llm_manager/

config.rs

mod model_config;
mod presets;
mod profiles;
mod store;

use std::collections::HashSet;
use std::path::PathBuf;

use chrono::Local;
use serde::{Deserialize, Serialize};

#[allow(unused_imports)]
pub use model_config::{ModelConfigStore, display_from_key, key_from_display};

pub use profiles::ProfileStore;

use crate::models::{
    Backend, CacheType, CacheTypeK, CacheTypeV, Mirostat, NumMode, RopeScaling, Samplers, SplitMode,
};
use crate::tui::app::ActivePanel;
pub use presets::PresetStore;

/// Default system prompt used when no preset is selected.
pub const DEFAULT_SYSTEM_PROMPT: &str = "You are an expert software developer. Write clean, well-documented code. Explain your reasoning and suggest improvements.";

/// Resolve the base config directory with a safe fallback chain.
///
/// Prefers `dirs::config_dir()` (XDG on Linux, ~/Library/Application Support on macOS,
/// etc.), falls back to `~/.config`, and lastly `./.llm-manager` if both fail.
pub fn config_base_dir() -> PathBuf {
    if let Some(d) = dirs::config_dir() {
        return d;
    }
    if let Some(home) = dirs::home_dir() {
        return home.join(".config");
    }
    PathBuf::from(".").join(".llm-manager")
}

/// Count physical CPU cores on Linux (ignores hyperthreading).
/// Falls back to 1 if the file can't be read or parsing fails.
pub fn physical_cores() -> u32 {
    let content = match std::fs::read_to_string("/proc/cpuinfo") {
        Ok(c) => c,
        Err(_) => {
            return std::thread::available_parallelism()
                .map(|p| p.get() as u32)
                .unwrap_or(1);
        }
    };
    let mut seen = HashSet::new();
    let mut cur_phys: Option<&str> = None;
    let mut cur_core: Option<&str> = None;
    for line in content.lines() {
        if let Some((key, val)) = line.split_once(':') {
            let key = key.trim();
            let val = val.trim();
            match key {
                "physical id" => cur_phys = Some(val),
                "core id" => cur_core = Some(val),
                _ => {}
            }
            if let (Some(phys), Some(core)) = (cur_phys, cur_core) {
                seen.insert((phys, core));
            }
        }
    }
    seen.len() as u32
}

/// A remote RPC worker for distributed inference.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct RpcWorker {
    #[serde(default)]
    pub selected: bool,
    #[serde(default)]
    pub name: String,
    pub ip: String,
    #[serde(default = "default_rpc_port")]
    pub port: u16,
}

fn default_rpc_port() -> u16 {
    50052
}

/// Global configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Config {
    pub models_dirs: Vec<PathBuf>,
    pub llama_server: PathBuf,
    pub default: DefaultParams,
    /// Per-model overrides (keyed by display_name/path relative to model dir, stored as YAML in models/).
    #[serde(default, skip)]
    pub model_overrides: ModelConfigStore,
    /// Named profiles of settings presets (stored as YAML in profiles/).
    #[serde(default, skip)]
    pub profiles: ProfileStore,
    /// System prompt presets (stored as YAML in presets/).
    #[serde(default, skip)]
    pub system_prompt_presets: PresetStore,
    /// RPC Workers for distributed inference.
    #[serde(default)]
    pub rpc_workers: Vec<RpcWorker>,
    /// Number of results per HuggingFace search query.
    #[serde(default = "default_search_limit")]
    pub search_limit: u32,
    /// The last focused panel position (for restoring on next launch).
    #[serde(default)]
    pub active_panel: crate::tui::app::ActivePanel,
    /// Left panel width percentage (20-80).
    #[serde(default = "default_left_pct")]
    pub left_pct: u16,
    /// Server Settings panel height in rows (3-20).
    #[serde(default = "default_server_settings_height")]
    pub server_settings_height: u16,
    /// UI language (en, fr, it). Falls back to en.
    #[serde(default = "default_language")]
    pub language: String,
    /// Whether the first-launch onboarding wizard has been completed.
    #[serde(default = "default_onboarding")]
    pub onboarding_complete: bool,
}

fn default_language() -> String {
    "en".to_string()
}

fn default_left_pct() -> u16 {
    55
}

fn default_server_settings_height() -> u16 {
    9
}

fn default_onboarding() -> bool {
    false
}

fn default_search_limit() -> u32 {
    50
}

/// A named profile of settings.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct Profile {
    pub name: String,
    /// Brief description shown in the profile list.
    pub description: String,
    /// The settings for this profile.
    #[serde(default)]
    pub settings: ModelOverride,
}

impl Profile {
    /// Apply this profile's settings to a base ModelSettings.
    pub fn apply(&self, mut base: crate::models::ModelSettings) -> crate::models::ModelSettings {
        self.settings.apply(&mut base);
        base
    }
}

impl crate::config::store::NamedItem for Profile {
    fn name(&self) -> &str {
        &self.name
    }
}

/// A named system prompt preset.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SystemPromptPreset {
    pub name: String,
    pub description: String,
    pub content: String,
}

impl crate::config::store::NamedItem for SystemPromptPreset {
    fn name(&self) -> &str {
        &self.name
    }
}

/// Built-in system prompt presets.
pub fn builtin_system_prompt_presets() -> Vec<SystemPromptPreset> {
    vec![
        SystemPromptPreset {
            name: "General".into(),
            description: "General-purpose assistant".into(),
            content: "You are a helpful assistant.".into(),
        },
        SystemPromptPreset {
            name: "Coder".into(),
            description: "Expert software developer".into(),
            content: "You are an expert software developer. Write clean, well-documented code. Explain your reasoning and suggest improvements. Split code to avoid too big files. Always try to re-use existing code, avoid duplicate function. In case of reviewing code: every claim must have file:line reference. For kernel Every Kconfig symbol must be verified against actual Kconfig files. Every XML example must be validated against Relax NG schema. Every limitation must be verified against source code. Do NOT include claims you cannot verify. Use tables for structured data. Use code blocks for commands. Cross-reference between components. Workflows using SVG.".into(),
        },
        SystemPromptPreset {
            name: "Thinker".into(),
            description: "Analytical and thoughtful".into(),
            content: "You are a thoughtful and analytical AI assistant. Think carefully before answering. Provide well-reasoned responses with clear explanations. Do NOT include claims you cannot verify".into(),
        },
        SystemPromptPreset {
            name: "Mathematician".into(),
            description: "Expert in mathematics".into(),
            content: "You are an expert in mathematics. Provide clear, step-by-step solutions to mathematical problems. Show your reasoning and explain key concepts.".into(),
        },
    ]
}

#[derive(Debug, Clone, Serialize, Deserialize, Default, PartialEq)]
pub struct ModelOverride {
    // Loading
    pub context_length: Option<u32>,
    pub batch_size: Option<u32>,
    pub ubatch_size: Option<u32>,
    pub cache_type_k: Option<CacheTypeK>,
    pub cache_type_v: Option<CacheTypeV>,
    pub keep: Option<i32>,
    pub swa_full: Option<bool>,
    pub mlock: Option<bool>,
    pub mmap: Option<bool>,
    pub numa: Option<NumMode>,
    pub uniform_cache: Option<bool>,
    pub system_prompt: Option<String>,
    pub system_prompt_preset_name: Option<String>,
    pub max_concurrent_predictions: Option<u32>,
    pub threads: Option<u32>,
    pub threads_batch: Option<u32>,
    pub parallel: Option<u32>,

