vtcode_config/core/

agent.rs

use crate::constants::{defaults, llm_generation, prompt_budget};
use crate::types::{
    ReasoningEffortLevel, SystemPromptMode, ToolDocumentationMode, UiSurfacePreference,
    VerbosityLevel,
};
use serde::{Deserialize, Serialize};
use std::collections::BTreeMap;

const DEFAULT_CHECKPOINTS_ENABLED: bool = true;
const DEFAULT_MAX_SNAPSHOTS: usize = 50;
const DEFAULT_MAX_AGE_DAYS: u64 = 30;

/// Agent-wide configuration
#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentConfig {
    /// AI provider for single agent mode (gemini, openai, anthropic, openrouter, zai)
    #[serde(default = "default_provider")]
    pub provider: String,

    /// Environment variable that stores the API key for the active provider
    #[serde(default = "default_api_key_env")]
    pub api_key_env: String,

    /// Default model to use
    #[serde(default = "default_model")]
    pub default_model: String,

    /// UI theme identifier controlling ANSI styling
    #[serde(default = "default_theme")]
    pub theme: String,

    /// System prompt mode controlling prompt verbosity and token overhead.
    /// Options target lean base prompts: minimal (~150-250 tokens), lightweight/default
    /// (~250-350 tokens), specialized (~350-500 tokens) before dynamic runtime addenda.
    #[serde(default)]
    pub system_prompt_mode: SystemPromptMode,

    /// Tool documentation mode controlling token overhead for tool definitions
    /// Options: minimal (~800 tokens), progressive (~1.2k), full (~3k current)
    /// Progressive: signatures upfront, detailed docs on-demand (recommended)
    /// Minimal: signatures only, pi-coding-agent style (power users)
    /// Full: all documentation upfront (current behavior, default)
    #[serde(default)]
    pub tool_documentation_mode: ToolDocumentationMode,

    /// Enable split tool results for massive token savings (Phase 4)
    /// When enabled, tools return dual-channel output:
    /// - llm_content: Concise summary sent to LLM (token-optimized, 53-95% reduction)
    /// - ui_content: Rich output displayed to user (full details preserved)
    ///   Applies to: unified_search, unified_file, unified_exec
    ///   Default: true (opt-out for compatibility), recommended for production use
    #[serde(default = "default_enable_split_tool_results")]
    pub enable_split_tool_results: bool,

    /// Enable TODO planning helper mode for structured task management
    #[serde(default = "default_todo_planning_mode")]
    pub todo_planning_mode: bool,

    /// Preferred rendering surface for the interactive chat UI (auto, alternate, inline)
    #[serde(default)]
    pub ui_surface: UiSurfacePreference,

    /// Maximum number of conversation turns before auto-termination
    #[serde(default = "default_max_conversation_turns")]
    pub max_conversation_turns: usize,

    /// Reasoning effort level for models that support it (none, minimal, low, medium, high, xhigh)
    /// Applies to: Claude, GPT-5 family, Gemini, Qwen3, DeepSeek with reasoning capability
    #[serde(default = "default_reasoning_effort")]
    pub reasoning_effort: ReasoningEffortLevel,

    /// Verbosity level for output text (low, medium, high)
    /// Applies to: GPT-5.4-family Responses workflows and other models that support verbosity control
    #[serde(default = "default_verbosity")]
    pub verbosity: VerbosityLevel,

    /// Temperature for main LLM responses (0.0-1.0)
    /// Lower values = more deterministic, higher values = more creative
    /// Recommended: 0.7 for balanced creativity and consistency
    /// Range: 0.0 (deterministic) to 1.0 (maximum randomness)
    #[serde(default = "default_temperature")]
    pub temperature: f32,

    /// Temperature for prompt refinement (0.0-1.0, default: 0.3)
    /// Lower values ensure prompt refinement is more deterministic/consistent
    /// Keep lower than main temperature for stable prompt improvement
    #[serde(default = "default_refine_temperature")]
    pub refine_temperature: f32,

    /// Enable an extra self-review pass to refine final responses
    #[serde(default = "default_enable_self_review")]
    pub enable_self_review: bool,

    /// Maximum number of self-review passes
    #[serde(default = "default_max_review_passes")]
    pub max_review_passes: usize,

    /// Enable prompt refinement pass before sending to LLM
    #[serde(default = "default_refine_prompts_enabled")]
    pub refine_prompts_enabled: bool,

    /// Max refinement passes for prompt writing
    #[serde(default = "default_refine_max_passes")]
    pub refine_prompts_max_passes: usize,

    /// Optional model override for the refiner (empty = auto pick efficient sibling)
    #[serde(default)]
    pub refine_prompts_model: String,

    /// Small/lightweight model configuration for efficient operations
    /// Used for tasks like large file reads, parsing, git history, conversation summarization
    /// Typically 70-80% cheaper than main model; ~50% of VT Code's calls use this tier
    #[serde(default)]
    pub small_model: AgentSmallModelConfig,

    /// Inline prompt suggestion configuration for the chat composer
    #[serde(default)]
    pub prompt_suggestions: AgentPromptSuggestionsConfig,

    /// Session onboarding and welcome message configuration
    #[serde(default)]
    pub onboarding: AgentOnboardingConfig,

    /// Maximum bytes of AGENTS.md content to load from project hierarchy
    #[serde(default = "default_project_doc_max_bytes")]
    pub project_doc_max_bytes: usize,

    /// Additional filenames to check when AGENTS.md is absent at a directory level.
    #[serde(default)]
    pub project_doc_fallback_filenames: Vec<String>,

    /// Maximum bytes of instruction content to load from AGENTS.md hierarchy
    #[serde(
        default = "default_instruction_max_bytes",
        alias = "rule_doc_max_bytes"
    )]
    pub instruction_max_bytes: usize,

    /// Additional instruction files or globs to merge into the hierarchy
    #[serde(default, alias = "instruction_paths", alias = "instructions")]
    pub instruction_files: Vec<String>,

    /// Instruction files or globs to exclude from AGENTS.md and rules discovery
    #[serde(default)]
    pub instruction_excludes: Vec<String>,

    /// Maximum recursive `@path` import depth for instruction and rule files
    #[serde(default = "default_instruction_import_max_depth")]
    pub instruction_import_max_depth: usize,

    /// Durable per-repository memory for main sessions
    #[serde(default)]
    pub persistent_memory: PersistentMemoryConfig,

