headquarter/apps/api/src/services/tool/instance_service.py

"""Tool instance service functions."""

import asyncio
import contextlib
import glob as glob_module
import logging
import os
import subprocess
import uuid
from datetime import datetime

from fastapi import HTTPException, status
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession

from src.models import (
    ConfigProfile,
    GitRepository,
    Project,
    SSHKey,
    ToolInstance,
    ToolType,
)
from src.schemas.tool import CreateInstanceRequest, StartInstanceRequest
from src.services.git.clone import check_dirty_state, clone_repository
from src.services.config.config_profile_resolver import (
    ConfigProfileCycleError,
    ResolvedProfile,
    apply_resolved_profile,
    expand_container_path,
    resolve_profile,
)
from src.services.docker import (
    connect_container_to_network,
    ensure_instance_directory,
    execute_compose_command,
    find_free_port,
    get_backend_network_name,
    get_container_id,
    get_container_ip_on_network,
    get_container_logs,
    get_container_status,
    is_container_on_network,
    render_compose_template,
    sort_volumes_by_specificity,
    wait_for_container_running,
    write_compose_file,
    write_config_files,
    write_env_file,
)
from src.services.shared.tunnel import (
    check_tunnel_health,
    recreate_tunnel,
    start_tunnel,
    stop_tunnel,
)
from src.services.build.docker_build import build_image
from src.services.build.manifest_compiler import (
    compile_compose,
    compile_dockerfile,
    compile_entrypoint,
    compute_image_tag,
    deep_merge,
    get_manifest_home_dir,
    merge_with_config,
    resolve_base,
)
from src.services.shared.permission_fixer import (
    apply_mount_permissions,
    apply_ssh_permissions,
)
from src.services.shared.readiness_probe import execute_probe
from src.services.shared.ssh_keys import prepare_ssh_key_files
from src.services.instance.event_bus import InstanceEventBus
from src.services.instance.lifecycle_hooks import publish_lifecycle_event

logger = logging.getLogger(__name__)
_event_bus = InstanceEventBus()


def _chown_path(path: str, uid: int, gid: int) -> None:
    """Recursively chown a path, suppressing permission errors."""
    try:
        if os.path.isdir(path):
            for root, dirs, files in os.walk(path):
                for name in dirs + files:
                    full = os.path.join(root, name)
                    with contextlib.suppress(OSError):
                        os.chown(full, uid, gid)
        with contextlib.suppress(OSError):
            os.chown(path, uid, gid)
    except Exception as exc:
        logger.warning("Failed to chown %s to %s:%s: %s", path, uid, gid, exc)


def _chown_staged_mounts(
    extra_volumes: list[dict],
    instance_dir: str,
    uid: int,
    gid: int,
) -> None:
    """Recursively chown staged mount sources to the container user.

    Config-profile mounts, git mounts, and SSH key mounts are staged under
    instance_dir by the API process (root). Without this, the container
    user cannot write into bind-mounted directories such as ~/.config.
    """
    for vol in extra_volumes:
        source = vol.get("source", "")
        if not source or not source.startswith(instance_dir):
            continue
        _chown_path(source, uid, gid)


async def resolve_git_mounts(
    session: AsyncSession,
    resolved: ResolvedProfile,
    instance_dir: str | None = None,
    working_directory: str | None = None,
    home_dir: str = "/root",
) -> list[dict]:
    """Convert git mounts from resolved profile to Docker volume mounts.

    Looks up repository paths, auto-clones if needed, handles branch checkout,
    expands glob patterns, and prepares bind mount entries.
    Logs warnings for missing repos or invalid paths (non-blocking).
    """
    if not resolved.git_mounts:
        return []

    # Process all git mounts concurrently
    tasks = []
    for git_mount in resolved.git_mounts:
        tasks.append(
            resolve_single_git_mount(
                session, git_mount, instance_dir, working_directory, home_dir
            )
        )

    results = await asyncio.gather(*tasks, return_exceptions=True)

    volume_mounts = []
    for result in results:
        if isinstance(result, Exception):
            logger.warning("Git mount failed: %s", result)
            continue
        if isinstance(result, list):
            volume_mounts.extend(result)

    return volume_mounts


def normalize_git_mount(entry: dict) -> dict:
    """Normalize a git mount entry to the unified mappings format.

    Converts legacy source_path + target_path into a single-entry mappings array.
    """
    entry = dict(entry)
    if "mappings" not in entry or not entry.get("mappings"):
        source = entry.get("source_path", ".")
        target = entry.get("target_path")
        if target is not None:
            entry["mappings"] = [{"source_path": source, "target_path": target}]
    entry.pop("source_path", None)
    entry.pop("target_path", None)
    return entry


def clone_git_repo(
    remote_url: str,
    branch: str | None,
    clone_parent: str,
) -> str:
    """Clone or pull a git repository.

    Returns the path to the cloned repo (repo-clone directory).
    """
    import hashlib

    url_hash = hashlib.md5(remote_url.encode()).hexdigest()[:12]
    repo_name = remote_url.split("/")[-1].replace(".git", "") or "repo"
    clone_dir = os.path.join(clone_parent, "git-mounts", f"{repo_name}-{url_hash}")
    repo_path = os.path.join(clone_dir, "repo-clone")

    if not os.path.exists(repo_path):
        try:
            os.makedirs(clone_dir, exist_ok=True)
            repo_path = clone_repository(
                remote_url,
                None,  # No SSH key for now - can be added later
                clone_dir,
                branch or "main",
            )
            logger.debug("Cloned git mount repository %s to %s", remote_url, repo_path)
        except Exception as exc:
            logger.warning("Clone failed for git mount %s: %s", remote_url, exc)
            raise
    else:
        # Repo exists - pull latest updates
        try:
            pull_repository_updates(repo_path, remote_url)
            logger.debug("Pulled updates for git mount %s", remote_url)
        except Exception as exc:
            logger.warning("Failed to pull updates for %s: %s", remote_url, exc)

    # Handle branch checkout if specified
    if branch and repo_path:
        success = checkout_branch(repo_path, branch)
        if success:
            logger.debug("Checked out branch %s for %s", branch, remote_url)
        else:
            logger.warning(
                "Branch %s not found in %s, using current branch", branch, remote_url
            )

    return repo_path


def resolve_git_mount_mappings(
    repo_path: str,
    mappings: list[dict],
    working_directory: str | None,
    home_dir: str = "/root",
) -> list[dict]:
    """Resolve mappings from an already-cloned repo to volume mount entries.

    Returns a flat list of volume mount dicts.
    """
    volume_mounts = []

    for mapping in mappings:
        source_path = mapping.get("source_path", ".")
        target_path = mapping.get("target_path")

        if not target_path:
            logger.warning("Invalid mapping skipped: missing target_path")
            continue

        # Expand ~ and $HOME in target path
        target_path = expand_container_path(target_path, home_dir)

        # Resolve relative target paths against working directory
        final_target = target_path
        if not target_path.startswith("/"):
            if not working_directory:
                logger.warning(
                    "Git mount skipped: target_path '%s' is relative but no working_directory is configured. "
                    "Set working_directory in the tool config or use an absolute path.",
                    target_path,
                )
                continue
            final_target = os.path.join(working_directory, target_path)

        # Build source path and expand globs
        if source_path and source_path != ".":
            source_full = os.path.join(repo_path, source_path)
        else:
            source_full = repo_path

        # Expand glob patterns
        matched_paths = expand_glob_source(source_full, repo_path)

        if not matched_paths:
            logger.warning(
                "Git mount skipped: no files matched source path %s in repo",
                source_path,
            )
            continue

        for matched_path in matched_paths:
            if not os.path.exists(matched_path):
                continue

            # Determine target path for this match
            if len(matched_paths) == 1:
                # Single match: mount directly to target_path
                mount_target = final_target
            else:
                # Multiple matches: append relative path to target
                rel_path = os.path.relpath(matched_path, repo_path)
                mount_target = os.path.join(final_target, rel_path)

            volume_mounts.append(
                {
                    "source": matched_path,
                    "target": mount_target,
                    "type": "bind",
                }
            )
            logger.debug(
                "Added git mount: %s -> %s",
                matched_path,
                mount_target,
            )

    return volume_mounts


async def resolve_single_git_mount(
    session: AsyncSession,
    git_mount: dict,
    instance_dir: str | None = None,
    working_directory: str | None = None,
    home_dir: str = "/root",
) -> list[dict]:
    """Resolve a single git mount to volume mount entries.

