (FileNotFoundError, UnicodeDecodeError):
-                        # File might not exist or not be text
                         continue
 
         if not file_to_diff:
             return "no per-file diffs produced"
 
-        # 2) Embed each file's diff with Voyage (preconfigured in voyage_service)
-        suggestions: List[Tuple[str, str]] = []  # (file_path, suggested_message)
+        suggestions: List[Tuple[str, str]] = []
 
-        # Connect Helix client - supports both local and cloud via environment variables
         use_local = os.getenv("HELIX_LOCAL", "false").lower() == "true"
         
         if use_local:
             db = helix.Client(local=True)
         else:
-            # Use cloud deployment from helix.toml (production.fly)
-            # Helix SDK automatically reads helix.toml and uses the configured deployment
             api_endpoint = os.getenv("HELIX_API_ENDPOINT", "")
-            if not HELIX_API_ENDPOINT:
-                return "error: HELIX API ENDPOINT is not set"
+            if not api_endpoint:
+                return "error: HELIX_API_ENDPOINT is not set"
             db = helix.Client(local=False, api_endpoint=api_endpoint)
 
         for file_path, diff_text in file_to_diff.items():
-            # 2a) Embed with timeout (5 seconds)
             try:
                 vec_batch = await asyncio.wait_for(
                     asyncio.to_thread(embed_code, diff_text, file_path=file_path),
                     timeout=5
                 )
             except asyncio.TimeoutError:
-                return f"error: embedding timed out for {file_path} (expected to always work)"
+                return f"error: embedding timed out for {file_path}"
             except Exception as embed_exc:
-                return f"error: embedding failed for {file_path}: {str(embed_exc)} (expected to always work)"
+                return f"error: embedding failed for {file_path}: {str(embed_exc)}"
             
             if not vec_batch:
                 return f"error: embedding returned empty result for {file_path}"
             vec = vec_batch[0]
 
             try:
-                # 3) ANN search for similar diffs; k kept small to keep it snappy
-                # Add timeout to database query (5 seconds)
                 res = await asyncio.wait_for(
                     asyncio.to_thread(db.query, "getSimilarDiffsByVector", {"vec": vec, "k": 8}),
                     timeout=5
                 )
-            except asyncio.TimeoutError:
-                # If database query times out, continue without examples
+            except (asyncio.TimeoutError, Exception):
                 res = []
-            except Exception as db_exc:
-                # If database query fails, continue without examples
-                res = []
-            # Result rows include commit_message, summary, file_path
+            
             examples = []
             if isinstance(res, list):
                 for row in res[:5]:
@@ -334,12 +270,18 @@ async def split_commit(workspace_root: str = None):
 
             example_block = "\n\n".join(examples) if examples else ""
             
-            # Helper function to detect and reject generic messages
             def is_generic_message(msg: str) -> bool:
                 """Check if a commit message is too generic."""
                 if not msg:
                     return True
                 msg_lower = msg.lower().strip()
+                
+                # Reject reasoning tag patterns
+                if ("redacted_reasoning" in msg_lower or 
+                    "<think>" in msg_lower or 
+                    "</think>" in msg_lower):
+                    return True
+                
                 generic_patterns = [
                     "update ",
                     "fix bug",
@@ -352,11 +294,9 @@ async def split_commit(workspace_root: str = None):
                     "minor",
                     "temporary",
                 ]
-                # Check if message starts with generic patterns
                 for pattern in generic_patterns:
                     if msg_lower.startswith(pattern):
                         return True
-                # Check if message is just a filename (e.g., "Update app.py")
                 if msg_lower.startswith("update ") and len(msg_lower.split()) <= 3:
                     return True
                 return False
@@ -397,47 +337,55 @@ EXAMPLES OF GOOD MESSAGES:
 - "docs(readme): add installation instructions"
 
 EXAMPLES OF BAD MESSAGES (DO NOT USE):
-- "Update app.py" ❌
-- "Fix bug" ❌
-- "Refactor code" ❌
-- "Changes" ❌
+- "Update app.py"
+- "Fix bug"
+- "Refactor code"
+- "Changes"
 