    /// Provider-specific API keys captured from interactive configuration flows
    ///
    /// Note: Actual API keys are stored securely in the OS keyring.
    /// This field only tracks which providers have keys stored (for UI/migration purposes).
    /// The keys themselves are NOT serialized to the config file for security.
    #[serde(default, skip_serializing)]
    pub custom_api_keys: BTreeMap<String, String>,

    /// Preferred storage backend for credentials (OAuth tokens, API keys, etc.)
    ///
    /// - `keyring`: Use OS-specific secure storage (macOS Keychain, Windows Credential
    ///   Manager, Linux Secret Service). This is the default as it's the most secure.
    /// - `file`: Use AES-256-GCM encrypted file with machine-derived key
    /// - `auto`: Try keyring first, fall back to file if unavailable
    #[serde(default)]
    pub credential_storage_mode: crate::auth::AuthCredentialsStoreMode,

    /// Checkpointing configuration for automatic turn snapshots
    #[serde(default)]
    pub checkpointing: AgentCheckpointingConfig,

    /// Vibe coding configuration for lazy or vague request support
    #[serde(default)]
    pub vibe_coding: AgentVibeCodingConfig,

    /// Maximum number of retries for agent task execution (default: 2)
    /// When an agent task fails due to retryable errors (timeout, network, 503, etc.),
    /// it will be retried up to this many times with exponential backoff
    #[serde(default = "default_max_task_retries")]
    pub max_task_retries: u32,

    /// Harness configuration for turn-level budgets, telemetry, and execution limits
    #[serde(default)]
    pub harness: AgentHarnessConfig,

    /// Experimental Codex app-server sidecar configuration.
    #[serde(default)]
    pub codex_app_server: AgentCodexAppServerConfig,

    /// Include current date/time in system prompt for temporal awareness
    /// Helps LLM understand context for time-sensitive tasks (default: true)
    #[serde(default = "default_include_temporal_context")]
    pub include_temporal_context: bool,

    /// Use UTC instead of local time for temporal context in system prompts
    #[serde(default)]
    pub temporal_context_use_utc: bool,

    /// Include current working directory in system prompt (default: true)
    #[serde(default = "default_include_working_directory")]
    pub include_working_directory: bool,

    /// Controls inclusion of the structured reasoning tag instructions block.
    ///
    /// Behavior:
    /// - `Some(true)`: always include structured reasoning instructions.
    /// - `Some(false)`: never include structured reasoning instructions.
    /// - `None` (default): include only for `default` and `specialized` prompt modes.
    ///
    /// This keeps lightweight/minimal prompts smaller by default while allowing
    /// explicit opt-in when users want tag-based reasoning guidance.
    #[serde(default)]
    pub include_structured_reasoning_tags: Option<bool>,

    /// Custom instructions provided by the user via configuration to guide agent behavior
    #[serde(default)]
    pub user_instructions: Option<String>,

    /// Require user confirmation before executing a plan generated in plan mode
    /// When true, exiting plan mode shows the implementation blueprint and
    /// requires explicit user approval before enabling edit tools.
    #[serde(default = "default_require_plan_confirmation")]
    pub require_plan_confirmation: bool,

    /// Circuit breaker configuration for resilient tool execution
    /// Controls when the agent should pause and ask for user guidance due to repeated failures
    #[serde(default)]
    pub circuit_breaker: CircuitBreakerConfig,

    /// Open Responses specification compliance configuration
    /// Enables vendor-neutral LLM API format for interoperable workflows
    #[serde(default)]
    pub open_responses: OpenResponsesConfig,
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[cfg_attr(feature = "schema", schemars(rename_all = "snake_case"))]
#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize)]
#[serde(rename_all = "snake_case")]
pub enum ContinuationPolicy {
    Off,
    ExecOnly,
    #[default]
    All,
}

impl ContinuationPolicy {
    pub fn as_str(&self) -> &'static str {
        match self {
            Self::Off => "off",
            Self::ExecOnly => "exec_only",
            Self::All => "all",
        }
    }

    pub fn parse(value: &str) -> Option<Self> {
        let normalized = value.trim();
        if normalized.eq_ignore_ascii_case("off") {
            Some(Self::Off)
        } else if normalized.eq_ignore_ascii_case("exec_only")
            || normalized.eq_ignore_ascii_case("exec-only")
        {
            Some(Self::ExecOnly)
        } else if normalized.eq_ignore_ascii_case("all") {
            Some(Self::All)
        } else {
            None
        }
    }
}

impl<'de> Deserialize<'de> for ContinuationPolicy {
    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
    where
        D: serde::Deserializer<'de>,
    {
        let raw = String::deserialize(deserializer)?;
        Ok(Self::parse(&raw).unwrap_or_default())
    }
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[cfg_attr(feature = "schema", schemars(rename_all = "snake_case"))]
#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize)]
#[serde(rename_all = "snake_case")]
pub enum HarnessOrchestrationMode {
    #[default]
    Single,
    PlanBuildEvaluate,
}

impl HarnessOrchestrationMode {
    pub fn as_str(&self) -> &'static str {
        match self {
            Self::Single => "single",
            Self::PlanBuildEvaluate => "plan_build_evaluate",
        }
    }

    pub fn parse(value: &str) -> Option<Self> {
        let normalized = value.trim();
        if normalized.eq_ignore_ascii_case("single") {
            Some(Self::Single)
        } else if normalized.eq_ignore_ascii_case("plan_build_evaluate")
            || normalized.eq_ignore_ascii_case("plan-build-evaluate")
            || normalized.eq_ignore_ascii_case("planner_generator_evaluator")
            || normalized.eq_ignore_ascii_case("planner-generator-evaluator")
        {
            Some(Self::PlanBuildEvaluate)
        } else {
            None
        }
    }
}