    Clones directly from remote_url, no database lookup needed.
    Returns a list of volume mounts (one for each matched file/directory).
    """
    git_mount = normalize_git_mount(git_mount)
    remote_url = git_mount.get("remote_url")
    branch = git_mount.get("branch")
    mappings = git_mount.get("mappings", [])

    if not remote_url:
        logger.warning("Invalid git mount skipped: missing remote_url")
        return []

    if not mappings:
        logger.warning("Invalid git mount skipped: no mappings")
        return []

    if not instance_dir:
        logger.warning("Git mount skipped: no instance_dir provided for cloning")
        return []

    # Clone or pull the repository
    try:
        repo_path = await asyncio.to_thread(
            clone_git_repo, remote_url, branch, instance_dir
        )
    except Exception:
        return []

    # Resolve all mappings from the cloned repo
    return resolve_git_mount_mappings(repo_path, mappings, working_directory, home_dir)


def checkout_branch(repo_path: str, branch: str) -> bool:
    """Checkout a specific branch in a git repository.

    Returns True if checkout succeeded, False if it failed.
    On failure, the repository remains on its current branch.
    """
    import subprocess

    # First try to checkout existing branch
    result = subprocess.run(
        ["git", "-C", repo_path, "checkout", branch],
        capture_output=True,
        text=True,
    )

    if result.returncode != 0:
        # Try fetching and checking out
        subprocess.run(
            ["git", "-C", repo_path, "fetch", "origin", branch],
            capture_output=True,
            text=True,
        )
        result = subprocess.run(
            ["git", "-C", repo_path, "checkout", "-b", branch, f"origin/{branch}"],
            capture_output=True,
            text=True,
        )

        if result.returncode != 0:
            logger.warning(
                "Failed to checkout branch %s in %s: %s",
                branch,
                repo_path,
                result.stderr.strip(),
            )
            return False

    return True


def pull_repository_updates(repo_path: str, remote_url: str) -> None:
    """Pull latest updates from remote repository.

    Used when starting a new container with an existing cloned repository
    to ensure the latest code is mounted.
    """
    import subprocess

    # Fetch latest changes
    result = subprocess.run(
        ["git", "-C", repo_path, "fetch", "origin"],
        capture_output=True,
        text=True,
    )

    if result.returncode != 0:
        raise RuntimeError(f"Failed to fetch updates: {result.stderr}")

    # Pull changes for current branch
    result = subprocess.run(
        ["git", "-C", repo_path, "pull", "origin"],
        capture_output=True,
        text=True,
    )

    if result.returncode != 0:
        raise RuntimeError(f"Failed to pull updates: {result.stderr}")


def expand_glob_source(source_path: str, repo_path: str) -> list[str]:
    """Expand glob patterns in source path.

    Returns a list of matched absolute paths.
    Limits results to prevent abuse.
    """
    MAX_GLOB_MATCHES = 100

    # Check if path contains glob characters
    if not any(c in source_path for c in "*?["):
        # No glob pattern: return single path if it exists
        return [source_path] if os.path.exists(source_path) else []

    # Expand glob pattern
    matched = glob_module.glob(source_path, recursive=True)
    total_matched = len(matched)

    # Filter to only paths within the repo and limit count
    results = []
    for path in matched:
        abs_path = os.path.abspath(path)
        if abs_path.startswith(os.path.abspath(repo_path)):
            results.append(abs_path)
        if len(results) >= MAX_GLOB_MATCHES:
            logger.warning(
                "Glob pattern matched %d files, limited to %d",
                total_matched,
                MAX_GLOB_MATCHES,
            )
            break

    return results


async def validate_config_profile(
    session: AsyncSession,
    profile_id: str | None,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    tool_type_id: uuid.UUID,
) -> uuid.UUID | None:
    """Validate a config profile selection.

    Args:
        session: Database session.
        profile_id: Profile ID string or None.
        user_id: Authenticated user ID.
        project_id: Project ID for compatibility check.
        tool_type_id: Tool type ID for compatibility check.

    Returns:
        Validated UUID or None.

    Raises:
        HTTPException: If profile is not found, not owned, or incompatible.
    """
    if not profile_id:
        return None

    try:
        profile_uuid = uuid.UUID(profile_id)
    except ValueError:
        raise HTTPException(
            status_code=status.HTTP_400_BAD_REQUEST,
            detail=f"Invalid config profile ID: {profile_id}",
        )

    profile = await session.get(ConfigProfile, profile_uuid)
    if profile is None:
        raise HTTPException(
            status_code=status.HTTP_404_NOT_FOUND,
            detail=f"Config profile not found: {profile_id}",
        )

    if profile.user_id != user_id:
        raise HTTPException(
            status_code=status.HTTP_403_FORBIDDEN,
            detail="Not authorized to use this config profile",
        )

    # Check compatibility: profile must be portable or match project/tool
    is_compatible = (
        (profile.project_id is None and profile.tool_type_id is None)
        or (profile.project_id == project_id)
        or (profile.tool_type_id == tool_type_id)
        or (profile.project_id == project_id and profile.tool_type_id == tool_type_id)
    )

    if not is_compatible:
        raise HTTPException(
            status_code=status.HTTP_400_BAD_REQUEST,
            detail="Selected config profile is not compatible with this project and tool type",
        )

    return profile_uuid


def sanitize_compose_file(compose_path: str) -> None:
    """Remove invalid port mappings (target port 0) from compose file."""
    import yaml
    from pathlib import Path

    compose_file = Path(compose_path)
    if not compose_file.exists():
        return

    content = compose_file.read_text()
    compose_data = yaml.safe_load(content)

    if not compose_data or "services" not in compose_data:
        return

    modified = False
    for service_name, service_config in compose_data["services"].items():
        if "ports" in service_config:
            valid_ports = []
            for port_mapping in service_config["ports"]:
                if isinstance(port_mapping, str) and ":" in port_mapping:
                    parts = port_mapping.split(":")
                    if len(parts) == 2:
                        host_port, container_port = parts
                        # Skip invalid mappings (target port 0 or empty)
                        if container_port == "0" or not container_port:
                            modified = True
                            continue
                valid_ports.append(port_mapping)

            if valid_ports:
                service_config["ports"] = valid_ports
            else:
                del service_config["ports"]
                modified = True
        break  # Only check first service

    if modified:
        compose_file.write_text(yaml.dump(compose_data, default_flow_style=False))


def modify_compose_file(
    compose_path: str,
    port_override: int | None = None,
    start_command: str | None = None,
    working_directory: str | None = None,
    extra_volumes: list[dict] | None = None,
    home_dir: str = "/root",
) -> None:
    """Modify compose file with runtime overrides."""
    import yaml
    from pathlib import Path

    compose_file = Path(compose_path)
    content = compose_file.read_text()
    compose_data = yaml.safe_load(content)

    if not compose_data or "services" not in compose_data:
        return

    # Apply modifications to the first service
    for service_name, service_config in compose_data["services"].items():
        if port_override and "ports" in service_config:
            # Update port mapping
            for i, port_mapping in enumerate(service_config["ports"]):
                if isinstance(port_mapping, str) and ":" in port_mapping:
                    host_port, container_port = port_mapping.split(":", 1)
                    service_config["ports"][i] = f"{port_override}:{container_port}"
                    break

        if start_command:
            service_config["command"] = start_command

        if working_directory:
            service_config["working_dir"] = expand_container_path(
                working_directory, home_dir
            )

        if extra_volumes:
            if "volumes" not in service_config:
                service_config["volumes"] = []
            for vol in extra_volumes:
                source = vol.get("source", "")
                target = expand_container_path(vol.get("target", ""), home_dir)
                vol_type = vol.get("type", "bind")
                readonly = ":ro" if vol.get("readonly") else ""
                if vol_type == "bind":
                    service_config["volumes"].append(f"{source}:{target}{readonly}")
                else:
                    service_config["volumes"].append(
                        f"{source}:{target}:{vol_type}{readonly}"
                    )

        # Sort volumes so parent paths come before child paths
        if service_config.get("volumes"):
            service_config["volumes"] = sort_volumes_by_specificity(
                service_config["volumes"]
            )

        break  # Only modify the first service

    # Write back
    compose_file.write_text(yaml.dump(compose_data, default_flow_style=False))


def ensure_container_name_in_compose(compose_path: str, container_name: str) -> None:
    """Ensure compose file has explicit container_name for predictable naming.

