spec-kit/src/specify_cli/agents.py

"""
Agent Command Registrar for Spec Kit

Shared infrastructure for registering commands with AI agents.
Used by both the extension system and the preset system to write
command files into agent-specific directories in the correct format.
"""

import os
import platform
import re
from copy import deepcopy
from pathlib import Path
from typing import Any, Dict, List, Optional

import yaml

from ._init_options import is_ai_skills_enabled, load_init_options
from ._toml_string import escape_toml_basic as _escape_toml_basic
from ._toml_string import has_illegal_toml_control as _has_illegal_toml_control
from ._utils import relative_extension_path_violation


def _build_agent_configs() -> dict[str, Any]:
    """Derive CommandRegistrar.AGENT_CONFIGS from INTEGRATION_REGISTRY."""
    from specify_cli.integrations import INTEGRATION_REGISTRY

    configs: dict[str, dict[str, Any]] = {}
    for key, integration in INTEGRATION_REGISTRY.items():
        if key == "generic":
            continue
        if integration.registrar_config:
            config = dict(integration.registrar_config)
            # Propagate invoke_separator from the integration class when the
            # registrar_config dict doesn't already declare it explicitly.
            # SkillsIntegration subclasses (claude, codex, …) set
            # invoke_separator="-" as a class attribute but omit it from
            # registrar_config, so without this they would fall back to "."
            # when register_commands() resolves __SPECKIT_COMMAND_*__ tokens.
            if "invoke_separator" not in config:
                config["invoke_separator"] = integration.invoke_separator
            if integration.dev_no_symlink:
                config["dev_no_symlink"] = True
            configs[key] = config
    return configs


class CommandRegistrar:
    """Handles registration of commands with AI agents.

    Supports writing command files in Markdown or TOML format to the
    appropriate agent directory, with correct argument placeholders
    and companion files (e.g. Copilot .prompt.md).
    """

    # Derived from INTEGRATION_REGISTRY — single source of truth.
    # Populated lazily via _ensure_configs() on first use.
    AGENT_CONFIGS: dict[str, dict[str, Any]] = {}
    _configs_loaded: bool = False

    def __init__(self) -> None:
        self._ensure_configs()

    def __init_subclass__(cls, **kwargs: Any) -> None:
        super().__init_subclass__(**kwargs)
        cls._ensure_configs()

    @classmethod
    def _ensure_configs(cls) -> None:
        if not cls._configs_loaded:
            try:
                cls.AGENT_CONFIGS = _build_agent_configs()
                cls._configs_loaded = True
            except ImportError:
                pass  # Circular import during module init; retry on next access

    @staticmethod
    def _hyphenate_frontmatter_refs(val: Any) -> Any:
        """Recursively find any dotted references starting with speckit. and hyphenate them."""
        if isinstance(val, dict):
            return {
                k: CommandRegistrar._hyphenate_frontmatter_refs(v)
                for k, v in val.items()
            }
        elif isinstance(val, list):
            return [CommandRegistrar._hyphenate_frontmatter_refs(x) for x in val]
        elif isinstance(val, str):
            return re.sub(
                r"\bspeckit\.[A-Za-z0-9-_]+(?:\.[A-Za-z0-9-_]+)*\b",
                lambda m: m.group(0).replace(".", "-"),
                val,
            )
        return val

    @staticmethod
    def _hyphenate_body_refs(body: str) -> str:
        """Hyphenate dotted speckit references in command body text."""
        return re.sub(
            r"\bspeckit\.[A-Za-z0-9-_]+(?:\.[A-Za-z0-9-_]+)*\b",
            lambda m: m.group(0).replace(".", "-"),
            body,
        )

    @staticmethod
    def parse_frontmatter(content: str) -> tuple[dict, str]:
        """Parse YAML frontmatter from Markdown content.

        Args:
            content: Markdown content with YAML frontmatter

        Returns:
            Tuple of (frontmatter_dict, body_content)
        """
        if not content.startswith("---"):
            return {}, content

        # Find second ---
        end_marker = content.find("---", 3)
        if end_marker == -1:
            return {}, content

        frontmatter_str = content[3:end_marker].strip()
        body = content[end_marker + 3 :].strip()

        try:
            frontmatter = yaml.safe_load(frontmatter_str) or {}
        except yaml.YAMLError:
            frontmatter = {}

        if not isinstance(frontmatter, dict):
            frontmatter = {}

        return frontmatter, body

    @staticmethod
    def render_frontmatter(fm: dict) -> str:
        """Render frontmatter dictionary as YAML.

        Args:
            fm: Frontmatter dictionary

        Returns:
            YAML-formatted frontmatter with delimiters
        """
        if not fm:
            return ""

        yaml_str = yaml.dump(
            fm, default_flow_style=False, sort_keys=False, allow_unicode=True
        )
        return f"---\n{yaml_str}---\n"

    def _adjust_script_paths(
        self, frontmatter: dict, extension_id: Optional[str] = None
    ) -> dict:
        """Normalize script paths in frontmatter to generated project locations.

        Rewrites known repo-relative and top-level script paths under the
        ``scripts`` key (for example ``../../scripts/``,
        ``../../templates/``, ``../../memory/``, ``scripts/``, ``templates/``, and
        ``memory/``) to the ``.specify/...`` paths used in generated projects.

        Args:
            frontmatter: Frontmatter dictionary
            extension_id: Extension id when rendering extension-owned commands.