    // GPU
    pub gpu_layers: Option<i32>,
    pub split_mode: Option<SplitMode>,
    pub tensor_split: Option<String>,
    pub main_gpu: Option<i32>,
    pub fit: Option<bool>,
    pub lora: Option<PathBuf>,
    pub lora_scaled: Option<(PathBuf, f32)>,
    pub rpc: Option<String>,
    pub embedding: Option<bool>,
    pub kv_cache_offload: Option<bool>,
    pub flash_attn: Option<bool>,
    pub jinja: Option<bool>,
    pub auto_chat_template: Option<bool>,
    pub chat_template: Option<String>,
    pub chat_template_kwargs: Option<String>,
    pub expert_count: Option<i32>,
    pub gpu_layers_mode: Option<crate::models::GpuLayersMode>,

    // Sampling
    pub seed: Option<i32>,
    pub temperature: Option<f32>,
    pub top_k: Option<i32>,
    pub top_p: Option<f32>,
    pub min_p: Option<f32>,
    pub typical_p: Option<f32>,
    pub mirostat: Option<Mirostat>,
    pub mirostat_lr: Option<f32>,
    pub mirostat_ent: Option<f32>,
    pub ignore_eos: Option<bool>,
    pub samplers: Option<Samplers>,

    // Repetition
    pub repeat_penalty: Option<f32>,
    pub repeat_last_n: Option<i32>,
    pub presence_penalty: Option<f32>,
    pub frequency_penalty: Option<f32>,
    pub dry_multiplier: Option<f32>,
    pub dry_base: Option<f32>,
    pub dry_allowed_length: Option<i32>,
    pub dry_penalty_last_n: Option<i32>,

    // RoPE
    pub rope_scaling: Option<RopeScaling>,
    pub rope_scale: Option<f32>,
    pub rope_freq_base: Option<f32>,
    pub rope_freq_scale: Option<f32>,
    pub rope_yarn_enabled: Option<bool>,

    // Server
    pub cache_prompt: Option<bool>,
    pub cache_reuse: Option<u32>,
    pub webui: Option<bool>,

    // Other
    pub max_tokens: Option<u32>,
    pub cache_type: Option<CacheType>,
    pub llama_cpp_version_cpu: Option<String>,
    pub llama_cpp_version_vulkan: Option<String>,
    pub llama_cpp_version_rocm: Option<String>,
    pub llama_cpp_version_rocm_lemonade: Option<String>,
    pub llama_cpp_version_cuda: Option<String>,
    pub spec_type: Option<String>,
    pub draft_tokens: Option<u32>,
    pub tags: Option<Vec<String>>,
}

/// Apply a scalar Copy field from override: `base.f = self.f.unwrap_or(base.f)`.
macro_rules! apply_scalar {
    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
        $(
            $base.$field = $self.$field.unwrap_or($base.$field);
        )+
    };
}

/// Apply a Clone field from override: `if let Some(v) = &self.f { base.f = v.clone(); }`.
macro_rules! apply_clone {
    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
        $(
            if let Some(v) = &$self.$field {
                $base.$field = v.clone();
            }
        )+
    };
}

/// Apply an Option<T> field from override: `if let Some(v) = &self.f { base.f = Some(v.clone()); }`.
macro_rules! apply_option {
    ($self:ident, $base:ident, $($field:ident),+ $(,)?) => {
        $(
            if let Some(v) = &$self.$field {
                $base.$field = Some(v.clone());
            }
        )+
    };
}

impl ModelOverride {
    pub fn from_settings(s: &crate::models::ModelSettings) -> Self {
        Self {
            context_length: Some(s.context_length),
            batch_size: Some(s.batch_size),
            ubatch_size: Some(s.ubatch_size),
            cache_type_k: s.cache_type_k,
            cache_type_v: s.cache_type_v,
            keep: Some(s.keep),
            swa_full: Some(s.swa_full),
            mlock: Some(s.mlock),
            mmap: Some(s.mmap),
            numa: Some(s.numa),
            uniform_cache: Some(s.uniform_cache),
            system_prompt: Some(s.system_prompt.clone()),
            system_prompt_preset_name: Some(s.system_prompt_preset_name.clone()),
            max_concurrent_predictions: s.max_concurrent_predictions,
            threads: Some(s.threads),
            threads_batch: Some(s.threads_batch),
            parallel: Some(s.parallel),
            gpu_layers: Some(match s.gpu_layers_mode {
                crate::models::GpuLayersMode::Auto => 0,
                crate::models::GpuLayersMode::Specific(n) => n as i32,
                crate::models::GpuLayersMode::All => -1,
            }),
            gpu_layers_mode: Some(s.gpu_layers_mode),
            split_mode: Some(s.split_mode),
            tensor_split: Some(s.tensor_split.clone()),
            main_gpu: Some(s.main_gpu),
            fit: Some(s.fit),
            lora: s.lora.clone(),
            lora_scaled: s.lora_scaled.clone(),
            rpc: Some(s.rpc.clone()),
            embedding: Some(s.embedding),
            kv_cache_offload: Some(s.kv_cache_offload),
            flash_attn: Some(s.flash_attn),
            jinja: Some(s.jinja),
            auto_chat_template: Some(s.auto_chat_template),
            chat_template: s.chat_template.clone(),
            chat_template_kwargs: s.chat_template_kwargs.clone(),
            expert_count: Some(s.expert_count),
            seed: Some(s.seed),
            temperature: Some(s.temperature),
            top_k: Some(s.top_k),
            top_p: Some(s.top_p),
            min_p: Some(s.min_p),
            typical_p: Some(s.typical_p),
            mirostat: Some(s.mirostat),
            mirostat_lr: Some(s.mirostat_lr),
            mirostat_ent: Some(s.mirostat_ent),
            ignore_eos: Some(s.ignore_eos),
            samplers: Some(s.samplers.clone()),
            repeat_penalty: Some(s.repeat_penalty),
            repeat_last_n: Some(s.repeat_last_n),
            presence_penalty: s.presence_penalty,
            frequency_penalty: s.frequency_penalty,
            dry_multiplier: Some(s.dry_multiplier),
            dry_base: Some(s.dry_base),
            dry_allowed_length: Some(s.dry_allowed_length),
            dry_penalty_last_n: Some(s.dry_penalty_last_n),
            rope_scaling: Some(s.rope_scaling),
            rope_scale: Some(s.rope_scale),
            rope_freq_base: Some(s.rope_freq_base),
            rope_freq_scale: Some(s.rope_freq_scale),
            rope_yarn_enabled: Some(s.rope_yarn_enabled),
            cache_prompt: Some(s.cache_prompt),
            cache_reuse: Some(s.cache_reuse),
            webui: Some(s.webui),
            max_tokens: s.max_tokens,
            cache_type: Some(s.cache_type),
            llama_cpp_version_cpu: s.llama_cpp_version_cpu.clone(),
            llama_cpp_version_vulkan: s.llama_cpp_version_vulkan.clone(),
            llama_cpp_version_rocm: s.llama_cpp_version_rocm.clone(),
            llama_cpp_version_rocm_lemonade: s.llama_cpp_version_rocm_lemonade.clone(),
            llama_cpp_version_cuda: s.llama_cpp_version_cuda.clone(),
            spec_type: Some(s.spec_type.clone()),
            draft_tokens: Some(s.draft_tokens),
            tags: Some(s.tags.clone()),
        }
    }

    /// Merge override into a base ModelSettings (in-place).
    pub fn apply(&self, base: &mut crate::models::ModelSettings) {
        // Override values always take precedence. For Option<T> fields,
        // the override value (even None) is explicitly set by the user.

