from pydantic import BaseModel, Field, field_validator, model_validator


class ToolScopeConfig(BaseModel):
    """Tool set for a single scope (agent or subagent)."""

    native: list[str] = Field(default_factory=list)
    mcp: dict[str, list[str]] = Field(default_factory=dict)


class ToolConfig(BaseModel):
    """Explicit per-profile tool configuration."""

    agent: ToolScopeConfig = Field(default_factory=ToolScopeConfig)
    subagent: ToolScopeConfig = Field(default_factory=ToolScopeConfig)


class AgentProfile(BaseModel):
    """
    Defines a complete agent configuration.
    A profile ties together a system prompt, an LLM backend, a model,
    and the set of tools the agent is allowed to use.
    """

    model_config = {"extra": "allow"}

    id: str
    name: str
    description: str
    system_prompt: str
    tools: ToolConfig = Field(default_factory=ToolConfig)

    # --- Deprecated: kept for auto-migration from old configs ------------------
    # These fields are automatically migrated into ``tools`` by the loader.
    # Do not set them in new configs; use ``tools.agent`` / ``tools.subagent``.
    enabled_tools: list[str] = Field(default_factory=list)
    mcp_servers: dict[str, list[str]] = Field(default_factory=dict)
    subagent_tools: list[str] = Field(default_factory=list)

    llm_backend: str = "ollama"  # backend key, e.g. "ollama", "openai"
    # Ordered list of preferred models; first available wins at runtime.
    # Accepts a plain string for backward compatibility (auto-wrapped in a list).
    model: list[str] = Field(default_factory=lambda: ["gemma4:31b-cloud"])

    max_iterations: int = 10
    temperature: float = 0.7
    top_k: int | None = None
    top_p: float | None = None
    # Number of CPU threads for local inference. None = Ollama default (physical cores).
    # Cloud models ignore this option.
    num_thread: int | None = None
    planning_enabled: bool = False  # if True, run a planning LLM call before the main loop

    # Profile discoverability — used for system prompt injection and list_profiles tool.
    # short_description: 1-line summary shown in every system prompt to all profiles.
    # full_description: structured dict with keys: specialization, when_to_use, key_tools.
    short_description: str = ""
    full_description: dict = Field(default_factory=dict)

    # Admin-only profiles are hidden from non-admin users in the profile list.
    is_admin_only: bool = False

    # Subagent-only profiles can only be used via spawn_agent — switch_profile
    # is blocked. Useful for narrow specialist agents that should never become
    # the main session profile.
    is_subagent_only: bool = False

    # ── Thinking mechanics ────────────────────────────────────────────────────
    # Each flag can be set per-profile in config.json to tune the balance
    # between reasoning depth and response latency.

    # Extended reasoning: passes think=True to the LLM on every main-loop call.
    # Disable for latency-sensitive profiles (e.g. smart_home).
    think_enabled: bool = True

    # Inject remaining iteration count at the end of every LLM context so the
    # model knows when to wrap up instead of hitting the limit blindly.
    iteration_budget_enabled: bool = True

    # ── Planning phases ───────────────────────────────────────────────────────
    # planning_mandatory: if True, the DIRECT shortcut is never offered to the
    #   model — planning always runs in full. If False, the model can output DIRECT
    #   to skip straight to execution (simple requests bypass planning).
    #   First-message planning is always forced regardless of this flag.
    planning_mandatory: bool = False

    # Individual phase switches — allow disabling expensive phases for profiles
    # that don't need them.
    # Phase 1: task analysis (TASK/GOAL/UNKNOWNS). Entry point for the pipeline.
    planning_phase1_enabled: bool = True
    # Phase 2: structured review (Critic/Pragmatist/Detailer + Plan Adjustments).
    #   Adds 1 LLM call only when Phase 1 outputs REFLECT: yes.
    planning_phase2_enabled: bool = False
    # Phase 3: structured execution plan (numbered steps with TOOL/AGENT/SELF).
    planning_phase3_enabled: bool = True

    # Inject a goal-reminder system message every N iterations to prevent drift
    # on long tasks. N = goal_anchoring_interval.
    goal_anchoring_enabled: bool = True
    goal_anchoring_interval: int = 5

    # Detect when the model is looping without todo progress for
    # anti_stall_threshold iterations and inject a hard warning.
    anti_stall_enabled: bool = True
    anti_stall_threshold: int = 8

    # After the model marks a todo step as done, run a lightweight LLM check:
    # "did the result actually satisfy the step goal?" Adds ~1 LLM call per step.
    step_validation_enabled: bool = False

    # When a todo step is marked as failed, trigger a lightweight re-planning
    # pass to adjust the remaining steps. Depends on step_validation.
    adaptive_replan_enabled: bool = False

    # Sub-agent configuration
    # subagent_system_prompt: injected as an additional system message for sub-agents,
    #   after the profile's main system_prompt. Loaded from subagent_system_prompt.txt if present.
    subagent_planning_enabled: bool = False
    subagent_think_enabled: bool | None = None
    subagent_system_prompt: str = ""

    # Extra context providers to inject for this profile (by name).
    # Global providers (global_provider=True) are always injected regardless of this list.
    context_providers: list[str] = Field(default_factory=list)

    @field_validator("model", mode="before")
    @classmethod
    def _coerce_model(cls, v):
        if isinstance(v, str):
            return [v] if v else ["gemma4:31b-cloud"]
        return v

    @model_validator(mode="after")
    def _migrate_tools(self):
        """Auto-migrate legacy fields into the explicit ``tools`` structure."""
        if self.enabled_tools or self.mcp_servers or self.subagent_tools:
            if not self.tools.agent.native:
                self.tools.agent.native = list(self.enabled_tools)
            if not self.tools.agent.mcp:
                self.tools.agent.mcp = dict(self.mcp_servers)
            if self.subagent_tools:
                if not self.tools.subagent.native:
                    self.tools.subagent.native = list(self.subagent_tools)
                if not self.tools.subagent.mcp:
                    # Infer MCP servers from subagent_tools if they contain mcp__ names
                    from navi.mcp.tools import is_mcp_tool
                    inferred: dict[str, list[str]] = {}
                    for name in self.subagent_tools:
                        if is_mcp_tool(name):
                            # e.g. mcp__navi_3d__compile_scad -> server navi_3d
                            parts = name.split("__", 2)
                            if len(parts) >= 2:
                                srv = parts[1]
                                inferred.setdefault(srv, []).append("*")
                    self.tools.subagent.mcp = inferred
        return self

    def get_agent_tools(self) -> ToolScopeConfig:
        """Return the resolved agent tool set."""
        return self.tools.agent

    def get_subagent_tools(self) -> ToolScopeConfig:
        """Return the resolved subagent tool set."""
        if self.tools.subagent.native or self.tools.subagent.mcp:
            return self.tools.subagent
        # Fallback: if subagent is empty but agent is not, use agent
        return self.tools.agent