adk_model/gemini/

client.rs

use crate::attachment;
use crate::retry::{RetryConfig, execute_with_retry, is_retryable_model_error};
use adk_core::{
    CacheCapable, CitationMetadata, CitationSource, Content, ErrorCategory, ErrorComponent,
    FinishReason, Llm, LlmRequest, LlmResponse, LlmResponseStream, Part, Result, SchemaAdapter,
    SchemaCache, UsageMetadata,
};
use adk_gemini::Gemini;
use adk_gemini::schema_adapter::GeminiSchemaAdapter;
use async_trait::async_trait;
use base64::{Engine as _, engine::general_purpose::STANDARD as BASE64_STANDARD};
use futures::TryStreamExt;

#[cfg(feature = "gemini-interactions")]
use super::interactions_target::InteractionTarget;

/// Which Gemini wire API a [`GeminiModel`] uses.
///
/// Defaults to [`GeminiTransport::GenerateContent`], the classic
/// `models/{model}:generateContent` endpoint. Selecting
/// [`GeminiTransport::Interactions`] (via [`GeminiModel::use_interactions_api`])
/// routes requests through the Interactions API (Beta): a stateful, step-based
/// transport that drives the same [`adk_core::Llm`] contract.
///
/// Only compiled when the `gemini-interactions` feature is enabled.
#[cfg(feature = "gemini-interactions")]
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum GeminiTransport {
    /// The classic `models/{model}:generateContent` API (default).
    #[default]
    GenerateContent,
    /// The Interactions API (Beta): stateful, step-based.
    Interactions,
}

/// Background-execution policy for the Interactions transport.
///
/// Controls whether interactions run with `background=true`. The default,
/// [`BackgroundMode::AgentTargetsOnly`], keeps low-latency chat turns
/// foreground while letting long-running agent targets (e.g. Deep Research)
/// run in the background.
///
/// Only compiled when the `gemini-interactions` feature is enabled.
#[cfg(feature = "gemini-interactions")]
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum BackgroundMode {
    /// `background=true` for agent targets, `false` for model targets
    /// (default). Keeps chat turns low-latency while letting Deep Research and
    /// other long-running agents run in the background.
    #[default]
    AgentTargetsOnly,
    /// Always run interactions with `background=true`.
    Always,
    /// Never run interactions in the background.
    Never,
}

/// Faithful-to-API options for the Interactions transport.
///
/// The defaults mirror the Interactions API's intended posture: interactions
/// are stored (`store = true`), stateful continuation via
/// `previous_interaction_id` is enabled (`stateful = true`), background
/// execution applies to agent targets only
/// ([`BackgroundMode::AgentTargetsOnly`]), and background interactions are
/// polled once per second.
///
/// Only compiled when the `gemini-interactions` feature is enabled.
#[cfg(feature = "gemini-interactions")]
#[derive(Debug, Clone)]
pub struct InteractionOptions {
    /// Whether interactions are stored server-side. Default: `true`.
    pub store: bool,
    /// Whether to continue conversations statefully via
    /// `previous_interaction_id`. Default: `true`.
    pub stateful: bool,
    /// Background-execution policy. Default: [`BackgroundMode::AgentTargetsOnly`].
    pub background: BackgroundMode,
    /// Poll interval for background interactions. Default: 1 second.
    pub poll_interval: std::time::Duration,
}

#[cfg(feature = "gemini-interactions")]
impl Default for InteractionOptions {
    fn default() -> Self {
        Self {
            store: true,
            stateful: true,
            background: BackgroundMode::AgentTargetsOnly,
            poll_interval: std::time::Duration::from_secs(1),
        }
    }
}

/// Gemini model client wrapping the `adk-gemini` crate for the `Llm` trait.
pub struct GeminiModel {
    client: Gemini,
    model_name: String,
    retry_config: RetryConfig,
    /// Default thinking configuration applied to every request.
    ///
    /// Controls the model's reasoning effort. For Gemini 3 series, use
    /// `ThinkingLevel` (Low/Medium/High). For Gemini 2.5 series, use
    /// `thinking_budget` (token count).
    thinking_config: Option<adk_gemini::ThinkingConfig>,
    /// Selected wire transport. Defaults to
    /// [`GeminiTransport::GenerateContent`]; set to
    /// [`GeminiTransport::Interactions`] via [`GeminiModel::use_interactions_api`].
    #[cfg(feature = "gemini-interactions")]
    transport: GeminiTransport,
    /// The validated Interactions destination, populated when the Interactions
    /// transport is enabled. `None` for the generateContent transport.
    #[cfg(feature = "gemini-interactions")]
    interaction_target: Option<InteractionTarget>,
    /// Faithful-to-API options for the Interactions transport.
    #[cfg(feature = "gemini-interactions")]
    interaction_options: InteractionOptions,
    /// Server-side RAG grounding declaration attached to every request.
    /// Vertex AI backend only; see [`GeminiModel::with_vertex_rag_store`].
    #[cfg(feature = "gemini-vertex")]
    vertex_rag_store: Option<adk_gemini::VertexRagStore>,
    /// Whether the wrapped client targets the Vertex AI backend (as opposed
    /// to AI Studio). Set by the `new_google_cloud*` constructors.
    #[cfg(feature = "gemini-vertex")]
    vertex_backend: bool,
}

/// Convert a Gemini client error to a structured `AdkError` with proper category and retry hints.
pub(crate) fn gemini_error_to_adk(e: &adk_gemini::ClientError) -> adk_core::AdkError {
    fn format_error_chain(e: &dyn std::error::Error) -> String {
        let mut msg = e.to_string();
        let mut source = e.source();
        while let Some(s) = source {
            msg.push_str(": ");
            msg.push_str(&s.to_string());
            source = s.source();
        }
        msg
    }

    let message = format_error_chain(e);
    // Vertex no longer replays a failed request over REST, so a transport failure
    // is left to the configured retry policy instead.
    #[cfg(feature = "gemini-vertex")]
    let transport_failure =
        adk_gemini::backend::vertex::VertexBackend::is_transport_error(&message);
    #[cfg(not(feature = "gemini-vertex"))]
    let transport_failure = false;

    // Extract status code from BadResponse variant via Display output
    // BadResponse format: "bad response from server; code {code}; description: ..."
    let (category, code, status_code) = if message.contains("code 429")
        || message.contains("RESOURCE_EXHAUSTED")
        || message.contains("rate limit")
    {
        (ErrorCategory::RateLimited, "model.gemini.rate_limited", Some(429u16))
    } else if message.contains("code 503") || message.contains("UNAVAILABLE") {
        (ErrorCategory::Unavailable, "model.gemini.unavailable", Some(503))
    } else if message.contains("code 529") || message.contains("OVERLOADED") {
        (ErrorCategory::Unavailable, "model.gemini.overloaded", Some(529))
    } else if message.contains("code 408")
        || message.contains("DEADLINE_EXCEEDED")
        || message.contains("TIMEOUT")
    {
        (ErrorCategory::Timeout, "model.gemini.timeout", Some(408))
    } else if message.contains("code 401") || message.contains("Invalid API key") {
        (ErrorCategory::Unauthorized, "model.gemini.unauthorized", Some(401))
    } else if message.contains("code 400") {
        (ErrorCategory::InvalidInput, "model.gemini.bad_request", Some(400))
    } else if message.contains("code 404") {
        (ErrorCategory::NotFound, "model.gemini.not_found", Some(404))
    } else if message.contains("invalid generation config") {
        (ErrorCategory::InvalidInput, "model.gemini.invalid_config", None)
    } else if transport_failure {
        (ErrorCategory::Unavailable, "model.gemini.unavailable", None)
    } else {
        (ErrorCategory::Internal, "model.gemini.internal", None)
    };

    let mut err = adk_core::AdkError::new(ErrorComponent::Model, category, code, message)
        .with_provider("gemini");
    if let Some(sc) = status_code {
        err = err.with_upstream_status(sc);
    }
    err
}

/// Maps a terminal [`InteractionStatus`](adk_gemini::interactions::InteractionStatus)
/// to a `Result`, surfacing the API's failure states as errors.
///
/// The Interactions API can terminate an interaction in `failed` or
/// `budget_exceeded` (Requirements 7.7 / 9.2). This helper converts those two
/// states into an [`adk_core::AdkError`] (provider `"gemini"`, category
/// [`ErrorCategory::Internal`]) and treats every other status as success — the
/// caller has already decided to read the interaction's content for non-failure
/// states. Factored out as a free function so the terminal-status mapping is
/// unit-testable without a network round-trip.
#[cfg(feature = "gemini-interactions")]
fn interaction_status_to_result(status: adk_gemini::interactions::InteractionStatus) -> Result<()> {
    use adk_gemini::interactions::InteractionStatus;
    match status {
        InteractionStatus::Failed => Err(interaction_terminal_error(
            "model.gemini.interactions.failed",
            "the Gemini interaction terminated with status `failed`",
        )),
        InteractionStatus::BudgetExceeded => Err(interaction_terminal_error(
            "model.gemini.interactions.budget_exceeded",
            "the Gemini interaction terminated with status `budget_exceeded`",
        )),
        InteractionStatus::InProgress
        | InteractionStatus::RequiresAction
        | InteractionStatus::Completed
        | InteractionStatus::Cancelled
        | InteractionStatus::Incomplete => Ok(()),
    }
}

/// Builds the [`adk_core::AdkError`] for a terminal Interactions failure status.
#[cfg(feature = "gemini-interactions")]
fn interaction_terminal_error(code: &'static str, message: &str) -> adk_core::AdkError {
    adk_core::AdkError::new(
        ErrorComponent::Model,
        ErrorCategory::Internal,
        code,
        message.to_string(),
    )
    .with_provider("gemini")
}

/// Returns the value of the environment variable `name` when it is set and
/// non-empty (after trimming whitespace).
fn env_non_empty(name: &str) -> Option<String> {
    std::env::var(name).ok().filter(|v| !v.trim().is_empty())
}

/// Returns `true` when an opt-in flag value selects the Vertex AI backend:
/// `1` or a case-insensitive `true`.
fn env_flag_truthy(value: &str) -> bool {
    let v = value.trim();
    v == "1" || v.eq_ignore_ascii_case("true")
}

/// Reports whether the environment opts in to the Vertex AI backend.
///
/// Consults `GOOGLE_GENAI_USE_ENTERPRISE` first and, when that variable is
/// unset or empty, the deprecated `GOOGLE_GENAI_USE_VERTEXAI`. A flag is
/// truthy when its value is `1` or a case-insensitive `true`.
/// `GOOGLE_GENAI_USE_ENTERPRISE` takes precedence when both are set.
///
/// # Example
///
/// ```rust,no_run
/// if adk_model::gemini::vertex_env_requested() {
///     // route Gemini traffic through Vertex AI
/// }
/// ```
pub fn vertex_env_requested() -> bool {
    // `GOOGLE_GENAI_USE_VERTEXAI` is the deprecated name adk-python still
    // honors. Deliberately no deprecation warning for it —
    // google/adk-python#6168 documents the churn such a warning caused.
    if let Some(v) = env_non_empty("GOOGLE_GENAI_USE_ENTERPRISE") {
        return env_flag_truthy(&v);
    }
    env_non_empty("GOOGLE_GENAI_USE_VERTEXAI").is_some_and(|v| env_flag_truthy(&v))
}