        Returns:
            Modified frontmatter with normalized project paths
        """
        frontmatter = deepcopy(frontmatter)

        scripts = frontmatter.get("scripts")
        if isinstance(scripts, dict):
            for key, script_path in scripts.items():
                if isinstance(script_path, str):
                    scripts[key] = self.rewrite_project_relative_paths(
                        script_path, extension_id=extension_id
                    )
        return frontmatter

    @staticmethod
    def rewrite_project_relative_paths(
        text: str, extension_id: Optional[str] = None
    ) -> str:
        """Rewrite repo-relative paths to their generated project locations."""
        if not isinstance(text, str) or not text:
            return text

        for old, new in (
            ("../../memory/", ".specify/memory/"),
            ("../../scripts/", ".specify/scripts/"),
            ("../../templates/", ".specify/templates/"),
        ):
            text = text.replace(old, new)

        # Only rewrite top-level style references so existing generated paths
        # like ".specify/extensions/<ext>/scripts/..." remain intact. When
        # rendering extension commands, top-level "scripts/" is extension-local.
        scripts_replacement = (
            f".specify/extensions/{extension_id}/scripts/"
            if extension_id
            else ".specify/scripts/"
        )
        text = re.sub(r'(^|[\s`"\'(])(?:\.?/)?memory/', r"\1.specify/memory/", text)
        text = re.sub(
            r'(^|[\s`"\'(])(?:\.?/)?scripts/', rf"\1{scripts_replacement}", text
        )
        text = re.sub(
            r'(^|[\s`"\'(])(?:\.?/)?templates/', r"\1.specify/templates/", text
        )

        return text.replace(".specify/.specify/", ".specify/").replace(
            ".specify.specify/", ".specify/"
        )

    @staticmethod
    def rewrite_extension_paths(
        text: str, extension_id: str, extension_dir: Path
    ) -> str:
        """Rewrite extension-relative paths to their installed locations.

        Extension command bodies reference bundled files relative to the
        extension root (e.g. ``agents/control/commander.md``). After install
        those files live under ``.specify/extensions/<id>/``, so bare
        references would resolve against the workspace root and never be
        found (#2101).

        Only directories that actually exist inside *extension_dir* are
        rewritten, keeping the behaviour conservative and avoiding false
        positives on prose. ``commands`` (slash-command sources), ``specs``
        (user project artifacts) and dot-directories are never rewritten.
        """
        if not isinstance(text, str) or not text:
            return text

        skip = {"commands", ".git", "specs"}
        try:
            subdirs = [
                entry.name
                for entry in extension_dir.iterdir()
                if entry.is_dir()
                and entry.name not in skip
                and not entry.name.startswith(".")
            ]
        except OSError:
            return text

        for subdir in subdirs:
            # Only rewrite relative references (subdir/... or ./subdir/...);
            # absolute paths like /subdir/... keep their meaning. Use a
            # callable replacement: subdir/extension_id come from the
            # filesystem and could contain backslashes or "\1"-like
            # sequences, which would corrupt a string replacement template.
            replacement = f".specify/extensions/{extension_id}/{subdir}/"
            text = re.sub(
                r'(^|[\s`"\'(])(?:\./)?' + re.escape(subdir) + "/",
                lambda m: m.group(1) + replacement,
                text,
            )
        return text

    def render_markdown_command(
        self, frontmatter: dict, body: str, source_id: str, context_note: str = None
    ) -> str:
        """Render command in Markdown format.

        Args:
            frontmatter: Command frontmatter
            body: Command body content
            source_id: Source identifier (extension or preset ID)
            context_note: Custom context comment (default: <!-- Source: {source_id} -->)

        Returns:
            Formatted Markdown command file content
        """
        if context_note is None:
            context_note = f"\n<!-- Source: {source_id} -->\n"
        return self.render_frontmatter(frontmatter) + "\n" + context_note + body

    def render_toml_command(self, frontmatter: dict, body: str, source_id: str) -> str:
        """Render command in TOML format.

        Args:
            frontmatter: Command frontmatter
            body: Command body content
            source_id: Source identifier (extension or preset ID)

        Returns:
            Formatted TOML command file content
        """
        toml_lines = []

        if "description" in frontmatter:
            toml_lines.append(
                f"description = {self._render_basic_toml_string(frontmatter['description'])}"
            )
            toml_lines.append("")

        toml_lines.append(f"# Source: {source_id}")
        toml_lines.append("")

        # Keep TOML output valid even when body contains triple-quote delimiters
        # or backslashes. Prefer multiline forms, then fall back to escaped basic
        # string. A multiline *basic* string ("""...""") processes backslash escape
        # sequences, so a body containing a backslash (e.g. a Windows path
        # ``C:\\Users\\...`` whose ``\\U`` reads as an invalid unicode escape) would
        # produce unparseable TOML — route those to the *literal* form ('''...'''),
        # which does not process escapes, or to the escaped basic string.
        # Control characters (U+0000–U+001F except tab/newline, U+007F) and a bare
        # CR are illegal in every TOML string form, so a body containing them must
        # go to the escaped basic string regardless of which delimiters it uses.
        if self._has_illegal_toml_control(body):
            toml_lines.append(f"prompt = {self._render_basic_toml_string(body)}")
        elif '"""' not in body and "\\" not in body:
            toml_lines.append('prompt = """')
            toml_lines.append(body)
            toml_lines.append('"""')
        elif "'''" not in body:
            toml_lines.append("prompt = '''")
            toml_lines.append(body)
            toml_lines.append("'''")
        else:
            toml_lines.append(f"prompt = {self._render_basic_toml_string(body)}")

        return "\n".join(toml_lines)

    # Control-char detection and basic-string escaping are shared with the
    # gemini/tabnine renderer in ``specify_cli.integrations.base`` via
    # ``specify_cli._toml_string`` so the two never drift apart.
    _has_illegal_toml_control = staticmethod(_has_illegal_toml_control)
    _render_basic_toml_string = staticmethod(_escape_toml_basic)

    def render_yaml_command(
        self,
        frontmatter: dict,
        body: str,
        source_id: str,
        cmd_name: str = "",
    ) -> str:
        """Render command in YAML recipe format for Goose.

