claude-mpm 5.0.2__py3-none-any.whl → 5.4.3__py3-none-any.whl

claude_mpm/services/agents/deployment/remote_agent_discovery_service.py

@@ -15,6 +15,8 @@ from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
+import yaml
+
 from claude_mpm.core.logging_config import get_logger
 
 logger = get_logger(__name__)
@@ -32,6 +34,9 @@ class RemoteAgentMetadata:
     routing_priority: int
     source_file: Path
     version: str  # SHA-256 hash from cache metadata
+    collection_id: Optional[str] = None  # Format: owner/repo-name
+    source_path: Optional[str] = None  # Relative path in repo
+    canonical_id: Optional[str] = None  # Format: collection_id:agent_id
 
 
 class RemoteAgentDiscoveryService:
@@ -65,6 +70,195 @@ class RemoteAgentDiscoveryService:
         self.remote_agents_dir = remote_agents_dir
         self.logger = get_logger(__name__)
 
+    def _extract_collection_id_from_path(self, file_path: Path) -> Optional[str]:
+        """Extract collection_id from repository path structure.
+
+        Collection ID is derived from the repository path structure:
+        ~/.claude-mpm/cache/remote-agents/{owner}/{repo}/agents/...
+
+        Args:
+            file_path: Absolute path to agent Markdown file
+
+        Returns:
+            Collection ID in format "owner/repo-name" or None if not found
+
+        Example:
+            Input:  ~/.claude-mpm/cache/remote-agents/bobmatnyc/claude-mpm-agents/agents/pm.md
+            Output: "bobmatnyc/claude-mpm-agents"
+        """
+        try:
+            # Find "remote-agents" in the path
+            path_parts = file_path.parts
+            remote_agents_idx = -1
+
+            for i, part in enumerate(path_parts):
+                if part == "remote-agents":
+                    remote_agents_idx = i
+                    break
+
+            if remote_agents_idx == -1 or remote_agents_idx + 2 >= len(path_parts):
+                self.logger.debug(
+                    f"Could not extract collection_id from path: {file_path}"
+                )
+                return None
+
+            # Extract owner and repo (next two parts after "remote-agents")
+            owner = path_parts[remote_agents_idx + 1]
+            repo = path_parts[remote_agents_idx + 2]
+
+            collection_id = f"{owner}/{repo}"
+            self.logger.debug(f"Extracted collection_id: {collection_id}")
+            return collection_id
+
+        except Exception as e:
+            self.logger.warning(
+                f"Failed to extract collection_id from {file_path}: {e}"
+            )
+            return None
+
+    def _extract_source_path_from_file(self, file_path: Path) -> Optional[str]:
+        """Extract relative source path within repository.
+
+        Source path is relative to the repository root (not the agents subdirectory).
+
+        Args:
+            file_path: Absolute path to agent Markdown file
+
+        Returns:
+            Relative path from repo root, or None if not found
+
+        Example:
+            Input:  ~/.claude-mpm/cache/remote-agents/bobmatnyc/claude-mpm-agents/agents/pm.md
+            Output: "agents/pm.md"
+        """
+        try:
+            # Find "remote-agents" in the path
+            path_parts = file_path.parts
+            remote_agents_idx = -1
+
+            for i, part in enumerate(path_parts):
+                if part == "remote-agents":
+                    remote_agents_idx = i
+                    break
+
+            if remote_agents_idx == -1 or remote_agents_idx + 3 >= len(path_parts):
+                return None
+
+            # Path after owner/repo is the source path
+            # remote-agents/{owner}/{repo}/{source_path}
+            repo_root_idx = remote_agents_idx + 3
+            source_parts = path_parts[repo_root_idx:]
+
+            return "/".join(source_parts)
+
+        except Exception as e:
+            self.logger.warning(f"Failed to extract source_path from {file_path}: {e}")
+            return None
+
+    def _parse_yaml_frontmatter(self, content: str) -> Optional[Dict[str, Any]]:
+        """Parse YAML frontmatter from Markdown content.
+