/// Reads `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION`, erroring with a
/// message that names whichever variable is missing.
#[cfg(feature = "gemini-vertex")]
fn vertex_target_from_env() -> Result<(String, String)> {
    let project = env_non_empty("GOOGLE_CLOUD_PROJECT");
    let location = env_non_empty("GOOGLE_CLOUD_LOCATION");
    if let (Some(project), Some(location)) = (&project, &location) {
        return Ok((project.clone(), location.clone()));
    }
    let mut missing = Vec::new();
    if project.is_none() {
        missing.push("GOOGLE_CLOUD_PROJECT");
    }
    if location.is_none() {
        missing.push("GOOGLE_CLOUD_LOCATION");
    }
    Err(adk_core::AdkError::new(
        ErrorComponent::Model,
        ErrorCategory::InvalidInput,
        "model.gemini.vertex_env_incomplete",
        format!(
            "Vertex AI backend selected via GOOGLE_GENAI_USE_ENTERPRISE/GOOGLE_GENAI_USE_VERTEXAI, but {} not set. Set the missing variable(s) to your Google Cloud project ID and region (e.g. us-central1).",
            missing.join(" and ")
        ),
    )
    .with_provider("gemini"))
}

impl GeminiModel {
    fn gemini_part_thought_signature(value: &serde_json::Value) -> Option<String> {
        value.get("thoughtSignature").and_then(serde_json::Value::as_str).map(str::to_string)
    }

    fn validate_request_contract(&self, req: &LlmRequest) -> Result<()> {
        // Server-side RAG grounding is a Vertex AI capability; reject it on
        // the AI Studio backend before any network IO (mirrors the request-
        // time gating the Vertex backend applies to Studio-only fields).
        #[cfg(feature = "gemini-vertex")]
        if self.vertex_rag_store.is_some() && !self.vertex_backend {
            return Err(adk_core::AdkError::new(
                ErrorComponent::Model,
                ErrorCategory::Unsupported,
                "model.gemini.rag_store_requires_vertex",
                "Vertex RAG store grounding requires the Vertex AI backend. Construct the model with new_google_cloud/new_google_cloud_adc (or set GOOGLE_GENAI_USE_VERTEXAI with GOOGLE_CLOUD_PROJECT/GOOGLE_CLOUD_LOCATION) instead of an AI Studio API key.",
            )
            .with_provider("gemini"));
        }

        let model = self.model_name.strip_prefix("models/").unwrap_or(&self.model_name);
        if !matches!(model, "gemini-3.6-flash" | "gemini-3.7-flash") {
            return Ok(());
        }

        if let Some(config) = &req.config
            && (config.temperature.is_some() || config.top_p.is_some() || config.top_k.is_some())
        {
            return Err(adk_core::AdkError::new(
                ErrorComponent::Model,
                ErrorCategory::InvalidInput,
                "model.gemini.sampling_unsupported",
                format!(
                    "{model} does not accept temperature, top_p, or top_k; remove explicit sampling parameters"
                ),
            )
            .with_provider("gemini"));
        }
        if self.thinking_config.as_ref().is_some_and(|config| config.thinking_budget.is_some()) {
            return Err(adk_core::AdkError::new(
                ErrorComponent::Model,
                ErrorCategory::InvalidInput,
                "model.gemini.thinking_budget_unsupported",
                format!(
                    "{model} uses thinking levels instead of token budgets; set thinking_level and clear thinking_budget"
                ),
            )
            .with_provider("gemini"));
        }
        Ok(())
    }

    /// Builds a `GeminiModel` from a constructed client and model name with all
    /// configurable fields defaulted.
    ///
    /// Centralizing struct construction here keeps the cfg-gated Interactions
    /// fields out of every public constructor's `Self { .. }` literal.
    pub(crate) fn from_client(client: Gemini, model_name: String) -> Self {
        Self {
            client,
            model_name,
            retry_config: RetryConfig::default(),
            thinking_config: None,
            #[cfg(feature = "gemini-interactions")]
            transport: GeminiTransport::GenerateContent,
            #[cfg(feature = "gemini-interactions")]
            interaction_target: None,
            #[cfg(feature = "gemini-interactions")]
            interaction_options: InteractionOptions::default(),
            #[cfg(feature = "gemini-vertex")]
            vertex_rag_store: None,
            #[cfg(feature = "gemini-vertex")]
            vertex_backend: false,
        }
    }

    /// Create a new Gemini model client with an API key and model name.
    pub fn new(api_key: impl Into<String>, model: impl Into<String>) -> Result<Self> {
        let model_name = model.into();
        crate::catalog::warn_if_obsolete("gemini", &model_name);
        let client = Gemini::with_model(api_key.into(), model_name.clone())
            .map_err(|e| adk_core::AdkError::model(e.to_string()))?;

        Ok(Self::from_client(client, model_name))
    }

    /// Create an API-key model using an explicit REST base URL.
    ///
    /// The base URL includes the API version and ends with a slash. No environment
    /// variable is read.
    ///
    /// # Errors
    ///
    /// Returns an error when `base_url` is not a valid URL or the client cannot be built.
    ///
    /// # Example
    ///
    /// ```
    /// use adk_model::GeminiModel;
    ///
    /// let model = GeminiModel::new_with_base_url(
    ///     "api-key",
    ///     "gemini-3.7-flash",
    ///     "https://gateway.example.com/v1beta/",
    /// )?;
    /// # Ok::<(), adk_core::AdkError>(())
    /// ```
    pub fn new_with_base_url(
        api_key: impl Into<String>,
        model: impl Into<String>,
        base_url: impl AsRef<str>,
    ) -> Result<Self> {
        let model_name = model.into();
        let base_url = base_url
            .as_ref()
            .parse()
            .map_err(|_| adk_core::AdkError::model("Invalid Gemini base URL"))?;
        let client = Gemini::with_model_and_base_url(api_key.into(), model_name.clone(), base_url)
            .map_err(|e| adk_core::AdkError::model(e.to_string()))?;
        Ok(Self::from_client(client, model_name))
    }

    /// Create a Gemini model via Vertex AI with API key auth.
    ///
    /// Requires `gemini-vertex` feature.
    #[cfg(feature = "gemini-vertex")]
    pub fn new_google_cloud(
        api_key: impl Into<String>,
        project_id: impl AsRef<str>,
        location: impl AsRef<str>,
        model: impl Into<String>,
    ) -> Result<Self> {
        let model_name = model.into();
        let client = Gemini::with_google_cloud_model(
            api_key.into(),
            project_id,
            location,
            model_name.clone(),
        )
        .map_err(|e| adk_core::AdkError::model(e.to_string()))?;

        let mut model = Self::from_client(client, model_name);
        model.vertex_backend = true;
        Ok(model)
    }

    /// Use an explicit Vertex endpoint and either a supplied key or execution-host ADC.
    ///
    /// Requires `gemini-vertex` feature.
    ///
    /// # Errors
    ///
    /// Returns an error when Application Default Credentials are requested
    /// (`api_key` is `None`) but unavailable, or when the client cannot be built.
    ///
    /// # Example
    ///
    /// ```no_run
    /// use adk_model::GeminiModel;
    ///
    /// let model = GeminiModel::new_google_cloud_endpoint(
    ///     None,
    ///     "my-project",
    ///     "us-central1",
    ///     "gemini-3.7-flash",
    ///     "https://us-central1-aiplatform.googleapis.com",
    /// )?;
    /// # Ok::<(), adk_core::AdkError>(())
    /// ```
    #[cfg(feature = "gemini-vertex")]
    pub fn new_google_cloud_endpoint(
        api_key: Option<&str>,
        project: &str,
        location: &str,
        model_name: &str,
        endpoint: &str,
    ) -> Result<Self> {
        let mut builder = adk_gemini::GeminiBuilder::new(api_key.unwrap_or_default())
            .with_model(model_name.to_owned())
            .with_google_cloud(project, location)
            .with_google_cloud_endpoint(endpoint);
        if api_key.is_none() {
            builder = builder.with_google_cloud_adc().map_err(|e| gemini_error_to_adk(&e))?;
        }
        let client = builder.build().map_err(|e| gemini_error_to_adk(&e))?;
        let mut model = Self::from_client(client, model_name.to_owned());
        model.vertex_backend = true;
        Ok(model)
    }

    /// Create a Gemini model via Vertex AI with service account JSON.
    ///
    /// Requires `gemini-vertex` feature.
    #[cfg(feature = "gemini-vertex")]
    pub fn new_google_cloud_service_account(
        service_account_json: &str,
        project_id: impl AsRef<str>,
        location: impl AsRef<str>,
        model: impl Into<String>,
    ) -> Result<Self> {
        let model_name = model.into();
        let client = Gemini::with_google_cloud_service_account_json(
            service_account_json,
            project_id.as_ref(),
            location.as_ref(),
            model_name.clone(),
        )
        .map_err(|e| adk_core::AdkError::model(e.to_string()))?;

        let mut model = Self::from_client(client, model_name);
        model.vertex_backend = true;
        Ok(model)
    }

    /// Create a Gemini model via Vertex AI with Application Default Credentials.
    ///
    /// Requires `gemini-vertex` feature.
    #[cfg(feature = "gemini-vertex")]
    pub fn new_google_cloud_adc(
        project_id: impl AsRef<str>,
        location: impl AsRef<str>,
        model: impl Into<String>,
    ) -> Result<Self> {
        let model_name = model.into();
        let client = Gemini::with_google_cloud_adc_model(
            project_id.as_ref(),
            location.as_ref(),
            model_name.clone(),
        )
        .map_err(|e| adk_core::AdkError::model(e.to_string()))?;

        let mut model = Self::from_client(client, model_name);
        model.vertex_backend = true;
        Ok(model)
    }

    /// Create a Gemini model via Vertex AI with Workload Identity Federation.
    ///
    /// Requires `gemini-vertex` feature.
    #[cfg(feature = "gemini-vertex")]
    pub fn new_google_cloud_wif(
        wif_json: &str,
        project_id: impl AsRef<str>,
        location: impl AsRef<str>,
        model: impl Into<String>,
    ) -> Result<Self> {
        let model_name = model.into();
        let client = Gemini::with_google_cloud_wif_json(
            wif_json,
            project_id.as_ref(),
            location.as_ref(),
            model_name.clone(),
        )
        .map_err(|e| adk_core::AdkError::model(e.to_string()))?;

        let mut model = Self::from_client(client, model_name);
        model.vertex_backend = true;
        Ok(model)
    }

    /// Create a Gemini model from environment variables.
    ///
    /// When `GOOGLE_GENAI_USE_ENTERPRISE` or `GOOGLE_GENAI_USE_VERTEXAI` is
    /// truthy (`1` or a case-insensitive `true`), builds a Vertex AI client
    /// with Application Default Credentials from `GOOGLE_CLOUD_PROJECT` and
    /// `GOOGLE_CLOUD_LOCATION` — this path requires the `gemini-vertex`
    /// feature. Otherwise builds a Gemini API (AI Studio) client from
    /// `GOOGLE_API_KEY` or `GEMINI_API_KEY`.
    ///
    /// `GOOGLE_GENAI_USE_ENTERPRISE` takes precedence when both flags are set.
    ///
    /// The Vertex path builds Application Default Credentials, which requires
    /// a Tokio runtime to be current — call from async code or under
    /// `#[tokio::main]`.
    ///
    /// # Errors
    ///
    /// - A Vertex flag is truthy but `GOOGLE_CLOUD_PROJECT` or
    ///   `GOOGLE_CLOUD_LOCATION` is not set.
    /// - A Vertex flag is truthy but the `gemini-vertex` feature is not
    ///   compiled.
    /// - No Vertex flag is truthy and neither `GOOGLE_API_KEY` nor
    ///   `GEMINI_API_KEY` is set.
    ///
    /// # Example
    ///
    /// ```rust,no_run
    /// use adk_model::gemini::GeminiModel;
    ///
    /// # #[tokio::main]
    /// # async fn main() -> adk_core::Result<()> {
    /// let model = GeminiModel::from_env("gemini-2.5-flash")?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn from_env(model: &str) -> Result<Self> {
        if vertex_env_requested() {
            #[cfg(feature = "gemini-vertex")]
            {
                let (project, location) = vertex_target_from_env()?;
                return Self::new_google_cloud_adc(project, location, model);
            }
            #[cfg(not(feature = "gemini-vertex"))]
            {
                return Err(adk_core::AdkError::new(
                    ErrorComponent::Model,
                    ErrorCategory::Unsupported,
                    "model.gemini.vertex_feature_missing",
                    "Vertex AI backend selected via GOOGLE_GENAI_USE_ENTERPRISE/GOOGLE_GENAI_USE_VERTEXAI, but the `gemini-vertex` feature is not compiled. Enable the `gemini-vertex` feature (or `gemini-agent-platform` on the adk-rust umbrella crate) to route Gemini traffic through Vertex AI.",
                )
                .with_provider("gemini"));
            }
        }

        let Some(api_key) =
            env_non_empty("GOOGLE_API_KEY").or_else(|| env_non_empty("GEMINI_API_KEY"))
        else {
            return Err(adk_core::AdkError::new(
                ErrorComponent::Model,
                ErrorCategory::InvalidInput,
                "model.gemini.api_key_missing",
                "No Gemini API key found. Set GOOGLE_API_KEY or GEMINI_API_KEY, or set GOOGLE_GENAI_USE_ENTERPRISE=true with GOOGLE_CLOUD_PROJECT/GOOGLE_CLOUD_LOCATION for Vertex AI.",
            )
            .with_provider("gemini"));
        };
        Self::new(api_key, model)
    }