        Args:
            frontmatter: Command frontmatter
            body: Command body content
            source_id: Source identifier (extension or preset ID)
            cmd_name: Command name used as title fallback

        Returns:
            Formatted YAML recipe file content
        """
        from specify_cli.integrations.base import YamlIntegration

        title = frontmatter.get("title", "") or frontmatter.get("name", "")
        if not isinstance(title, str):
            title = str(title) if title is not None else ""
        if not title and cmd_name:
            title = YamlIntegration._human_title(cmd_name)
        if not title and source_id:
            title = YamlIntegration._human_title(Path(str(source_id)).stem)
        if not title:
            title = "Command"

        description = frontmatter.get("description", "")
        if not isinstance(description, str):
            description = str(description) if description is not None else ""
        return YamlIntegration._render_yaml(title, description, body, source_id)

    def render_skill_command(
        self,
        agent_name: str,
        skill_name: str,
        frontmatter: dict,
        body: str,
        source_id: str,
        source_file: str,
        project_root: Path,
        extension_id: Optional[str] = None,
    ) -> str:
        """Render a command override as a SKILL.md file.

        SKILL-target agents should receive the same skills-oriented
        frontmatter shape used elsewhere in the project instead of the
        original command frontmatter.

        Technical debt note:
        Spec-kit currently has multiple SKILL.md generators (template packaging,
        init-time conversion, and extension/preset overrides). Keep the skill
        frontmatter keys aligned (name/description/compatibility/metadata, with
        metadata.author and metadata.source subkeys) to avoid drift across agents.
        """
        if not isinstance(frontmatter, dict):
            frontmatter = {}

        agent_config = self.AGENT_CONFIGS.get(agent_name, {})
        if agent_config.get("extension") == "/SKILL.md":
            body = self.resolve_skill_placeholders(
                agent_name, frontmatter, body, project_root, extension_id=extension_id
            )

        description = frontmatter.get(
            "description", f"Spec-kit workflow command: {skill_name}"
        )
        skill_frontmatter = self.build_skill_frontmatter(
            agent_name,
            skill_name,
            description,
            f"{source_id}:{source_file}",
        )
        return self.render_frontmatter(skill_frontmatter) + "\n" + body

    @staticmethod
    def build_skill_frontmatter(
        agent_name: str,
        skill_name: str,
        description: str,
        source: str,
    ) -> dict:
        """Build consistent SKILL.md frontmatter across all skill generators."""
        skill_frontmatter = {
            "name": skill_name,
            "description": description,
            "compatibility": "Requires spec-kit project structure with .specify/ directory",
            "metadata": {
                "author": "github-spec-kit",
                "source": source,
            },
        }
        return skill_frontmatter

    @staticmethod
    def apply_argument_hint(
        source_frontmatter: Dict[str, Any],
        skill_frontmatter: Dict[str, Any],
        integration: Optional[object] = None,
    ) -> None:
        """Carry a command's ``argument-hint`` into its generated skill frontmatter.

        Copies ``argument-hint`` from the parsed source command frontmatter into
        *skill_frontmatter* (mutated in place) before serialization, so that a
        folded multi-line ``description`` cannot be split into invalid YAML. Only
        integrations that support the field — those exposing
        ``inject_argument_hint`` (currently Claude) — receive the key, leaving
        :meth:`build_skill_frontmatter`'s shared shape unchanged for every other
        agent. Built-in templates carry no ``argument-hint``, so this is a no-op
        for the core path.
        """
        if not isinstance(source_frontmatter, dict) or not isinstance(skill_frontmatter, dict):
            return
        argument_hint = source_frontmatter.get("argument-hint")
        if (
            argument_hint
            and integration is not None
            and hasattr(integration, "inject_argument_hint")
        ):
            skill_frontmatter["argument-hint"] = str(argument_hint)

    @staticmethod
    def resolve_skill_placeholders(
        agent_name: str,
        frontmatter: dict,
        body: str,
        project_root: Path,
        extension_id: Optional[str] = None,
    ) -> str:
        """Resolve script placeholders for skills-backed agents."""
        if not isinstance(frontmatter, dict):
            frontmatter = {}

        scripts = frontmatter.get("scripts", {}) or {}
        if not isinstance(scripts, dict):
            scripts = {}

        init_opts = load_init_options(project_root)
        if not isinstance(init_opts, dict):
            init_opts = {}

        script_variant = init_opts.get("script")
        if script_variant not in {"sh", "ps"}:
            fallback_order = []
            default_variant = (
                "ps" if platform.system().lower().startswith("win") else "sh"
            )
            secondary_variant = "sh" if default_variant == "ps" else "ps"

            if default_variant in scripts:
                fallback_order.append(default_variant)
            if secondary_variant in scripts:
                fallback_order.append(secondary_variant)

            for key in scripts:
                if key not in fallback_order:
                    fallback_order.append(key)

            script_variant = fallback_order[0] if fallback_order else None

        script_command = scripts.get(script_variant) if script_variant else None
        if script_command:
            script_command = script_command.replace("{ARGS}", "$ARGUMENTS")
            body = body.replace("{SCRIPT}", script_command)

        body = body.replace("{ARGS}", "$ARGUMENTS").replace("__AGENT__", agent_name)

        return CommandRegistrar.rewrite_project_relative_paths(
            body, extension_id=extension_id
        )

    def _convert_argument_placeholder(
        self, content: str, from_placeholder: str, to_placeholder: str
    ) -> str:
        """Convert argument placeholder format.