 Remember: Your output must be SPECIFIC and describe WHAT changed, not generic file operations."""
             )
             user_prompt = (
-                "Generate a commit message for this diff. Consider similar past changes if given.\n\n"
+                "/no_think\n\nGenerate a commit message for this diff. Consider similar past changes if given.\n\n"
                 f"DIFF (truncated if long):\n{diff_text}\n\n"
                 f"SIMILAR EXAMPLES:\n{example_block}\n\n"
                 "Output ONLY the commit message title, nothing else."
             )
             
-            # Call Cerebras inference - should always work
             try:
                 raw_response = await asyncio.wait_for(
-                    complete(user_prompt, system=system_prompt, max_tokens=100),
+                    complete(user_prompt, system=system_prompt, temperature=0.0),
                     timeout=30.0
                 )
             except asyncio.TimeoutError:
-                return f"error: Cerebras inference timed out for {file_path} (expected to always work)"
+                return f"error: Cerebras inference timed out for {file_path}"
             except Exception as llm_exc:
-                return f"error: Cerebras inference failed for {file_path}: {str(llm_exc)} (expected to always work)"
+                return f"error: Cerebras inference failed for {file_path}: {str(llm_exc)}"
             
             if not raw_response:
                 return f"error: Cerebras inference returned empty response for {file_path}"
             
-            commit_message = raw_response.strip().splitlines()[0].strip()
+            # Strip reasoning tags from response (e.g., <think>, </think>, <think>, etc.)
+            cleaned_response = raw_response.strip()
+            # Remove XML-like reasoning tags
+            cleaned_response = re.sub(r'<[^>]*think[^>]*>', '', cleaned_response, flags=re.IGNORECASE)
+            cleaned_response = re.sub(r'<[^>]*reasoning[^>]*>', '', cleaned_response, flags=re.IGNORECASE)
+            cleaned_response = re.sub(r'<[^>]*redacted[^>]*>', '', cleaned_response, flags=re.IGNORECASE)
+            
+            # Extract first non-empty line after cleaning
+            lines = [line.strip() for line in cleaned_response.splitlines() if line.strip()]
+            if not lines:
+                return f"error: No valid commit message found in response for {file_path} after cleaning reasoning tags"
+            
+            commit_message = lines[0]
             
-            # Remove quotes if present
             if commit_message.startswith('"') and commit_message.endswith('"'):
                 commit_message = commit_message[1:-1]
             if commit_message.startswith("'") and commit_message.endswith("'"):
                 commit_message = commit_message[1:-1]
             
-            # Validate the message is not generic - fail if it is
             if not commit_message or is_generic_message(commit_message):
                 return (
-                    f"error: Cerebras inference generated generic message '{commit_message}' for {file_path}. "
-                    f"Please improve the system prompt or check the inference output."
+                    f"error: Cerebras inference generated generic message '{commit_message}' for {file_path}"
                 )
 
             suggestions.append((file_path, commit_message))
@@ -445,7 +393,6 @@ Remember: Your output must be SPECIFIC and describe WHAT changed, not generic fi
         if not suggestions:
             return "no commit suggestions could be generated"
 
-        # 4) Commit each file separately with its suggested message
         for file_path, message in suggestions:
             try:
                 await run_subprocess(
@@ -465,21 +412,456 @@ Remember: Your output must be SPECIFIC and describe WHAT changed, not generic fi
                     "Ensure the file exists, is not conflicted, and git is functioning properly."
                 )
 