    /// Set the retry configuration (builder pattern).
    #[must_use]
    pub fn with_retry_config(mut self, retry_config: RetryConfig) -> Self {
        self.retry_config = retry_config;
        self
    }

    /// Set the retry configuration (mutable reference).
    pub fn set_retry_config(&mut self, retry_config: RetryConfig) {
        self.retry_config = retry_config;
    }

    /// Returns the current retry configuration.
    pub fn retry_config(&self) -> &RetryConfig {
        &self.retry_config
    }

    /// Set the default thinking configuration applied to every request.
    ///
    /// Controls the model's reasoning effort. For Gemini 3 series models,
    /// use `ThinkingLevel` (Low/Medium/High). For Gemini 2.5 series, use
    /// `thinking_budget` (token count).
    ///
    /// # Example
    ///
    /// ```rust,ignore
    /// use adk_gemini::{ThinkingConfig, ThinkingLevel};
    ///
    /// // Gemini 3 — level-based thinking
    /// let model = GeminiModel::new(api_key, "gemini-3.1-pro-preview")?
    ///     .with_thinking_config(
    ///         ThinkingConfig::new().with_thinking_level(ThinkingLevel::Low)
    ///     );
    ///
    /// // Gemini 2.5 — budget-based thinking
    /// let model = GeminiModel::new(api_key, "gemini-2.5-flash")?
    ///     .with_thinking_config(
    ///         ThinkingConfig::new().with_thinking_budget(2048)
    ///     );
    /// ```
    #[must_use]
    pub fn with_thinking_config(mut self, thinking_config: adk_gemini::ThinkingConfig) -> Self {
        self.thinking_config = Some(thinking_config);
        self
    }

    /// Set the thinking configuration (mutable reference variant).
    pub fn set_thinking_config(&mut self, thinking_config: adk_gemini::ThinkingConfig) {
        self.thinking_config = Some(thinking_config);
    }

    /// Returns the current thinking configuration, if set.
    pub fn thinking_config(&self) -> Option<&adk_gemini::ThinkingConfig> {
        self.thinking_config.as_ref()
    }

    /// Ground every request in a Vertex AI RAG store (server-side retrieval).
    ///
    /// Declares a `retrieval` tool on each outgoing `generateContent` request
    /// so Vertex AI retrieves grounding context from the referenced RAG
    /// corpora before generating. Requires a Vertex AI backend — requests on
    /// an AI Studio model fail with the structured error
    /// `model.gemini.rag_store_requires_vertex` before any network IO.
    ///
    /// # Example
    ///
    /// ```rust,no_run
    /// use adk_gemini::VertexRagStore;
    /// use adk_model::gemini::GeminiModel;
    ///
    /// # fn main() -> adk_core::Result<()> {
    /// let store = VertexRagStore::corpus("projects/p/locations/us-central1/ragCorpora/123")
    ///     .with_top_k(5);
    /// let model =
    ///     GeminiModel::new_google_cloud("api-key", "p", "us-central1", "gemini-3.7-flash")?
    ///         .with_vertex_rag_store(store);
    /// # Ok(())
    /// # }
    /// ```
    #[cfg(feature = "gemini-vertex")]
    #[must_use]
    pub fn with_vertex_rag_store(mut self, store: adk_gemini::VertexRagStore) -> Self {
        self.vertex_rag_store = Some(store);
        self
    }

    /// Returns the configured Vertex RAG store, if set.
    #[cfg(feature = "gemini-vertex")]
    pub fn vertex_rag_store(&self) -> Option<&adk_gemini::VertexRagStore> {
        self.vertex_rag_store.as_ref()
    }

    /// Enable (or disable) the Interactions API transport.
    ///
    /// When enabling, the model's configured model id is validated against the
    /// Interactions allowlist via [`InteractionTarget::parse`]. On success the
    /// transport switches to [`GeminiTransport::Interactions`] and the
    /// validated target is stored. Disabling reverts to
    /// [`GeminiTransport::GenerateContent`] and clears the stored target.
    ///
    /// Requires the `gemini-interactions` feature.
    ///
    /// # Errors
    ///
    /// Returns an [`adk_core::AdkError`] with category `InvalidInput` when
    /// enabling the transport for a model id outside the Interactions
    /// allowlist (Requirement 2.4). The error message names the supported
    /// model and agent targets.
    ///
    /// # Example
    ///
    /// ```rust,ignore
    /// let model = GeminiModel::new(api_key, "gemini-2.5-flash")?
    ///     .use_interactions_api(true)?;
    /// ```
    #[cfg(feature = "gemini-interactions")]
    pub fn use_interactions_api(mut self, enabled: bool) -> Result<Self> {
        if enabled {
            let target = InteractionTarget::parse(&self.model_name)?;
            self.transport = GeminiTransport::Interactions;
            self.interaction_target = Some(target);
        } else {
            self.transport = GeminiTransport::GenerateContent;
            self.interaction_target = None;
        }
        Ok(self)
    }

    /// Override the Interactions transport options (store, stateful, background
    /// mode, poll interval).
    ///
    /// Requires the `gemini-interactions` feature. The defaults (see
    /// [`InteractionOptions::default`]) mirror the Interactions API's intended
    /// posture; override only when a different behavior is required.
    #[cfg(feature = "gemini-interactions")]
    #[must_use]
    pub fn interaction_options(mut self, opts: InteractionOptions) -> Self {
        self.interaction_options = opts;
        self
    }

    /// Returns the currently selected wire transport.
    ///
    /// Requires the `gemini-interactions` feature.
    #[cfg(feature = "gemini-interactions")]
    pub fn transport(&self) -> GeminiTransport {
        self.transport
    }

    /// Returns the configured Interactions options.
    ///
    /// Requires the `gemini-interactions` feature.
    #[cfg(feature = "gemini-interactions")]
    pub fn interaction_options_ref(&self) -> &InteractionOptions {
        &self.interaction_options
    }

    /// Resolves whether background execution should be used for the configured
    /// Interactions target, honoring [`BackgroundMode`].
    ///
    /// Returns `true` when the mode is [`BackgroundMode::Always`], `false` when
    /// [`BackgroundMode::Never`], and (for [`BackgroundMode::AgentTargetsOnly`])
    /// `true` only when the configured target is an agent target. When no
    /// Interactions target is configured, agent-targets-only resolves to
    /// `false`.
    ///
    /// Used by the non-streaming/background Interactions path (task 7.3); kept
    /// here so the transport state and its policy resolution live together.
    #[cfg(feature = "gemini-interactions")]
    fn resolve_background(&self) -> bool {
        match self.interaction_options.background {
            BackgroundMode::Always => true,
            BackgroundMode::Never => false,
            BackgroundMode::AgentTargetsOnly => {
                self.interaction_target.as_ref().is_some_and(InteractionTarget::is_agent)
            }
        }
    }

    fn convert_response(resp: &adk_gemini::GenerationResponse) -> Result<LlmResponse> {
        let mut converted_parts: Vec<Part> = Vec::new();

        // Convert content parts
        if let Some(parts) = resp.candidates.first().and_then(|c| c.content.parts.as_ref()) {
            for p in parts {
                match p {
                    adk_gemini::Part::Text { text, thought, thought_signature } => {
                        if thought == &Some(true) {
                            converted_parts.push(Part::Thinking {
                                thinking: text.clone(),
                                signature: thought_signature.clone(),
                            });
                        } else {
                            converted_parts.push(Part::Text { text: text.clone() });
                        }
                    }
                    adk_gemini::Part::InlineData { inline_data } => {
                        let decoded =
                            BASE64_STANDARD.decode(&inline_data.data).map_err(|error| {
                                adk_core::AdkError::model(format!(
                                    "failed to decode inline data from gemini response: {error}"
                                ))
                            })?;
                        converted_parts.push(Part::InlineData {
                            mime_type: inline_data.mime_type.clone(),
                            data: decoded,
                            uri: None,
                            annotations: None,
                        });
                    }
                    adk_gemini::Part::FunctionCall { function_call, thought_signature } => {
                        converted_parts.push(Part::FunctionCall {
                            name: function_call.name.clone(),
                            args: function_call.args.clone(),
                            id: function_call.id.clone(),
                            thought_signature: thought_signature.clone(),
                        });
                    }
                    adk_gemini::Part::FunctionResponse { function_response, .. } => {
                        converted_parts.push(Part::FunctionResponse {
                            function_response: adk_core::FunctionResponseData::new(
                                function_response.name.clone(),
                                function_response
                                    .response
                                    .clone()
                                    .unwrap_or(serde_json::Value::Null),
                            ),
                            id: None,
                            annotations: None,
                        });
                    }
                    adk_gemini::Part::ToolCall { .. } | adk_gemini::Part::ExecutableCode { .. } => {
                        if let Ok(value) = serde_json::to_value(p) {
                            converted_parts.push(Part::ServerToolCall { server_tool_call: value });
                        }
                    }
                    adk_gemini::Part::ToolResponse { .. }
                    | adk_gemini::Part::CodeExecutionResult { .. } => {
                        let value = serde_json::to_value(p).unwrap_or(serde_json::Value::Null);
                        converted_parts
                            .push(Part::ServerToolResponse { server_tool_response: value });
                    }
                    adk_gemini::Part::FileData { file_data } => {
                        converted_parts.push(Part::FileData {
                            mime_type: file_data.mime_type.clone(),
                            file_uri: file_data.file_uri.clone(),
                            annotations: None,
                        });
                    }
                }
            }
        }

        let content = if converted_parts.is_empty() {
            None
        } else {
            Some(Content { role: "model".to_string(), parts: converted_parts })
        };
        let tool_call_turn = content.as_ref().is_some_and(Content::has_function_calls);

        let usage_metadata = resp.usage_metadata.as_ref().map(|u| UsageMetadata {
            prompt_token_count: u.prompt_token_count.unwrap_or(0),
            // Gemini's candidate count excludes the separately reported thoughts.
            candidates_token_count: u
                .candidates_token_count
                .unwrap_or(0)
                .saturating_add(u.thoughts_token_count.unwrap_or(0)),
            total_token_count: u.total_token_count.unwrap_or(0),
            thinking_token_count: u.thoughts_token_count,
            cache_read_input_token_count: u.cached_content_token_count,
            provider_usage: serde_json::to_value(u).ok(),
            ..Default::default()
        });

        let finish_reason =
            resp.candidates.first().and_then(|c| c.finish_reason.as_ref()).map(|fr| match fr {
                adk_gemini::FinishReason::Stop => FinishReason::Stop,
                adk_gemini::FinishReason::MaxTokens => FinishReason::MaxTokens,
                adk_gemini::FinishReason::Safety => FinishReason::Safety,
                adk_gemini::FinishReason::Recitation => FinishReason::Recitation,
                _ => FinishReason::Other,
            });

        let citation_metadata =
            resp.candidates.first().and_then(|c| c.citation_metadata.as_ref()).map(|meta| {
                CitationMetadata {
                    citation_sources: meta
                        .citation_sources
                        .iter()
                        .map(|source| CitationSource {
                            uri: source.uri.clone(),
                            title: source.title.clone(),
                            start_index: source.start_index,
                            end_index: source.end_index,
                            license: source.license.clone(),
                            publication_date: source.publication_date.map(|d| d.to_string()),
                        })
                        .collect(),
                }
            });

        // Serialize grounding metadata into provider_metadata so consumers
        // can display search queries and sources without changing the answer.
        let provider_metadata = resp
            .candidates
            .first()
            .and_then(|c| c.grounding_metadata.as_ref())
            .and_then(|g| serde_json::to_value(g).ok());