    Docker Compose auto-generates container names from the project directory
    when container_name is absent. This breaks tunnel connectivity because
    get_container_name(instance.name) cannot find the container. We inject
    container_name into every service so the container has a predictable name.
    """
    import yaml
    from pathlib import Path

    compose_file = Path(compose_path)
    if not compose_file.exists():
        return

    content = compose_file.read_text()
    compose_data = yaml.safe_load(content)

    if not compose_data or "services" not in compose_data:
        return

    modified = False
    for svc_name, svc_config in compose_data["services"].items():
        if "container_name" not in svc_config:
            svc_config["container_name"] = container_name.lower()
            modified = True

    if modified:
        compose_file.write_text(yaml.dump(compose_data, default_flow_style=False))
        logger.info(
            "Injected container_name '%s' into compose file",
            container_name.lower(),
        )


def ensure_web_bind_address(
    compose_path: str, tool_type_name: str, default_port: int
) -> None:
    """Auto-inject bind address for known web tools that default to 127.0.0.1.

    Many web tools (code-server, jupyter) bind to localhost by default,
    making them inaccessible from the Docker network. This function detects
    known tool images and injects the correct --bind-addr or --ip flag.
    """
    import yaml
    from pathlib import Path

    if default_port <= 0:
        return

    KNOWN_BIND_FIXES: dict[str, str] = {
        "code-server": f"--bind-addr 0.0.0.0:{default_port}",
        "jupyter-notebook": f"start-notebook.sh --ip=0.0.0.0 --port={default_port} --no-browser",
    }

    bind_command = KNOWN_BIND_FIXES.get(tool_type_name)
    if not bind_command:
        return

    compose_file = Path(compose_path)
    if not compose_file.exists():
        return

    content = compose_file.read_text()
    compose_data = yaml.safe_load(content)

    if not compose_data or "services" not in compose_data:
        return

    for service_config in compose_data["services"].values():
        image = service_config.get("image", "")
        if not image:
            continue

        # LSIO images already bind to 0.0.0.0 — command override breaks s6 init
        if "linuxserver" in image:
            existing_command = service_config.get("command", "")
            if "--bind-addr" in existing_command or "--host" in existing_command:
                del service_config["command"]
                compose_file.write_text(
                    yaml.dump(compose_data, default_flow_style=False)
                )
                logger.warning(
                    "Removed broken command override from LSIO image: %s",
                    existing_command,
                )
                return
            return

        # Check if the image matches a known tool
        is_code_server = tool_type_name == "code-server" and (
            "code-server" in image or "coder" in image
        )
        is_jupyter = tool_type_name == "jupyter-notebook" and (
            "jupyter" in image or "notebook" in image
        )
        if not is_code_server and not is_jupyter:
            continue

        existing_command = service_config.get("command", "")
        if existing_command:
            # Already correct — nothing to do
            if bind_command in existing_command:
                return
            # Fix broken or outdated bind flags
            if (
                "--bind-addr" in existing_command
                or "--host" in existing_command
                or "--ip=" in existing_command
            ):
                service_config["command"] = bind_command
                compose_file.write_text(
                    yaml.dump(compose_data, default_flow_style=False)
                )
                logger.warning(
                    "Replaced broken bind address for %s: %s → %s",
                    tool_type_name,
                    existing_command,
                    bind_command,
                )
                return
            # Some other command override exists — don't touch it
            return

        # No command yet — inject the correct bind address
        service_config["command"] = bind_command
        compose_file.write_text(yaml.dump(compose_data, default_flow_style=False))
        logger.info("Injected bind address for %s: %s", tool_type_name, bind_command)
        return


def ensure_backend_network_in_compose(compose_path: str) -> None:
    """Inject the backend network into the compose file so compose up attaches it.

    Instead of running 'docker network connect' after container creation (which
    is prone to race conditions and silent failures), we declare the network in
    the compose file itself. Docker Compose then connects the container to the
    network atomically during 'docker compose up'.
    """
    import yaml
    from pathlib import Path

    compose_file = Path(compose_path)
    if not compose_file.exists():
        return

    content = compose_file.read_text()
    compose_data = yaml.safe_load(content)

    if not compose_data or "services" not in compose_data:
        return

    network_name = get_backend_network_name()
    modified = False

    for svc_config in compose_data["services"].values():
        existing = svc_config.get("networks", [])
        if network_name not in existing:
            svc_config["networks"] = existing + [network_name]
            modified = True
        break  # Only modify first service

    # Declare the network as external at the top level
    if "networks" not in compose_data:
        compose_data["networks"] = {}
    if network_name not in compose_data["networks"]:
        compose_data["networks"][network_name] = {"external": True}
        modified = True

    if modified:
        compose_file.write_text(yaml.dump(compose_data, default_flow_style=False))
        logger.info("Injected backend network '%s' into compose file", network_name)


async def prepare_manifest_instance(
    session: AsyncSession,
    instance: ToolInstance,
    instance_dir: str,
    repo_path: str,
    env_vars: dict,
    extra_volumes: list,
    working_directory: str | None,
) -> tuple[str, str, dict, str]:
    """Build image and generate compose from a manifest-based tool type.

    Returns:
        Tuple of (image_tag, compose_content, resolved_manifest, home_dir)
    """
    from src.models import ToolDefinitionManifest

    tool_type = await session.get(ToolType, instance.tool_type_id)
    if not tool_type:
        raise RuntimeError(f"Tool type not found for instance {instance.id}")

    manifest_def = await session.get(ToolDefinitionManifest, tool_type.manifest_id)
    if not manifest_def:
        raise RuntimeError(f"Manifest not found for tool type {tool_type.id}")

    manifest = dict(manifest_def.manifest)

    # Resolve base if referenced
    if manifest_def.base_definition_id:
        base_def = await session.get(
            ToolDefinitionManifest, manifest_def.base_definition_id
        )
        if base_def:
            base_manifest = dict(base_def.manifest)
            manifest = resolve_base(deep_merge(base_manifest, manifest))
        else:
            logger.warning(
                "Base definition %s not found for manifest %s",
                manifest_def.base_definition_id,
                manifest_def.id,
            )

    manifest = merge_with_config(manifest)

    # Resolve extra env and volumes from merge_with_config
    extra_env = manifest.pop("_extra_env", {})
    extra_cfg_volumes = manifest.pop("_extra_volumes", [])
    env_vars.update(extra_env)
    extra_volumes.extend(extra_cfg_volumes)