impl<'de> Deserialize<'de> for HarnessOrchestrationMode {
    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
    where
        D: serde::Deserializer<'de>,
    {
        let raw = String::deserialize(deserializer)?;
        Ok(Self::parse(&raw).unwrap_or_default())
    }
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentHarnessConfig {
    /// Maximum number of tool calls allowed per turn. Set to `0` to disable the cap.
    #[serde(default = "default_harness_max_tool_calls_per_turn")]
    pub max_tool_calls_per_turn: usize,
    /// Maximum wall clock time (seconds) for tool execution in a turn
    #[serde(default = "default_harness_max_tool_wall_clock_secs")]
    pub max_tool_wall_clock_secs: u64,
    /// Maximum retries for retryable tool errors
    #[serde(default = "default_harness_max_tool_retries")]
    pub max_tool_retries: u32,
    /// Enable automatic context compaction when token pressure crosses threshold.
    ///
    /// Disabled by default. When disabled, no automatic compaction is triggered.
    #[serde(default = "default_harness_auto_compaction_enabled")]
    pub auto_compaction_enabled: bool,
    /// Optional absolute compact threshold (tokens) for Responses server-side compaction.
    ///
    /// When unset, VT Code derives a threshold from the provider context window.
    #[serde(default)]
    pub auto_compaction_threshold_tokens: Option<u64>,
    /// Provider-native tool-result clearing policy.
    #[serde(default)]
    pub tool_result_clearing: ToolResultClearingConfig,
    /// Optional maximum estimated API cost in USD before VT Code stops the session.
    #[serde(default)]
    pub max_budget_usd: Option<f64>,
    /// Controls whether harness-managed continuation loops are enabled.
    #[serde(default)]
    pub continuation_policy: ContinuationPolicy,
    /// Optional JSONL event log path for harness events.
    /// Defaults to `~/.vtcode/sessions/` when unset.
    #[serde(default)]
    pub event_log_path: Option<String>,
    /// Select the exec/full-auto harness orchestration path.
    #[serde(default)]
    pub orchestration_mode: HarnessOrchestrationMode,
    /// Maximum generator revision rounds after evaluator rejection.
    #[serde(default = "default_harness_max_revision_rounds")]
    pub max_revision_rounds: usize,
}

impl Default for AgentHarnessConfig {
    fn default() -> Self {
        Self {
            max_tool_calls_per_turn: default_harness_max_tool_calls_per_turn(),
            max_tool_wall_clock_secs: default_harness_max_tool_wall_clock_secs(),
            max_tool_retries: default_harness_max_tool_retries(),
            auto_compaction_enabled: default_harness_auto_compaction_enabled(),
            auto_compaction_threshold_tokens: None,
            tool_result_clearing: ToolResultClearingConfig::default(),
            max_budget_usd: None,
            continuation_policy: ContinuationPolicy::default(),
            event_log_path: None,
            orchestration_mode: HarnessOrchestrationMode::default(),
            max_revision_rounds: default_harness_max_revision_rounds(),
        }
    }
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct ToolResultClearingConfig {
    #[serde(default = "default_tool_result_clearing_enabled")]
    pub enabled: bool,
    #[serde(default = "default_tool_result_clearing_trigger_tokens")]
    pub trigger_tokens: u64,
    #[serde(default = "default_tool_result_clearing_keep_tool_uses")]
    pub keep_tool_uses: u32,
    #[serde(default = "default_tool_result_clearing_clear_at_least_tokens")]
    pub clear_at_least_tokens: u64,
    #[serde(default)]
    pub clear_tool_inputs: bool,
}

impl Default for ToolResultClearingConfig {
    fn default() -> Self {
        Self {
            enabled: default_tool_result_clearing_enabled(),
            trigger_tokens: default_tool_result_clearing_trigger_tokens(),
            keep_tool_uses: default_tool_result_clearing_keep_tool_uses(),
            clear_at_least_tokens: default_tool_result_clearing_clear_at_least_tokens(),
            clear_tool_inputs: false,
        }
    }
}

impl ToolResultClearingConfig {
    pub fn validate(&self) -> Result<(), String> {
        if self.trigger_tokens == 0 {
            return Err("tool_result_clearing.trigger_tokens must be greater than 0".to_string());
        }
        if self.keep_tool_uses == 0 {
            return Err("tool_result_clearing.keep_tool_uses must be greater than 0".to_string());
        }
        if self.clear_at_least_tokens == 0 {
            return Err(
                "tool_result_clearing.clear_at_least_tokens must be greater than 0".to_string(),
            );
        }
        Ok(())
    }
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentCodexAppServerConfig {
    /// Executable used to launch the official Codex app-server sidecar.
    #[serde(default = "default_codex_app_server_command")]
    pub command: String,
    /// Arguments passed before VT Code appends `--listen stdio://`.
    #[serde(default = "default_codex_app_server_args")]
    pub args: Vec<String>,
    /// Maximum startup handshake time when launching the sidecar.
    #[serde(default = "default_codex_app_server_startup_timeout_secs")]
    pub startup_timeout_secs: u64,
    /// Enable experimental Codex app-server features such as collaboration modes
    /// and native review routing.
    #[serde(default = "default_codex_app_server_experimental_features")]
    pub experimental_features: bool,
}

impl Default for AgentCodexAppServerConfig {
    fn default() -> Self {
        Self {
            command: default_codex_app_server_command(),
            args: default_codex_app_server_args(),
            startup_timeout_secs: default_codex_app_server_startup_timeout_secs(),
            experimental_features: default_codex_app_server_experimental_features(),
        }
    }
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct CircuitBreakerConfig {
    /// Enable circuit breaker functionality
    #[serde(default = "default_circuit_breaker_enabled")]
    pub enabled: bool,

    /// Number of consecutive failures before opening circuit
    #[serde(default = "default_failure_threshold")]
    pub failure_threshold: u32,

    /// Pause and ask user when circuit opens (vs auto-backoff)
    #[serde(default = "default_pause_on_open")]
    pub pause_on_open: bool,

    /// Number of open circuits before triggering pause
    #[serde(default = "default_max_open_circuits")]
    pub max_open_circuits: usize,

    /// Cooldown period between recovery prompts (seconds)
    #[serde(default = "default_recovery_cooldown")]
    pub recovery_cooldown: u64,
}

impl Default for CircuitBreakerConfig {
    fn default() -> Self {
        Self {
            enabled: default_circuit_breaker_enabled(),
            failure_threshold: default_failure_threshold(),
            pause_on_open: default_pause_on_open(),
            max_open_circuits: default_max_open_circuits(),
            recovery_cooldown: default_recovery_cooldown(),
        }
    }
}

/// Open Responses specification compliance configuration
///
/// Enables vendor-neutral LLM API format per the Open Responses specification
/// (<https://www.openresponses.org/>). When enabled, VT Code emits semantic
/// streaming events and uses standardized response/item structures.
#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct OpenResponsesConfig {
    /// Enable Open Responses specification compliance layer
    /// When true, VT Code emits semantic streaming events alongside internal events
    /// Default: false (opt-in feature)
    #[serde(default)]
    pub enabled: bool,

    /// Emit Open Responses events to the event sink
    /// When true, streaming events follow Open Responses format
    /// (response.created, response.output_item.added, response.output_text.delta, etc.)
    #[serde(default = "default_open_responses_emit_events")]
    pub emit_events: bool,