        Args:
            content: Command content
            from_placeholder: Source placeholder (e.g., "$ARGUMENTS")
            to_placeholder: Target placeholder (e.g., "{{args}}")

        Returns:
            Content with converted placeholders
        """
        return content.replace(from_placeholder, to_placeholder)

    @staticmethod
    def _compute_output_name(
        agent_name: str, cmd_name: str, agent_config: Dict[str, Any]
    ) -> str:
        """Compute the on-disk command or skill name for an agent."""
        if agent_config["extension"] != "/SKILL.md":
            format_name = agent_config.get("format_name")
            if format_name:
                return format_name(cmd_name)
            return cmd_name

        short_name = cmd_name
        if short_name.startswith("speckit."):
            short_name = short_name[len("speckit.") :]
        short_name = short_name.replace(".", "-")

        return f"speckit-{short_name}"

    @staticmethod
    def _ensure_inside(candidate: Path, base: Path) -> None:
        """Validate that a write target stays within the expected base directory.

        Uses lexical normalization so traversal via ``..`` or absolute paths is
        rejected while intentionally symlinked sub-directories remain
        supported.

        Args:
            candidate: Path that will be written.
            base: Directory the write must remain within.

        Raises:
            ValueError: If the normalized candidate path escapes ``base``.
        """
        normalized = Path(os.path.normpath(candidate))
        base_normalized = Path(os.path.normpath(base))
        if not normalized.is_relative_to(base_normalized):
            raise ValueError(f"Output path {candidate!r} escapes directory {base!r}")

    @staticmethod
    def _is_safe_command_name(name: str) -> bool:
        """Reject names that could escape the commands directory via path traversal."""
        if os.path.sep in name or "/" in name or "\\" in name:
            return False
        return os.path.normpath(name) == name

    @staticmethod
    def _same_lexical_path(left: Path, right: Path) -> bool:
        """Compare paths after lexical normalization without resolving symlinks."""
        return os.path.normcase(os.path.normpath(os.fspath(left))) == os.path.normcase(
            os.path.normpath(os.fspath(right))
        )

    @staticmethod
    def _active_skills_agent(project_root: Path) -> Optional[str]:
        """Return the initialized skills-backed agent, if skills mode is active."""
        opts = load_init_options(project_root)
        if not isinstance(opts, dict):
            return None

        agent = opts.get("ai")
        if not isinstance(agent, str) or not agent:
            return None
        # Kimi is a native skills integration; when ai_skills is not boolean
        # True, Kimi still uses its existing SKILL.md layout.
        if not is_ai_skills_enabled(opts) and agent != "kimi":
            return None
        return agent

    def register_commands(
        self,
        agent_name: str,
        commands: List[Dict[str, Any]],
        source_id: str,
        source_dir: Path,
        project_root: Path,
        context_note: str = None,
        _resolved_dir: Path = None,
        link_outputs: bool = False,
        extension_id: Optional[str] = None,
    ) -> List[str]:
        """Register commands for a specific agent.

        Args:
            agent_name: Agent name (claude, gemini, copilot, etc.)
            commands: List of command info dicts with 'name', 'file', and optional 'aliases'
            source_id: Identifier of the source (extension or preset ID)
            source_dir: Directory containing command source files
            project_root: Path to project root
            context_note: Custom context comment for markdown output
            _resolved_dir: Pre-resolved command directory (internal use
                only — avoids a second ``_resolve_agent_dir`` call and
                duplicate deprecation warnings when invoked from
                ``register_commands_for_all_agents``).
            link_outputs: If True, write rendered output to a source-local
                dev cache and symlink the agent command file to it. Falls back
                to a normal file write when symlinks are unavailable.
            extension_id: Extension id when rendering extension-owned commands.

        Returns:
            List of registered command names

        Raises:
            ValueError: If agent is not supported
        """
        self._ensure_configs()
        if agent_name not in self.AGENT_CONFIGS:
            raise ValueError(f"Unsupported agent: {agent_name}")

        agent_config = self.AGENT_CONFIGS[agent_name]
        commands_dir = _resolved_dir or self._resolve_agent_dir(
            agent_name, agent_config, project_root,
        )
        commands_dir.mkdir(parents=True, exist_ok=True)

        registered = []
        is_cline_ext = agent_name == "cline" and source_id != "core"
        source_root = source_dir.resolve()