        // Scalar Copy fields: base.f = self.f.unwrap_or(base.f)
        apply_scalar!(
            self,
            base,
            context_length,
            batch_size,
            ubatch_size,
            keep,
            swa_full,
            mlock,
            mmap,
            numa,
            uniform_cache,
            kv_cache_offload,
            threads,
            threads_batch,
            parallel,
            split_mode,
            main_gpu,
            fit,
            embedding,
            flash_attn,
            jinja,
            auto_chat_template,
            expert_count,
            seed,
            temperature,
            top_k,
            top_p,
            min_p,
            typical_p,
            mirostat,
            mirostat_lr,
            mirostat_ent,
            ignore_eos,
            repeat_penalty,
            repeat_last_n,
            dry_multiplier,
            dry_base,
            dry_allowed_length,
            dry_penalty_last_n,
            rope_scaling,
            rope_scale,
            rope_freq_base,
            rope_freq_scale,
            rope_yarn_enabled,
            cache_prompt,
            cache_reuse,
            webui,
            cache_type,
            draft_tokens,
            gpu_layers_mode,
        );

        // Cloneable fields: if let Some(v) = &self.f { base.f = v.clone(); }
        apply_clone!(
            self,
            base,
            system_prompt,
            system_prompt_preset_name,
            tensor_split,
            rpc,
            samplers,
            spec_type,
            tags,
        );

        // Option<T> fields: if let Some(v) = &self.f { base.f = Some(v.clone()); }
        apply_option!(
            self,
            base,
            lora,
            lora_scaled,
            chat_template,
            chat_template_kwargs,
            llama_cpp_version_cpu,
            llama_cpp_version_vulkan,
            llama_cpp_version_rocm,
            llama_cpp_version_rocm_lemonade,
            llama_cpp_version_cuda,
        );

        // Direct Option<T> assignment (same type in both structs) — only apply if Some
        if let Some(v) = self.cache_type_k {
            base.cache_type_k = Some(v);
        }
        if let Some(v) = self.cache_type_v {
            base.cache_type_v = Some(v);
        }
        if let Some(v) = self.presence_penalty {
            base.presence_penalty = Some(v);
        }
        if let Some(v) = self.frequency_penalty {
            base.frequency_penalty = Some(v);
        }
        base.max_tokens = self.max_tokens.or(base.max_tokens);

        // Special: max_concurrent_predictions uses or() for Option chaining
        base.max_concurrent_predictions = self
            .max_concurrent_predictions
            .or(base.max_concurrent_predictions);

        // Special: gpu_layers converts i32 legacy field to GpuLayersMode enum
        // Only applies when gpu_layers is explicitly set in the override.
        if let Some(n) = self.gpu_layers {
            base.gpu_layers_mode = match n {
                n if n < 0 => crate::models::GpuLayersMode::All,
                n => crate::models::GpuLayersMode::Specific(n as u32),
            };
        }

        // FIELD ACCOUNTING (ModelOverride: 87 fields):
        // - apply_scalar: 53 fields
        // - apply_clone: 7 fields
        // - apply_option: 10 fields
        // - direct Option assign: 5 fields (cache_type_k, cache_type_v, presence_penalty,
        //   frequency_penalty, max_tokens)
        // - special: 1 field (max_concurrent_predictions)
        // - conditional: gpu_layers overrides gpu_layers_mode only when Some
        // - NOT in ModelSettings: 0 (all ModelOverride fields mapped above)
        //
        // ModelSettings fields NOT in ModelOverride (not overridable):
        // host, port, timeout, backend, platform, router_max_models, server_mode,
        // api_endpoint_enabled, api_endpoint_port
        //
        // When adding a field: ensure it appears in exactly one category above.
    }
}

/// Built-in profiles with sensible defaults for popular model families.
pub fn builtin_profiles() -> Vec<Profile> {
    vec![
        Profile {
            name: "Qwen".into(),
            description: "Optimized for Qwen models (dense)".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                temperature: Some(0.7),
                top_k: Some(20),
                top_p: Some(0.95),
                max_tokens: Some(4096),
                presence_penalty: Some(0.0),
                uniform_cache: Some(true),
                jinja: Some(true),
                ..Default::default()
            },
        },
        Profile {
            name: "Qwen-MoE".into(),
            description: "Optimized for Qwen MoE models (35B-A3B)".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                temperature: Some(0.8),
                top_k: Some(20),
                top_p: Some(0.95),
                max_tokens: Some(4096),
                presence_penalty: Some(1.5),
                uniform_cache: Some(true),
                jinja: Some(true),
                ..Default::default()
            },
        },
        Profile {
            name: "Qwen-Coding".into(),
            description: "Optimized for Qwen models in coding mode".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                temperature: Some(0.6),
                top_k: Some(20),
                top_p: Some(0.95),
                max_tokens: Some(4096),
                presence_penalty: Some(0.0),
                uniform_cache: Some(true),
                jinja: Some(true),
                ..Default::default()
            },
        },
        Profile {
            name: "Gemma".into(),
            description: "Optimized for Gemma 2/4 models".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                min_p: Some(0.1),
                temperature: Some(1.0),
                top_k: Some(65),
                top_p: Some(0.95),
                max_tokens: Some(4096),
                uniform_cache: Some(true),
                jinja: Some(true),
                ..Default::default()
            },
        },
        Profile {
            name: "Llama".into(),
            description: "Optimized for Llama 3.1/3.3 models".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                temperature: Some(0.7),
                top_p: Some(0.9),
                repeat_penalty: Some(1.1),
                max_tokens: Some(4096),
                uniform_cache: Some(true),
                jinja: Some(true),
                ..Default::default()
            },
        },
        Profile {
            name: "Mistral".into(),
            description: "Optimized for Mistral 7B/NeMo models".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                temperature: Some(0.7),
                top_k: Some(50),
                top_p: Some(0.9),
                max_tokens: Some(4096),
                uniform_cache: Some(true),
                jinja: Some(true),
                ..Default::default()
            },
        },
        Profile {
            name: "Phi".into(),
            description: "Optimized for Phi 3.5 Mini models".into(),
            settings: ModelOverride {
                context_length: Some(131072),
                temperature: Some(0.7),
                top_k: Some(50),
                top_p: Some(0.9),
                repeat_penalty: Some(1.1),
                max_tokens: Some(4096),
                uniform_cache: Some(true),
                ..Default::default()
            },
        },
    ]
}