        Ok(LlmResponse {
            content,
            usage_metadata,
            finish_reason,
            citation_metadata,
            partial: false,
            turn_complete: !tool_call_turn,
            interrupted: false,
            error_code: None,
            error_message: None,
            provider_metadata,
            interaction_id: None,
        })
    }

    fn gemini_function_response_payload(response: serde_json::Value) -> serde_json::Value {
        match response {
            // Gemini functionResponse.response must be a JSON object.
            serde_json::Value::Object(_) => response,
            other => serde_json::json!({ "result": other }),
        }
    }

    fn merge_object_value(
        target: &mut serde_json::Map<String, serde_json::Value>,
        value: serde_json::Value,
    ) {
        if let serde_json::Value::Object(object) = value {
            for (key, value) in object {
                target.insert(key, value);
            }
        }
    }

    fn build_gemini_tools(
        tools: &std::collections::HashMap<String, serde_json::Value>,
        adapter: &dyn SchemaAdapter,
        cache: &SchemaCache,
    ) -> Result<(Vec<adk_gemini::Tool>, adk_gemini::ToolConfig)> {
        let mut gemini_tools = Vec::new();
        let mut function_declarations = Vec::new();
        let mut has_provider_native_tools = false;
        let mut tool_config_json = serde_json::Map::new();

        for (name, tool_decl) in tools {
            if let Some(provider_tool) = tool_decl.get("x-adk-gemini-tool") {
                let tool = serde_json::from_value::<adk_gemini::Tool>(provider_tool.clone())
                    .map_err(|error| {
                        adk_core::AdkError::model(format!(
                            "failed to deserialize Gemini native tool '{name}': {error}"
                        ))
                    })?;
                has_provider_native_tools = true;
                gemini_tools.push(tool);
            } else {
                // Normalize tool name via the schema adapter
                let normalized_name = adapter.normalize_tool_name(name);

                // Get the parameters schema from the declaration, or use the
                // adapter's empty_schema fallback when none is provided.
                let schema =
                    tool_decl.get("parameters").cloned().unwrap_or_else(|| adapter.empty_schema());
                let normalized_schema = cache.normalize(&schema);

                // Build the FunctionDeclaration with normalized values
                let description =
                    tool_decl.get("description").and_then(|v| v.as_str()).unwrap_or("").to_string();

                let mut func_decl_json = serde_json::json!({
                    "name": normalized_name.as_ref(),
                    "description": description,
                    "parameters": normalized_schema,
                });

                // Preserve response schema if present (normalized like parameters)
                if let Some(response) = tool_decl.get("response") {
                    func_decl_json["response"] = cache.normalize(response);
                }

                // Preserve behavior if present
                if let Some(behavior) = tool_decl.get("behavior") {
                    func_decl_json["behavior"] = behavior.clone();
                }

                let func_decl =
                    serde_json::from_value::<adk_gemini::FunctionDeclaration>(func_decl_json)
                        .map_err(|error| {
                            adk_core::AdkError::model(format!(
                                "failed to build Gemini function declaration for '{name}': {error}"
                            ))
                        })?;
                function_declarations.push(func_decl);
            }

            if let Some(tool_config) = tool_decl.get("x-adk-gemini-tool-config") {
                Self::merge_object_value(&mut tool_config_json, tool_config.clone());
            }
        }

        let has_function_declarations = !function_declarations.is_empty();
        if has_function_declarations {
            gemini_tools.push(adk_gemini::Tool::with_functions(function_declarations));
        }

        if has_provider_native_tools {
            tool_config_json.insert(
                "includeServerSideToolInvocations".to_string(),
                serde_json::Value::Bool(true),
            );
        }

        let tool_config = if tool_config_json.is_empty() {
            adk_gemini::ToolConfig::default()
        } else {
            serde_json::from_value::<adk_gemini::ToolConfig>(serde_json::Value::Object(
                tool_config_json,
            ))
            .map_err(|error| {
                adk_core::AdkError::model(format!(
                    "failed to deserialize Gemini tool configuration: {error}"
                ))
            })?
        };

        Ok((gemini_tools, tool_config))
    }

    fn stream_chunks_from_response(
        mut response: LlmResponse,
        saw_partial_chunk: bool,
    ) -> (Vec<LlmResponse>, bool) {
        let is_final = response.finish_reason.is_some();

        if !is_final {
            response.partial = true;
            response.turn_complete = false;
            return (vec![response], true);
        }

        response.partial = false;
        response.turn_complete =
            response.content.as_ref().is_none_or(|content| !content.has_function_calls());

        if saw_partial_chunk {
            return (vec![response], true);
        }

        let synthetic_partial = LlmResponse {
            content: None,
            usage_metadata: None,
            finish_reason: None,
            citation_metadata: None,
            partial: true,
            turn_complete: false,
            interrupted: false,
            error_code: None,
            error_message: None,
            provider_metadata: None,
            interaction_id: None,
        };

        (vec![synthetic_partial, response], true)
    }

    async fn generate_content_internal(
        &self,
        req: LlmRequest,
        stream: bool,
    ) -> Result<LlmResponseStream> {
        let mut builder = self.client.generate_content();

        // Build a map of function_name → thought_signature from FunctionCall parts
        // in model content. Gemini 3.x requires thought_signature on FunctionResponse
        // parts when thinking is active, but adk_core::Part::FunctionResponse doesn't
        // carry it (it's Gemini-specific). We recover it here at the provider boundary.
        let mut fn_call_signatures: std::collections::HashMap<String, String> =
            std::collections::HashMap::new();
        for content in &req.contents {
            if content.role == "model" {
                for part in &content.parts {
                    if let Part::FunctionCall { name, thought_signature: Some(sig), .. } = part {
                        fn_call_signatures.insert(name.clone(), sig.clone());
                    }
                }
            }
        }

        // Add contents using proper builder methods
        for content in &req.contents {
            match content.role.as_str() {
                "user" => {
                    // For user messages, build gemini Content with potentially multiple parts
                    let mut gemini_parts = Vec::new();
                    for part in &content.parts {
                        match part {
                            Part::Text { text } => {
                                gemini_parts.push(adk_gemini::Part::Text {
                                    text: text.clone(),
                                    thought: None,
                                    thought_signature: None,
                                });
                            }
                            Part::Thinking { thinking, signature } => {
                                gemini_parts.push(adk_gemini::Part::Text {
                                    text: thinking.clone(),
                                    thought: Some(true),
                                    thought_signature: signature.clone(),
                                });
                            }
                            Part::InlineData { data, mime_type, .. } => {
                                let encoded = attachment::encode_base64(data);
                                gemini_parts.push(adk_gemini::Part::InlineData {
                                    inline_data: adk_gemini::Blob {
                                        mime_type: mime_type.clone(),
                                        data: encoded,
                                    },
                                });
                            }
                            Part::FileData { mime_type, file_uri, .. } => {
                                gemini_parts.push(adk_gemini::Part::Text {
                                    text: attachment::file_attachment_to_text(mime_type, file_uri),
                                    thought: None,
                                    thought_signature: None,
                                });
                            }
                            _ => {}
                        }
                    }
                    if !gemini_parts.is_empty() {
                        let user_content = adk_gemini::Content {
                            role: Some(adk_gemini::Role::User),
                            parts: Some(gemini_parts),
                        };
                        builder = builder.with_message(adk_gemini::Message {
                            content: user_content,
                            role: adk_gemini::Role::User,
                        });
                    }
                }
                "model" => {
                    // For model messages, build gemini Content
                    let mut gemini_parts = Vec::new();
                    for part in &content.parts {
                        match part {
                            Part::Text { text } => {
                                gemini_parts.push(adk_gemini::Part::Text {
                                    text: text.clone(),
                                    thought: None,
                                    thought_signature: None,
                                });
                            }
                            Part::Thinking { thinking, signature } => {
                                gemini_parts.push(adk_gemini::Part::Text {
                                    text: thinking.clone(),
                                    thought: Some(true),
                                    thought_signature: signature.clone(),
                                });
                            }
                            Part::FunctionCall { name, args, thought_signature, id } => {
                                gemini_parts.push(adk_gemini::Part::FunctionCall {
                                    function_call: adk_gemini::FunctionCall {
                                        name: name.clone(),
                                        args: args.clone(),
                                        id: id.clone(),
                                        thought_signature: None,
                                    },
                                    thought_signature: thought_signature.clone(),
                                });
                            }
                            Part::ServerToolCall { server_tool_call } => {
                                if let Ok(native_part) = serde_json::from_value::<adk_gemini::Part>(
                                    server_tool_call.clone(),
                                ) {
                                    match native_part {
                                        adk_gemini::Part::ToolCall { .. }
                                        | adk_gemini::Part::ExecutableCode { .. } => {
                                            gemini_parts.push(native_part);
                                            continue;
                                        }
                                        _ => {}
                                    }
                                }

                                gemini_parts.push(adk_gemini::Part::ToolCall {
                                    tool_call: server_tool_call.clone(),
                                    thought_signature: Self::gemini_part_thought_signature(
                                        server_tool_call,
                                    ),
                                });
                            }
                            Part::ServerToolResponse { server_tool_response } => {
                                if let Ok(native_part) = serde_json::from_value::<adk_gemini::Part>(
                                    server_tool_response.clone(),
                                ) {
                                    match native_part {
                                        adk_gemini::Part::ToolResponse { .. }
                                        | adk_gemini::Part::CodeExecutionResult { .. } => {
                                            gemini_parts.push(native_part);
                                            continue;
                                        }
                                        _ => {}
                                    }
                                }

                                gemini_parts.push(adk_gemini::Part::ToolResponse {
                                    tool_response: server_tool_response.clone(),
                                    thought_signature: Self::gemini_part_thought_signature(
                                        server_tool_response,
                                    ),
                                });
                            }
                            _ => {}
                        }
                    }
                    if !gemini_parts.is_empty() {
                        let model_content = adk_gemini::Content {
                            role: Some(adk_gemini::Role::Model),
                            parts: Some(gemini_parts),
                        };
                        builder = builder.with_message(adk_gemini::Message {
                            content: model_content,
                            role: adk_gemini::Role::Model,
                        });
                    }
                }
                "function" => {
                    // For function responses, build content directly to attach thought_signature
                    // recovered from the preceding FunctionCall (Gemini 3.x requirement)
                    let mut gemini_parts = Vec::new();
                    for part in &content.parts {
                        if let Part::FunctionResponse { function_response, id, .. } = part {
                            let sig = fn_call_signatures.get(&function_response.name).cloned();

                            // Build nested FunctionResponsePart entries for multimodal data
                            let mut fr_parts = Vec::new();
                            for inline in &function_response.inline_data {
                                let encoded = attachment::encode_base64(&inline.data);
                                fr_parts.push(adk_gemini::FunctionResponsePart::InlineData {
                                    inline_data: adk_gemini::Blob {
                                        mime_type: inline.mime_type.clone(),
                                        data: encoded,
                                    },
                                });
                            }
                            for file in &function_response.file_data {
                                fr_parts.push(adk_gemini::FunctionResponsePart::FileData {
                                    file_data: adk_gemini::FileDataRef {
                                        mime_type: file.mime_type.clone(),
                                        file_uri: file.file_uri.clone(),
                                    },
                                });
                            }

                            let mut gemini_fr = adk_gemini::tools::FunctionResponse::new(
                                &function_response.name,
                                Self::gemini_function_response_payload(
                                    function_response.response.clone(),
                                ),
                            );
                            gemini_fr.parts = fr_parts;
                            // Echo the call id so Gemini 3.x strict response matching
                            // (id + name + count) can correlate this response.
                            gemini_fr.id = id.clone();

                            gemini_parts.push(adk_gemini::Part::FunctionResponse {
                                function_response: gemini_fr,
                                thought_signature: sig,
                            });
                        }
                    }
                    if !gemini_parts.is_empty() {
                        let fn_content = adk_gemini::Content {
                            role: Some(adk_gemini::Role::User),
                            parts: Some(gemini_parts),
                        };
                        builder = builder.with_message(adk_gemini::Message {
                            content: fn_content,
                            role: adk_gemini::Role::User,
                        });
                    }
                }
                _ => {}
            }
        }