    /// Include VT Code extension items (vtcode:file_change, vtcode:web_search, etc.)
    /// When false, extension items are omitted from the Open Responses output
    #[serde(default = "default_open_responses_include_extensions")]
    pub include_extensions: bool,

    /// Map internal tool calls to Open Responses function_call items
    /// When true, command executions and MCP tool calls are represented as function_call items
    #[serde(default = "default_open_responses_map_tool_calls")]
    pub map_tool_calls: bool,

    /// Include reasoning items in Open Responses output
    /// When true, model reasoning/thinking is exposed as reasoning items
    #[serde(default = "default_open_responses_include_reasoning")]
    pub include_reasoning: bool,
}

impl Default for OpenResponsesConfig {
    fn default() -> Self {
        Self {
            enabled: false, // Opt-in by default
            emit_events: default_open_responses_emit_events(),
            include_extensions: default_open_responses_include_extensions(),
            map_tool_calls: default_open_responses_map_tool_calls(),
            include_reasoning: default_open_responses_include_reasoning(),
        }
    }
}

#[inline]
const fn default_open_responses_emit_events() -> bool {
    true // When enabled, emit events by default
}

#[inline]
const fn default_open_responses_include_extensions() -> bool {
    true // Include VT Code-specific extensions by default
}

#[inline]
const fn default_open_responses_map_tool_calls() -> bool {
    true // Map tool calls to function_call items by default
}

#[inline]
const fn default_open_responses_include_reasoning() -> bool {
    true // Include reasoning items by default
}

#[inline]
fn default_codex_app_server_command() -> String {
    "codex".to_string()
}

#[inline]
fn default_codex_app_server_args() -> Vec<String> {
    vec!["app-server".to_string()]
}

#[inline]
const fn default_codex_app_server_startup_timeout_secs() -> u64 {
    10
}

#[inline]
const fn default_codex_app_server_experimental_features() -> bool {
    false
}

impl Default for AgentConfig {
    fn default() -> Self {
        Self {
            provider: default_provider(),
            api_key_env: default_api_key_env(),
            default_model: default_model(),
            theme: default_theme(),
            system_prompt_mode: SystemPromptMode::default(),
            tool_documentation_mode: ToolDocumentationMode::default(),
            enable_split_tool_results: default_enable_split_tool_results(),
            todo_planning_mode: default_todo_planning_mode(),
            ui_surface: UiSurfacePreference::default(),
            max_conversation_turns: default_max_conversation_turns(),
            reasoning_effort: default_reasoning_effort(),
            verbosity: default_verbosity(),
            temperature: default_temperature(),
            refine_temperature: default_refine_temperature(),
            enable_self_review: default_enable_self_review(),
            max_review_passes: default_max_review_passes(),
            refine_prompts_enabled: default_refine_prompts_enabled(),
            refine_prompts_max_passes: default_refine_max_passes(),
            refine_prompts_model: String::new(),
            small_model: AgentSmallModelConfig::default(),
            prompt_suggestions: AgentPromptSuggestionsConfig::default(),
            onboarding: AgentOnboardingConfig::default(),
            project_doc_max_bytes: default_project_doc_max_bytes(),
            project_doc_fallback_filenames: Vec::new(),
            instruction_max_bytes: default_instruction_max_bytes(),
            instruction_files: Vec::new(),
            instruction_excludes: Vec::new(),
            instruction_import_max_depth: default_instruction_import_max_depth(),
            persistent_memory: PersistentMemoryConfig::default(),
            custom_api_keys: BTreeMap::new(),
            credential_storage_mode: crate::auth::AuthCredentialsStoreMode::default(),
            checkpointing: AgentCheckpointingConfig::default(),
            vibe_coding: AgentVibeCodingConfig::default(),
            max_task_retries: default_max_task_retries(),
            harness: AgentHarnessConfig::default(),
            codex_app_server: AgentCodexAppServerConfig::default(),
            include_temporal_context: default_include_temporal_context(),
            temporal_context_use_utc: false, // Default to local time
            include_working_directory: default_include_working_directory(),
            include_structured_reasoning_tags: None,
            user_instructions: None,
            require_plan_confirmation: default_require_plan_confirmation(),
            circuit_breaker: CircuitBreakerConfig::default(),
            open_responses: OpenResponsesConfig::default(),
        }
    }
}

impl AgentConfig {
    /// Determine whether structured reasoning tag instructions should be included.
    pub fn should_include_structured_reasoning_tags(&self) -> bool {
        self.include_structured_reasoning_tags.unwrap_or(matches!(
            self.system_prompt_mode,
            SystemPromptMode::Specialized
        ))
    }

    /// Validate LLM generation parameters
    pub fn validate_llm_params(&self) -> Result<(), String> {
        // Validate temperature range
        if !(0.0..=1.0).contains(&self.temperature) {
            return Err(format!(
                "temperature must be between 0.0 and 1.0, got {}",
                self.temperature
            ));
        }

        if !(0.0..=1.0).contains(&self.refine_temperature) {
            return Err(format!(
                "refine_temperature must be between 0.0 and 1.0, got {}",
                self.refine_temperature
            ));
        }

        if self.instruction_import_max_depth == 0 {
            return Err("instruction_import_max_depth must be greater than 0".to_string());
        }

        self.persistent_memory.validate()?;
        self.harness.tool_result_clearing.validate()?;

        Ok(())
    }
}

// Optimized: Use inline defaults with constants to reduce function call overhead
#[inline]
fn default_provider() -> String {
    defaults::DEFAULT_PROVIDER.into()
}

#[inline]
fn default_api_key_env() -> String {
    defaults::DEFAULT_API_KEY_ENV.into()
}

#[inline]
fn default_model() -> String {
    defaults::DEFAULT_MODEL.into()
}

#[inline]
fn default_theme() -> String {
    defaults::DEFAULT_THEME.into()
}

#[inline]
const fn default_todo_planning_mode() -> bool {
    true
}

#[inline]
const fn default_enable_split_tool_results() -> bool {
    true // Default: enabled for production use (84% token savings)
}

#[inline]
const fn default_max_conversation_turns() -> usize {
    defaults::DEFAULT_MAX_CONVERSATION_TURNS
}

#[inline]
fn default_reasoning_effort() -> ReasoningEffortLevel {
    ReasoningEffortLevel::None
}

#[inline]
fn default_verbosity() -> VerbosityLevel {
    VerbosityLevel::default()
}

#[inline]
const fn default_temperature() -> f32 {
    llm_generation::DEFAULT_TEMPERATURE
}