+        Extracts YAML frontmatter delimited by --- markers at the start of the file.
+        Uses a tolerant approach: attempts full YAML parsing first, falls back to
+        simple key-value extraction for malformed YAML.
+
+        Design Decision: Tolerant YAML Parsing
+
+        Rationale: Some agent markdown files have malformed YAML (incorrect indentation
+        in nested structures). Rather than failing completely, we:
+        1. Try full YAML parsing first (handles well-formed YAML)
+        2. Fall back to regex extraction for critical fields (agent_id, name, etc.)
+        3. Log warnings but continue processing
+
+        This ensures we can still extract agent_id even if complex nested structures
+        (like template_changelog) have indentation issues.
+
+        Args:
+            content: Full Markdown file content
+
+        Returns:
+            Dictionary of parsed YAML frontmatter, or None if not found
+
+        Example:
+            Input:
+                ---
+                agent_id: python-engineer
+                name: Python Engineer
+                version: 2.3.0
+                ---
+                # Agent content...
+
+            Output:
+                {"agent_id": "python-engineer", "name": "Python Engineer", "version": "2.3.0"}
+        """
+        try:
+            # Check if content starts with YAML frontmatter
+            if not content.startswith("---"):
+                self.logger.debug("No YAML frontmatter found (doesn't start with ---)")
+                return None
+
+            # Extract frontmatter content between --- markers
+            frontmatter_match = re.match(r"^---\n(.*?)\n---\s*\n", content, re.DOTALL)
+            if not frontmatter_match:
+                self.logger.debug("No closing --- marker found for YAML frontmatter")
+                return None
+
+            yaml_content = frontmatter_match.group(1)
+
+            # Try full YAML parsing first
+            try:
+                parsed = yaml.safe_load(yaml_content)
+                if isinstance(parsed, dict):
+                    return parsed
+                self.logger.warning(
+                    f"YAML frontmatter is not a dictionary: {type(parsed)}"
+                )
+            except yaml.YAMLError as e:
+                # Malformed YAML (e.g., indentation errors) - fall back to regex extraction
+                self.logger.debug(
+                    f"Full YAML parse failed, using fallback extraction: {e}"
+                )
+
+                # Extract key fields using regex (tolerant of malformed nested structures)
+                result = {}
+
+                # Extract simple key-value pairs (no nested structures)
+                simple_keys = [
+                    "agent_id",
+                    "name",
+                    "description",
+                    "version",
+                    "model",
+                    "agent_type",
+                    "category",
+                    "author",
+                    "schema_version",
+                ]
+
+                for key in simple_keys:
+                    # Match key: value on a line (not indented, so it's top-level)
+                    pattern = rf"^{key}:\s*(.+?)$"
+                    match = re.search(pattern, yaml_content, re.MULTILINE)
+                    if match:
+                        value = match.group(1).strip()
+                        # Remove quotes if present
+                        if value.startswith(("'", '"')) and value.endswith(("'", '"')):
+                            value = value[1:-1]
+                        result[key] = value
+
+                if result:
+                    self.logger.debug(
+                        f"Extracted {len(result)} fields using fallback method"
+                    )
+                    return result
+                return None
+
+        except Exception as e:
+            self.logger.warning(f"Unexpected error parsing frontmatter: {e}")
+            return None
+
+        return None
+
     def _generate_hierarchical_id(self, file_path: Path) -> str:
         """Generate hierarchical agent ID from file path.
 