    # Compute image tag
    image_tag = compute_image_tag(tool_type.name, manifest)

    # Check if image already exists
    check = subprocess.run(
        ["docker", "images", "-q", image_tag],
        capture_output=True,
        text=True,
    )
    image_exists = check.returncode == 0 and check.stdout.strip()

    if not image_exists:
        # Compile and build
        dockerfile = compile_dockerfile(manifest)
        entrypoint = compile_entrypoint(manifest)

        logger.debug(
            "Compiled Dockerfile for instance %s (%d chars)",
            instance.id,
            len(dockerfile),
        )

        build_ctx = {
            "Dockerfile": dockerfile,
            ".headquarter/entrypoint.sh": entrypoint,
        }

        returncode, stdout, stderr = await asyncio.to_thread(
            build_image,
            instance_dir=instance_dir,
            dockerfile=dockerfile,
            tag=image_tag,
            build_context=build_ctx,
        )

        if returncode != 0:
            raise RuntimeError(f"Docker build failed: {stderr}")

        logger.info("Built image %s for instance %s", image_tag, instance.id)
    else:
        logger.info("Reusing existing image %s for instance %s", image_tag, instance.id)

    # Prepare SSH path for mount resolution
    ssh_path = ""

    # Resolve git mount variables from config profile
    git_mount_vars = {}
    if instance.selected_config_profile_id:
        resolved_profile = await resolve_profile(
            session, instance.selected_config_profile_id
        )
        for gm in resolved_profile.git_mounts or []:
            ref = gm.get("git_mount_ref", "default")
            # The actual resolution happens in resolve_git_mounts; we store placeholder
            git_mount_vars[f"GIT_MOUNT_{ref}"] = ""

    repo_name = os.path.basename(os.path.normpath(repo_path))
    variables = {
        "IMAGE_TAG": image_tag,
        "INSTANCE_NAME": instance.name.lower(),
        "INSTANCE_DIR": instance_dir,
        "WORKSPACE_PATH": repo_path,
        "REPO_PATH": repo_path,
        "REPO_NAME": repo_name,
        "WORKSPACE_NAME": repo_name,
        "SSH_PATH": ssh_path,
        "TOOL_PORT": instance.port or 0,
        "EXTRA_ENV": env_vars,
        "EXTRA_VOLUMES": extra_volumes,
        **git_mount_vars,
    }

    compose_content = compile_compose(manifest, variables)

    logger.debug(
        "_prepare_manifest_instance for %s: repo_path=%s compose_volumes=%s",
        instance.id,
        repo_path or "<empty>",
        manifest.get("mounts", []),
    )
    logger.debug(
        "Generated compose for %s:\n%s",
        instance.id,
        compose_content,
    )

    # Cache
    instance.image_tag = image_tag
    instance.manifest_compiled_at = datetime.now()

    home_dir = get_manifest_home_dir(manifest)
    return image_tag, compose_content, manifest, home_dir


async def create_tool_instance(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    data: "CreateInstanceRequest",
) -> ToolInstance:
    """Create a new tool instance for a repository.

    Returns the created ToolInstance.
    Raises ValueError for invalid input, RuntimeError for internal failures.
    """
    repo = await session.get(GitRepository, repo_id)
    if repo is None or repo.project_id != project_id:
        raise ValueError("repository not found")

    tool_type_id = uuid.UUID(data.tool_type_id)
    tool_type = await session.get(ToolType, tool_type_id)
    if tool_type is None:
        raise ValueError("tool type not found")

    # Validate config profile if provided
    selected_profile_id = await validate_config_profile(
        session, data.config_profile_id, user_id, project_id, tool_type_id
    )

    # Resolve workspace if provided
    workspace = None
    workspace_id = None
    if data.workspace_id:
        from src.models import Workspace as WorkspaceModel

        try:
            workspace_id = uuid.UUID(data.workspace_id)
        except ValueError:
            raise ValueError("Invalid workspace_id format")
        workspace = await session.get(WorkspaceModel, workspace_id)
        if workspace is None:
            raise ValueError("workspace not found")
        if workspace.repo_id != repo_id:
            raise ValueError("workspace does not belong to this repository")

    # Generate unique name
    instance_name = f"{tool_type.name}-{repo.name}-{uuid.uuid4().hex[:8]}"

    # Auto-generate display name with scoped numbering.
    if data.display_name:
        instance_display = data.display_name
    else:
        project = await session.get(Project, project_id)
        project_name = project.name if project else "Unknown"
        scope_name = workspace.name if workspace else repo.name
        auto_name = f"{project_name} / {scope_name} / {tool_type.display_name}"

        if workspace:
            count_query = (
                select(ToolInstance)
                .where(ToolInstance.workspace_id == workspace_id)
                .where(ToolInstance.tool_type_id == tool_type_id)
                .where(ToolInstance.owner_id == user_id)
            )
        else:
            count_query = (
                select(ToolInstance)
                .where(ToolInstance.repository_id == repo_id)
                .where(ToolInstance.tool_type_id == tool_type_id)
                .where(ToolInstance.owner_id == user_id)
            )

        result = await session.execute(count_query)
        existing_count = len(result.scalars().all())
        if existing_count > 0:
            instance_display = f"{auto_name} #{existing_count + 1}"
        else:
            instance_display = auto_name

    # Create instance directory
    instance_dir = ensure_instance_directory(instance_name)
    compose_path = os.path.join(instance_dir, "docker-compose.yml")

    # Find free port
    tool_port = find_free_port()

    # Determine the source path to mount. Workspaces are the canonical path;
    # legacy instances without a workspace fall back to the repository path.
    if workspace:
        repo_path = workspace.path
    else:
        repo_path = repo.path

    # Handle based on definition type
    if tool_type.definition_type == "dockerfile":
        image_tag = f"headquarter/{instance_name}:latest".lower()

        if tool_type.dockerfile_template:
            returncode, stdout, stderr = await asyncio.to_thread(
                build_image,
                instance_dir=instance_dir,
                dockerfile=tool_type.dockerfile_template,
                tag=image_tag,
                build_context=tool_type.build_context,
            )

            if returncode != 0:
                logger.error(
                    "Failed to build image for instance %s: %s",
                    instance_name,
                    stderr,
                )
                raise RuntimeError(f"Failed to build Docker image: {stderr[:500]}")

            logger.info(
                "Successfully built image %s for instance %s",
                image_tag,
                instance_name,
            )

        ports_section = (
            f"""    ports:\n      - "{tool_port}:{tool_type.default_port}"\n"""
            if tool_type.default_port and tool_type.default_port > 0
            else ""
        )

        home_dir = tool_type.home_directory or "/home/user"
        repo_name = os.path.basename(os.path.normpath(repo_path))
        workspace_target = f"{home_dir}/{repo_name}"

        compose_content = f"""version: "3.8"\nservices:
  app:
    image: {image_tag}
    container_name: {instance_name.lower()}
    stdin_open: true
    tty: true
{ports_section}    environment:
      - HOME={home_dir}
    volumes:
      - {repo_path}:{workspace_target}
    working_dir: {workspace_target}
    restart: "no"
"""
        write_compose_file(instance_dir, compose_content)

    elif tool_type.definition_type == "manifest":
        from src.models import ToolDefinitionManifest

        manifest_def = await session.get(ToolDefinitionManifest, tool_type.manifest_id)
        if not manifest_def:
            raise RuntimeError("Manifest definition not found for this tool type")

        manifest = dict(manifest_def.manifest)
        if manifest_def.base_definition_id:
            base_def = await session.get(
                ToolDefinitionManifest, manifest_def.base_definition_id
            )
            if base_def:
                manifest = resolve_base(deep_merge(dict(base_def.manifest), manifest))

        image_tag = compute_image_tag(tool_type.name, manifest)