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
#[serde(default)]
pub struct DefaultParams {
    // Loading
    #[serde(default)]
    pub context_length: u32,
    #[serde(default)]
    pub threads: u32,
    #[serde(default)]
    pub threads_batch: u32,
    #[serde(default)]
    pub batch_size: u32,
    #[serde(default)]
    pub ubatch_size: u32,
    #[serde(default = "default_cache_type_k")]
    pub cache_type_k: Option<CacheTypeK>,
    #[serde(default = "default_cache_type_v")]
    pub cache_type_v: Option<CacheTypeV>,
    #[serde(default)]
    pub keep: i32,
    #[serde(default)]
    pub swa_full: bool,
    #[serde(default)]
    pub mlock: bool,
    #[serde(default)]
    pub mmap: bool,
    #[serde(default)]
    pub numa: NumMode,
    #[serde(default)]
    pub uniform_cache: bool,
    #[serde(default)]
    pub kv_cache_offload: bool,
    #[serde(default)]
    pub parallel: u32,
    #[serde(default)]
    pub max_concurrent_predictions: Option<u32>,
    #[serde(default)]
    pub system_prompt: String,
    #[serde(default = "default_system_prompt_preset_name")]
    pub system_prompt_preset_name: String,
    // GPU
    #[serde(default)]
    pub gpu_layers: i32,
    #[serde(default = "default_gpu_layers_mode")]
    pub gpu_layers_mode: crate::models::GpuLayersMode,
    #[serde(default)]
    pub split_mode: SplitMode,
    #[serde(default)]
    pub tensor_split: String,
    #[serde(default)]
    pub main_gpu: i32,
    #[serde(default)]
    pub fit: bool,
    #[serde(default)]
    pub lora: Option<PathBuf>,
    #[serde(default)]
    pub lora_scaled: Option<(PathBuf, f32)>,
    #[serde(default)]
    pub rpc: String,
    #[serde(default)]
    pub embedding: bool,
    #[serde(default)]
    pub flash_attn: bool,
    #[serde(default)]
    pub jinja: bool,
    #[serde(default)]
    pub auto_chat_template: bool,
    #[serde(default)]
    pub chat_template: Option<String>,
    #[serde(default)]
    pub chat_template_kwargs: Option<String>,
    #[serde(default)]
    pub expert_count: i32,

    // Sampling
    #[serde(default)]
    pub seed: i32,
    #[serde(default)]
    pub temperature: f32,
    #[serde(default)]
    pub top_k: i32,
    #[serde(default)]
    pub top_p: f32,
    #[serde(default)]
    pub min_p: f32,
    #[serde(default)]
    pub typical_p: f32,
    #[serde(default)]
    pub mirostat: Mirostat,
    #[serde(default)]
    pub mirostat_lr: f32,
    #[serde(default)]
    pub mirostat_ent: f32,
    #[serde(default)]
    pub ignore_eos: bool,
    #[serde(default)]
    pub samplers: Samplers,

    // Repetition
    #[serde(default)]
    pub repeat_penalty: f32,
    #[serde(default)]
    pub repeat_last_n: i32,
    #[serde(default = "default_presence_penalty")]
    pub presence_penalty: Option<f32>,
    #[serde(default = "default_frequency_penalty")]
    pub frequency_penalty: Option<f32>,
    #[serde(default)]
    pub dry_multiplier: f32,
    #[serde(default)]
    pub dry_base: f32,
    #[serde(default)]
    pub dry_allowed_length: i32,
    #[serde(default)]
    pub dry_penalty_last_n: i32,

    // RoPE
    #[serde(default)]
    pub rope_scaling: RopeScaling,
    #[serde(default)]
    pub rope_scale: f32,
    #[serde(default)]
    pub rope_freq_base: f32,
    #[serde(default)]
    pub rope_freq_scale: f32,
    #[serde(default)]
    pub rope_yarn_enabled: bool,

    // Server
    #[serde(default)]
    pub host: String,
    #[serde(default)]
    pub port: u16,
    #[serde(default)]
    pub timeout: u32,
    #[serde(default = "default_cache_prompt")]
    pub cache_prompt: bool,
    #[serde(default)]
    pub cache_reuse: u32,
    #[serde(default)]
    pub webui: bool,
    #[serde(default)]
    pub ws_server_enabled: bool,
    #[serde(default = "default_ws_server_port")]
    pub ws_server_port: u16,
    #[serde(default = "default_server_tls_enabled")]
    pub server_tls_enabled: bool,
    #[serde(default)]
    pub server_tls_cert: Option<String>,
    #[serde(default)]
    pub server_tls_key: Option<String>,
    #[serde(default)]
    pub router_max_models: u32,
    #[serde(default)]
    pub server_mode: crate::models::ServerMode,
    #[serde(default = "default_log_level")]
    pub log_level: String,

    // Other
    #[serde(default = "default_max_tokens")]
    pub max_tokens: Option<u32>,
    #[serde(default)]
    pub cache_type: CacheType,
    #[serde(default)]
    pub backend: Backend,
    /// Platform override: "linux", "windows", or "macos". If None, auto-detected.
    #[serde(default)]
    pub platform: Option<String>,
    #[serde(default)]
    pub llama_cpp_version_cpu: Option<String>,
    #[serde(default)]
    pub llama_cpp_version_vulkan: Option<String>,
    #[serde(default)]
    pub llama_cpp_version_rocm: Option<String>,
    #[serde(default)]
    pub llama_cpp_version_rocm_lemonade: Option<String>,
    #[serde(default)]
    pub llama_cpp_version_cuda: Option<String>,

    // API
    #[serde(default)]
    pub api_endpoint_enabled: bool,
    #[serde(default = "default_api_endpoint_port")]
    pub api_endpoint_port: u16,
    #[serde(default = "default_web_search_engine")]
    pub web_search_engine: String,
    #[serde(default)]
    pub web_search_engine_url: String,
    #[serde(default = "default_web_search_enabled")]
    pub web_search_enabled: bool,
    #[serde(default)]
    pub web_search_api_key: Option<String>,
    #[serde(default)]
    pub api_endpoint_key: Option<String>,
    #[serde(default)]
    pub spec_type: String,
    #[serde(default)]
    pub draft_tokens: u32,
    #[serde(default)]
    pub tags: Vec<String>,
}

fn default_api_endpoint_port() -> u16 {
    49222
}

fn default_web_search_engine() -> String {
    "searxng".to_string()
}

fn default_web_search_enabled() -> bool {
    false
}

fn default_system_prompt_preset_name() -> String {
    "General".to_string()
}

fn default_cache_type_k() -> Option<CacheTypeK> {
    None
}
fn default_cache_type_v() -> Option<CacheTypeV> {
    None
}
fn default_presence_penalty() -> Option<f32> {
    None
}
fn default_frequency_penalty() -> Option<f32> {
    None
}
fn default_max_tokens() -> Option<u32> {
    None
}
fn default_cache_prompt() -> bool {
    true
}
fn default_ws_server_port() -> u16 {
    49223
}

fn default_server_tls_enabled() -> bool {
    true
}

fn default_log_level() -> String {
    "trace".to_string()
}

fn default_gpu_layers_mode() -> crate::models::GpuLayersMode {
    crate::models::GpuLayersMode::Auto
}

impl Default for DefaultParams {
    fn default() -> Self {
        Self {
            // Loading
            context_length: 131072,
            threads: physical_cores(),
            threads_batch: 8,
            batch_size: 512,
            ubatch_size: 512,
            cache_type_k: None,
            cache_type_v: None,
            keep: 0,
            swa_full: false,
            mlock: false,
            mmap: true,
            numa: NumMode::None,
            uniform_cache: true,
            kv_cache_offload: true,
            parallel: 1,
            max_concurrent_predictions: None,
            system_prompt: DEFAULT_SYSTEM_PROMPT.to_string(),
            system_prompt_preset_name: "Coder".to_string(),