@@ -77,14 +271,20 @@ class RemoteAgentDiscoveryService:
         - Preset matching against AUTO-DEPLOY-INDEX.md
         - Multi-level organization without name collisions
 
-        Bug #4 Fix: Calculate relative to /agents/ subdirectory, not repository root
-        This ensures agent IDs are based on their position within the agents directory.
+        Supports both cache structures:
+        1. Git repo: Calculate relative to /agents/ subdirectory
+        2. Flattened cache: Calculate relative to remote_agents_dir directly
 
-        Example:
+        Example (Git repo):
             Input:  /cache/bobmatnyc/claude-mpm-agents/agents/engineer/backend/python-engineer.md
             Root:   /cache/bobmatnyc/claude-mpm-agents/agents
             Output: engineer/backend/python-engineer
 
+        Example (Flattened cache):
+            Input:  /cache/remote-agents/engineer/python-engineer.md
+            Root:   /cache/remote-agents
+            Output: engineer/python-engineer
+
         Args:
             file_path: Absolute path to agent Markdown file
 
@@ -92,16 +292,30 @@ class RemoteAgentDiscoveryService:
             Hierarchical agent ID with forward slashes
         """
         try:
-            # Calculate relative to /agents/ subdirectory (Bug #4 fix)
+            # Try git repo structure first: /agents/ subdirectory
             agents_dir = self.remote_agents_dir / "agents"
-            relative_path = file_path.relative_to(agents_dir)
+            if agents_dir.exists():
+                try:
+                    relative_path = file_path.relative_to(agents_dir)
+                    return str(relative_path.with_suffix("")).replace("\\", "/")
+                except ValueError:
+                    pass  # Not under agents_dir, try flattened structure
+
+            # Try flattened cache structure: calculate relative to remote_agents_dir
+            try:
+                relative_path = file_path.relative_to(self.remote_agents_dir)
+                return str(relative_path.with_suffix("")).replace("\\", "/")
+            except ValueError:
+                pass  # Not under remote_agents_dir either
 
-            # Remove .md extension and convert to forward slashes
-            return str(relative_path.with_suffix("")).replace("\\", "/")
-        except ValueError:
-            # File is not under agents subdirectory, fall back to filename
+            # Fall back to filename
             self.logger.warning(
-                f"File {file_path} not under agents directory, using filename"
+                f"File {file_path} not under expected directories, using filename"
+            )
+            return file_path.stem
+        except Exception as e:
+            self.logger.warning(
+                f"Error generating hierarchical ID for {file_path}: {e}"
             )
             return file_path.stem
 
@@ -111,14 +325,20 @@ class RemoteAgentDiscoveryService:
         Extracts category from directory structure. Category is the path
         from agents subdirectory to the file, excluding the filename.
 
-        Bug #4 Fix: Calculate relative to /agents/ subdirectory, not repository root
-        This ensures categories are based on agent organization within /agents/.
+        Supports both cache structures:
+        1. Git repo: Calculate relative to /agents/ subdirectory
+        2. Flattened cache: Calculate relative to remote_agents_dir directly
 
-        Example:
+        Example (Git repo):
             Input:  /cache/bobmatnyc/claude-mpm-agents/agents/engineer/backend/python-engineer.md
             Root:   /cache/bobmatnyc/claude-mpm-agents/agents
             Output: engineer/backend
 
+        Example (Flattened cache):
+            Input:  /cache/remote-agents/engineer/python-engineer.md
+            Root:   /cache/remote-agents
+            Output: engineer
+
         Args:
             file_path: Absolute path to agent Markdown file
 
@@ -126,12 +346,26 @@ class RemoteAgentDiscoveryService:
             Category path with forward slashes, or "universal" if in root
         """
         try:
-            # Calculate relative to /agents/ subdirectory (Bug #4 fix)
+            # Try git repo structure first: /agents/ subdirectory
             agents_dir = self.remote_agents_dir / "agents"
-            relative_path = file_path.relative_to(agents_dir)
-            parts = relative_path.parts[:-1]  # Exclude filename
-            return "/".join(parts) if parts else "universal"
-        except ValueError:
+            if agents_dir.exists():
+                try:
+                    relative_path = file_path.relative_to(agents_dir)
+                    parts = relative_path.parts[:-1]  # Exclude filename
+                    return "/".join(parts) if parts else "universal"
+                except ValueError:
+                    pass  # Not under agents_dir, try flattened structure
+
+            # Try flattened cache structure: calculate relative to remote_agents_dir
+            try:
+                relative_path = file_path.relative_to(self.remote_agents_dir)
+                parts = relative_path.parts[:-1]  # Exclude filename
+                return "/".join(parts) if parts else "universal"
+            except ValueError:
+                pass  # Not under remote_agents_dir either
+
+            return "universal"
+        except Exception:
             return "universal"
 
     def discover_remote_agents(self) -> List[Dict[str, Any]]:
@@ -140,8 +374,12 @@ class RemoteAgentDiscoveryService:
         Scans the remote agents directory for *.md files recursively and converts each
         to JSON template format. Skips files that can't be parsed.
 