        # Manifest templates use WORKSPACE_PATH; REPO_PATH is retained as a
        # deprecated alias for backward compatibility with older templates.
        repo_name = os.path.basename(os.path.normpath(repo_path))
        variables = {
            "IMAGE_TAG": image_tag,
            "INSTANCE_NAME": instance_name.lower(),
            "INSTANCE_DIR": instance_dir,
            "WORKSPACE_PATH": repo_path,
            "REPO_PATH": repo_path,
            "REPO_NAME": repo_name,
            "WORKSPACE_NAME": repo_name,
            "SSH_PATH": "",
            "TOOL_PORT": tool_port,
            "EXTRA_ENV": {},
            "EXTRA_VOLUMES": [],
        }
        compose_content = compile_compose(manifest, variables)
        write_compose_file(instance_dir, compose_content)

    else:
        if not tool_type.compose_template:
            raise ValueError("Tool type has no compose template configured")
        variables = {
            "WORKSPACE_PATH": repo_path,
            "REPO_PATH": repo_path,
            "INSTANCE_NAME": instance_name,
            "INSTANCE_ID": instance_name,
            "TOOL_NAME": instance_name,
            "TOOL_PORT": tool_port,
            "USER_ID": str(user_id),
            "PROJECT_ID": str(project_id),
            "WORKSPACE_NAME": os.path.basename(os.path.normpath(repo_path)),
            "HOME_DIRECTORY": tool_type.home_directory or "/home/user",
        }
        compose_content = render_compose_template(tool_type.compose_template, variables)

        write_compose_file(instance_dir, compose_content)

    # Create database record
    instance = ToolInstance(
        name=instance_name,
        display_name=instance_display,
        tool_type_id=tool_type_id,
        repository_id=repo_id,
        project_id=project_id,
        owner_id=user_id,
        status="pending",
        compose_path=compose_path,
        port=tool_port,
        workspace_id=workspace_id,
        clone_mode=None,
        branch=None,
        selected_config_profile_id=selected_profile_id,
        ssh_key_ids=data.ssh_key_ids or None,
    )
    session.add(instance)
    await session.commit()
    await session.refresh(instance)
    await publish_lifecycle_event(
        event_bus=_event_bus,
        session=session,
        instance=instance,
        event_type="instance.created",
        created_by=user_id,
        status="pending",
        message="Instance created",
    )

    return instance


async def start_tool_instance(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    instance_id: uuid.UUID,
    data: "StartInstanceRequest | None",
) -> dict:
    """Start a tool instance.

    Returns a dict with status and url.
    Raises ValueError for invalid input, RuntimeError for internal failures.
    """
    instance = await session.get(ToolInstance, instance_id)
    if instance is None or instance.repository_id != repo_id:
        raise ValueError("instance not found")

    # Validate and store config profile selection
    if data and data.config_profile_id is not None:
        selected_profile_id = await validate_config_profile(
            session, data.config_profile_id, user_id, project_id, instance.tool_type_id
        )
        instance.selected_config_profile_id = selected_profile_id
        await session.commit()

    # Store SSH key selection if provided
    if data and data.ssh_key_ids is not None:
        instance.ssh_key_ids = data.ssh_key_ids or None
        await session.commit()

    if not instance.compose_path or not os.path.exists(instance.compose_path):
        raise ValueError("compose file not found")

    instance.status = "building"
    await session.commit()
    logger.info("Starting instance %s (name=%s)", instance.id, instance.name)

    # Runtime overrides populated by config profiles
    env_vars = {}
    config_files = {}
    port_override = None
    start_command = None
    working_directory = None
    extra_volumes = []

    # Fetch tool type early to determine home directory and container user
    tool_type = await session.get(ToolType, instance.tool_type_id)
    home_dir = "/root"
    container_uid = 0
    container_gid = 0
    if tool_type and tool_type.definition_type == "manifest" and tool_type.manifest_id:
        from src.models import ToolDefinitionManifest

        manifest_def = await session.get(ToolDefinitionManifest, tool_type.manifest_id)
        if manifest_def:
            manifest = dict(manifest_def.manifest)
            if manifest_def.base_definition_id:
                base_def = await session.get(
                    ToolDefinitionManifest, manifest_def.base_definition_id
                )
                if base_def:
                    manifest = resolve_base(
                        deep_merge(dict(base_def.manifest), manifest)
                    )
            home_dir = get_manifest_home_dir(manifest)
            user_cfg = manifest.get("user")
            if user_cfg:
                container_uid = user_cfg.get("uid", 0)
                container_gid = user_cfg.get("gid", 0)
            logger.debug(
                "Manifest user resolved for instance %s: uid=%s, gid=%s, home=%s",
                instance.id,
                container_uid,
                container_gid,
                home_dir,
            )

    # Apply selected config profile if any
    instance_dir = os.path.dirname(instance.compose_path)
    if instance.selected_config_profile_id is not None:
        try:
            resolved = await resolve_profile(
                session, instance.selected_config_profile_id
            )
            profile_env, profile_files, profile_mounts, profile_hints = (
                apply_resolved_profile(instance_dir, resolved, home_dir)
            )
            env_vars.update(profile_env)
            config_files.update(profile_files)
            extra_volumes.extend(profile_mounts)
            git_mount_volumes = await resolve_git_mounts(
                session, resolved, instance_dir, working_directory, home_dir
            )
            extra_volumes.extend(git_mount_volumes)
            if profile_hints.get("start_command"):
                start_command = profile_hints["start_command"]
            if profile_hints.get("working_directory"):
                working_directory = profile_hints["working_directory"]
            if profile_hints.get("port_override"):
                port_override = profile_hints["port_override"]
            logger.debug(
                "Applied config profile %s to instance %s (env=%d, files=%d, mounts=%d, git_mounts=%d)",
                resolved.profile_name,
                instance.id,
                len(profile_env),
                len(profile_files),
                len(profile_mounts),
                len(git_mount_volumes),
            )
        except ConfigProfileCycleError as exc:
            logger.error(
                "Cycle detected in config profile for instance %s: %s", instance.id, exc
            )
            raise ValueError(f"Config profile cycle detected: {exc}")
    else:
        logger.debug("No config profile selected for instance %s", instance.id)

    # Write env file and config files
    env_file_path = None

    if env_vars:
        env_file_path = write_env_file(instance_dir, env_vars)
        logger.debug("Wrote env file for instance %s: %s", instance.id, env_file_path)

    if config_files:
        write_config_files(instance_dir, config_files)
        logger.debug(
            "Wrote %d config files for instance %s", len(config_files), instance.id
        )

    # Mount selected SSH keys into container home dir
    if instance.ssh_key_ids:
        from src.services.shared.ssh_keys import write_ssh_config, _sanitize_filename

        ssh_keys_to_mount = []
        for key_id in instance.ssh_key_ids:
            ssh_key = await session.get(SSHKey, uuid.UUID(key_id))
            if ssh_key and ssh_key.user_id == user_id:
                ssh_keys_to_mount.append(ssh_key)
            else:
                logger.warning(
                    "SSH key %s not found or not authorized for user %s",
                    key_id,
                    user_id,
                )

        if ssh_keys_to_mount:
            ssh_dir = os.path.join(instance_dir, "mounts", "ssh", ".ssh")
            os.makedirs(ssh_dir, exist_ok=True)

            key_filenames = []
            for ssh_key in ssh_keys_to_mount:
                key_name = _sanitize_filename(ssh_key.name)
                base_filename = f"id_ed25519_{key_name}"
                filename = base_filename
                counter = 1
                while filename in key_filenames:
                    filename = f"{base_filename}_{counter}"
                    counter += 1
                key_filenames.append(filename)

                try:
                    prepare_ssh_key_files(
                        instance_dir,
                        ssh_key,
                        subdir="mounts/ssh/.ssh",
                        uid=container_uid,
                        gid=container_gid,
                        key_filename=filename,
                        write_config=False,
                    )
                    logger.debug(
                        "Prepared SSH key %s as %s for instance %s",
                        ssh_key.name,
                        filename,
                        instance.id,
                    )
                except Exception as exc:
                    logger.error(
                        "Failed to prepare SSH key %s for instance %s: %s",
                        ssh_key.id,
                        instance.id,
                        exc,
                    )

            try:
                write_ssh_config(
                    ssh_dir,
                    key_filenames,
                    uid=container_uid,
                    gid=container_gid,
                )
            except Exception as exc:
                logger.error(
                    "Failed to write SSH config for instance %s: %s",
                    instance.id,
                    exc,
                )

            ssh_target = os.path.join(home_dir, ".ssh")
            extra_volumes.append(
                {
                    "source": ssh_dir,
                    "target": ssh_target,
                    "type": "bind",
                }
            )
            logger.debug(
                "Mounted %d SSH key(s) for instance %s to %s",
                len(ssh_keys_to_mount),
                instance.id,
                ssh_target,
            )