            // GPU
            gpu_layers: -1,
            gpu_layers_mode: crate::models::GpuLayersMode::Auto,
            split_mode: SplitMode::Layer,
            tensor_split: String::new(),
            main_gpu: 0,
            fit: true,
            lora: None,
            lora_scaled: None,
            rpc: String::new(),
            embedding: false,
            flash_attn: true,
            jinja: true,
            auto_chat_template: false,
            chat_template: None,
            chat_template_kwargs: None,
            expert_count: -1,

            // Sampling
            seed: -1,
            temperature: 0.8,
            top_k: 40,
            top_p: 0.95,
            min_p: 0.0,
            typical_p: 1.0,
            mirostat: Mirostat::Off,
            mirostat_lr: 0.1,
            mirostat_ent: 5.0,
            ignore_eos: false,
            samplers: Samplers::default(),

            // Repetition
            repeat_penalty: 1.1,
            repeat_last_n: 64,
            presence_penalty: None,
            frequency_penalty: None,
            dry_multiplier: 0.0,
            dry_base: 1.75,
            dry_allowed_length: 2,
            dry_penalty_last_n: -1,

            // RoPE
            rope_scaling: RopeScaling::None,
            rope_scale: 1.0,
            rope_freq_base: 0.0,
            rope_freq_scale: 1.0,
            rope_yarn_enabled: false,

            // Server
            host: "127.0.0.1".to_string(),
            port: 8080,
            timeout: 600,
            cache_prompt: true,
            cache_reuse: 0,
            webui: false,
            ws_server_enabled: false,
            ws_server_port: 49223,
            server_tls_enabled: true,
            server_tls_cert: None,
            server_tls_key: None,
            router_max_models: 4,
            server_mode: crate::models::ServerMode::Normal,
            log_level: "trace".to_string(),

            // Other
            max_tokens: None,
            cache_type: CacheType::F16,
            backend: {
                use crate::backend::hardware::{GpuVendor, detect_gpu_vendors};
                let vendors = detect_gpu_vendors();
                let mut result = Backend::Cpu;
                for v in &vendors {
                    if matches!(v, GpuVendor::Nvidia) {
                        result = Backend::Cuda;
                        break;
                    }
                    if matches!(v, GpuVendor::Amd) {
                        result = Backend::Rocm;
                        break;
                    }
                    if matches!(v, GpuVendor::Intel) {
                        result = Backend::Vulkan;
                        break;
                    }
                }
                result
            },
            platform: None,
            llama_cpp_version_cpu: None,
            llama_cpp_version_vulkan: None,
            llama_cpp_version_rocm: None,
            llama_cpp_version_rocm_lemonade: None,
            llama_cpp_version_cuda: None,
            api_endpoint_enabled: false,
            api_endpoint_port: 49222,
            api_endpoint_key: None,
            web_search_engine: "searxng".to_string(),
            web_search_engine_url: String::new(),
            web_search_enabled: false,
            web_search_api_key: None,
            spec_type: String::new(),
            draft_tokens: 0,
            tags: Vec::new(),
        }
    }
}

impl Default for Config {
    fn default() -> Self {
        Self {
            models_dirs: vec![
                dirs::data_dir()
                    .unwrap_or_default()
                    .join("llm-manager")
                    .join("models"),
            ],
            llama_server: "llama-server".into(),
            default: DefaultParams::default(),
            model_overrides: ModelConfigStore::new(),
            profiles: Default::default(),
            system_prompt_presets: Default::default(),
            rpc_workers: Vec::new(),
            search_limit: default_search_limit(),
            active_panel: ActivePanel::Models,
            left_pct: 55,
            server_settings_height: 9,
            language: default_language(),
            onboarding_complete: false,
        }
    }
}

impl Config {
    pub fn config_path() -> PathBuf {
        config_base_dir().join("llm-manager").join("config.yaml")
    }

    /// Known top-level config keys.
    fn config_keys() -> &'static [&'static str] {
        &[
            "models_dirs",
            "llama_server",
            "default",
            "model_overrides",
            "profiles",
            "system_prompt_presets",
            "rpc_workers",
            "search_limit",
            "active_panel",
            "left_pct",
            "server_settings_height",
            "language",
            "onboarding_complete",
        ]
    }

    /// Known DefaultParams keys.
    fn default_params_keys() -> &'static [&'static str] {
        &[
            "context_length",
            "threads",
            "threads_batch",
            "batch_size",
            "ubatch_size",
            "cache_type_k",
            "cache_type_v",
            "keep",
            "swa_full",
            "mlock",
            "mmap",
            "numa",
            "uniform_cache",
            "kv_cache_offload",
            "parallel",
            "max_concurrent_predictions",
            "system_prompt",
            "system_prompt_preset_name",
            "gpu_layers",
            "gpu_layers_mode",
            "split_mode",
            "tensor_split",
            "main_gpu",
            "fit",
            "lora",
            "lora_scaled",
            "rpc",
            "embedding",
            "flash_attn",
            "jinja",
            "chat_template",
            "chat_template_kwargs",
            "expert_count",
            "seed",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
            "typical_p",
            "mirostat",
            "mirostat_lr",
            "mirostat_ent",
            "ignore_eos",
            "samplers",
            "repeat_penalty",
            "repeat_last_n",
            "presence_penalty",
            "frequency_penalty",
            "dry_multiplier",
            "dry_base",
            "dry_allowed_length",
            "dry_penalty_last_n",
            "rope_scaling",
            "rope_scale",
            "rope_freq_base",
            "rope_freq_scale",
            "rope_yarn_enabled",
            "host",
            "port",
            "timeout",
            "cache_prompt",
            "cache_reuse",
            "webui",
            "ws_server_enabled",
            "ws_server_port",
            "server_tls_enabled",
            "server_tls_cert",
            "server_tls_key",
            "router_max_models",
            "server_mode",
            "log_level",
            "max_tokens",
            "cache_type",
            "backend",
            "platform",
            "llama_cpp_version_cpu",
            "llama_cpp_version_vulkan",
            "llama_cpp_version_rocm",
            "llama_cpp_version_rocm_lemonade",
            "llama_cpp_version_cuda",
            "api_endpoint_enabled",
            "api_endpoint_port",
            "api_endpoint_key",
            "web_search_engine",
            "web_search_engine_url",
            "web_search_enabled",
            "web_search_api_key",
            "spec_type",
            "draft_tokens",
            "tags",
            "auto_chat_template",
        ]
    }