        # Resolve the command-reference separator for the file THIS registrar
        # is about to write.  The separator must match the *output layout* the
        # registrar produces for this agent — not the project's persisted
        # ``ai_skills`` flag, and not unrelated sibling directories on disk.  A
        # skill scaffold ("/SKILL.md") uses the skills separator; any
        # command-layout output (".md", ".agent.md", ".toml", …) uses the
        # command separator.
        #
        # This holds for the *active* agent too.  Dual-layout agents (Bob,
        # Copilot) write their skills via their own setup()/skills path, so
        # ``register_commands`` only ever emits their command-layout files.
        # Deriving the separator from ``ai_skills`` would render such a
        # ``.bob/commands/*.md`` (or ``.github/agents/*.agent.md``) file with
        # ``/speckit-*`` whenever that agent is active in skills mode — even
        # though a command-layout file must use ``/speckit.*``.  Deriving it
        # from the agent's static output config avoids that mismatch and stays
        # correct when a stale ``.bob/skills`` directory coexists with
        # ``.bob/commands``.
        _sep = agent_config.get("invoke_separator", ".")
        try:
            from specify_cli.integrations import get_integration  # noqa: PLC0415

            _integ = get_integration(agent_name)
            if _integ is not None:
                registrar_writes_skills = (
                    agent_config.get("extension") == "/SKILL.md"
                )
                _sep = _integ.invoke_separator_for_mode(registrar_writes_skills)
        except Exception:
            pass

        for cmd_info in commands:
            cmd_name = cmd_info["name"]
            aliases = cmd_info.get("aliases", [])
            cmd_file = cmd_info["file"]

            # Guard against path traversal using the single shared policy in
            # relative_extension_path_violation(), so the runtime guard stays
            # aligned with ExtensionManifest._validate() and the skill/preset
            # readers. Skip a malformed/unsafe ``file`` (non-string, empty,
            # whitespace, absolute/anchored, or ``..`` traversal); the
            # resolve()/relative_to() check below is the final containment
            # backstop.
            if relative_extension_path_violation(cmd_file):
                continue
            try:
                source_file = (source_root / cmd_file).resolve()
                source_file.relative_to(source_root)  # raises ValueError if outside
            except (OSError, ValueError):
                continue

            if not source_file.is_file():
                continue

            try:
                content = source_file.read_text(encoding="utf-8")
            except (OSError, UnicodeDecodeError) as exc:
                import warnings

                warnings.warn(
                    f"Skipping command '{cmd_name}': could not read source file "
                    f"'{cmd_file}' ({exc.__class__.__name__}: {exc}).",
                    stacklevel=2,
                )
                continue
            frontmatter, body = self.parse_frontmatter(content)

            if frontmatter.get("strategy") == "wrap":
                from .presets import _substitute_core_template

                body, core_frontmatter = _substitute_core_template(
                    body, cmd_name, project_root, self
                )
                frontmatter = dict(frontmatter)
                for key in ("scripts", "agent_scripts"):
                    if key not in frontmatter and key in core_frontmatter:
                        frontmatter[key] = core_frontmatter[key]
                frontmatter.pop("strategy", None)

            if extension_id:
                body = self.rewrite_extension_paths(body, extension_id, source_root)

            frontmatter = self._adjust_script_paths(
                frontmatter, extension_id=extension_id
            )

            for key in agent_config.get("strip_frontmatter_keys", []):
                frontmatter.pop(key, None)

            if agent_config.get("inject_name") and not frontmatter.get("name"):
                # Use custom name formatter if provided (e.g., Forge's hyphenated format)
                format_name = agent_config.get("format_name")
                frontmatter["name"] = format_name(cmd_name) if format_name else cmd_name

            if is_cline_ext:
                frontmatter = self._hyphenate_frontmatter_refs(frontmatter)
                body = self._hyphenate_body_refs(body)

            body = self._convert_argument_placeholder(
                body, "$ARGUMENTS", agent_config["args"]
            )

            # Resolve __SPECKIT_COMMAND_*__ tokens using the agent's invoke separator.
            # For dual-layout agents (e.g. Bob) the separator differs between the
            # skills and command layouts, so a single static AGENT_CONFIGS value is
            # insufficient. ``_sep`` (resolved above) is derived from the *output
            # layout* this registrar writes — a "/SKILL.md" scaffold uses the skills
            # separator, any command-layout file uses the command separator — not
            # the project's persisted ai_skills state. Single-layout agents fall back
            # to the static AGENT_CONFIGS value unchanged (invoke_separator_for_mode
            # default).
            # Deferred import of IntegrationBase avoids a circular import at module load
            # (base.py itself imports CommandRegistrar lazily).
            from specify_cli.integrations.base import IntegrationBase  # noqa: PLC0415

            body = IntegrationBase.resolve_command_refs(body, _sep)

            output_name = self._compute_output_name(agent_name, cmd_name, agent_config)

            if agent_config["extension"] == "/SKILL.md":
                output = self.render_skill_command(
                    agent_name,
                    output_name,
                    frontmatter,
                    body,
                    source_id,
                    cmd_file,
                    project_root,
                    extension_id=extension_id,
                )
            elif agent_config["format"] == "markdown":
                body = self.resolve_skill_placeholders(
                    agent_name, frontmatter, body, project_root, extension_id=extension_id
                )
                body = self._convert_argument_placeholder(
                    body, "$ARGUMENTS", agent_config["args"]
                )
                output = self.render_markdown_command(
                    frontmatter, body, source_id, context_note
                )
            elif agent_config["format"] == "toml":
                body = self.resolve_skill_placeholders(
                    agent_name, frontmatter, body, project_root, extension_id=extension_id
                )
                body = self._convert_argument_placeholder(
                    body, "$ARGUMENTS", agent_config["args"]
                )
                output = self.render_toml_command(frontmatter, body, source_id)
            elif agent_config["format"] == "yaml":
                body = self.resolve_skill_placeholders(
                    agent_name, frontmatter, body, project_root
                )
                body = self._convert_argument_placeholder(
                    body, "$ARGUMENTS", agent_config["args"]
                )
                output = self.render_yaml_command(
                    frontmatter, body, source_id, cmd_name
                )
            else:
                raise ValueError(f"Unsupported format: {agent_config['format']}")