        // Add generation config
        if let Some(config) = req.config {
            let has_schema = config.response_schema.is_some();
            let gen_config = adk_gemini::GenerationConfig {
                temperature: config.temperature,
                top_p: config.top_p,
                top_k: config.top_k,
                max_output_tokens: config.max_output_tokens,
                response_schema: config.response_schema,
                response_mime_type: if has_schema {
                    Some("application/json".to_string())
                } else {
                    None
                },
                thinking_config: self.thinking_config.clone(),
                ..Default::default()
            };
            builder = builder.with_generation_config(gen_config);

            // Attach cached content reference if provided
            if let Some(ref name) = config.cached_content {
                let handle = self.client.get_cached_content(name);
                builder = builder.with_cached_content(&handle);
            }
        } else if self.thinking_config.is_some() {
            // No generation config from the request, but we have a default
            // thinking config — apply it in an otherwise-default gen config.
            let gen_config = adk_gemini::GenerationConfig {
                thinking_config: self.thinking_config.clone(),
                ..Default::default()
            };
            builder = builder.with_generation_config(gen_config);
        }

        // Add tools
        if !req.tools.is_empty() {
            let adapter = self.schema_adapter();
            use std::sync::LazyLock;
            static SCHEMA_CACHE: LazyLock<SchemaCache> = LazyLock::new(|| {
                SchemaCache::for_adapter(std::sync::Arc::new(GeminiSchemaAdapter::new()))
            });
            let (gemini_tools, tool_config) =
                Self::build_gemini_tools(&req.tools, adapter, &SCHEMA_CACHE)?;
            for tool in gemini_tools {
                builder = builder.with_tool(tool);
            }
            if tool_config != adk_gemini::ToolConfig::default() {
                builder = builder.with_tool_config(tool_config);
            }
        }

        // Declare server-side RAG grounding; the Vertex backend injects it
        // as a `retrieval` tool at send time (validate_request_contract has
        // already rejected the Studio backend at this point).
        #[cfg(feature = "gemini-vertex")]
        if let Some(store) = &self.vertex_rag_store {
            builder = builder.with_vertex_rag_store(store.clone());
        }

        if stream {
            adk_telemetry::debug!("Executing streaming request");
            let response_stream = builder.execute_stream().await.map_err(|e| {
                adk_telemetry::error!(error = %e, "Model request failed");
                gemini_error_to_adk(&e)
            })?;

            let mapped_stream = async_stream::stream! {
                let mut stream = response_stream;
                let mut saw_partial_chunk = false;
                while let Some(result) = stream.try_next().await.transpose() {
                    match result {
                        Ok(resp) => {
                            match Self::convert_response(&resp) {
                                Ok(llm_resp) => {
                                    let (chunks, next_saw_partial) =
                                        Self::stream_chunks_from_response(llm_resp, saw_partial_chunk);
                                    saw_partial_chunk = next_saw_partial;
                                    for chunk in chunks {
                                        yield Ok(chunk);
                                    }
                                }
                                Err(e) => {
                                    adk_telemetry::error!(error = %e, "Failed to convert response");
                                    yield Err(e);
                                }
                            }
                        }
                        Err(e) => {
                            adk_telemetry::error!(error = %e, "Stream error");
                            yield Err(gemini_error_to_adk(&e));
                        }
                    }
                }
            };

            Ok(Box::pin(mapped_stream))
        } else {
            adk_telemetry::debug!("Executing blocking request");
            let response = builder.execute().await.map_err(|e| {
                adk_telemetry::error!(error = %e, "Model request failed");
                gemini_error_to_adk(&e)
            })?;

            let llm_response = Self::convert_response(&response)?;

            let stream = async_stream::stream! {
                yield Ok(llm_response);
            };

            Ok(Box::pin(stream))
        }
    }

    /// Create a cached content resource with the given system instruction, tools, and TTL.
    ///
    /// Returns the cache name (e.g., "cachedContents/abc123") on success.
    /// The cache is created using the model configured on this `GeminiModel` instance.
    pub async fn create_cached_content(
        &self,
        system_instruction: &str,
        tools: &std::collections::HashMap<String, serde_json::Value>,
        ttl_seconds: u32,
    ) -> Result<String> {
        let mut cache_builder = self
            .client
            .create_cache()
            .with_system_instruction(system_instruction)
            .with_ttl(std::time::Duration::from_secs(u64::from(ttl_seconds)));

        let adapter = self.schema_adapter();
        use std::sync::LazyLock;
        static SCHEMA_CACHE: LazyLock<SchemaCache> = LazyLock::new(|| {
            SchemaCache::for_adapter(std::sync::Arc::new(GeminiSchemaAdapter::new()))
        });
        let (gemini_tools, tool_config) = Self::build_gemini_tools(tools, adapter, &SCHEMA_CACHE)?;
        if !gemini_tools.is_empty() {
            cache_builder = cache_builder.with_tools(gemini_tools);
        }
        if tool_config != adk_gemini::ToolConfig::default() {
            cache_builder = cache_builder.with_tool_config(tool_config);
        }

        let handle = cache_builder
            .execute()
            .await
            .map_err(|e| adk_core::AdkError::model(format!("cache creation failed: {e}")))?;

        Ok(handle.name().to_string())
    }

    /// Delete a cached content resource by name.
    pub async fn delete_cached_content(&self, name: &str) -> Result<()> {
        let handle = self.client.get_cached_content(name);
        handle
            .delete()
            .await
            .map_err(|(_, e)| adk_core::AdkError::model(format!("cache deletion failed: {e}")))?;
        Ok(())
    }

    /// Drives a single turn through the Interactions API (Beta), non-streaming.
    ///
    /// This is the Interactions counterpart to
    /// [`generate_content_internal`](Self::generate_content_internal). It builds
    /// a [`CreateInteractionRequest`](adk_gemini::interactions::CreateInteractionRequest)
    /// from `req` via [`interactions_convert::build_request`], sends it, polls to
    /// completion when running in the background, maps terminal failure statuses
    /// to errors, and converts the final interaction into a single
    /// [`LlmResponse`].
    ///
    /// Behavior of note:
    ///
    /// - **Background completion (Requirement 7.5).** When `background` is set
    ///   and the first response is neither terminal nor awaiting a tool result,
    ///   the interaction is polled via
    ///   [`get_interaction`](adk_gemini::Gemini::get_interaction) every
    ///   `poll_interval` until it reaches a terminal or `requires_action` state.
    /// - **Stale continuation fallback (Requirement 4.4).** If the initial send
    ///   fails with a `NotFound` error *and* the request carried a
    ///   `previous_response_id`, the request is transparently rebuilt without
    ///   stateful continuation (full transcript, no `previous_interaction_id`)
    ///   and re-sent once. The original `NotFound` is not surfaced.
    /// - **Terminal failure (Requirements 7.7 / 9.2).** A final `failed` /
    ///   `budget_exceeded` status becomes an [`adk_core::AdkError`].
    ///
    /// The streaming counterpart is
    /// [`generate_interactions_stream`](Self::generate_interactions_stream).
    #[cfg(feature = "gemini-interactions")]
    async fn generate_interactions_once(&self, req: LlmRequest) -> Result<LlmResponse> {
        use super::interactions_convert;

        // The Interactions target is always populated when the transport is
        // active (set by `use_interactions_api`); guard defensively.
        let target = self.interaction_target.as_ref().ok_or_else(|| {
            adk_core::AdkError::new(
                ErrorComponent::Model,
                ErrorCategory::InvalidInput,
                "model.gemini.interactions.missing_target",
                "the Interactions transport is active but no validated target is configured",
            )
            .with_provider("gemini")
        })?;

        // Resolve the thinking level (Gemini 3 level-based reasoning). Budget-only
        // configs (Gemini 2.5) carry no level, so this stays `None` for them.
        let thinking_level = self.thinking_config.as_ref().and_then(|c| c.thinking_level);

        let stateful = self.interaction_options.stateful;
        let store = self.interaction_options.store;
        let background = self.resolve_background();

        // Build the request and stamp the background flag.
        let mut request =
            interactions_convert::build_request(&req, target, thinking_level, stateful, store)?;
        request.background = Some(background);

        // Send, with a transparent transcript fallback for a stale continuation id.
        let interaction = match self.client.send_interaction(request.clone()).await {
            Ok(interaction) => interaction,
            Err(error) => {
                let mapped = gemini_error_to_adk(&error);
                // Requirement 4.4: a rejected `previous_interaction_id` (retention
                // expiry) maps to NotFound. Rebuild statelessly (full transcript,
                // no continuation id) and retry once, without surfacing the error.
                if mapped.category == ErrorCategory::NotFound && req.previous_response_id.is_some()
                {
                    let mut fallback = interactions_convert::build_request(
                        &req,
                        target,
                        thinking_level,
                        /* stateful */ false,
                        store,
                    )?;
                    fallback.background = Some(background);
                    self.client
                        .send_interaction(fallback)
                        .await
                        .map_err(|e| gemini_error_to_adk(&e))?
                } else {
                    return Err(mapped);
                }
            }
        };

        // Background completion: poll until terminal or awaiting a tool result.
        let final_interaction = if background {
            self.poll_interaction_to_completion(interaction).await?
        } else {
            interaction
        };

        // Requirements 7.7 / 9.2: surface terminal failure statuses as errors.
        interaction_status_to_result(final_interaction.status)?;

        Ok(interactions_convert::to_llm_response(&final_interaction))
    }

    /// Polls a background interaction until it reaches a terminal or
    /// `requires_action` state, honoring the configured `poll_interval`
    /// (Requirement 7.5).
    ///
    /// Returns immediately when the interaction is already terminal or awaiting
    /// a tool result. Otherwise it sleeps for `poll_interval` and re-fetches via
    /// [`get_interaction`](adk_gemini::Gemini::get_interaction) (with
    /// `include_input = false`) until one of those states is reached, or until a
    /// bounded safeguard (`MAX_POLL_ATTEMPTS`) trips.
    ///
    /// ## Cancellation (Requirement 7.6)
    ///
    /// True invocation-driven cancellation (calling
    /// [`cancel_interaction`](adk_gemini::Gemini::cancel_interaction) when the
    /// caller cancels) is **not reachable from this layer**: the [`Llm`] trait's
    /// [`generate_content`](Llm::generate_content) signature receives only an
    /// [`LlmRequest`] and a `stream` flag — it has no `InvocationContext` or
    /// cancellation token. Cancellation is handled by the runner at the
    /// event-stream boundary: when a run is cancelled the runner stops consuming
    /// the agent's event stream, which drops this future and cancels its
    /// `await` points (the in-flight `sleep` / `get_interaction`) cooperatively.
    /// The polled interaction is left running server-side; reviving it to issue
    /// an explicit `cancel_interaction` would require threading the cancellation
    /// token through the trait, which is intentionally out of scope here (the
    /// trait is transport-only and shared by every provider).
    ///
    /// The `MAX_POLL_ATTEMPTS` bound is a safeguard against an interaction that
    /// never reaches a terminal/`requires_action` state (e.g. a server-side
    /// stall): rather than looping forever it returns a [`ErrorCategory::Timeout`]
    /// error so the call fails fast instead of hanging.
    #[cfg(feature = "gemini-interactions")]
    async fn poll_interaction_to_completion(
        &self,
        interaction: adk_gemini::interactions::Interaction,
    ) -> Result<adk_gemini::interactions::Interaction> {
        /// Upper bound on poll iterations before giving up, guarding against an
        /// interaction that never settles. With the default 1s `poll_interval`
        /// this is ~10 minutes; shorter intervals trade latency for a tighter
        /// wall-clock cap. Deep Research agents complete well within this.
        const MAX_POLL_ATTEMPTS: u32 = 600;

        let mut current = interaction;
        let mut attempts: u32 = 0;
        while !current.status.is_terminal() && !current.status.requires_action() {
            if attempts >= MAX_POLL_ATTEMPTS {
                return Err(adk_core::AdkError::new(
                    ErrorComponent::Model,
                    ErrorCategory::Timeout,
                    "model.gemini.interactions.poll_timeout",
                    format!(
                        "the Gemini interaction did not reach a terminal or requires_action \
                         state after {MAX_POLL_ATTEMPTS} poll attempts"
                    ),
                )
                .with_provider("gemini"));
            }
            attempts += 1;
            tokio::time::sleep(self.interaction_options.poll_interval).await;
            current = self
                .client
                .get_interaction(&current.id, false)
                .await
                .map_err(|e| gemini_error_to_adk(&e))?;
        }
        Ok(current)
    }