    /// Known ModelOverride keys.
    fn model_override_keys() -> &'static [&'static str] {
        &[
            "context_length",
            "batch_size",
            "ubatch_size",
            "cache_type_k",
            "cache_type_v",
            "keep",
            "swa_full",
            "mlock",
            "mmap",
            "numa",
            "uniform_cache",
            "system_prompt",
            "system_prompt_preset_name",
            "max_concurrent_predictions",
            "threads",
            "threads_batch",
            "parallel",
            "gpu_layers",
            "split_mode",
            "tensor_split",
            "main_gpu",
            "fit",
            "lora",
            "lora_scaled",
            "rpc",
            "embedding",
            "kv_cache_offload",
            "flash_attn",
            "jinja",
            "chat_template",
            "chat_template_kwargs",
            "expert_count",
            "gpu_layers_mode",
            "seed",
            "temperature",
            "top_k",
            "top_p",
            "min_p",
            "typical_p",
            "mirostat",
            "mirostat_lr",
            "mirostat_ent",
            "ignore_eos",
            "samplers",
            "repeat_penalty",
            "repeat_last_n",
            "presence_penalty",
            "frequency_penalty",
            "dry_multiplier",
            "dry_base",
            "dry_allowed_length",
            "dry_penalty_last_n",
            "rope_scaling",
            "rope_scale",
            "rope_freq_base",
            "rope_freq_scale",
            "rope_yarn_enabled",
            "cache_prompt",
            "cache_reuse",
            "webui",
            "max_tokens",
            "cache_type",
            "llama_cpp_version_cpu",
            "llama_cpp_version_vulkan",
            "llama_cpp_version_rocm",
            "llama_cpp_version_rocm_lemonade",
            "llama_cpp_version_cuda",
            "spec_type",
            "draft_tokens",
            "tags",
        ]
    }

    /// Known RpcWorker keys.
    fn rpc_worker_keys() -> &'static [&'static str] {
        &["selected", "name", "ip", "port"]
    }

    /// Validate unknown YAML keys against known field lists.
    pub fn validate_unknown_fields(value: &serde_yml::Value) -> Vec<ValidationWarning> {
        let mut warnings = Vec::new();
        Self::check_unknown_fields(value, "", Self::config_keys(), &mut warnings);
        warnings
    }

    fn check_unknown_fields(
        value: &serde_yml::Value,
        prefix: &str,
        known_keys: &'static [&'static str],
        warnings: &mut Vec<ValidationWarning>,
    ) {
        if let serde_yml::Value::Mapping(map) = value {
            for key in map.keys() {
                let key_str = key.as_str();
                if !known_keys.contains(&key_str) {
                    let field = if prefix.is_empty() {
                        key_str.to_string()
                    } else {
                        format!("{}.{}", prefix, key_str)
                    };
                    warnings.push(ValidationWarning {
                        field,
                        message: format!("Unknown config field: {}", key_str),
                        severity: ValidationSeverity::Warning,
                    });
                }
            }

            // Recurse into nested structures
            for (key, val) in map {
                let key_str = key.as_str();
                let new_prefix = if prefix.is_empty() {
                    key_str.to_string()
                } else {
                    format!("{}.{}", prefix, key_str)
                };

                match key_str {
                    "default" => {
                        Self::check_unknown_fields(
                            val,
                            &new_prefix,
                            Self::default_params_keys(),
                            warnings,
                        );
                    }
                    "model_overrides" => {
                        if let serde_yml::Value::Mapping(overrides) = val {
                            for (override_key, override_val) in overrides {
                                let k = override_key.as_str();
                                let override_prefix = format!("{}.{}", new_prefix, k);
                                Self::check_unknown_fields(
                                    override_val,
                                    &override_prefix,
                                    Self::model_override_keys(),
                                    warnings,
                                );
                            }
                        }
                    }
                    "rpc_workers" => {
                        if let serde_yml::Value::Sequence(items) = val {
                            for item in items {
                                Self::check_unknown_fields(
                                    item,
                                    &new_prefix,
                                    Self::rpc_worker_keys(),
                                    warnings,
                                );
                            }
                        }
                    }
                    _ => {
                        Self::check_unknown_fields(val, &new_prefix, known_keys, warnings);
                    }
                }
            }
        }
    }

    /// Validate config values and return a list of warnings for invalid entries.
    pub fn validate(&self) -> Vec<ValidationWarning> {
        let mut warnings = Vec::new();
        let default = &self.default;

        // Numeric range checks
        if default.context_length < 512 || default.context_length > 1048576 {
            warnings.push(ValidationWarning {
                field: "default.context_length".to_string(),
                message: format!(
                    "context_length {} is outside recommended range 512-1048576",
                    default.context_length
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.threads == 0 {
            warnings.push(ValidationWarning {
                field: "default.threads".to_string(),
                message: "threads must be greater than 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.temperature < 0.0 || default.temperature > 2.0 {
            warnings.push(ValidationWarning {
                field: "default.temperature".to_string(),
                message: format!(
                    "temperature {} is outside recommended range 0.0-2.0",
                    default.temperature
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.top_k < 0 {
            warnings.push(ValidationWarning {
                field: "default.top_k".to_string(),
                message: "top_k must be >= 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.top_p < 0.0 || default.top_p > 1.0 {
            warnings.push(ValidationWarning {
                field: "default.top_p".to_string(),
                message: format!(
                    "top_p {} is outside recommended range 0.0-1.0",
                    default.top_p
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.min_p < 0.0 || default.min_p > 1.0 {
            warnings.push(ValidationWarning {
                field: "default.min_p".to_string(),
                message: format!(
                    "min_p {} is outside recommended range 0.0-1.0",
                    default.min_p
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.typical_p < 0.0 || default.typical_p > 1.0 {
            warnings.push(ValidationWarning {
                field: "default.typical_p".to_string(),
                message: format!(
                    "typical_p {} is outside recommended range 0.0-1.0",
                    default.typical_p
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.repeat_penalty < 0.0 || default.repeat_penalty > 3.0 {
            warnings.push(ValidationWarning {
                field: "default.repeat_penalty".to_string(),
                message: format!(
                    "repeat_penalty {} is outside recommended range 0.0-3.0",
                    default.repeat_penalty
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.mirostat_lr < 0.0 || default.mirostat_lr > 1.0 {
            warnings.push(ValidationWarning {
                field: "default.mirostat_lr".to_string(),
                message: format!(
                    "mirostat_lr {} is outside recommended range 0.0-1.0",
                    default.mirostat_lr
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.mirostat_ent < 0.0 || default.mirostat_ent > 10.0 {
            warnings.push(ValidationWarning {
                field: "default.mirostat_ent".to_string(),
                message: format!(
                    "mirostat_ent {} is outside recommended range 0.0-10.0",
                    default.mirostat_ent
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.dry_multiplier < 0.0 {
            warnings.push(ValidationWarning {
                field: "default.dry_multiplier".to_string(),
                message: "dry_multiplier must be >= 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.dry_base <= 0.0 {
            warnings.push(ValidationWarning {
                field: "default.dry_base".to_string(),
                message: "dry_base must be > 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.rope_scale <= 0.0 {
            warnings.push(ValidationWarning {
                field: "default.rope_scale".to_string(),
                message: "rope_scale must be > 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.rope_freq_scale <= 0.0 {
            warnings.push(ValidationWarning {
                field: "default.rope_freq_scale".to_string(),
                message: "rope_freq_scale must be > 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.port == 0 {
            warnings.push(ValidationWarning {
                field: "default.port".to_string(),
                message: "port must be between 1 and 65535".to_string(),
                severity: ValidationSeverity::Error,
            });
        }
        if default.port < 1024 {
            warnings.push(ValidationWarning {
                field: "default.port".to_string(),
                message: format!("port {} is a privileged port (< 1024)", default.port),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.api_endpoint_port == 0 {
            warnings.push(ValidationWarning {
                field: "default.api_endpoint_port".to_string(),
                message: "api_endpoint_port must be between 1 and 65535".to_string(),
                severity: ValidationSeverity::Error,
            });
        }
        if default.ws_server_port == 0 {
            warnings.push(ValidationWarning {
                field: "default.ws_server_port".to_string(),
                message: "ws_server_port must be between 1 and 65535".to_string(),
                severity: ValidationSeverity::Error,
            });
        }
        if default.timeout < 1 {
            warnings.push(ValidationWarning {
                field: "default.timeout".to_string(),
                message: "timeout must be at least 1 second".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.router_max_models == 0 {
            warnings.push(ValidationWarning {
                field: "default.router_max_models".to_string(),
                message: "router_max_models must be >= 1".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if let Some(max_tokens) = default.max_tokens {
            if max_tokens == 0 {
                warnings.push(ValidationWarning {
                    field: "default.max_tokens".to_string(),
                    message: "max_tokens must be > 0".to_string(),
                    severity: ValidationSeverity::Warning,
                });
            } else if max_tokens > 65536 {
                warnings.push(ValidationWarning {
                    field: "default.max_tokens".to_string(),
                    message: format!("max_tokens {} is very large (> 65536)", max_tokens),
                    severity: ValidationSeverity::Warning,
                });
            }
        }
        if default.context_length > 262144 {
            warnings.push(ValidationWarning {
                field: "default.context_length".to_string(),
                message: format!(
                    "context_length {} may cause high memory usage",
                    default.context_length
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.threads_batch > default.batch_size && default.batch_size > 0 {
            warnings.push(ValidationWarning {
                field: "default.threads_batch".to_string(),
                message: "threads_batch should not exceed batch_size".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }

        // Path validation
        if let Some(lora) = &default.lora
            && !lora.exists()
        {
            warnings.push(ValidationWarning {
                field: "default.lora".to_string(),
                message: format!("lora path does not exist: {}", lora.display()),
                severity: ValidationSeverity::Warning,
            });
        }
        if let Some((lora, _)) = &default.lora_scaled
            && !lora.exists()
        {
            warnings.push(ValidationWarning {
                field: "default.lora_scaled".to_string(),
                message: format!("lora path does not exist: {}", lora.display()),
                severity: ValidationSeverity::Warning,
            });
        }
        if let Some(cert) = &default.server_tls_cert
            && !cert.is_empty()
            && !std::path::Path::new(cert).exists()
        {
            warnings.push(ValidationWarning {
                field: "default.server_tls_cert".to_string(),
                message: format!("TLS cert path does not exist: {}", cert),
                severity: ValidationSeverity::Warning,
            });
        }
        if let Some(key) = &default.server_tls_key
            && !key.is_empty()
            && !std::path::Path::new(key).exists()
        {
            warnings.push(ValidationWarning {
                field: "default.server_tls_key".to_string(),
                message: format!("TLS key path does not exist: {}", key),
                severity: ValidationSeverity::Warning,
            });
        }
        if self.llama_server.is_absolute() && !self.llama_server.exists() {
            warnings.push(ValidationWarning {
                field: "llama_server".to_string(),
                message: format!(
                    "llama_server binary not found: {}",
                    self.llama_server.display()
                ),
                severity: ValidationSeverity::Warning,
            });
        }

        // Model override validation
        for model_name in self.model_overrides.keys() {
            if let Some(override_settings) = self.model_overrides.get(model_name.as_str()) {
                if let Some(lora) = &override_settings.lora
                    && !lora.exists()
                {
                    warnings.push(ValidationWarning {
                        field: format!("model_overrides.{}.lora", model_name),
                        message: format!("lora path does not exist: {}", lora.display()),
                        severity: ValidationSeverity::Warning,
                    });
                }
                if let Some((lora, _)) = &override_settings.lora_scaled
                    && !lora.exists()
                {
                    warnings.push(ValidationWarning {
                        field: format!("model_overrides.{}.lora_scaled", model_name),
                        message: format!("lora path does not exist: {}", lora.display()),
                        severity: ValidationSeverity::Warning,
                    });
                }
            }
        }

        // Cross-field validation
        if default.port == default.api_endpoint_port {
            warnings.push(ValidationWarning {
                field: "default.port".to_string(),
                message: format!(
                    "port conflict: server port and API endpoint port both set to {}",
                    default.port
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.port == default.ws_server_port {
            warnings.push(ValidationWarning {
                field: "default.port".to_string(),
                message: format!(
                    "port conflict: server port and WS server port both set to {}",
                    default.port
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.api_endpoint_port == default.ws_server_port {
            warnings.push(ValidationWarning {
                field: "default.api_endpoint_port".to_string(),
                message: format!(
                    "port conflict: API endpoint port and WS server port both set to {}",
                    default.api_endpoint_port
                ),
                severity: ValidationSeverity::Warning,
            });
        }
        if matches!(default.server_mode, crate::models::ServerMode::Router)
            && default.router_max_models < 2
        {
            warnings.push(ValidationWarning {
                field: "default.router_max_models".to_string(),
                message: "router_max_models should be >= 2 for Router mode".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if !default.spec_type.is_empty() && default.draft_tokens == 0 {
            warnings.push(ValidationWarning {
                field: "default.spec_type".to_string(),
                message: "spec_type is set but draft_tokens is 0".to_string(),
                severity: ValidationSeverity::Warning,
            });
        }
        if default.server_tls_enabled {
            if let Some(cert) = &default.server_tls_cert
                && !cert.is_empty()
                && !std::path::Path::new(cert).exists()
            {
                warnings.push(ValidationWarning {
                    field: "default.server_tls_cert".to_string(),
                    message: format!("TLS cert path does not exist: {}", cert),
                    severity: ValidationSeverity::Warning,
                });
            }
            if let Some(key) = &default.server_tls_key
                && !key.is_empty()
                && !std::path::Path::new(key).exists()
            {
                warnings.push(ValidationWarning {
                    field: "default.server_tls_key".to_string(),
                    message: format!("TLS key path does not exist: {}", key),
                    severity: ValidationSeverity::Warning,
                });
            }
        }

        warnings
    }

    /// Resolve settings for a specific model and profile.
    pub fn resolve_settings(
        &self,
        model_name: Option<&str>,
        profile_name: Option<&str>,
    ) -> crate::models::ModelSettings {
        let mut settings = crate::models::ModelSettings::from_config(self);

        // Apply model-specific override
        if let Some(name) = model_name
            && let Some(override_settings) = self.model_overrides.get(name)
        {
            override_settings.apply(&mut settings);
        }

        // Apply profile override if specified
        if let Some(p_name) = profile_name {
            if let Some(profile) = self.profiles.get(p_name) {
                profile.settings.apply(&mut settings);
            } else if let Some(profile) = builtin_profiles().iter().find(|p| p.name == p_name) {
                profile.settings.apply(&mut settings);
            }
        }