            # -- Post-process for non-skills agents -----------------------
            _integration = None
            if agent_config["extension"] != "/SKILL.md":
                from specify_cli.integrations import (  # noqa: PLC0415
                    get_integration,
                )

                _integration = get_integration(agent_name)
                if _integration is not None:
                    output = _integration.post_process_command_content(output)

            dest_file = commands_dir / f"{output_name}{agent_config['extension']}"
            self._ensure_inside(dest_file, commands_dir)
            dest_file.parent.mkdir(parents=True, exist_ok=True)
            self._write_registered_output(
                dest_file,
                output,
                source_dir,
                agent_name,
                output_name,
                agent_config["extension"],
                link_outputs,
                agent_config,
            )

            if agent_name == "copilot":
                self.write_copilot_prompt(project_root, cmd_name)

            registered.append(cmd_name)

            for alias in aliases:
                alias_output_name = self._compute_output_name(
                    agent_name, alias, agent_config
                )

                # For agents with inject_name, render with alias-specific frontmatter
                if agent_config.get("inject_name"):
                    alias_frontmatter = deepcopy(frontmatter)
                    # Use custom name formatter if provided (e.g., Forge's hyphenated format)
                    format_name = agent_config.get("format_name")
                    alias_frontmatter["name"] = (
                        format_name(alias) if format_name else alias
                    )

                    if agent_config["extension"] == "/SKILL.md":
                        alias_output = self.render_skill_command(
                            agent_name,
                            alias_output_name,
                            alias_frontmatter,
                            body,
                            source_id,
                            cmd_file,
                            project_root,
                            extension_id=extension_id,
                        )
                    elif agent_config["format"] == "markdown":
                        alias_output = self.render_markdown_command(
                            alias_frontmatter, body, source_id, context_note
                        )
                    elif agent_config["format"] == "toml":
                        alias_output = self.render_toml_command(
                            alias_frontmatter, body, source_id
                        )
                    elif agent_config["format"] == "yaml":
                        alias_output = self.render_yaml_command(
                            alias_frontmatter, body, source_id, alias
                        )
                    else:
                        raise ValueError(
                            f"Unsupported format: {agent_config['format']}"
                        )

                    if agent_config["extension"] != "/SKILL.md" and _integration is not None:
                        alias_output = _integration.post_process_command_content(alias_output)
                else:
                    # For other agents, reuse the primary output
                    alias_output = output
                    if agent_config["extension"] == "/SKILL.md":
                        alias_output = self.render_skill_command(
                            agent_name,
                            alias_output_name,
                            frontmatter,
                            body,
                            source_id,
                            cmd_file,
                            project_root,
                            extension_id=extension_id,
                        )

                alias_file = (
                    commands_dir / f"{alias_output_name}{agent_config['extension']}"
                )
                self._ensure_inside(alias_file, commands_dir)
                alias_file.parent.mkdir(parents=True, exist_ok=True)
                self._write_registered_output(
                    alias_file,
                    alias_output,
                    source_dir,
                    agent_name,
                    alias_output_name,
                    agent_config["extension"],
                    link_outputs,
                    agent_config,
                )
                if agent_name == "copilot":
                    self.write_copilot_prompt(project_root, alias)
                registered.append(alias)

        return registered

    @staticmethod
    def _write_registered_output(
        dest_file: Path,
        content: str,
        source_dir: Path,
        agent_name: str,
        output_name: str,
        extension: str,
        link_outputs: bool,
        agent_config: dict[str, Any] | None = None,
    ) -> None:
        """Write a rendered agent artifact, optionally as a dev-mode symlink."""
        if not link_outputs or (agent_config or {}).get("dev_no_symlink"):
            if dest_file.is_symlink():
                dest_file.unlink()
            dest_file.write_text(content, encoding="utf-8")
            return

        rel_output = Path(f"{output_name}{extension}")
        cache_root = source_dir / ".specify-dev" / "agent-commands" / agent_name
        cache_file = cache_root / rel_output
        CommandRegistrar._ensure_inside(cache_file, cache_root)

        try:
            cache_file.parent.mkdir(parents=True, exist_ok=True)
            cache_file.write_text(content, encoding="utf-8")
            if dest_file.exists() or dest_file.is_symlink():
                dest_file.unlink()
            target = os.path.relpath(cache_file, dest_file.parent)
            os.symlink(target, dest_file)
        except (OSError, ValueError):
            # Windows often requires Developer Mode or admin privileges for
            # symlinks, and relpath can fail across drives. Keep dev installs
            # functional by falling back to a copy.
            if dest_file.is_symlink():
                dest_file.unlink()
            dest_file.write_text(content, encoding="utf-8")

    @staticmethod
    def write_copilot_prompt(project_root: Path, cmd_name: str) -> None:
        """Generate a companion .prompt.md file for a Copilot agent command.

        Args:
            project_root: Path to project root
            cmd_name: Command name (e.g. 'speckit.my-ext.example')
        """
        prompts_dir = project_root / ".github" / "prompts"
        prompts_dir.mkdir(parents=True, exist_ok=True)
        prompt_file = prompts_dir / f"{cmd_name}.prompt.md"
        CommandRegistrar._ensure_inside(prompt_file, prompts_dir)
        prompt_file.write_text(f"---\nagent: {cmd_name}\n---\n", encoding="utf-8")

    @staticmethod
    def _resolve_agent_dir(
        agent_name: str,
        agent_config: dict[str, Any],
        project_root: Path,
    ) -> Path:
        """Return the agent command directory, falling back to legacy_dir.