-        Bug #4 Fix: Only scan /agents/ subdirectory, not entire repository
-        This prevents README.md, CHANGELOG.md, etc. from being treated as agents.
+        Supports two cache structures:
+        1. Git repo path: {path}/agents/ - has /agents/ subdirectory
+        2. Flattened cache: {path}/ - directly contains category directories
+
+        Bug #4 Fix: Only scan /agents/ subdirectory when it exists to prevent
+        README.md, CHANGELOG.md, etc. from being treated as agents.
 
         Returns:
             List of agent dictionaries in JSON template format
@@ -162,20 +400,94 @@ class RemoteAgentDiscoveryService:
             )
             return agents
 
-        # Bug #4 Fix: Only scan /agents/ subdirectory, not entire repository
-        # This prevents non-agent files (README.md, CHANGELOG.md, etc.) from polluting results
+        # Support three cache structures (PRIORITY ORDER):
+        # 1. Built output: {path}/dist/agents/ - PREFERRED (built with BASE-AGENT composition)
+        # 2. Git repo path: {path}/agents/ - source files (fallback)
+        # 3. Flattened cache: {path}/ - directly contains category directories (legacy)
+
+        # Priority 1: Check for dist/agents/ (built output with BASE-AGENT composition)
+        dist_agents_dir = self.remote_agents_dir / "dist" / "agents"
         agents_dir = self.remote_agents_dir / "agents"
 
-        if not agents_dir.exists():
-            self.logger.warning(
-                f"Agents subdirectory not found: {agents_dir}. "
-                f"Expected agents to be in /agents/ subdirectory."
+        if dist_agents_dir.exists():
+            # PREFERRED: Use built agents from dist/agents/
+            # These have BASE-AGENT.md files properly composed by build-agent.py
+            self.logger.debug(f"Using built agents from dist: {dist_agents_dir}")
+            scan_dir = dist_agents_dir
+        elif agents_dir.exists():
+            # FALLBACK: Git repo structure - scan /agents/ subdirectory (source files)
+            # This path is used when dist/agents/ hasn't been built yet
+            self.logger.debug(f"Using source agents (no dist/ found): {agents_dir}")
+            scan_dir = agents_dir
+        else:
+            # LEGACY: Flattened cache structure - scan root directly
+            # Check if this looks like the flattened cache (has category subdirectories)
+            category_dirs = [
+                "universal",
+                "engineer",
+                "ops",
+                "qa",
+                "security",
+                "documentation",
+            ]
+            has_categories = any(
+                (self.remote_agents_dir / cat).exists() for cat in category_dirs
             )
-            return agents
 