-        # 5) Return a compact report of what was committed
         report = {"commits": [{"file": f, "message": m} for f, m in suggestions]}
         return json.dumps(report, indent=2)
 
     except Exception as e:
-        return (
-            f"failed to split commit: {str(e)}\n"
-            f"Exception type: {type(e).__name__}\n"
-            "Ensure git is available and HelixDB is reachable on localhost:6969."
+        return f"failed to split commit: {str(e)}"
+
+
+_resolved_conflicts: Dict[str, Dict[str, str]] = {}
+
+
+async def get_conflicted_files(workspace_root: str) -> List[str]:
+    """
+    Find all files with merge conflicts in the git repository.
+    
+    Checks both git's merge state and files with conflict markers directly.
+    
+    Args:
+        workspace_root: Path to the git repository root
+        
+    Returns:
+        List of file paths with merge conflicts
+    """
+    conflicted_files = set()
+    
+    # Method 1: Check git's merge state (for active merges)
+    try:
+        proc = await run_subprocess(
+            ["git", "ls-files", "-u"],
+            capture_output=True,
+            text=True,
+            cwd=workspace_root
         )
+        
+        if proc.returncode == 0 and proc.stdout.strip():
+            # Extract unique file paths (git ls-files -u shows multiple entries per stage)
+            for line in proc.stdout.splitlines():
+                if line.strip():
+                    # Format: stage_number mode hash filename
+                    parts = line.split()
+                    if len(parts) >= 4:
+                        file_path = ' '.join(parts[3:])  # Handle filenames with spaces
+                        conflicted_files.add(file_path)
+    except Exception:
+        pass
+    
+    # Method 2: Scan files for conflict markers (works even if not in active merge)
+    try:
+        # Get all tracked files (or modified files)
+        proc = await run_subprocess(
+            ["git", "ls-files"],
+            capture_output=True,
+            text=True,
+            cwd=workspace_root
+        )
+        
+        if proc.returncode == 0:
+            all_files = [line.strip() for line in proc.stdout.splitlines() if line.strip()]
+            
+            # Also check untracked/modified files
+            proc_modified = await run_subprocess(
+                ["git", "ls-files", "--others", "--exclude-standard"],
+                capture_output=True,
+                text=True,
+                cwd=workspace_root
+            )
+            if proc_modified.returncode == 0:
+                all_files.extend([line.strip() for line in proc_modified.stdout.splitlines() if line.strip()])
+            
+            # Pattern to match conflict markers
+            conflict_pattern = re.compile(r'^<<<<<<<', re.MULTILINE)
+            
+            for file_path in all_files:
+                full_path = os.path.join(workspace_root, file_path) if not os.path.isabs(file_path) else file_path
+                
+                # Skip if already found via git ls-files -u
+                if file_path in conflicted_files:
+                    continue
+                
+                try:
+                    # Only check text files (skip binary files)
+                    with open(full_path, "r", encoding="utf-8", errors="ignore") as f:
+                        content = f.read()
+                        # Check for conflict markers
+                        if conflict_pattern.search(content):
+                            conflicted_files.add(file_path)
+                except (FileNotFoundError, UnicodeDecodeError, PermissionError):
+                    # Skip files we can't read
+                    continue
+    except Exception:
+        pass
+    
+    return sorted(list(conflicted_files))
+
+
+def extract_conflict_content(file_path: str, workspace_root: str) -> Tuple[str, str]:
+    """
+    Extract conflict content from a file.
+    
+    Args:
+        file_path: Relative path to the conflicted file
+        workspace_root: Path to the git repository root
+        
+    Returns:
+        Tuple of (original_file_content, conflict_text_with_markers)
+    """
+    full_path = os.path.join(workspace_root, file_path) if not os.path.isabs(file_path) else file_path
+    
+    try:
+        with open(full_path, "r", encoding="utf-8") as f:
+            original_content = f.read()
+        
+        # Extract conflict sections (between <<<<<<< and >>>>>>> markers)
+        conflict_pattern = re.compile(
+            r'<<<<<<<[^\n]*\n(.*?)\n=======\n(.*?)\n>>>>>>>[^\n]*',
+            re.DOTALL
+        )
+        
+        conflicts = conflict_pattern.findall(original_content)
+        if not conflicts:
+            # If no conflicts found with standard pattern, return the whole file
+            # as the conflict (might be a different conflict format)
+            return original_content, original_content
+        
+        # Combine all conflict sections
+        conflict_texts = []
+        for match in conflict_pattern.finditer(original_content):
+            conflict_texts.append(match.group(0))
+        
+        conflict_text = "\n\n".join(conflict_texts)
+        return original_content, conflict_text
+    except (FileNotFoundError, UnicodeDecodeError) as e:
+        raise RuntimeError(f"Failed to read file {file_path}: {str(e)}")
+
+
+def format_resolution_as_edit_snippet(original_content: str, conflict_text: str, resolution: str) -> Tuple[str, str]:
+    """
+    Format the resolution as an edit snippet for morphllm.
+    
+    Args:
+        original_content: Original file content with conflicts
+        conflict_text: The conflict section with markers
+        resolution: The resolved content from breeze model
+        
+    Returns:
+        Tuple of (instructions, edit_snippet)
+    """
+    # Find the conflict section in the original content
+    conflict_start = original_content.find(conflict_text)
+    if conflict_start == -1:
+        # If exact match not found, try to find by markers
+        conflict_start = original_content.find("<<<<<<<")
+    
+    if conflict_start == -1:
+        # Fallback: replace the entire conflict text
+        instructions = "Replace the merge conflict section with the resolved code."
+        edit_snippet = resolution
+        return instructions, edit_snippet
+    
+    # Find lines before and after conflict
+    lines_before = original_content[:conflict_start].splitlines()
+    lines_after = original_content[conflict_start + len(conflict_text):].splitlines()
+    
+    # Get context lines (last 3 lines before, first 3 lines after)
+    context_before = "\n".join(lines_before[-3:]) if lines_before else ""
+    context_after = "\n".join(lines_after[:3]) if lines_after else ""
+    
+    # Determine comment style based on file extension (simple heuristic)
+    # Default to // for most languages
+    comment_style = "//"
+    
+    # Build edit snippet
+    edit_lines = []
+    if context_before:
+        edit_lines.append(context_before)
+    edit_lines.append(f"{comment_style} ... existing code ...")
+    edit_lines.append(resolution)
+    edit_lines.append(f"{comment_style} ... existing code ...")
+    if context_after:
+        edit_lines.append(context_after)
+    
+    edit_snippet = "\n".join(edit_lines)
+    instructions = "Replace the merge conflict markers and conflicting code sections with the resolved code."
+    
+    return instructions, edit_snippet
 