#[inline]
const fn default_refine_temperature() -> f32 {
    llm_generation::DEFAULT_REFINE_TEMPERATURE
}

#[inline]
const fn default_enable_self_review() -> bool {
    false
}

#[inline]
const fn default_max_review_passes() -> usize {
    1
}

#[inline]
const fn default_refine_prompts_enabled() -> bool {
    false
}

#[inline]
const fn default_refine_max_passes() -> usize {
    1
}

#[inline]
const fn default_project_doc_max_bytes() -> usize {
    prompt_budget::DEFAULT_MAX_BYTES
}

#[inline]
const fn default_instruction_max_bytes() -> usize {
    prompt_budget::DEFAULT_MAX_BYTES
}

#[inline]
const fn default_instruction_import_max_depth() -> usize {
    5
}

#[inline]
const fn default_max_task_retries() -> u32 {
    2 // Retry twice on transient failures
}

#[inline]
const fn default_harness_max_tool_calls_per_turn() -> usize {
    defaults::DEFAULT_MAX_TOOL_CALLS_PER_TURN
}

#[inline]
const fn default_harness_max_tool_wall_clock_secs() -> u64 {
    defaults::DEFAULT_MAX_TOOL_WALL_CLOCK_SECS
}

#[inline]
const fn default_harness_max_tool_retries() -> u32 {
    defaults::DEFAULT_MAX_TOOL_RETRIES
}

#[inline]
const fn default_harness_auto_compaction_enabled() -> bool {
    false
}

#[inline]
const fn default_tool_result_clearing_enabled() -> bool {
    false
}

#[inline]
const fn default_tool_result_clearing_trigger_tokens() -> u64 {
    100_000
}

#[inline]
const fn default_tool_result_clearing_keep_tool_uses() -> u32 {
    3
}

#[inline]
const fn default_tool_result_clearing_clear_at_least_tokens() -> u64 {
    30_000
}

#[inline]
const fn default_harness_max_revision_rounds() -> usize {
    2
}

#[inline]
const fn default_include_temporal_context() -> bool {
    true // Enable by default - minimal overhead (~20 tokens)
}

#[inline]
const fn default_include_working_directory() -> bool {
    true // Enable by default - minimal overhead (~10 tokens)
}

#[inline]
const fn default_require_plan_confirmation() -> bool {
    true // Default: require confirmation (HITL pattern)
}

#[inline]
const fn default_circuit_breaker_enabled() -> bool {
    true // Default: enabled for resilient execution
}

#[inline]
const fn default_failure_threshold() -> u32 {
    7 // Open circuit after 7 consecutive failures
}

#[inline]
const fn default_pause_on_open() -> bool {
    true // Default: ask user for guidance on circuit breaker
}

#[inline]
const fn default_max_open_circuits() -> usize {
    3 // Pause when 3+ tools have open circuits
}

#[inline]
const fn default_recovery_cooldown() -> u64 {
    60 // Cooldown between recovery prompts (seconds)
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentCheckpointingConfig {
    /// Enable automatic checkpoints after each successful turn
    #[serde(default = "default_checkpointing_enabled")]
    pub enabled: bool,

    /// Optional custom directory for storing checkpoints (relative to workspace or absolute)
    #[serde(default)]
    pub storage_dir: Option<String>,

    /// Maximum number of checkpoints to retain on disk
    #[serde(default = "default_checkpointing_max_snapshots")]
    pub max_snapshots: usize,

    /// Maximum age in days before checkpoints are removed automatically (None disables)
    #[serde(default = "default_checkpointing_max_age_days")]
    pub max_age_days: Option<u64>,
}

impl Default for AgentCheckpointingConfig {
    fn default() -> Self {
        Self {
            enabled: default_checkpointing_enabled(),
            storage_dir: None,
            max_snapshots: default_checkpointing_max_snapshots(),
            max_age_days: default_checkpointing_max_age_days(),
        }
    }
}

#[inline]
const fn default_checkpointing_enabled() -> bool {
    DEFAULT_CHECKPOINTS_ENABLED
}

#[inline]
const fn default_checkpointing_max_snapshots() -> usize {
    DEFAULT_MAX_SNAPSHOTS
}

#[inline]
const fn default_checkpointing_max_age_days() -> Option<u64> {
    Some(DEFAULT_MAX_AGE_DAYS)
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct PersistentMemoryConfig {
    /// Toggle main-session persistent memory for this repository
    #[serde(default = "default_persistent_memory_enabled")]
    pub enabled: bool,

    /// Write durable memory after completed turns and session finalization
    #[serde(default = "default_persistent_memory_auto_write")]
    pub auto_write: bool,

    /// Optional user-local directory override for persistent memory storage
    #[serde(default)]
    pub directory_override: Option<String>,

    /// Startup line budget scanned from memory_summary.md before VT Code renders a compact startup summary
    #[serde(default = "default_persistent_memory_startup_line_limit")]
    pub startup_line_limit: usize,

    /// Startup byte budget scanned from memory_summary.md before VT Code renders a compact startup summary
    #[serde(default = "default_persistent_memory_startup_byte_limit")]
    pub startup_byte_limit: usize,
}

impl Default for PersistentMemoryConfig {
    fn default() -> Self {
        Self {
            enabled: default_persistent_memory_enabled(),
            auto_write: default_persistent_memory_auto_write(),
            directory_override: None,
            startup_line_limit: default_persistent_memory_startup_line_limit(),
            startup_byte_limit: default_persistent_memory_startup_byte_limit(),
        }
    }
}

impl PersistentMemoryConfig {
    pub fn validate(&self) -> Result<(), String> {
        if self.startup_line_limit == 0 {
            return Err("persistent_memory.startup_line_limit must be greater than 0".to_string());
        }

        if self.startup_byte_limit == 0 {
            return Err("persistent_memory.startup_byte_limit must be greater than 0".to_string());
        }

        Ok(())
    }
}

#[inline]
const fn default_persistent_memory_enabled() -> bool {
    false
}

#[inline]
const fn default_persistent_memory_auto_write() -> bool {
    true
}

#[inline]
const fn default_persistent_memory_startup_line_limit() -> usize {
    200
}

#[inline]
const fn default_persistent_memory_startup_byte_limit() -> usize {
    25 * 1024
}

#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentOnboardingConfig {
    /// Toggle onboarding message rendering
    #[serde(default = "default_onboarding_enabled")]
    pub enabled: bool,

    /// Introductory text shown at session start
    #[serde(default = "default_intro_text")]
    pub intro_text: String,