-        # Find all Markdown files recursively in /agents/ subdirectory only
-        md_files = list(agents_dir.rglob("*.md"))
-        self.logger.debug(f"Found {len(md_files)} Markdown files in {agents_dir}")
+            if has_categories:
+                self.logger.debug(
+                    f"Using flattened cache structure: {self.remote_agents_dir}"
+                )
+                scan_dir = self.remote_agents_dir
+            else:
+                self.logger.warning(
+                    f"No agent directories found. Checked: {dist_agents_dir}, {agents_dir}, "
+                    f"and category directories in {self.remote_agents_dir}. "
+                    f"Expected agents in /dist/agents/, /agents/, or category directories."
+                )
+                return agents
+
+        # Find all Markdown files recursively
+        md_files = list(scan_dir.rglob("*.md"))
+
+        # Filter out non-agent files and git repository files
+        excluded_files = {
+            "README.md",
+            "CHANGELOG.md",
+            "CONTRIBUTING.md",
+            "LICENSE.md",
+            "BASE-AGENT.md",
+            "SUMMARY.md",
+            "IMPLEMENTATION-SUMMARY.md",
+            "REFACTORING_REPORT.md",
+            "REORGANIZATION-PLAN.md",
+            "AUTO-DEPLOY-INDEX.md",
+            "PHASE1_COMPLETE.md",
+            "AGENTS.md",
+        }
+        md_files = [f for f in md_files if f.name not in excluded_files]
+
+        # In flattened cache mode, also exclude files from git repository subdirectories
+        # (files under directories that contain .git folder)
+        if scan_dir == self.remote_agents_dir:
+            filtered_files = []
+            for f in md_files:
+                # Check if this file is inside a git repository (has .git in path)
+                # Git repos are at {remote_agents_dir}/{owner}/{repo}/.git
+                path_parts = f.relative_to(self.remote_agents_dir).parts
+                if len(path_parts) >= 2:
+                    # Check if this looks like a git repo path (owner/repo)
+                    potential_repo = (
+                        self.remote_agents_dir / path_parts[0] / path_parts[1]
+                    )
+                    if (potential_repo / ".git").exists():
+                        # This file is in a git repo, skip it (we'll handle git repos separately)
+                        self.logger.debug(f"Skipping file in git repo: {f}")
+                        continue
+                filtered_files.append(f)
+            md_files = filtered_files
+
+        self.logger.debug(f"Found {len(md_files)} Markdown files in {scan_dir}")
 
         for md_file in md_files:
             try:
@@ -200,8 +512,14 @@ class RemoteAgentDiscoveryService:
     def _parse_markdown_agent(self, md_file: Path) -> Optional[Dict[str, Any]]:
         """Parse Markdown agent file and convert to JSON template format.
 
-        Expected Markdown format:
+        Expected Markdown format with YAML frontmatter:
         ```markdown
+        ---
+        agent_id: python-engineer
+        name: Python Engineer
+        version: 2.3.0
+        model: sonnet
+        ---
         # Agent Name
 
         Description paragraph (first paragraph after heading)
@@ -215,6 +533,11 @@ class RemoteAgentDiscoveryService:
         - Paths: /path1/, /path2/
         ```
 
+        Agent ID Priority (Mismatch Fix):
+        1. Use agent_id from YAML frontmatter if present (e.g., "python-engineer")
+        2. Fall back to leaf filename if no YAML frontmatter (e.g., "python-engineer.md" -> "python-engineer")
+        3. Store hierarchical path separately as category_path for categorization
+
         Args:
             md_file: Path to Markdown agent file
 
@@ -232,22 +555,54 @@ class RemoteAgentDiscoveryService:
             self.logger.error(f"Failed to read file {md_file}: {e}")
             return None
 