 
-@mcp.tool
-async def resolve_conflict():
-    return "resolve conflict ran successfully"
+@mcp.tool(
+    name="resolve_conflict",
+    description="Detects merge conflicts in the repository, resolves them using AI, and applies the changes with MorphLLM. Changes are written to files immediately. Use revert_conflict_resolution to undo changes if needed.",
+)
+async def resolve_conflict(workspace_root: Optional[str] = None):
+    """
+    Detect and resolve merge conflicts using AI. Changes are applied immediately to files.
+    Review the previews and confirm or revert as needed.
+    
+    Args:
+        workspace_root: Optional path to the workspace root directory.
+                        If not provided, will attempt to detect from environment variables or current directory.
+    
+    Returns:
+        JSON string with applied resolutions and previews for review
+    """
+    global _resolved_conflicts
+    
+    try:
+        if workspace_root:
+            detected_root = await find_git_root(workspace_root)
+            if detected_root:
+                workspace_root = detected_root
+            elif not os.path.isdir(workspace_root):
+                return json.dumps({"error": f"provided workspace_root '{workspace_root}' does not exist or is not a directory."})
+        else:
+            workspace_root = await find_git_root()
+            if not workspace_root:
+                cwd = os.getcwd()
+                return json.dumps({
+                    "error": "could not detect git repository root.",
+                    "current_directory": cwd,
+                    "message": "Please either run this tool from within a git repository, or provide the workspace_root parameter."
+                })
+        
+        # Find conflicted files
+        conflicted_files = await get_conflicted_files(workspace_root)
+        
+        if not conflicted_files:
+            return json.dumps({
+                "message": "No merge conflicts detected.",
+                "resolved_files": []
+            })
+        
+        resolved_files = []
+        _resolved_conflicts = {}  # Store for potential revert
+        
+        for file_path in conflicted_files:
+            try:
+                # Extract conflict content
+                original_content, conflict_text = extract_conflict_content(file_path, workspace_root)
+                
+                # Get resolution from breeze model
+                try:
+                    resolution = await asyncio.wait_for(
+                        resolve_merge_conflict(conflict_text),
+                        timeout=60.0
+                    )
+                except asyncio.TimeoutError:
+                    resolved_files.append({
+                        "file": file_path,
+                        "status": "error",
+                        "error": "Breeze model timeout"
+                    })
+                    continue
+                except Exception as e:
+                    resolved_files.append({
+                        "file": file_path,
+                        "status": "error",
+                        "error": f"Breeze model failed: {str(e)}"
+                    })
+                    continue
+                
+                if not resolution:
+                    resolved_files.append({
+                        "file": file_path,
+                        "status": "error",
+                        "error": "Breeze model returned empty resolution"
+                    })
+                    continue
+                
+                # Format resolution as edit snippet
+                instructions, edit_snippet = format_resolution_as_edit_snippet(
+                    original_content, conflict_text, resolution
+                )
+                
+                # Apply via morphllm
+                try:
+                    final_content = await asyncio.wait_for(
+                        apply_code_edit(original_content, instructions, edit_snippet),
+                        timeout=60.0
+                    )
+                except asyncio.TimeoutError:
+                    resolved_files.append({
+                        "file": file_path,
+                        "status": "error",
+                        "error": "MorphLLM timeout"
+                    })
+                    continue
+                except Exception as e:
+                    resolved_files.append({
+                        "file": file_path,
+                        "status": "error",
+                        "error": f"MorphLLM failed: {str(e)}"
+                    })
+                    continue
+                
+                # Apply the change immediately by writing to file
+                full_path = os.path.join(workspace_root, file_path) if not os.path.isabs(file_path) else file_path
+                try:
+                    with open(full_path, "w", encoding="utf-8") as f:
+                        f.write(final_content)
+                except Exception as e:
+                    resolved_files.append({
+                        "file": file_path,
+                        "status": "error",
+                        "error": f"Failed to write resolved content: {str(e)}"
+                    })
+                    continue
+                
+                # Store original content for potential revert
+                _resolved_conflicts[file_path] = {
+                    "workspace_root": workspace_root,
+                    "original_content": original_content,
+                    "resolved_content": final_content,
+                }
+                
+                # Create a simple diff preview (show before/after around conflict)
+                conflict_lines = conflict_text.splitlines()
+                conflict_preview = "\n".join(conflict_lines[:10])
+                if len(conflict_lines) > 10:
+                    conflict_preview += f"\n... ({len(conflict_lines) - 10} more lines) ..."
+                
+                resolved_lines = final_content.splitlines()