    # Ensure staged bind-mount sources are owned by the container user so
    # directories like ~/.config remain writable inside the container.
    _chown_staged_mounts(extra_volumes, instance_dir, container_uid, container_gid)

    # ── MANIFEST-BASED FLOW ──────────────────────────────────────
    resolved_manifest = None

    if tool_type and tool_type.definition_type == "manifest" and tool_type.manifest_id:
        logger.info("Using manifest-based startup for instance %s", instance.id)

        repo_path = ""
        if instance.workspace_id:
            from src.models import Workspace as WorkspaceModel

            workspace = await session.get(WorkspaceModel, instance.workspace_id)
            if workspace:
                repo_path = workspace.path
        else:
            repo = await session.get(GitRepository, instance.repository_id)
            repo_path = repo.path if repo else ""
            if instance.clone_mode == "clone":
                repo_path = os.path.join(instance_dir, "repo-clone")

        try:
            (
                image_tag,
                compose_content,
                resolved_manifest,
                _home_dir,
            ) = await prepare_manifest_instance(
                session=session,
                instance=instance,
                instance_dir=instance_dir,
                repo_path=repo_path,
                env_vars=env_vars,
                extra_volumes=extra_volumes,
                working_directory=working_directory,
            )
            write_compose_file(instance_dir, compose_content)
            logger.debug(
                "Generated manifest-based compose for instance %s", instance.id
            )
        except Exception as exc:
            logger.exception(
                "Manifest compilation failed for instance %s: %s", instance.id, exc
            )
            instance.status = "error"
            await session.commit()
            raise RuntimeError(f"Manifest compilation failed: {exc}")
    else:
        # ── LEGACY FLOW ──────────────────────────────────────────
        if instance.clone_mode == "clone" and not instance.workspace_id:
            repo = await session.get(GitRepository, instance.repository_id)
            if repo and repo.ssh_key_id:
                ssh_key = await session.get(SSHKey, repo.ssh_key_id)
                if ssh_key:
                    try:
                        ssh_dir = prepare_ssh_key_files(
                            instance_dir, ssh_key, uid=0, gid=0
                        )
                        extra_volumes.append(
                            {
                                "source": ssh_dir,
                                "target": "/root/.ssh",
                                "type": "bind",
                            }
                        )
                        logger.debug(
                            "Mounted SSH key for clone-mode instance %s", instance.id
                        )
                    except Exception as exc:
                        logger.error(
                            "Failed to prepare SSH key for instance %s: %s",
                            instance.id,
                            exc,
                        )

        if port_override or start_command or working_directory or extra_volumes:
            modify_compose_file(
                instance.compose_path,
                port_override,
                start_command,
                working_directory,
                extra_volumes,
                home_dir,
            )
            logger.debug("Modified compose file for instance %s", instance.id)

    # Sanitize compose file
    sanitize_compose_file(instance.compose_path)

    # Auto-fix bind address for known web tools
    if tool_type and tool_type.interface_type == "web":
        ensure_web_bind_address(
            instance.compose_path, tool_type.name, tool_type.default_port
        )

    # Ensure predictable container name
    ensure_container_name_in_compose(instance.compose_path, instance.name)
    ensure_backend_network_in_compose(instance.compose_path)

    # Execute docker compose up
    logger.debug(
        "Running docker compose up for instance %s (compose_path=%s)",
        instance.id,
        instance.compose_path,
    )
    returncode, stdout, stderr = execute_compose_command(
        instance.compose_path, "up", env_file=env_file_path
    )
    logger.debug(
        "Docker compose up completed for instance %s: returncode=%d, stdout=%s, stderr=%s",
        instance.id,
        returncode,
        stdout[:200] if stdout else "",
        stderr[:500] if stderr else "",
    )

    if returncode != 0:
        instance.status = "error"
        await session.commit()
        logger.error("Failed to start instance %s: %s", instance.id, stderr)
        raise RuntimeError(f"failed to start instance: {stderr}")

    # Get container ID and name
    expected_container_name = instance.name.lower()
    container_id = get_container_id(expected_container_name)
    if container_id:
        instance.container_id = container_id
        logger.debug("Container ID for instance %s: %s", instance.id, container_id)

    instance.container_name = expected_container_name
    logger.debug(
        "Container name for instance %s: %s", instance.id, expected_container_name
    )

    # Verify container reached running state
    if instance.container_id:
        instance.status = "starting"
        instance.last_started_at = datetime.now()
        await session.commit()
        await publish_lifecycle_event(
            event_bus=_event_bus,
            session=session,
            instance=instance,
            event_type="instance.started",
            created_by=user_id,
            status="starting",
            message="Container starting...",
        )
        logger.debug("Instance %s: verifying container startup...", instance.id)

        startup_result = wait_for_container_running(
            instance.container_id, timeout=30, interval=2.0
        )

        if not startup_result["success"]:
            error_msg = f"Container failed to start: status={startup_result['status']}"
            if startup_result["exit_code"] is not None:
                error_msg += f", exit_code={startup_result['exit_code']}"

            logs = get_container_logs(instance.container_id, tail=50)

            instance.status = "error"
            await session.commit()
            await publish_lifecycle_event(
                event_bus=_event_bus,
                session=session,
                instance=instance,
                event_type="instance.error",
                created_by=user_id,
                status="error",
                message=error_msg,
                metadata={
                    "exit_code": startup_result["exit_code"],
                    "error_type": "container",
                },
            )
            logger.error(
                "Instance %s container startup failed after %.1fs: %s\nLogs:\n%s",
                instance.id,
                startup_result["waited_seconds"],
                error_msg,
                logs,
            )
            return {
                "status": "error",
                "error": error_msg,
                "logs": logs,
            }

        logger.debug(
            "Instance %s container started successfully after %.1fs",
            instance.id,
            startup_result["waited_seconds"],
        )

        # Apply mount permission fixes for manifest-based instances
        if resolved_manifest and instance.container_id:
            mounts = resolved_manifest.get("mounts", [])
            if mounts:
                logger.debug(
                    "Applying permission fixes for instance %s (%d mounts)",
                    instance.id,
                    len(mounts),
                )
                permission_results = apply_mount_permissions(
                    instance.container_id,
                    mounts,
                )
                for result in permission_results:
                    if not result["success"]:
                        logger.warning(
                            "Permission fix failed for mount %s on instance %s: %s",
                            result["mount_name"],
                            instance.id,
                            result["error"],
                        )

        # Fix SSH key ownership/permissions inside the container
        if instance.ssh_key_ids and instance.container_id:
            container_user = (
                "root"
                if home_dir == "/root"
                else home_dir[6:]
                if home_dir.startswith("/home/")
                else "root"
            )
            ssh_target = os.path.join(home_dir, ".ssh")
            logger.debug(
                "Applying SSH permissions for user %s on %s in instance %s",
                container_user,
                ssh_target,
                instance.id,
            )
            ssh_perm_result = apply_ssh_permissions(
                instance.container_id,
                ssh_target,
                container_user,
            )
            if not ssh_perm_result["success"]:
                logger.warning(
                    "SSH permission fix failed for instance %s: %s",
                    instance.id,
                    ssh_perm_result["error"],
                )