-        # Extract agent name from first heading
-        name_match = re.search(r"^#\s+(.+?)$", content, re.MULTILINE)
-        if not name_match:
-            self.logger.debug(f"No agent name heading found in {md_file.name}")
-            return None
-        name = name_match.group(1).strip()
-
-        # Extract description (first paragraph after heading, before next heading)
-        desc_match = re.search(
-            r"^#.+?\n\n(.+?)(?:\n\n##|\Z)", content, re.DOTALL | re.MULTILINE
-        )
-        description = desc_match.group(1).strip() if desc_match else ""
-
-        # Extract model from Configuration section
-        model_match = re.search(r"Model:\s*(\w+)", content, re.IGNORECASE)
-        model = model_match.group(1) if model_match else "sonnet"
+        # MISMATCH FIX: Parse YAML frontmatter to extract agent_id
+        frontmatter = self._parse_yaml_frontmatter(content)
+
+        # MISMATCH FIX: Use agent_id from YAML frontmatter if present, otherwise fall back to filename
+        if frontmatter and "agent_id" in frontmatter:
+            agent_id = frontmatter["agent_id"]
+            self.logger.debug(f"Using agent_id from YAML frontmatter: {agent_id}")
+        else:
+            # Fallback: Use leaf filename without extension
+            agent_id = md_file.stem
+            self.logger.debug(f"No agent_id in YAML, using filename: {agent_id}")
+
+        # Store hierarchical path separately for categorization (not as primary ID)
+        hierarchical_path = self._generate_hierarchical_id(md_file)
+
+        # Extract agent name - prioritize frontmatter over markdown heading
+        # Frontmatter is intentional metadata, headings may be arbitrary content
+        if frontmatter and "name" in frontmatter:
+            name = frontmatter["name"]
+        else:
+            # Fallback to first heading or filename
+            name_match = re.search(r"^#\s+(.+?)$", content, re.MULTILINE)
+            if name_match:
+                name = name_match.group(1).strip()
+            else:
+                # Last resort: derive from filename
+                name = md_file.stem.replace("-", " ").replace("_", " ").title()
+
+        # Extract description - prioritize frontmatter over markdown content
+        # Frontmatter is intentional metadata, paragraphs may be arbitrary content
+        if frontmatter and "description" in frontmatter:
+            description = frontmatter["description"]
+        else:
+            # Fallback to first paragraph after heading
+            desc_match = re.search(
+                r"^#.+?\n\n(.+?)(?:\n\n##|\Z)", content, re.DOTALL | re.MULTILINE
+            )
+            if desc_match:
+                description = desc_match.group(1).strip()
+            else:
+                description = ""
+
+        # Extract model from YAML frontmatter or Configuration section
+        if frontmatter and "model" in frontmatter:
+            model = frontmatter["model"]
+        else:
+            model_match = re.search(r"Model:\s*(\w+)", content, re.IGNORECASE)
+            model = model_match.group(1) if model_match else "sonnet"
 
         # Extract priority from Configuration section
         priority_match = re.search(r"Priority:\s*(\d+)", content, re.IGNORECASE)
@@ -265,27 +620,46 @@ class RemoteAgentDiscoveryService:
         if paths_match:
             paths = [p.strip() for p in paths_match.group(1).split(",")]
 
-        # Get version (SHA-256 hash) from cache metadata
-        version = self._get_agent_version(md_file)
-
-        # Bug #3 fix: Generate hierarchical agent_id from file path
-        # This preserves directory structure for category filtering and preset matching
-        agent_id = self._generate_hierarchical_id(md_file)
+        # Get version (SHA-256 hash) from cache metadata or YAML frontmatter
+        if frontmatter and "version" in frontmatter:
+            version = frontmatter["version"]
+        else:
+            version = self._get_agent_version(md_file)
 