    /// Whether to include project overview in onboarding message
    #[serde(default = "default_show_project_overview")]
    pub include_project_overview: bool,

    /// Whether to include language summary in onboarding message
    #[serde(default = "default_show_language_summary")]
    pub include_language_summary: bool,

    /// Whether to include AGENTS.md highlights in onboarding message
    #[serde(default = "default_show_guideline_highlights")]
    pub include_guideline_highlights: bool,

    /// Whether to surface usage tips inside the welcome text banner
    #[serde(default = "default_show_usage_tips_in_welcome")]
    pub include_usage_tips_in_welcome: bool,

    /// Whether to surface suggested actions inside the welcome text banner
    #[serde(default = "default_show_recommended_actions_in_welcome")]
    pub include_recommended_actions_in_welcome: bool,

    /// Maximum number of guideline bullets to surface
    #[serde(default = "default_guideline_highlight_limit")]
    pub guideline_highlight_limit: usize,

    /// Tips for collaborating with the agent effectively
    #[serde(default = "default_usage_tips")]
    pub usage_tips: Vec<String>,

    /// Recommended follow-up actions to display
    #[serde(default = "default_recommended_actions")]
    pub recommended_actions: Vec<String>,

    /// Placeholder suggestion for the chat input bar
    #[serde(default)]
    pub chat_placeholder: Option<String>,
}

impl Default for AgentOnboardingConfig {
    fn default() -> Self {
        Self {
            enabled: default_onboarding_enabled(),
            intro_text: default_intro_text(),
            include_project_overview: default_show_project_overview(),
            include_language_summary: default_show_language_summary(),
            include_guideline_highlights: default_show_guideline_highlights(),
            include_usage_tips_in_welcome: default_show_usage_tips_in_welcome(),
            include_recommended_actions_in_welcome: default_show_recommended_actions_in_welcome(),
            guideline_highlight_limit: default_guideline_highlight_limit(),
            usage_tips: default_usage_tips(),
            recommended_actions: default_recommended_actions(),
            chat_placeholder: None,
        }
    }
}

#[inline]
const fn default_onboarding_enabled() -> bool {
    true
}

const DEFAULT_INTRO_TEXT: &str =
    "Let's get oriented. I preloaded workspace context so we can move fast.";

#[inline]
fn default_intro_text() -> String {
    DEFAULT_INTRO_TEXT.into()
}

#[inline]
const fn default_show_project_overview() -> bool {
    true
}

#[inline]
const fn default_show_language_summary() -> bool {
    false
}

#[inline]
const fn default_show_guideline_highlights() -> bool {
    true
}

#[inline]
const fn default_show_usage_tips_in_welcome() -> bool {
    false
}

#[inline]
const fn default_show_recommended_actions_in_welcome() -> bool {
    false
}

#[inline]
const fn default_guideline_highlight_limit() -> usize {
    3
}

const DEFAULT_USAGE_TIPS: &[&str] = &[
    "Describe your current coding goal or ask for a quick status overview.",
    "Reference AGENTS.md guidelines when proposing changes.",
    "Prefer asking for targeted file reads or diffs before editing.",
];

const DEFAULT_RECOMMENDED_ACTIONS: &[&str] = &[
    "Review the highlighted guidelines and share the task you want to tackle.",
    "Ask for a workspace tour if you need more context.",
];

fn default_usage_tips() -> Vec<String> {
    DEFAULT_USAGE_TIPS.iter().map(|s| (*s).into()).collect()
}

fn default_recommended_actions() -> Vec<String> {
    DEFAULT_RECOMMENDED_ACTIONS
        .iter()
        .map(|s| (*s).into())
        .collect()
}

/// Small/lightweight model configuration for efficient operations
///
/// Following VT Code's pattern, use a smaller model (e.g., Haiku, GPT-4 Mini) for 50%+ of calls:
/// - Large file reads and parsing (>50KB)
/// - Web page summarization and analysis
/// - Git history and commit message processing
/// - One-word processing labels and simple classifications
///
/// Typically 70-80% cheaper than the main model while maintaining quality for these tasks.
#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentSmallModelConfig {
    /// Enable small model tier for efficient operations
    #[serde(default = "default_small_model_enabled")]
    pub enabled: bool,

    /// Small model to use (e.g., claude-4-5-haiku, "gpt-4-mini", "gemini-2.0-flash")
    /// Leave empty to auto-select a lightweight sibling of the main model
    #[serde(default)]
    pub model: String,

    /// Temperature for small model responses
    #[serde(default = "default_small_model_temperature")]
    pub temperature: f32,

    /// Enable small model for large file reads (>50KB)
    #[serde(default = "default_small_model_for_large_reads")]
    pub use_for_large_reads: bool,

    /// Enable small model for web content summarization
    #[serde(default = "default_small_model_for_web_summary")]
    pub use_for_web_summary: bool,

    /// Enable small model for git history processing
    #[serde(default = "default_small_model_for_git_history")]
    pub use_for_git_history: bool,

    /// Enable small model for persistent memory classification and summary refresh
    #[serde(default = "default_small_model_for_memory")]
    pub use_for_memory: bool,
}

impl Default for AgentSmallModelConfig {
    fn default() -> Self {
        Self {
            enabled: default_small_model_enabled(),
            model: String::new(),
            temperature: default_small_model_temperature(),
            use_for_large_reads: default_small_model_for_large_reads(),
            use_for_web_summary: default_small_model_for_web_summary(),
            use_for_git_history: default_small_model_for_git_history(),
            use_for_memory: default_small_model_for_memory(),
        }
    }
}

#[inline]
const fn default_small_model_enabled() -> bool {
    true // Enable by default following VT Code pattern
}

#[inline]
const fn default_small_model_temperature() -> f32 {
    0.3 // More deterministic for parsing/summarization
}

#[inline]
const fn default_small_model_for_large_reads() -> bool {
    true
}

#[inline]
const fn default_small_model_for_web_summary() -> bool {
    true
}

#[inline]
const fn default_small_model_for_git_history() -> bool {
    true
}

#[inline]
const fn default_small_model_for_memory() -> bool {
    true
}

/// Inline prompt suggestion configuration for the chat composer.
#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentPromptSuggestionsConfig {
    /// Enable inline prompt suggestions in the chat composer.
    #[serde(default = "default_prompt_suggestions_enabled")]
    pub enabled: bool,

    /// Lightweight model to use for suggestions.
    /// Leave empty to auto-select an efficient sibling of the main model.
    #[serde(default)]
    pub model: String,