    # Execute readiness probe if configured
    tool_type = await session.get(ToolType, instance.tool_type_id)
    if tool_type and instance.container_id:
        probe_command = None
        probe_timeout = 30
        probe_interval = 2

        if tool_type.readiness_probe:
            probe_config = tool_type.readiness_probe
            probe_command = probe_config.get("command", "")
            probe_timeout = probe_config.get("timeout", 30)
            probe_interval = probe_config.get("interval", 2)
        elif tool_type.interface_type == "web":
            probe_command = f"curl -f http://localhost:{tool_type.default_port or 8080}"
            probe_timeout = 30
            probe_interval = 2

        if probe_command:
            instance.status = "probing"
            await session.commit()
            logger.debug(
                "Executing readiness probe for instance %s: command='%s', timeout=%d, interval=%d",
                instance.id,
                probe_command,
                probe_timeout,
                probe_interval,
            )

            success, probe_logs = await execute_probe(
                container_id=instance.container_id,
                command=probe_command,
                timeout=probe_timeout,
                interval=probe_interval,
            )

            instance.probe_result = {
                "success": success,
                "command": probe_command,
                "logs": probe_logs,
                "timestamp": datetime.now().isoformat(),
            }

            if not success:
                instance.status = "unhealthy"
                await session.commit()
                await publish_lifecycle_event(
                    event_bus=_event_bus,
                    session=session,
                    instance=instance,
                    event_type="instance.health_changed",
                    created_by=user_id,
                    status="unhealthy",
                    message="Readiness probe failed",
                    metadata={"probe_output": "\n".join(probe_logs)},
                )
                logger.error(
                    "Readiness probe failed for instance %s after %ds: %s",
                    instance.id,
                    probe_timeout,
                    "\n".join(probe_logs),
                )
                return {
                    "status": "unhealthy",
                    "error": f"Readiness probe failed after {probe_timeout}s",
                    "probe_logs": probe_logs,
                }

            logger.info("Readiness probe succeeded for instance %s", instance.id)

    # Final stability check: the container must still be running after all
    # post-start setup. If it has already exited/restarted, mark it failed now
    # instead of optimistically reporting "running".
    if instance.container_id:
        final_check = get_container_status(instance.container_id)
        if final_check["status"] != "running":
            error_msg = (
                f"Container stopped during startup: status={final_check['status']}"
            )
            if final_check["exit_code"] is not None:
                error_msg += f", exit_code={final_check['exit_code']}"
            logs = get_container_logs(instance.container_id, tail=50)
            instance.status = "error"
            await session.commit()
            await publish_lifecycle_event(
                event_bus=_event_bus,
                session=session,
                instance=instance,
                event_type="instance.error",
                created_by=user_id,
                status="error",
                message=error_msg,
                metadata={
                    "exit_code": final_check["exit_code"],
                    "error_type": "container",
                },
            )
            logger.error(
                "Instance %s container stopped during startup: %s\nLogs:\n%s",
                instance.id,
                error_msg,
                logs,
            )
            return {"status": "error", "error": error_msg, "logs": logs}

    instance.status = "running"
    await session.commit()
    await publish_lifecycle_event(
        event_bus=_event_bus,
        session=session,
        instance=instance,
        event_type="instance.health_changed",
        created_by=user_id,
        status="running",
        message="Container running",
        metadata={"previous_status": "starting"},
    )
    logger.info("Instance %s is now running", instance.id)

    # Get tool type for default port
    tool_type = await session.get(ToolType, instance.tool_type_id)
    if not tool_type:
        logger.error("Tool type %s not found", instance.tool_type_id)
        instance.status = "error"
        await session.commit()
        raise RuntimeError(f"Tool type '{instance.tool_type_id}' not found")

    logger.debug(
        "Tool type for instance %s: name=%s, container_port=%s, interface_type=%s",
        instance.id,
        tool_type.name,
        tool_type.default_port or 0,
        tool_type.interface_type,
    )

    # Only create Cloudflare tunnel for web-enabled tools
    if tool_type.interface_type == "web":
        try:
            logger.debug(
                "Creating tunnel for instance %s (container_port=%d)",
                instance.id,
                tool_type.default_port or 0,
            )
            tunnel_info = start_tunnel(
                instance_name=instance.name,
                container_port=tool_type.default_port or 0,
            )
            instance.tunnel_id = tunnel_info["container_name"]
            instance.public_url = tunnel_info["url"]
            instance.url = tunnel_info["url"]
            await session.commit()
            logger.debug(
                "Created tunnel for instance %s: container=%s, url=%s",
                instance.id,
                tunnel_info["container_name"],
                tunnel_info["url"],
            )
        except Exception as exc:
            import traceback

            error_msg = str(exc)
            error_trace = traceback.format_exc()
            logger.error(
                "Failed to create tunnel for instance %s: %s\nTraceback:\n%s",
                instance.id,
                error_msg,
                error_trace,
            )
            instance.status = "error"
            instance.url = None
            await session.commit()
            raise RuntimeError(f"Failed to create tunnel: {error_msg}")
    else:
        logger.info(
            "Instance %s is terminal-only (no web interface), skipping tunnel creation",
            instance.id,
        )
        instance.url = None
        instance.public_url = None
        await session.commit()

    return {"status": instance.status, "url": instance.url}


async def restart_tool_instance(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    instance_id: uuid.UUID,
) -> dict:
    """Restart a tool instance.

    Returns a dict with status and url.
    Raises ValueError for invalid input, RuntimeError for internal failures.
    """
    instance = await session.get(ToolInstance, instance_id)
    if instance is None or instance.repository_id != repo_id:
        raise ValueError("instance not found")

    # Stop old tunnel if exists
    if instance.tunnel_id:
        try:
            stop_tunnel(instance.name)
            logger.debug(
                "Stopped old tunnel for instance %s (container=%s)",
                instance.id,
                instance.tunnel_id,
            )
        except Exception as exc:
            logger.warning(
                "Failed to stop old tunnel for instance %s: %s", instance.id, exc
            )

    # Re-apply stored config profile on restart
    if instance.compose_path and os.path.exists(instance.compose_path):
        instance_dir = os.path.dirname(instance.compose_path)
        if instance.selected_config_profile_id is not None:
            try:
                resolved = await resolve_profile(
                    session, instance.selected_config_profile_id
                )
                profile_env, profile_files, profile_mounts, profile_hints = (
                    apply_resolved_profile(instance_dir, resolved)
                )
                if profile_env:
                    write_env_file(instance_dir, profile_env)
                    logger.debug(
                        "Re-applied config profile %s on restart for instance %s",
                        resolved.profile_name,
                        instance.id,
                    )
            except ConfigProfileCycleError as exc:
                logger.error(
                    "Cycle detected in stored config profile for instance %s: %s",
                    instance.id,
                    exc,
                )

        # Re-apply compose fixes
        sanitize_compose_file(instance.compose_path)
        tool_type = await session.get(ToolType, instance.tool_type_id)
        if tool_type and tool_type.interface_type == "web":
            ensure_web_bind_address(
                instance.compose_path, tool_type.name, tool_type.default_port
            )
        ensure_container_name_in_compose(instance.compose_path, instance.name)
        ensure_backend_network_in_compose(instance.compose_path)

        returncode, stdout, stderr = execute_compose_command(
            instance.compose_path, "restart"
        )

        if returncode == 0:
            instance.status = "running"
            instance.last_started_at = datetime.now()

            tool_type = await session.get(ToolType, instance.tool_type_id)
            if not tool_type or not tool_type.default_port:
                logger.error(
                    "Tool type %s has no default_port configured. Cannot create tunnel.",
                    instance.tool_type_id,
                )
                instance.status = "error"
                await session.commit()
                raise RuntimeError(
                    f"Tool type '{tool_type.name if tool_type else 'unknown'}' has no port configured"
                )

            if tool_type.interface_type == "web":
                try:
                    tunnel_info = start_tunnel(
                        instance_name=instance.name,
                        container_port=tool_type.default_port or 0,
                    )
                    instance.tunnel_id = tunnel_info["container_name"]
                    instance.public_url = tunnel_info["url"]
                    instance.url = tunnel_info["url"]
                    logger.debug(
                        "Created new tunnel for instance %s: %s",
                        instance.id,
                        tunnel_info["url"],
                    )
                except Exception as exc:
                    logger.warning(
                        "Failed to create tunnel for instance %s: %s",
                        instance.id,
                        exc,
                    )
                    instance.status = "error"
                    instance.url = None
                    await session.commit()
                    raise RuntimeError(f"Failed to create tunnel: {exc}")
            else:
                instance.url = None
                instance.public_url = None

            await session.commit()
            await publish_lifecycle_event(
                event_bus=_event_bus,
                session=session,
                instance=instance,
                event_type="instance.restarted",
                created_by=user_id,
                status="running",
                message="Instance restarted",
            )
            return {"status": instance.status, "url": instance.url}

    instance.status = "error"
    await session.commit()
    return {"status": instance.status}


async def delete_tool_instance(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    instance_id: uuid.UUID,
    force: bool = False,
) -> None:
    """Delete a tool instance.