         # Bug #1 fix: Detect category from directory path
         category = self._detect_category_from_path(md_file)
 
+        # NEW: Extract collection metadata from path
+        collection_id = self._extract_collection_id_from_path(md_file)
+        source_path = self._extract_source_path_from_file(md_file)
+
+        # NEW: Generate canonical_id (collection_id:agent_id)
+        # Use leaf agent_id (not hierarchical path) for canonical_id
+        if collection_id:
+            canonical_id = f"{collection_id}:{agent_id}"
+        else:
+            # Fallback for legacy agents without collection
+            canonical_id = f"legacy:{agent_id}"
+
         # Convert to JSON template format and return
         # IMPORTANT: Include 'path' field for compatibility with deployment validation (ticket 1M-480)
         # Git-sourced agents must have 'path' field to match structure from AgentDiscoveryService
         return {
-            "agent_id": agent_id,
+            "agent_id": agent_id,  # MISMATCH FIX: Use leaf name from YAML, not hierarchical path
+            "hierarchical_path": hierarchical_path,  # Store hierarchical path separately
+            "canonical_id": canonical_id,  # NEW: Primary matching key (uses leaf agent_id)
+            "collection_id": collection_id,  # NEW: Collection identifier
+            "source_path": source_path,  # NEW: Path within repository
             "metadata": {
                 "name": name,
                 "description": description,
                 "version": version,
                 "author": "remote",  # Mark as remote agent
                 "category": category,  # Use detected category from path
+                "hierarchical_path": hierarchical_path,  # For categorization/filtering
+                "collection_id": collection_id,  # NEW: Also in metadata
+                "source_path": source_path,  # NEW: Also in metadata
+                "canonical_id": canonical_id,  # NEW: Also in metadata
             },
             "model": model,
             "source": "remote",  # Mark as remote agent
@@ -347,7 +721,12 @@ class RemoteAgentDiscoveryService:
         Returns:
             RemoteAgentMetadata if found, None otherwise
         """
-        for md_file in self.remote_agents_dir.glob("*.md"):
+        # Bug #4 fix: Search in /agents/ subdirectory, not root directory
+        agents_dir = self.remote_agents_dir / "agents"
+        if not agents_dir.exists():
+            return None
+
+        for md_file in agents_dir.rglob("*.md"):
             agent_dict = self._parse_markdown_agent(md_file)
             if agent_dict and agent_dict["metadata"]["name"] == agent_name:
                 return RemoteAgentMetadata(
@@ -359,5 +738,89 @@ class RemoteAgentDiscoveryService:
                     routing_priority=agent_dict["routing"]["priority"],
                     source_file=Path(agent_dict["source_file"]),
                     version=agent_dict["version"],
+                    collection_id=agent_dict.get("collection_id"),
+                    source_path=agent_dict.get("source_path"),
+                    canonical_id=agent_dict.get("canonical_id"),
                 )
         return None
+
+    def get_agents_by_collection(self, collection_id: str) -> List[Dict[str, Any]]:
+        """Get all agents belonging to a specific collection.
+
+        Args:
+            collection_id: Collection identifier in format "owner/repo-name"
+
+        Returns:
+            List of agent dictionaries from the specified collection
+
+        Example:
+            >>> service = RemoteAgentDiscoveryService(Path("~/.claude-mpm/cache/remote-agents"))
+            >>> agents = service.get_agents_by_collection("bobmatnyc/claude-mpm-agents")
+            >>> len(agents)
+            45
+        """
+        all_agents = self.discover_remote_agents()
+
+        # Filter by collection_id
+        collection_agents = [
+            agent for agent in all_agents if agent.get("collection_id") == collection_id
+        ]
+
+        self.logger.info(
+            f"Found {len(collection_agents)} agents in collection '{collection_id}'"
+        )
+
+        return collection_agents
+
+    def list_collections(self) -> List[Dict[str, Any]]:
+        """List all available collections with agent counts.
+
+        Returns:
+            List of collection info dictionaries with:
+            - collection_id: Collection identifier
+            - agent_count: Number of agents in collection
+            - agents: List of agent IDs in collection
+
+        Example:
+            >>> service = RemoteAgentDiscoveryService(Path("~/.claude-mpm/cache/remote-agents"))
+            >>> collections = service.list_collections()
+            >>> collections
+            [
+                {
+                    "collection_id": "bobmatnyc/claude-mpm-agents",
+                    "agent_count": 45,
+                    "agents": ["pm", "engineer", "qa", ...]
+                }
+            ]
+        """
+        all_agents = self.discover_remote_agents()
+
+        # Group by collection_id
+        collections_map: Dict[str, List[str]] = {}
+
+        for agent in all_agents:
+            collection_id = agent.get("collection_id")
+            if not collection_id:
+                # Skip agents without collection (legacy)
+                continue
+
+            if collection_id not in collections_map:
+                collections_map[collection_id] = []
+
+            agent_id = agent.get("agent_id", agent.get("metadata", {}).get("name"))
+            if agent_id:
+                collections_map[collection_id].append(agent_id)
+
+        # Convert to list format
+        collections = [
+            {
+                "collection_id": coll_id,
+                "agent_count": len(agent_ids),
+                "agents": sorted(agent_ids),
+            }
+            for coll_id, agent_ids in collections_map.items()
+        ]
+
+        self.logger.info(f"Found {len(collections)} collections")
+
+        return collections