    /// Drives a single turn through the Interactions API (Beta) as an SSE
    /// stream, yielding partial→final [`LlmResponse`] chunks (Requirement 7.4).
    ///
    /// This is the streaming counterpart to
    /// [`generate_interactions_once`](Self::generate_interactions_once). It
    /// builds the request the same way, forces non-background completion
    /// (streaming and background polling are mutually exclusive completion
    /// modes — SSE delivers the turn incrementally, so `background` is set to
    /// `false`), opens the SSE stream via
    /// [`send_interaction_stream`](adk_gemini::Gemini::send_interaction_stream),
    /// and folds each [`InteractionSseEvent`](adk_gemini::interactions::InteractionSseEvent)
    /// into chunks via [`interactions_convert::sse_event_to_chunk`].
    ///
    /// Stream setup (target resolution, request building, opening the SSE
    /// connection) is fallible and returns `Err` synchronously, so
    /// `execute_with_retry` (which wraps this in `generate_content`) can retry
    /// transient setup failures. Errors that occur *after* the stream starts
    /// are yielded into the stream and not retried, mirroring the
    /// generateContent streaming path.
    ///
    /// The stale-continuation fallback used by the non-streaming path is not
    /// applied here: a `NotFound` on stream setup surfaces as a normal error
    /// (the streaming path is opt-in and lower-level; callers that need
    /// transparent retention fallback use the default non-streaming path).
    #[cfg(feature = "gemini-interactions")]
    async fn generate_interactions_stream(&self, req: LlmRequest) -> Result<LlmResponseStream> {
        use super::interactions_convert::{self, SseAccumulator, sse_event_to_chunk};

        let target = self.interaction_target.as_ref().ok_or_else(|| {
            adk_core::AdkError::new(
                ErrorComponent::Model,
                ErrorCategory::InvalidInput,
                "model.gemini.interactions.missing_target",
                "the Interactions transport is active but no validated target is configured",
            )
            .with_provider("gemini")
        })?;

        let thinking_level = self.thinking_config.as_ref().and_then(|c| c.thinking_level);
        let stateful = self.interaction_options.stateful;
        let store = self.interaction_options.store;

        let mut request =
            interactions_convert::build_request(&req, target, thinking_level, stateful, store)?;
        // Streaming uses SSE for incremental completion, not background polling;
        // the two are distinct completion modes (Requirement 7.4 vs 7.5). Force
        // foreground so the server streams the turn rather than returning a
        // background handle.
        request.background = Some(false);

        let sse_stream = self
            .client
            .send_interaction_stream(request)
            .await
            .map_err(|e| gemini_error_to_adk(&e))?;

        let mapped = async_stream::stream! {
            let mut sse_stream = sse_stream;
            let mut acc = SseAccumulator::new();
            while let Some(result) = sse_stream.try_next().await.transpose() {
                match result {
                    Ok(event) => {
                        if let Some(chunk) = sse_event_to_chunk(event, &mut acc) {
                            yield chunk;
                        }
                    }
                    Err(e) => {
                        adk_telemetry::error!(error = %e, "Interaction stream error");
                        yield Err(gemini_error_to_adk(&e));
                    }
                }
            }
        };

        Ok(Box::pin(mapped))
    }
}

#[async_trait]
impl Llm for GeminiModel {
    fn name(&self) -> &str {
        &self.model_name
    }

    fn schema_adapter(&self) -> &dyn SchemaAdapter {
        use std::sync::LazyLock;
        static ADAPTER: LazyLock<GeminiSchemaAdapter> = LazyLock::new(GeminiSchemaAdapter::new);
        &*ADAPTER
    }

    #[cfg(feature = "gemini-interactions")]
    fn uses_interactions_api(&self) -> bool {
        self.transport == GeminiTransport::Interactions
    }

    // Named distinctly from the agent layer's `call_llm` span (which carries
    // the gcp.vertex.agent.* attributes and is what trace exporters capture):
    // this is the model-transport layer, so traces show one `call_llm` per
    // LLM call instead of an identical nested pair.
    #[adk_telemetry::instrument(
        name = "model.generate_content",
        skip(self, req),
        fields(
            model.name = %self.model_name,
            stream = %stream,
            request.contents_count = %req.contents.len(),
            request.tools_count = %req.tools.len()
        )
    )]
    async fn generate_content(&self, req: LlmRequest, stream: bool) -> Result<LlmResponseStream> {
        adk_telemetry::info!("Generating content");
        self.validate_request_contract(&req)?;
        let usage_span = adk_telemetry::llm_generate_span("gemini", &self.model_name, stream);

        // Dispatch on the configured transport. The default `GenerateContent`
        // path is unchanged; the Interactions path (task 7.3/7.4) is only
        // reachable when `use_interactions_api` switched the transport.
        //
        // Retries only cover request setup/execution. Stream failures after the
        // stream starts are yielded to the caller and are not replayed
        // automatically.
        #[cfg(feature = "gemini-interactions")]
        if self.transport == GeminiTransport::Interactions {
            // Streaming and non-streaming are distinct completion modes. The
            // streaming path consumes the Interactions SSE stream and yields
            // partial→final chunks; the non-streaming path sends a single
            // request and (optionally) polls a background interaction to
            // completion. `execute_with_retry` wraps only request *setup* in
            // both cases — once a stream starts, its mid-flight errors are
            // surfaced to the caller rather than replayed (mirroring the
            // generateContent streaming path).
            if stream {
                let mapped =
                    execute_with_retry(&self.retry_config, is_retryable_model_error, || {
                        self.generate_interactions_stream(req.clone())
                    })
                    .await?;
                return Ok(crate::usage_tracking::with_usage_tracking(mapped, usage_span));
            }
            let response = execute_with_retry(&self.retry_config, is_retryable_model_error, || {
                self.generate_interactions_once(req.clone())
            })
            .await?;
            let single = async_stream::stream! {
                yield Ok(response);
            };
            return Ok(crate::usage_tracking::with_usage_tracking(Box::pin(single), usage_span));
        }

        let result = execute_with_retry(&self.retry_config, is_retryable_model_error, || {
            self.generate_content_internal(req.clone(), stream)
        })
        .await?;
        Ok(crate::usage_tracking::with_usage_tracking(result, usage_span))
    }
}

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

    fn test_adapter() -> GeminiSchemaAdapter {
        GeminiSchemaAdapter::new()
    }

    fn test_cache() -> SchemaCache {
        SchemaCache::for_adapter(std::sync::Arc::new(GeminiSchemaAdapter::new()))
    }

    #[test]
    fn test_build_gemini_tools_supports_native_tool_metadata() {
        let mut tools = std::collections::HashMap::new();
        tools.insert(
            "google_search".to_string(),
            serde_json::json!({
                "x-adk-gemini-tool": {
                    "google_search": {}
                }
            }),
        );
        tools.insert(
            "lookup_weather".to_string(),
            serde_json::json!({
                "name": "lookup_weather",
                "description": "lookup weather",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "city": { "type": "string" }
                    }
                }
            }),
        );

        let adapter = test_adapter();
        let cache = test_cache();
        let (gemini_tools, tool_config) = GeminiModel::build_gemini_tools(&tools, &adapter, &cache)
            .expect("tool conversion should succeed");

        assert_eq!(gemini_tools.len(), 2);
        assert_eq!(tool_config.include_server_side_tool_invocations, Some(true));
    }

    #[test]
    fn test_build_gemini_tools_sets_flag_for_builtin_only() {
        let mut tools = std::collections::HashMap::new();
        tools.insert(
            "google_search".to_string(),
            serde_json::json!({
                "x-adk-gemini-tool": {
                    "google_search": {}
                }
            }),
        );

        let adapter = test_adapter();
        let cache = test_cache();
        let (_gemini_tools, tool_config) =
            GeminiModel::build_gemini_tools(&tools, &adapter, &cache)
                .expect("tool conversion should succeed");

        assert_eq!(
            tool_config.include_server_side_tool_invocations,
            Some(true),
            "includeServerSideToolInvocations should be set even with only built-in tools"
        );
    }

    #[test]
    fn test_build_gemini_tools_no_flag_for_function_only() {
        let mut tools = std::collections::HashMap::new();
        tools.insert(
            "lookup_weather".to_string(),
            serde_json::json!({
                "name": "lookup_weather",
                "description": "lookup weather",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "city": { "type": "string" }
                    }
                }
            }),
        );

        let adapter = test_adapter();
        let cache = test_cache();
        let (_gemini_tools, tool_config) =
            GeminiModel::build_gemini_tools(&tools, &adapter, &cache)
                .expect("tool conversion should succeed");

        assert_eq!(
            tool_config.include_server_side_tool_invocations, None,
            "includeServerSideToolInvocations should NOT be set for function-only tools"
        );
    }

    #[test]
    fn test_build_gemini_tools_merges_native_tool_config() {
        let mut tools = std::collections::HashMap::new();
        tools.insert(
            "google_maps".to_string(),
            serde_json::json!({
                "x-adk-gemini-tool": {
                    "google_maps": {
                        "enable_widget": true
                    }
                },
                "x-adk-gemini-tool-config": {
                    "retrievalConfig": {
                        "latLng": {
                            "latitude": 1.23,
                            "longitude": 4.56
                        }
                    }
                }
            }),
        );

        let adapter = test_adapter();
        let cache = test_cache();
        let (_gemini_tools, tool_config) =
            GeminiModel::build_gemini_tools(&tools, &adapter, &cache)
                .expect("tool conversion should succeed");

        assert_eq!(
            tool_config.retrieval_config,
            Some(serde_json::json!({
                "latLng": {
                    "latitude": 1.23,
                    "longitude": 4.56
                }
            }))
        );
    }

    #[test]
    fn test_response_schema_is_normalized_like_parameters() {
        // Regression test: MCP tools provide an output_schema that becomes the
        // `response` field in the tool declaration. Gemini rejects JSON-Schema
        // dialect fields ($schema, additionalProperties) in the response schema.
        // This test verifies that the response schema is normalized the same way
        // parameters are — stripping unsupported keywords.
        let mut tools = std::collections::HashMap::new();
        tools.insert(
            "read_file".to_string(),
            serde_json::json!({
                "name": "read_file",
                "description": "Read a file from the filesystem",
                "parameters": {
                    "$schema": "http://json-schema.org/draft-07/schema#",
                    "type": "object",
                    "properties": {
                        "path": { "type": "string", "description": "File path" }
                    },
                    "required": ["path"],
                    "additionalProperties": false
                },
                "response": {
                    "$schema": "http://json-schema.org/draft-07/schema#",
                    "type": "object",
                    "properties": {
                        "content": { "type": "string", "description": "File contents" }
                    },
                    "required": ["content"],
                    "additionalProperties": false
                }
            }),
        );

        let adapter = test_adapter();
        let cache = test_cache();
        let (gemini_tools, _) = GeminiModel::build_gemini_tools(&tools, &adapter, &cache)
            .expect("tool conversion should succeed");

        // Find the function declaration
        let func_tool = gemini_tools
            .iter()
            .find(|t| matches!(t, adk_gemini::Tool::Function { .. }))
            .expect("should have function declarations");

        let decls = match func_tool {
            adk_gemini::Tool::Function { function_declarations } => function_declarations,
            _ => panic!("expected Function tool"),
        };

        let decl = &decls[0];
        let decl_json = serde_json::to_value(decl).unwrap();

        // Parameters should be normalized (no $schema, no additionalProperties)
        let params = &decl_json["parameters"];
        assert!(params.get("$schema").is_none(), "parameters.$schema should be stripped");
        assert!(
            params.get("additionalProperties").is_none(),
            "parameters.additionalProperties should be stripped"
        );

        // Response should ALSO be normalized (this was the bug)
        let response = &decl_json["response"];
        assert!(
            response.get("$schema").is_none(),
            "response.$schema should be stripped (was the bug: copied raw)"
        );
        assert!(
            response.get("additionalProperties").is_none(),
            "response.additionalProperties should be stripped (was the bug: copied raw)"
        );
    }
}