    /// Temperature for inline prompt suggestion generation.
    #[serde(default = "default_prompt_suggestions_temperature")]
    pub temperature: f32,

    /// Whether VT Code should remind users that LLM-backed suggestions consume tokens.
    #[serde(default = "default_prompt_suggestions_show_cost_notice")]
    pub show_cost_notice: bool,
}

impl Default for AgentPromptSuggestionsConfig {
    fn default() -> Self {
        Self {
            enabled: default_prompt_suggestions_enabled(),
            model: String::new(),
            temperature: default_prompt_suggestions_temperature(),
            show_cost_notice: default_prompt_suggestions_show_cost_notice(),
        }
    }
}

#[inline]
const fn default_prompt_suggestions_enabled() -> bool {
    true
}

#[inline]
const fn default_prompt_suggestions_temperature() -> f32 {
    0.3
}

#[inline]
const fn default_prompt_suggestions_show_cost_notice() -> bool {
    true
}

/// Vibe coding configuration for lazy/vague request support
///
/// Enables intelligent context gathering and entity resolution to support
/// casual, imprecise requests like "make it blue" or "decrease by half".
#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct AgentVibeCodingConfig {
    /// Enable vibe coding support
    #[serde(default = "default_vibe_coding_enabled")]
    pub enabled: bool,

    /// Minimum prompt length for refinement (default: 5 chars)
    #[serde(default = "default_vibe_min_prompt_length")]
    pub min_prompt_length: usize,

    /// Minimum prompt words for refinement (default: 2 words)
    #[serde(default = "default_vibe_min_prompt_words")]
    pub min_prompt_words: usize,

    /// Enable fuzzy entity resolution
    #[serde(default = "default_vibe_entity_resolution")]
    pub enable_entity_resolution: bool,

    /// Entity index cache file path (relative to workspace)
    #[serde(default = "default_vibe_entity_cache")]
    pub entity_index_cache: String,

    /// Maximum entity matches to return (default: 5)
    #[serde(default = "default_vibe_max_entity_matches")]
    pub max_entity_matches: usize,

    /// Track workspace state (file activity, value changes)
    #[serde(default = "default_vibe_track_workspace")]
    pub track_workspace_state: bool,

    /// Maximum recent files to track (default: 20)
    #[serde(default = "default_vibe_max_recent_files")]
    pub max_recent_files: usize,

    /// Track value history for inference
    #[serde(default = "default_vibe_track_values")]
    pub track_value_history: bool,

    /// Enable conversation memory for pronoun resolution
    #[serde(default = "default_vibe_conversation_memory")]
    pub enable_conversation_memory: bool,

    /// Maximum conversation turns to remember (default: 50)
    #[serde(default = "default_vibe_max_memory_turns")]
    pub max_memory_turns: usize,

    /// Enable pronoun resolution (it, that, this)
    #[serde(default = "default_vibe_pronoun_resolution")]
    pub enable_pronoun_resolution: bool,

    /// Enable proactive context gathering
    #[serde(default = "default_vibe_proactive_context")]
    pub enable_proactive_context: bool,

    /// Maximum files to gather for context (default: 3)
    #[serde(default = "default_vibe_max_context_files")]
    pub max_context_files: usize,

    /// Maximum code snippets per file (default: 20 lines)
    #[serde(default = "default_vibe_max_snippets_per_file")]
    pub max_context_snippets_per_file: usize,

    /// Maximum search results to include (default: 5)
    #[serde(default = "default_vibe_max_search_results")]
    pub max_search_results: usize,

    /// Enable relative value inference (by half, double, etc.)
    #[serde(default = "default_vibe_value_inference")]
    pub enable_relative_value_inference: bool,
}

impl Default for AgentVibeCodingConfig {
    fn default() -> Self {
        Self {
            enabled: default_vibe_coding_enabled(),
            min_prompt_length: default_vibe_min_prompt_length(),
            min_prompt_words: default_vibe_min_prompt_words(),
            enable_entity_resolution: default_vibe_entity_resolution(),
            entity_index_cache: default_vibe_entity_cache(),
            max_entity_matches: default_vibe_max_entity_matches(),
            track_workspace_state: default_vibe_track_workspace(),
            max_recent_files: default_vibe_max_recent_files(),
            track_value_history: default_vibe_track_values(),
            enable_conversation_memory: default_vibe_conversation_memory(),
            max_memory_turns: default_vibe_max_memory_turns(),
            enable_pronoun_resolution: default_vibe_pronoun_resolution(),
            enable_proactive_context: default_vibe_proactive_context(),
            max_context_files: default_vibe_max_context_files(),
            max_context_snippets_per_file: default_vibe_max_snippets_per_file(),
            max_search_results: default_vibe_max_search_results(),
            enable_relative_value_inference: default_vibe_value_inference(),
        }
    }
}

// Vibe coding default functions
#[inline]
const fn default_vibe_coding_enabled() -> bool {
    false // Conservative default, opt-in
}

#[inline]
const fn default_vibe_min_prompt_length() -> usize {
    5
}

#[inline]
const fn default_vibe_min_prompt_words() -> usize {
    2
}

#[inline]
const fn default_vibe_entity_resolution() -> bool {
    true
}

#[inline]
fn default_vibe_entity_cache() -> String {
    ".vtcode/entity_index.json".into()
}

#[inline]
const fn default_vibe_max_entity_matches() -> usize {
    5
}

#[inline]
const fn default_vibe_track_workspace() -> bool {
    true
}

#[inline]
const fn default_vibe_max_recent_files() -> usize {
    20
}

#[inline]
const fn default_vibe_track_values() -> bool {
    true
}

#[inline]
const fn default_vibe_conversation_memory() -> bool {
    true
}

#[inline]
const fn default_vibe_max_memory_turns() -> usize {
    50
}

#[inline]
const fn default_vibe_pronoun_resolution() -> bool {
    true
}

#[inline]
const fn default_vibe_proactive_context() -> bool {
    true
}

#[inline]
const fn default_vibe_max_context_files() -> usize {
    3
}

#[inline]
const fn default_vibe_max_snippets_per_file() -> usize {
    20
}

#[inline]
const fn default_vibe_max_search_results() -> usize {
    5
}

#[inline]
const fn default_vibe_value_inference() -> bool {
    true
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_continuation_policy_defaults_and_parses() {
        assert_eq!(ContinuationPolicy::default(), ContinuationPolicy::All);
        assert_eq!(
            ContinuationPolicy::parse("off"),
            Some(ContinuationPolicy::Off)
        );
        assert_eq!(
            ContinuationPolicy::parse("exec-only"),
            Some(ContinuationPolicy::ExecOnly)
        );
        assert_eq!(
            ContinuationPolicy::parse("all"),
            Some(ContinuationPolicy::All)
        );
        assert_eq!(ContinuationPolicy::parse("invalid"), None);
    }