        // Resolve system_prompt from preset name (after all overrides)
        if let Some(preset) = self
            .system_prompt_presets
            .get(&settings.system_prompt_preset_name)
        {
            settings.system_prompt = preset.content.clone();
        }

        settings
    }

    /// Get a system prompt preset content by name.
    pub fn get_preset_content(&self, name: &str) -> Option<String> {
        self.system_prompt_presets
            .get(name)
            .map(|p| p.content.clone())
    }

    fn normalize_config(mut config: Config) -> Config {
        // normalize models_dirs
        for path in &mut config.models_dirs {
            let path_str = path.to_string_lossy();
            if let Some(stripped) = path_str.strip_prefix("~/") {
                let home = dirs::home_dir().unwrap_or_default();
                *path = home.join(stripped);
            } else if !path.is_absolute() {
                let home = dirs::home_dir().unwrap_or_default();
                *path = home.join(path_str.as_ref());
            }
        }

        // Merge built-in profiles into in-memory cache (do not persist to disk)
        for p in builtin_profiles() {
            if config.profiles.get(&p.name).is_none() {
                config.profiles.insert_builtin(p);
            }
        }

        // Merge built-in system prompt presets into in-memory cache (do not persist to disk)
        for p in builtin_system_prompt_presets() {
            if config.system_prompt_presets.get(&p.name).is_none() {
                config.system_prompt_presets.insert_builtin(p);
            }
        }
        config
    }

    fn load_impl(path: &PathBuf) -> Result<Self, Box<dyn std::error::Error>> {
        let content = std::fs::read_to_string(path)?;

        // Phase 1: Parse as Value to check for unknown fields
        let parsed: serde_yml::Value = serde_yml::from_str(&content)
            .map_err(|e| format!("Failed to parse config file {}: {}", path.display(), e))?;
        let mut warnings = Self::validate_unknown_fields(&parsed);

        // Phase 2: Deserialize normally (serde_yaml ignores unknown fields)
        let config: Config = serde_yml::from_str(&content)
            .map_err(|e| format!("Failed to parse config file {}: {}", path.display(), e))?;
        let config = Self::normalize_config(config);
        let config = config.auto_detect_platform();

        // Phase 3: Value validation
        warnings.extend(config.validate());

        // Log warnings
        if !warnings.is_empty() {
            eprintln!("Config validation warnings:");
            for warning in &warnings {
                eprintln!(
                    "  [{}] {}: {}",
                    match warning.severity {
                        ValidationSeverity::Warning => "WARN",
                        ValidationSeverity::Error => "ERROR",
                    },
                    warning.field,
                    warning.message
                );
            }
        }
        Ok(config)
    }

    pub fn load() -> Result<Self, Box<dyn std::error::Error>> {
        let path = Self::config_path();
        if path.exists() {
            Self::load_impl(&path)
        } else {
            let mut config = Config::default();
            config.save()?;
            Ok(config)
        }
    }

    pub fn load_from(path: PathBuf) -> Result<Self, Box<dyn std::error::Error>> {
        if path.exists() {
            let content = std::fs::read_to_string(&path)?;

            // Phase 1: Parse as Value to check for unknown fields
            let parsed: serde_yml::Value = serde_yml::from_str(&content)
                .map_err(|e| format!("Failed to parse config file {}: {}", path.display(), e))?;
            let mut warnings = Self::validate_unknown_fields(&parsed);

            // Phase 2: Deserialize normally
            let config: Config = serde_yml::from_str(&content)
                .map_err(|e| format!("Failed to parse config file {}: {}", path.display(), e))?;
            let config = Self::normalize_config(config);
            let config = config.auto_detect_platform();

            // Phase 3: Value validation
            warnings.extend(config.validate());

            if !warnings.is_empty() {
                eprintln!("Config validation warnings:");
                for warning in &warnings {
                    eprintln!(
                        "  [{}] {}: {}",
                        match warning.severity {
                            ValidationSeverity::Warning => "WARN",
                            ValidationSeverity::Error => "ERROR",
                        },
                        warning.field,
                        warning.message
                    );
                }
            }
            Ok(config)
        } else {
            Err(format!("Config file not found: {}", path.display()).into())
        }
    }

    /// Auto-detect the platform if not explicitly set in config.
    fn auto_detect_platform(mut self) -> Self {
        if self.default.platform.is_none() {
            self.default.platform =
                Some(
                    crate::backend::hardware::platform_name(
                        crate::backend::hardware::detect_platform(),
                    )
                    .to_string(),
                );
        }
        self
    }

    pub fn save(&mut self) -> Result<(), Box<dyn std::error::Error>> {
        let path = Self::config_path();
        if let Some(parent) = path.parent() {
            std::fs::create_dir_all(parent)?;
        }
        let content = serde_yml::to_string(self)?;
        std::fs::write(&path, content)?;
        #[cfg(unix)]
        {
            use std::os::unix::fs::PermissionsExt;
            std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o600))?;
        }
        // Persist model configs to individual YAML files
        let entries: Vec<(String, ModelOverride)> = self
            .model_overrides
            .keys()
            .iter()
            .filter_map(|k| self.model_overrides.get(k).map(|v| (k.clone(), v.clone())))
            .collect();
        for (name, cfg) in entries {
            self.model_overrides.save(&name, &cfg);
        }
        // Persist user profiles to individual YAML files (skip built-ins)
        for profile in self.profiles.user_profiles() {
            self.profiles.save(&profile);
        }
        // Persist user presets to individual YAML files (skip built-ins)
        for preset in self.system_prompt_presets.user_presets() {
            self.system_prompt_presets.save(&preset);
        }
        Ok(())
    }

    pub fn merged_profiles(&self) -> Vec<Profile> {
        self.profiles.all()
    }

    pub fn merged_presets(&self) -> Vec<SystemPromptPreset> {
        self.system_prompt_presets.all()
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum LogLevel {
    Info,
    Warning,
    Error,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ValidationSeverity {
    Warning,
    Error,
}

#[derive(Debug, Clone)]
pub struct ValidationWarning {
    pub field: String,
    pub message: String,
    pub severity: ValidationSeverity,
}

impl LogLevel {
    pub fn label(&self) -> &'static str {
        match self {
            LogLevel::Info => "INFO",
            LogLevel::Warning => "WARNING",
            LogLevel::Error => "ERROR",
        }
    }
}

#[derive(Debug, Clone)]
pub struct LogEntry {
    pub timestamp: String,
    pub level: LogLevel,
    pub message: String,
}

impl LogEntry {
    pub fn new(message: impl Into<String>, level: LogLevel) -> Self {
        let timestamp = Local::now().format("%H:%M:%S").to_string();
        let message = sanitize_log(&message.into());
        Self {
            timestamp,
            level,
            message,
        }
    }
}

/// Sanitize log messages to prevent TUI layout breakages.
/// Strips non-printable characters and control sequences, and limits length.
fn sanitize_log(input: &str) -> String {
    // Limit length to avoid layout/perf issues with massive lines
    let max_len = 2000;
    let chars: Vec<char> = input.chars().collect();
    let truncated = chars.len() > max_len;
    let chars = if truncated {
        chars[..max_len].to_vec()
    } else {
        chars
    };

    let mut output = String::with_capacity(chars.len());
    for c in chars {
        // Strip ALL control characters except newline and tab.
        // Critically: strip \r (carriage return) as it breaks TUI rendering.
        if c.is_control() && c != '\n' && c != '\t' {
            continue;
        }
        output.push(c);
    }

    // Replace tabs with spaces for consistent rendering
    let output = output.replace('\t', "    ");

    // Final trim to remove trailing junk
    let mut result = output.trim_end().to_string();
    if truncated {
        result.push_str("... (truncated)");
    }
    result
}