#[async_trait]
impl CacheCapable for GeminiModel {
    async fn create_cache(
        &self,
        system_instruction: &str,
        tools: &std::collections::HashMap<String, serde_json::Value>,
        ttl_seconds: u32,
    ) -> Result<String> {
        self.create_cached_content(system_instruction, tools, ttl_seconds).await
    }

    async fn delete_cache(&self, name: &str) -> Result<()> {
        self.delete_cached_content(name).await
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use adk_core::AdkError;
    use std::{
        sync::{
            Arc,
            atomic::{AtomicU32, Ordering},
        },
        time::Duration,
    };

    #[test]
    fn env_flag_truthy_accepts_one_and_case_insensitive_true() {
        for truthy in ["1", "true", "TRUE", "True", " true ", "tRuE"] {
            assert!(env_flag_truthy(truthy), "expected {truthy:?} to be truthy");
        }
        for falsy in ["0", "false", "FALSE", "yes", "on", "", "2", "vertex"] {
            assert!(!env_flag_truthy(falsy), "expected {falsy:?} to be falsy");
        }
    }

    #[test]
    fn constructor_is_backward_compatible_and_sync() {
        fn accepts_sync_constructor<F>(_f: F)
        where
            F: Fn(&str, &str) -> Result<GeminiModel>,
        {
        }

        accepts_sync_constructor(|api_key, model| GeminiModel::new(api_key, model));
    }

    #[test]
    fn stream_chunks_from_response_injects_partial_before_lone_final_chunk() {
        let response = LlmResponse {
            content: Some(Content::new("model").with_text("hello")),
            usage_metadata: None,
            finish_reason: Some(FinishReason::Stop),
            citation_metadata: None,
            partial: false,
            turn_complete: true,
            interrupted: false,
            error_code: None,
            error_message: None,
            provider_metadata: None,
            interaction_id: None,
        };

        let (chunks, saw_partial) = GeminiModel::stream_chunks_from_response(response, false);
        assert!(saw_partial);
        assert_eq!(chunks.len(), 2);
        assert!(chunks[0].partial);
        assert!(!chunks[0].turn_complete);
        assert!(chunks[0].content.is_none());
        assert!(!chunks[1].partial);
        assert!(chunks[1].turn_complete);
    }

    #[test]
    fn stream_chunks_from_response_keeps_final_only_when_partial_already_seen() {
        let response = LlmResponse {
            content: Some(Content::new("model").with_text("done")),
            usage_metadata: None,
            finish_reason: Some(FinishReason::Stop),
            citation_metadata: None,
            partial: false,
            turn_complete: true,
            interrupted: false,
            error_code: None,
            error_message: None,
            provider_metadata: None,
            interaction_id: None,
        };

        let (chunks, saw_partial) = GeminiModel::stream_chunks_from_response(response, true);
        assert!(saw_partial);
        assert_eq!(chunks.len(), 1);
        assert!(!chunks[0].partial);
        assert!(chunks[0].turn_complete);
    }

    #[test]
    fn stream_chunks_from_response_keeps_tool_call_turn_open() {
        let response = LlmResponse {
            content: Some(Content {
                role: "model".to_string(),
                parts: vec![Part::FunctionCall {
                    name: "get_weather".to_string(),
                    args: serde_json::json!({"city": "Boston"}),
                    id: Some("call-1".to_string()),
                    thought_signature: None,
                }],
            }),
            finish_reason: Some(FinishReason::Stop),
            turn_complete: true,
            ..Default::default()
        };

        let (chunks, _) = GeminiModel::stream_chunks_from_response(response, true);

        assert_eq!(chunks.len(), 1);
        assert!(!chunks[0].partial);
        assert!(!chunks[0].turn_complete);
    }

    #[tokio::test]
    async fn execute_with_retry_retries_retryable_errors() {
        let retry_config = RetryConfig::default()
            .with_max_retries(2)
            .with_initial_delay(Duration::from_millis(0))
            .with_max_delay(Duration::from_millis(0));
        let attempts = Arc::new(AtomicU32::new(0));

        let result = execute_with_retry(&retry_config, is_retryable_model_error, || {
            let attempts = Arc::clone(&attempts);
            async move {
                let attempt = attempts.fetch_add(1, Ordering::SeqCst);
                if attempt < 2 {
                    return Err(AdkError::model("code 429 RESOURCE_EXHAUSTED"));
                }
                Ok("ok")
            }
        })
        .await
        .expect("retry should eventually succeed");

        assert_eq!(result, "ok");
        assert_eq!(attempts.load(Ordering::SeqCst), 3);
    }

    #[tokio::test]
    async fn execute_with_retry_does_not_retry_non_retryable_errors() {
        let retry_config = RetryConfig::default()
            .with_max_retries(3)
            .with_initial_delay(Duration::from_millis(0))
            .with_max_delay(Duration::from_millis(0));
        let attempts = Arc::new(AtomicU32::new(0));

        let error = execute_with_retry(&retry_config, is_retryable_model_error, || {
            let attempts = Arc::clone(&attempts);
            async move {
                attempts.fetch_add(1, Ordering::SeqCst);
                Err::<(), _>(AdkError::model("code 400 invalid request"))
            }
        })
        .await
        .expect_err("non-retryable error should return immediately");

        assert!(error.is_model());
        assert_eq!(attempts.load(Ordering::SeqCst), 1);
    }

    #[tokio::test]
    async fn execute_with_retry_respects_disabled_config() {
        let retry_config = RetryConfig::disabled().with_max_retries(10);
        let attempts = Arc::new(AtomicU32::new(0));

        let error = execute_with_retry(&retry_config, is_retryable_model_error, || {
            let attempts = Arc::clone(&attempts);
            async move {
                attempts.fetch_add(1, Ordering::SeqCst);
                Err::<(), _>(AdkError::model("code 429 RESOURCE_EXHAUSTED"))
            }
        })
        .await
        .expect_err("disabled retries should return first error");

        assert!(error.is_model());
        assert_eq!(attempts.load(Ordering::SeqCst), 1);
    }

    #[test]
    fn convert_response_preserves_citation_metadata() {
        let response = adk_gemini::GenerationResponse {
            candidates: vec![adk_gemini::Candidate {
                content: adk_gemini::Content {
                    role: Some(adk_gemini::Role::Model),
                    parts: Some(vec![adk_gemini::Part::Text {
                        text: "hello world".to_string(),
                        thought: None,
                        thought_signature: None,
                    }]),
                },
                safety_ratings: None,
                citation_metadata: Some(adk_gemini::CitationMetadata {
                    citation_sources: vec![adk_gemini::CitationSource {
                        uri: Some("https://example.com".to_string()),
                        title: Some("Example".to_string()),
                        start_index: Some(0),
                        end_index: Some(5),
                        license: Some("CC-BY".to_string()),
                        publication_date: None,
                    }],
                }),
                grounding_metadata: None,
                finish_reason: Some(adk_gemini::FinishReason::Stop),
                index: Some(0),
            }],
            prompt_feedback: None,
            usage_metadata: None,
            model_version: None,
            response_id: None,
        };

        let converted =
            GeminiModel::convert_response(&response).expect("conversion should succeed");
        let metadata = converted.citation_metadata.expect("citation metadata should be mapped");
        assert_eq!(metadata.citation_sources.len(), 1);
        assert_eq!(metadata.citation_sources[0].uri.as_deref(), Some("https://example.com"));
        assert_eq!(metadata.citation_sources[0].start_index, Some(0));
        assert_eq!(metadata.citation_sources[0].end_index, Some(5));
    }

    #[test]
    fn convert_response_keeps_function_call_turn_open() {
        let response = adk_gemini::GenerationResponse {
            candidates: vec![adk_gemini::Candidate {
                content: adk_gemini::Content {
                    role: Some(adk_gemini::Role::Model),
                    parts: Some(vec![adk_gemini::Part::FunctionCall {
                        function_call: adk_gemini::FunctionCall {
                            name: "get_weather".to_string(),
                            args: serde_json::json!({"city": "Boston"}),
                            id: Some("call-1".to_string()),
                            thought_signature: None,
                        },
                        thought_signature: None,
                    }]),
                },
                safety_ratings: None,
                citation_metadata: None,
                grounding_metadata: None,
                finish_reason: Some(adk_gemini::FinishReason::Stop),
                index: Some(0),
            }],
            prompt_feedback: None,
            usage_metadata: None,
            model_version: None,
            response_id: None,
        };

        let converted =
            GeminiModel::convert_response(&response).expect("conversion should succeed");

        assert!(!converted.turn_complete);
        assert!(converted.content.is_some_and(|content| content.has_function_calls()));
    }

    #[test]
    fn convert_response_handles_inline_data_from_model() {
        let image_bytes = vec![0x89, 0x50, 0x4E, 0x47];
        let encoded = crate::attachment::encode_base64(&image_bytes);

        let response = adk_gemini::GenerationResponse {
            candidates: vec![adk_gemini::Candidate {
                content: adk_gemini::Content {
                    role: Some(adk_gemini::Role::Model),
                    parts: Some(vec![
                        adk_gemini::Part::Text {
                            text: "Here is the image".to_string(),
                            thought: None,
                            thought_signature: None,
                        },
                        adk_gemini::Part::InlineData {
                            inline_data: adk_gemini::Blob {
                                mime_type: "image/png".to_string(),
                                data: encoded,
                            },
                        },
                    ]),
                },
                safety_ratings: None,
                citation_metadata: None,
                grounding_metadata: None,
                finish_reason: Some(adk_gemini::FinishReason::Stop),
                index: Some(0),
            }],
            prompt_feedback: None,
            usage_metadata: None,
            model_version: None,
            response_id: None,
        };

        let converted =
            GeminiModel::convert_response(&response).expect("conversion should succeed");
        let content = converted.content.expect("should have content");
        assert!(
            content
                .parts
                .iter()
                .any(|part| matches!(part, Part::Text { text } if text == "Here is the image"))
        );
        assert!(content.parts.iter().any(|part| {
            matches!(
                part,
                Part::InlineData { mime_type, data, .. }
                    if mime_type == "image/png" && data.as_slice() == image_bytes.as_slice()
            )
        }));
    }

    #[test]
    fn gemini_function_response_payload_preserves_objects() {
        let value = serde_json::json!({
            "documents": [
                { "id": "pricing", "score": 0.91 }
            ]
        });

        let payload = GeminiModel::gemini_function_response_payload(value.clone());

        assert_eq!(payload, value);
    }

    #[test]
    fn gemini_function_response_payload_wraps_arrays() {
        let payload =
            GeminiModel::gemini_function_response_payload(serde_json::json!([{ "id": "pricing" }]));

        assert_eq!(payload, serde_json::json!({ "result": [{ "id": "pricing" }] }));
    }

    // ===== Multimodal function response conversion tests =====

    /// Helper to build a FunctionResponse with nested multimodal parts
    /// simulating the conversion logic from generate_content_internal.
    fn convert_function_response_to_gemini_fr(
        frd: &adk_core::FunctionResponseData,
    ) -> adk_gemini::tools::FunctionResponse {
        let mut fr_parts = Vec::new();

        for inline in &frd.inline_data {
            let encoded = crate::attachment::encode_base64(&inline.data);
            fr_parts.push(adk_gemini::FunctionResponsePart::InlineData {
                inline_data: adk_gemini::Blob {
                    mime_type: inline.mime_type.clone(),
                    data: encoded,
                },
            });
        }

        for file in &frd.file_data {
            fr_parts.push(adk_gemini::FunctionResponsePart::FileData {
                file_data: adk_gemini::FileDataRef {
                    mime_type: file.mime_type.clone(),
                    file_uri: file.file_uri.clone(),
                },
            });
        }

        let mut gemini_fr = adk_gemini::tools::FunctionResponse::new(
            &frd.name,
            GeminiModel::gemini_function_response_payload(frd.response.clone()),
        );
        gemini_fr.parts = fr_parts;
        gemini_fr
    }

    #[test]
    fn json_only_function_response_has_no_nested_parts() {
        let frd = adk_core::FunctionResponseData::new("tool", serde_json::json!({"ok": true}));
        let gemini_fr = convert_function_response_to_gemini_fr(&frd);
        assert!(gemini_fr.parts.is_empty());
        // Serialized JSON should have name and response but no parts key
        let json = serde_json::to_string(&gemini_fr).unwrap();
        assert!(!json.contains("\"parts\""));
    }