    #[test]
    fn test_harness_config_continuation_policy_deserializes_with_fallback() {
        let parsed: AgentHarnessConfig =
            toml::from_str("continuation_policy = \"all\"").expect("valid harness config");
        assert_eq!(parsed.continuation_policy, ContinuationPolicy::All);

        let fallback: AgentHarnessConfig =
            toml::from_str("continuation_policy = \"unexpected\"").expect("fallback config");
        assert_eq!(fallback.continuation_policy, ContinuationPolicy::All);
    }

    #[test]
    fn test_harness_orchestration_mode_defaults_and_parses() {
        assert_eq!(
            HarnessOrchestrationMode::default(),
            HarnessOrchestrationMode::Single
        );
        assert_eq!(
            HarnessOrchestrationMode::parse("single"),
            Some(HarnessOrchestrationMode::Single)
        );
        assert_eq!(
            HarnessOrchestrationMode::parse("plan_build_evaluate"),
            Some(HarnessOrchestrationMode::PlanBuildEvaluate)
        );
        assert_eq!(
            HarnessOrchestrationMode::parse("planner-generator-evaluator"),
            Some(HarnessOrchestrationMode::PlanBuildEvaluate)
        );
        assert_eq!(HarnessOrchestrationMode::parse("unexpected"), None);
    }

    #[test]
    fn test_harness_config_orchestration_deserializes_with_fallback() {
        let parsed: AgentHarnessConfig =
            toml::from_str("orchestration_mode = \"plan_build_evaluate\"")
                .expect("valid harness config");
        assert_eq!(
            parsed.orchestration_mode,
            HarnessOrchestrationMode::PlanBuildEvaluate
        );
        assert_eq!(parsed.max_revision_rounds, 2);

        let fallback: AgentHarnessConfig =
            toml::from_str("orchestration_mode = \"unexpected\"").expect("fallback config");
        assert_eq!(
            fallback.orchestration_mode,
            HarnessOrchestrationMode::Single
        );
    }

    #[test]
    fn test_plan_confirmation_config_default() {
        let config = AgentConfig::default();
        assert!(config.require_plan_confirmation);
    }

    #[test]
    fn test_persistent_memory_is_disabled_by_default() {
        let config = AgentConfig::default();
        assert!(!config.persistent_memory.enabled);
        assert!(config.persistent_memory.auto_write);
    }

    #[test]
    fn test_tool_result_clearing_defaults() {
        let config = AgentConfig::default();
        let clearing = config.harness.tool_result_clearing;

        assert!(!clearing.enabled);
        assert_eq!(clearing.trigger_tokens, 100_000);
        assert_eq!(clearing.keep_tool_uses, 3);
        assert_eq!(clearing.clear_at_least_tokens, 30_000);
        assert!(!clearing.clear_tool_inputs);
    }

    #[test]
    fn test_codex_app_server_experimental_features_default_to_disabled() {
        let config = AgentConfig::default();

        assert!(!config.codex_app_server.experimental_features);
    }

    #[test]
    fn test_codex_app_server_experimental_features_parse_from_toml() {
        let parsed: AgentCodexAppServerConfig = toml::from_str(
            r#"
                command = "codex"
                args = ["app-server"]
                startup_timeout_secs = 15
                experimental_features = true
            "#,
        )
        .expect("valid codex app-server config");

        assert!(parsed.experimental_features);
        assert_eq!(parsed.startup_timeout_secs, 15);
    }

    #[test]
    fn test_tool_result_clearing_parses_and_validates() {
        let parsed: AgentHarnessConfig = toml::from_str(
            r#"
                [tool_result_clearing]
                enabled = true
                trigger_tokens = 123456
                keep_tool_uses = 6
                clear_at_least_tokens = 4096
                clear_tool_inputs = true
            "#,
        )
        .expect("valid harness config");

        assert!(parsed.tool_result_clearing.enabled);
        assert_eq!(parsed.tool_result_clearing.trigger_tokens, 123_456);
        assert_eq!(parsed.tool_result_clearing.keep_tool_uses, 6);
        assert_eq!(parsed.tool_result_clearing.clear_at_least_tokens, 4_096);
        assert!(parsed.tool_result_clearing.clear_tool_inputs);
        assert!(parsed.tool_result_clearing.validate().is_ok());
    }

    #[test]
    fn test_tool_result_clearing_rejects_zero_values() {
        let clearing = ToolResultClearingConfig {
            trigger_tokens: 0,
            ..ToolResultClearingConfig::default()
        };
        assert!(clearing.validate().is_err());

        let clearing = ToolResultClearingConfig {
            keep_tool_uses: 0,
            ..ToolResultClearingConfig::default()
        };
        assert!(clearing.validate().is_err());

        let clearing = ToolResultClearingConfig {
            clear_at_least_tokens: 0,
            ..ToolResultClearingConfig::default()
        };
        assert!(clearing.validate().is_err());
    }

    #[test]
    fn test_structured_reasoning_defaults_follow_prompt_mode() {
        let default_mode = AgentConfig {
            system_prompt_mode: SystemPromptMode::Default,
            ..Default::default()
        };
        assert!(!default_mode.should_include_structured_reasoning_tags());

        let specialized_mode = AgentConfig {
            system_prompt_mode: SystemPromptMode::Specialized,
            ..Default::default()
        };
        assert!(specialized_mode.should_include_structured_reasoning_tags());

        let minimal_mode = AgentConfig {
            system_prompt_mode: SystemPromptMode::Minimal,
            ..Default::default()
        };
        assert!(!minimal_mode.should_include_structured_reasoning_tags());

        let lightweight_mode = AgentConfig {
            system_prompt_mode: SystemPromptMode::Lightweight,
            ..Default::default()
        };
        assert!(!lightweight_mode.should_include_structured_reasoning_tags());
    }

    #[test]
    fn test_structured_reasoning_explicit_override() {
        let mut config = AgentConfig {
            system_prompt_mode: SystemPromptMode::Minimal,
            include_structured_reasoning_tags: Some(true),
            ..AgentConfig::default()
        };
        assert!(config.should_include_structured_reasoning_tags());

        config.include_structured_reasoning_tags = Some(false);
        assert!(!config.should_include_structured_reasoning_tags());
    }
}