        Supports project-relative paths (e.g. ``.claude/skills/``),
        home-relative paths (e.g. ``~/.hermes/skills``), and absolute
        paths — the ``agent_config["dir"]`` value is resolved verbatim
        when absolute or starting with ``~/``, or joined with
        ``project_root`` when relative.

        When the canonical directory does not exist but a ``legacy_dir``
        is configured and present on disk, returns the legacy path and
        emits a deprecation warning advising the user to upgrade.

        Integrations that do not declare ``legacy_dir`` get the canonical
        path unconditionally — no fallback, no warning.
        """
        dir_str = agent_config["dir"]
        if dir_str.startswith("~"):
            # Use Path.home() + remainder instead of expanduser() so tests
            # that monkeypatch Path.home() can properly isolate the home dir.
            # expanduser() uses OS env/user lookup and ignores monkeypatches.
            agent_dir = Path.home() / dir_str[1:].lstrip("/")
        else:
            p = Path(dir_str)
            agent_dir = p if p.is_absolute() else project_root / p
        if not agent_dir.exists():
            legacy = agent_config.get("legacy_dir")
            if legacy:
                legacy_dir = project_root / legacy
                if legacy_dir.exists():
                    import warnings

                    warnings.warn(
                        f"Found legacy '{legacy}' directory for "
                        f"{agent_name}. Run 'specify integration "
                        f"upgrade {agent_name}' to migrate to "
                        f"'{agent_config['dir']}'.",
                        stacklevel=3,
                    )
                    return legacy_dir
        return agent_dir

    def register_commands_for_all_agents(
        self,
        commands: List[Dict[str, Any]],
        source_id: str,
        source_dir: Path,
        project_root: Path,
        context_note: str = None,
        link_outputs: bool = False,
        create_missing_active_skills_dir: bool = False,
        extension_id: Optional[str] = None,
    ) -> Dict[str, List[str]]:
        """Register commands for all detected agents in the project.

        Args:
            commands: List of command info dicts
            source_id: Identifier of the source (extension or preset ID)
            source_dir: Directory containing command source files
            project_root: Path to project root
            context_note: Custom context comment for markdown output
            link_outputs: If True, create dev-mode symlinks for rendered
                command files when supported by the OS.
            create_missing_active_skills_dir: If True, attempt missing-dir
                recovery only for the active initialized skills-backed agent.
                Recovery requires active skills mode (or Kimi's existing native
                skills directory) and is skipped when safe resolution or
                creation fails.
            extension_id: Extension id when rendering extension-owned commands.

        Returns:
            Dictionary mapping agent names to list of registered commands
        """
        results = {}

        self._ensure_configs()
        active_skills_agent = (
            self._active_skills_agent(project_root)
            if create_missing_active_skills_dir else None
        )
        active_skills_dir: Optional[Path] = None
        if active_skills_agent:
            active_skills_config = self.AGENT_CONFIGS.get(active_skills_agent)
            if (
                active_skills_config
                and active_skills_config.get("extension") == "/SKILL.md"
            ):
                active_skills_dir = self._resolve_agent_dir(
                    active_skills_agent, active_skills_config, project_root,
                )
        active_created_skills_dir: Optional[Path] = None
        for agent_name, agent_config in self.AGENT_CONFIGS.items():
            active_skills_output = (
                agent_name == active_skills_agent
                and agent_config.get("extension") == "/SKILL.md"
            )
            recovered_active_skills_dir: Optional[Path] = None
            # Check detect_dir first (project-local marker) if configured,
            # falling back to the resolved dir for output.  This prevents
            # global dirs (e.g. ~/.hermes/skills) from causing false
            # detection in every project.
            detect_dir_str = agent_config.get("detect_dir")
            if detect_dir_str:
                detect_path = project_root / detect_dir_str
                if not detect_path.is_dir():
                    if not active_skills_output:
                        continue
                    try:
                        from . import resolve_active_skills_dir

                        recovered_active_skills_dir = (
                            resolve_active_skills_dir(project_root)
                        )
                    except (ValueError, OSError):
                        continue
                    if recovered_active_skills_dir is None or not detect_path.is_dir():
                        continue
                    active_created_skills_dir = recovered_active_skills_dir
            agent_dir = self._resolve_agent_dir(
                agent_name, agent_config, project_root,
            )
            shares_active_skills_dir = (
                active_skills_dir is not None
                and agent_name != active_skills_agent
                and agent_config.get("extension") == "/SKILL.md"
                and self._same_lexical_path(agent_dir, active_skills_dir)
            )
            if shares_active_skills_dir:
                continue

            agent_dir_existed = agent_dir.is_dir()
            register_missing_active_skills_agent = (
                not agent_dir_existed
                and active_skills_output
            )
            if register_missing_active_skills_agent:
                if recovered_active_skills_dir is None:
                    try:
                        from . import resolve_active_skills_dir