    Raises ValueError for invalid input.
    """
    instance = await session.get(ToolInstance, instance_id)
    if instance is None or instance.repository_id != repo_id:
        raise ValueError("instance not found")

    # Check dirty state for clone-mode instances
    if instance.clone_mode == "clone" and not force:
        instance_dir = (
            os.path.dirname(instance.compose_path) if instance.compose_path else None
        )
        if instance_dir:
            clone_path = os.path.join(instance_dir, "repo-clone")
            if os.path.exists(clone_path):
                is_dirty, changed_files = check_dirty_state(clone_path)
                if is_dirty:
                    raise RuntimeError(
                        f"Repository has uncommitted changes: {changed_files}"
                    )

    # Stop Cloudflare tunnel if exists
    if instance.tunnel_id:
        try:
            stop_tunnel(instance.name)
            logger.debug(
                "Stopped tunnel for instance %s (container=%s)",
                instance.id,
                instance.tunnel_id,
            )
        except Exception as exc:
            logger.warning(
                "Failed to stop tunnel for instance %s: %s", instance.id, exc
            )

    # Stop and remove container
    if instance.compose_path and os.path.exists(instance.compose_path):
        execute_compose_command(instance.compose_path, "down")

    # Remove instance directory
    if instance.compose_path:
        instance_dir = os.path.dirname(instance.compose_path)
        if os.path.exists(instance_dir):
            import shutil

            shutil.rmtree(instance_dir)

    await publish_lifecycle_event(
        event_bus=_event_bus,
        session=session,
        instance=instance,
        event_type="instance.deleted",
        created_by=user_id,
        status="deleted",
        message="Instance deleted",
    )
    await session.delete(instance)
    await session.commit()


async def recreate_instance_tunnel(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    instance_id: uuid.UUID,
) -> dict:
    """Recreate the temporary tunnel for an instance.

    Returns a dict with status and url.
    Raises ValueError for invalid input, RuntimeError for internal failures.
    """
    instance = await session.get(ToolInstance, instance_id)
    if instance is None or instance.repository_id != repo_id:
        raise ValueError("instance not found")

    if instance.status != "running":
        raise ValueError("instance must be running to recreate tunnel")

    tool_type = await session.get(ToolType, instance.tool_type_id)
    if not tool_type:
        raise ValueError("Tool type not found for this instance")

    expected_name = instance.name.lower()
    logger.info(
        "Recreate tunnel for instance %s (expected container name: %s, default_port: %s)",
        instance.id,
        expected_name,
        tool_type.default_port,
    )

    # Find the tool container
    tool_container_id = instance.container_id
    if tool_container_id:
        logger.info("Using stored container_id: %s", tool_container_id)
    else:
        tool_container_id = get_container_id(expected_name)
        if tool_container_id:
            logger.info("Found container by name: %s", tool_container_id)
        else:
            logger.error("Container %s not found", expected_name)
            raise ValueError("Could not find running container for this instance")

    # Ensure the tool container is on the backend network
    network_name = get_backend_network_name()
    on_network = is_container_on_network(tool_container_id, network_name)
    logger.info(
        "Container %s on network %s: %s",
        tool_container_id,
        network_name,
        on_network,
    )
    if not on_network:
        logger.info(
            "Connecting container %s to network %s",
            tool_container_id,
            network_name,
        )
        connected = connect_container_to_network(tool_container_id, network_name)
        logger.info("Network connect result: %s", connected)

    # Get the container's IP on the backend network
    target_ip = get_container_ip_on_network(tool_container_id, network_name)
    if target_ip:
        target_url = f"http://{target_ip}:{tool_type.default_port or 0}"
        logger.info(
            "Tunnel target for instance %s: %s (IP %s on %s)",
            instance.id,
            target_url,
            target_ip,
            network_name,
        )
    else:
        target_url = f"http://{expected_name}:{tool_type.default_port or 0}"
        logger.warning(
            "Could not get container IP, falling back to name-based target: %s",
            target_url,
        )

    try:
        tunnel_info = recreate_tunnel(
            instance_name=instance.name,
            container_port=tool_type.default_port or 0,
            target_url=target_url,
        )
        logger.info(
            "Tunnel recreated: container=%s, url=%s",
            tunnel_info["container_name"],
            tunnel_info["url"],
        )

        # Verify the tunnel can actually reach the origin
        health = check_tunnel_health(tunnel_info["url"], timeout=10)
        logger.info(
            "Tunnel health check: status=%s, code=%s, error=%s",
            health.get("tunnel_status"),
            health.get("status_code"),
            health.get("error"),
        )

        # Also probe from inside the API container directly to the target
        probe = subprocess.run(
            [
                "curl",
                "-s",
                "-o",
                "/dev/null",
                "-w",
                "%{http_code}",
                "--max-time",
                "5",
                target_url,
            ],
            capture_output=True,
            text=True,
        )
        logger.info(
            "Direct probe from API to %s: HTTP %s", target_url, probe.stdout.strip()
        )

        instance.tunnel_id = tunnel_info["container_name"]
        instance.public_url = tunnel_info["url"]
        instance.url = tunnel_info["url"]
        await session.commit()
        return {"status": "healthy", "url": instance.url}
    except Exception as exc:
        logger.exception("Failed to recreate tunnel for instance %s", instance.id)
        raise RuntimeError(f"Failed to recreate tunnel: {str(exc)}")


async def stop_tool_instance(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    instance_id: uuid.UUID,
) -> dict:
    """Stop a tool instance.

    Returns a dict with the stopped status.
    Raises ValueError for invalid input.
    """
    instance = await session.get(ToolInstance, instance_id)
    if instance is None or instance.repository_id != repo_id:
        raise ValueError("instance not found")

    if instance.tunnel_id:
        try:
            stop_tunnel(instance.name)
        except Exception as exc:
            logger.warning(
                "Failed to stop tunnel for instance %s: %s", instance.id, exc
            )

    if instance.compose_path and os.path.exists(instance.compose_path):
        execute_compose_command(instance.compose_path, "stop")

    instance.status = "stopped"
    instance.last_stopped_at = datetime.now()
    instance.url = None
    instance.public_url = None
    instance.tunnel_id = None
    await session.commit()
    await publish_lifecycle_event(
        event_bus=_event_bus,
        session=session,
        instance=instance,
        event_type="instance.stopped",
        created_by=user_id,
        status="stopped",
        message="Instance stopped",
    )
    return {"status": instance.status}


async def rename_tool_instance(
    session: AsyncSession,
    user_id: uuid.UUID,
    project_id: uuid.UUID,
    repo_id: uuid.UUID,
    instance_id: uuid.UUID,
    display_name: str,
) -> ToolInstance:
    """Rename a tool instance (update display_name only)."""
    instance = await session.get(ToolInstance, instance_id)
    if instance is None or instance.repository_id != repo_id:
        raise ValueError("instance not found")
    if instance.owner_id != user_id:
        raise ValueError("not authorized")

    instance.display_name = display_name.strip()
    await session.commit()
    await session.refresh(instance)
    return instance