+                resolved_preview = "\n".join(resolved_lines[:30])
+                if len(resolved_lines) > 30:
+                    resolved_preview += f"\n... ({len(resolved_lines) - 30} more lines) ..."
+                
+                resolved_files.append({
+                    "file": file_path,
+                    "status": "applied",
+                    "conflict_preview": conflict_preview,
+                    "resolved_preview": resolved_preview,
+                    "message": "Changes have been applied to the file. Review and confirm to stage, or revert if needed."
+                })
+                
+            except Exception as e:
+                resolved_files.append({
+                    "file": file_path,
+                    "status": "error",
+                    "error": str(e)
+                })
+        
+        successful_count = len([f for f in resolved_files if f.get('status') == 'applied'])
+        result = {
+            "message": f"Applied resolutions to {successful_count} file(s). Changes have been written to files.",
+            "instruction": "Review the resolved files above. You can stage them yourself or use the commit splitter. If you want to undo the changes, call revert_conflict_resolution.",
+            "resolved_files": resolved_files
+        }
+        
+        return json.dumps(result, indent=2)
+        
+    except Exception as e:
+        return json.dumps({"error": f"failed to resolve conflict: {str(e)}"})
+
+
+@mcp.tool(
+    name="revert_conflict_resolution",
+    description="Reverts the resolved conflict files back to their original state with conflicts. Use if you want to undo the changes applied by resolve_conflict.",
+)
+async def revert_conflict_resolution(file_path: Optional[str] = None, workspace_root: Optional[str] = None, revert_all: bool = False):
+    """
+    Revert resolved conflict files back to their original conflicted state.
+    
+    Args:
+        file_path: Optional path to a specific file to revert (relative to workspace root).
+                   If not provided and revert_all=False, reverts all pending resolutions.
+        workspace_root: Optional path to the workspace root directory.
+                        If not provided, will attempt to detect from environment variables or current directory.
+        revert_all: If True, revert all pending resolutions at once
+    
+    Returns:
+        JSON string with success/error status
+    """
+    global _resolved_conflicts
+    
+    try:
+        if workspace_root:
+            detected_root = await find_git_root(workspace_root)
+            if detected_root:
+                workspace_root = detected_root
+            elif not os.path.isdir(workspace_root):
+                return json.dumps({"error": f"provided workspace_root '{workspace_root}' does not exist or is not a directory."})
+        else:
+            workspace_root = await find_git_root()
+            if not workspace_root:
+                return json.dumps({"error": "could not detect git repository root."})
+        
+        if revert_all or not file_path:
+            # Revert all pending resolutions
+            if not _resolved_conflicts:
+                return json.dumps({
+                    "message": "No pending resolutions to revert."
+                })
+            
+            reverted_files = []
+            errors = []
+            
+            for path, resolution_data in list(_resolved_conflicts.items()):
+                try:
+                    # Verify workspace root matches
+                    if resolution_data["workspace_root"] != workspace_root:
+                        errors.append({"file": path, "error": "Workspace root mismatch"})
+                        continue
+                    
+                    full_path = os.path.join(workspace_root, path) if not os.path.isabs(path) else path
+                    
+                    # Write original content back to file
+                    with open(full_path, "w", encoding="utf-8") as f:
+                        f.write(resolution_data["original_content"])
+                    
+                    reverted_files.append(path)
+                    del _resolved_conflicts[path]
+                    
+                except Exception as e:
+                    errors.append({"file": path, "error": str(e)})
+            
+            result = {
+                "message": f"Reverted {len(reverted_files)} file(s) back to conflicted state.",
+                "reverted_files": reverted_files,
+                "errors": errors if errors else None
+            }
+            return json.dumps(result, indent=2)
+        else:
+            # Revert single file
+            if file_path not in _resolved_conflicts:
+                return json.dumps({
+                    "error": f"No pending resolution found for '{file_path}'."
+                })
+            
+            resolution_data = _resolved_conflicts[file_path]
+            
+            # Verify workspace root matches
+            if resolution_data["workspace_root"] != workspace_root:
+                return json.dumps({
+                    "error": f"Workspace root mismatch. Expected '{resolution_data['workspace_root']}', got '{workspace_root}'."
+                })
+            
+            full_path = os.path.join(workspace_root, file_path) if not os.path.isabs(file_path) else file_path
+            
+            # Write original content back to file
+            with open(full_path, "w", encoding="utf-8") as f:
+                f.write(resolution_data["original_content"])
+            
+            # Remove from pending resolutions
+            del _resolved_conflicts[file_path]
+            
+            return json.dumps({
+                "message": f"Reverted '{file_path}' back to conflicted state.",
+                "file": file_path
+            })
+            
+    except Exception as e:
+        return json.dumps({"error": f"failed to revert conflict resolution: {str(e)}"})
 
 
 def main():