                        recovered_active_skills_dir = (
                            resolve_active_skills_dir(project_root)
                        )
                    except (ValueError, OSError):
                        continue
                    if recovered_active_skills_dir is None:
                        continue
                active_created_skills_dir = recovered_active_skills_dir
            # Shared skill dirs such as .agents/skills should not make
            # later integrations look detected when the active agent just
            # recreated the directory during this registration pass.
            created_by_active_agent = (
                active_created_skills_dir is not None
                and self._same_lexical_path(agent_dir, active_created_skills_dir)
                and agent_name != active_skills_agent
            )
            should_register = (
                agent_dir_existed and not created_by_active_agent
            ) or register_missing_active_skills_agent

            if should_register:
                try:
                    registered = self.register_commands(
                        agent_name,
                        commands,
                        source_id,
                        source_dir,
                        project_root,
                        context_note=context_note,
                        _resolved_dir=agent_dir,
                        link_outputs=link_outputs,
                        extension_id=extension_id,
                    )
                    if registered:
                        results[agent_name] = registered
                    if register_missing_active_skills_agent:
                        active_created_skills_dir = (
                            recovered_active_skills_dir or agent_dir
                        )
                except ValueError:
                    continue
                except OSError:
                    if register_missing_active_skills_agent:
                        continue
                    raise

        return results

    def register_commands_for_non_skill_agents(
        self,
        commands: List[Dict[str, Any]],
        source_id: str,
        source_dir: Path,
        project_root: Path,
        context_note: Optional[str] = None,
        link_outputs: bool = False,
        extension_id: Optional[str] = None,
    ) -> Dict[str, List[str]]:
        """Register commands for all non-skill agents in the project.

        Like register_commands_for_all_agents but skips skill-based agents
        (those with extension '/SKILL.md'). Used by reconciliation to avoid
        overwriting properly formatted SKILL.md files.

        Args:
            commands: List of command info dicts
            source_id: Identifier of the source
            source_dir: Directory containing command source files
            project_root: Path to project root
            context_note: Custom context comment for markdown output
            link_outputs: If True, create dev-mode symlinks for rendered
                command files when supported by the OS.
            extension_id: Extension id when rendering extension-owned commands.

        Returns:
            Dictionary mapping agent names to list of registered commands
        """
        results = {}
        self._ensure_configs()
        for agent_name, agent_config in self.AGENT_CONFIGS.items():
            if agent_config.get("extension") == "/SKILL.md":
                continue
            detect_dir_str = agent_config.get("detect_dir")
            if detect_dir_str:
                detect_path = project_root / detect_dir_str
                if not detect_path.is_dir():
                    continue
            agent_dir = self._resolve_agent_dir(
                agent_name, agent_config, project_root,
            )
            if agent_dir.is_dir():
                try:
                    registered = self.register_commands(
                        agent_name,
                        commands,
                        source_id,
                        source_dir,
                        project_root,
                        context_note=context_note,
                        _resolved_dir=agent_dir,
                        link_outputs=link_outputs,
                        extension_id=extension_id,
                    )
                    if registered:
                        results[agent_name] = registered
                except ValueError:
                    continue
        return results

    def unregister_commands(
        self, registered_commands: Dict[str, List[str]], project_root: Path
    ) -> None:
        """Remove previously registered command files from agent directories.

        When a ``legacy_dir`` is configured, files are removed from
        *both* the canonical and the legacy directory so that orphaned
        commands left behind after an ``integration upgrade`` are
        cleaned up as well.

        Args:
            registered_commands: Dict mapping agent names to command name lists
            project_root: Path to project root
        """
        self._ensure_configs()
        for agent_name, cmd_names in registered_commands.items():
            if agent_name not in self.AGENT_CONFIGS:
                continue

            agent_config = self.AGENT_CONFIGS[agent_name]
            commands_dir = self._resolve_agent_dir(
                agent_name, agent_config, project_root,
            )

            # Collect all directories to clean: canonical (or resolved
            # legacy) plus the legacy dir if it exists separately.
            dirs_to_clean = [commands_dir]
            legacy = agent_config.get("legacy_dir")
            if legacy:
                legacy_dir = project_root / legacy
                if legacy_dir.exists() and legacy_dir != commands_dir:
                    dirs_to_clean.append(legacy_dir)

            for cmd_name in cmd_names:
                output_name = self._compute_output_name(
                    agent_name, cmd_name, agent_config
                )

                names_to_clean = [output_name]
                if output_name != cmd_name and self._is_safe_command_name(cmd_name):
                    names_to_clean.append(cmd_name)

                for target_dir in dirs_to_clean:
                    for name in names_to_clean:
                        cmd_file = (
                            target_dir / f"{name}{agent_config['extension']}"
                        )
                        try:
                            self._ensure_inside(cmd_file, target_dir)
                        except ValueError:
                            continue
                        if cmd_file.exists() or cmd_file.is_symlink():
                            cmd_file.unlink()
                            # For SKILL.md agents each command lives in its own
                            # subdirectory (e.g. .agents/skills/speckit-ext-cmd/
                            # SKILL.md).  Remove the parent dir when it becomes
                            # empty to avoid orphaned directories.
                            parent = cmd_file.parent
                            if parent != target_dir and parent.exists():
                                try:
                                    parent.rmdir()
                                except OSError:
                                    pass

                if agent_name == "copilot":
                    prompt_file = (
                        project_root / ".github" / "prompts" / f"{cmd_name}.prompt.md"
                    )
                    if prompt_file.exists():
                        prompt_file.unlink()


# Populate AGENT_CONFIGS after class definition.
# Catches ImportError from circular imports during module loading;
# _configs_loaded stays False so the next explicit access retries.
try:
    CommandRegistrar._ensure_configs()
except ImportError:
    pass