    #[test]
    fn function_response_with_inline_data_has_nested_parts() {
        let frd = adk_core::FunctionResponseData::with_inline_data(
            "chart",
            serde_json::json!({"status": "ok"}),
            vec![adk_core::InlineDataPart {
                mime_type: "image/png".to_string(),
                data: vec![0x89, 0x50, 0x4E, 0x47],
                uri: None,
                annotations: None,
            }],
        );
        let gemini_fr = convert_function_response_to_gemini_fr(&frd);
        assert_eq!(gemini_fr.parts.len(), 1);
        match &gemini_fr.parts[0] {
            adk_gemini::FunctionResponsePart::InlineData { inline_data } => {
                assert_eq!(inline_data.mime_type, "image/png");
                let decoded = BASE64_STANDARD.decode(&inline_data.data).unwrap();
                assert_eq!(decoded, vec![0x89, 0x50, 0x4E, 0x47]);
            }
            other => panic!("expected InlineData, got {other:?}"),
        }
    }

    #[test]
    fn function_response_with_file_data_has_nested_parts() {
        let frd = adk_core::FunctionResponseData::with_file_data(
            "doc",
            serde_json::json!({"ok": true}),
            vec![adk_core::FileDataPart {
                mime_type: "application/pdf".to_string(),
                file_uri: "gs://bucket/report.pdf".to_string(),
                annotations: None,
            }],
        );
        let gemini_fr = convert_function_response_to_gemini_fr(&frd);
        assert_eq!(gemini_fr.parts.len(), 1);
        match &gemini_fr.parts[0] {
            adk_gemini::FunctionResponsePart::FileData { file_data } => {
                assert_eq!(file_data.mime_type, "application/pdf");
                assert_eq!(file_data.file_uri, "gs://bucket/report.pdf");
            }
            other => panic!("expected FileData, got {other:?}"),
        }
    }

    #[test]
    fn function_response_with_both_inline_and_file_data_ordering() {
        let frd = adk_core::FunctionResponseData::with_multimodal(
            "multi",
            serde_json::json!({}),
            vec![
                adk_core::InlineDataPart {
                    mime_type: "image/png".to_string(),
                    data: vec![1, 2],
                    uri: None,
                    annotations: None,
                },
                adk_core::InlineDataPart {
                    mime_type: "image/jpeg".to_string(),
                    data: vec![3, 4],
                    uri: None,
                    annotations: None,
                },
            ],
            vec![adk_core::FileDataPart {
                mime_type: "application/pdf".to_string(),
                file_uri: "gs://b/f.pdf".to_string(),
                annotations: None,
            }],
        );
        let gemini_fr = convert_function_response_to_gemini_fr(&frd);
        // 2 inline + 1 file = 3 nested parts
        assert_eq!(gemini_fr.parts.len(), 3);
        assert!(matches!(&gemini_fr.parts[0], adk_gemini::FunctionResponsePart::InlineData { .. }));
        assert!(matches!(&gemini_fr.parts[1], adk_gemini::FunctionResponsePart::InlineData { .. }));
        assert!(matches!(&gemini_fr.parts[2], adk_gemini::FunctionResponsePart::FileData { .. }));
    }
}

#[cfg(all(test, feature = "gemini-interactions"))]
mod interactions_transport_tests {
    use super::*;
    use adk_core::GenerateContentConfig;

    /// **Feature: gemini-interactions-runtime, Property 2: Default options match the API**
    /// *For any* default `InteractionOptions`, `store == true`, `stateful == true`,
    /// `background` is `AgentTargetsOnly`, and `poll_interval` is 1 second.
    /// **Validates: Requirements 3.1, 3.2, 3.3**
    #[test]
    fn default_interaction_options_match_api_posture() {
        let opts = InteractionOptions::default();
        assert!(opts.store, "store should default to true");
        assert!(opts.stateful, "stateful should default to true");
        assert_eq!(opts.background, BackgroundMode::AgentTargetsOnly);
        assert_eq!(opts.poll_interval, std::time::Duration::from_secs(1));
    }

    #[test]
    fn new_model_defaults_to_generate_content_transport() {
        let model = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("constructing a Gemini model should not require network");
        assert_eq!(model.transport(), GeminiTransport::GenerateContent);
    }

    #[test]
    fn use_interactions_api_enables_transport_for_allowlisted_model() {
        let model = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("construct model")
            .use_interactions_api(true)
            .expect("allowlisted model should enable the Interactions transport");

        assert_eq!(model.transport(), GeminiTransport::Interactions);
        assert_eq!(
            model.interaction_target,
            Some(InteractionTarget::Model("gemini-2.5-flash".to_string()))
        );
    }

    #[test]
    fn use_interactions_api_rejects_unsupported_model_with_invalid_input() {
        let result = GeminiModel::new("test-key", "gemini-2.0-flash")
            .expect("construct model")
            .use_interactions_api(true);

        let err = match result {
            Ok(_) => panic!("unsupported model should be rejected"),
            Err(err) => err,
        };
        assert_eq!(err.category, adk_core::ErrorCategory::InvalidInput);
        assert_eq!(err.details.provider.as_deref(), Some("gemini"));
    }

    #[test]
    fn use_interactions_api_false_reverts_to_generate_content() {
        let model = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("construct model")
            .use_interactions_api(true)
            .expect("enable interactions")
            .use_interactions_api(false)
            .expect("disabling should always succeed");

        assert_eq!(model.transport(), GeminiTransport::GenerateContent);
        assert_eq!(model.interaction_target, None);
    }

    #[test]
    fn interaction_options_override_is_stored() {
        let opts = InteractionOptions {
            store: false,
            stateful: false,
            background: BackgroundMode::Always,
            poll_interval: std::time::Duration::from_millis(250),
        };
        let model = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("construct model")
            .interaction_options(opts.clone());

        let stored = model.interaction_options_ref();
        assert_eq!(stored.store, opts.store);
        assert_eq!(stored.stateful, opts.stateful);
        assert_eq!(stored.background, opts.background);
        assert_eq!(stored.poll_interval, opts.poll_interval);
    }

    #[test]
    fn resolve_background_honors_background_mode() {
        // Always → true regardless of target.
        let always = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("construct model")
            .use_interactions_api(true)
            .expect("enable")
            .interaction_options(InteractionOptions {
                background: BackgroundMode::Always,
                ..InteractionOptions::default()
            });
        assert!(always.resolve_background());

        // Never → false regardless of target.
        let never = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("construct model")
            .use_interactions_api(true)
            .expect("enable")
            .interaction_options(InteractionOptions {
                background: BackgroundMode::Never,
                ..InteractionOptions::default()
            });
        assert!(!never.resolve_background());

        // AgentTargetsOnly → false for a model target.
        let model_target = GeminiModel::new("test-key", "gemini-2.5-flash")
            .expect("construct model")
            .use_interactions_api(true)
            .expect("enable");
        assert!(!model_target.resolve_background());

        // AgentTargetsOnly → true for an agent target.
        let agent_target = GeminiModel::new("test-key", "deep-research-preview-04-2026")
            .expect("construct model")
            .use_interactions_api(true)
            .expect("enable");
        assert!(agent_target.resolve_background());
    }

    /// Requirements 7.7 / 9.2: a terminal `failed` status maps to an
    /// `Internal` `AdkError` (provider `"gemini"`).
    #[test]
    fn terminal_failed_status_maps_to_error() {
        use adk_gemini::interactions::InteractionStatus;

        let err = interaction_status_to_result(InteractionStatus::Failed)
            .expect_err("a failed interaction must surface an error");
        assert_eq!(err.category, adk_core::ErrorCategory::Internal);
        assert_eq!(err.details.provider.as_deref(), Some("gemini"));
        assert_eq!(err.code, "model.gemini.interactions.failed");
    }

    /// Requirements 7.7 / 9.2: a terminal `budget_exceeded` status maps to an
    /// `Internal` `AdkError` (provider `"gemini"`).
    #[test]
    fn terminal_budget_exceeded_status_maps_to_error() {
        use adk_gemini::interactions::InteractionStatus;

        let err = interaction_status_to_result(InteractionStatus::BudgetExceeded)
            .expect_err("a budget_exceeded interaction must surface an error");
        assert_eq!(err.category, adk_core::ErrorCategory::Internal);
        assert_eq!(err.details.provider.as_deref(), Some("gemini"));
        assert_eq!(err.code, "model.gemini.interactions.budget_exceeded");
    }

    /// Non-failure statuses (including `requires_action` and the other terminal
    /// states the transport reads content from) do not produce an error.
    #[test]
    fn non_failure_statuses_are_ok() {
        use adk_gemini::interactions::InteractionStatus;

        for status in [
            InteractionStatus::InProgress,
            InteractionStatus::RequiresAction,
            InteractionStatus::Completed,
            InteractionStatus::Cancelled,
            InteractionStatus::Incomplete,
        ] {
            assert!(
                interaction_status_to_result(status).is_ok(),
                "status {status:?} should not map to an error"
            );
        }
    }

    /// A constructed `failed` `Interaction` flows through the same status check
    /// the transport uses, surfacing an error after conversion.
    #[test]
    fn failed_interaction_resource_surfaces_error() {
        use adk_gemini::interactions::{Interaction, InteractionStatus};

        let interaction = Interaction {
            id: "v1_failed".to_string(),
            model: Some("gemini-2.5-flash".to_string()),
            agent: None,
            status: InteractionStatus::Failed,
            steps: Vec::new(),
            usage: None,
            created: None,
            updated: None,
            environment_id: None,
        };

        let err = interaction_status_to_result(interaction.status)
            .expect_err("failed interaction must surface an error");
        assert_eq!(err.category, adk_core::ErrorCategory::Internal);
        assert_eq!(err.details.provider.as_deref(), Some("gemini"));
    }
    #[test]
    fn gemini_37_rejects_sampling_before_network_io() {
        let model = GeminiModel::new("test-key", "gemini-3.7-flash").expect("construct model");
        let mut request = LlmRequest::new("gemini-3.7-flash", Vec::new());
        request.config =
            Some(GenerateContentConfig { temperature: Some(0.2), ..Default::default() });

        let error = model
            .validate_request_contract(&request)
            .expect_err("sampling must be rejected locally");
        assert_eq!(error.code, "model.gemini.sampling_unsupported");
    }

    #[test]
    fn gemini_37_rejects_token_budget_thinking() {
        let model = GeminiModel::new("test-key", "gemini-3.7-flash")
            .expect("construct model")
            .with_thinking_config(adk_gemini::ThinkingConfig {
                thinking_budget: Some(1024),
                include_thoughts: None,
                thinking_level: None,
            });

        let error = model
            .validate_request_contract(&LlmRequest::new("gemini-3.7-flash", Vec::new()))
            .expect_err("token budgets must be rejected locally");
        assert_eq!(error.code, "model.gemini.thinking_budget_unsupported");
    }
}

#[cfg(all(test, feature = "gemini-vertex"))]
mod vertex_rag_tests {
    use super::*;
    use adk_gemini::VertexRagStore;

    fn test_store() -> VertexRagStore {
        VertexRagStore::corpus("projects/p/locations/us-central1/ragCorpora/123").with_top_k(5)
    }

    fn vertex_model() -> GeminiModel {
        GeminiModel::new_google_cloud("test-key", "p", "us-central1", "gemini-3.7-flash")
            .expect("construct vertex model")
    }

    // The Vertex constructor builds credential plumbing that requires a
    // current Tokio runtime, so this test is async.
    #[tokio::test]
    async fn vertex_backend_accepts_and_stores_the_rag_store() {
        let model = vertex_model().with_vertex_rag_store(test_store());

        model
            .validate_request_contract(&LlmRequest::new("gemini-3.7-flash", Vec::new()))
            .expect("vertex backend must accept the rag store");

        // The stored store is handed to adk-gemini's ContentBuilder per
        // request; the exact wire-shape injection is asserted in adk-gemini's
        // backend tests.
        assert_eq!(model.vertex_rag_store(), Some(&test_store()));
    }

    #[test]
    fn studio_backend_rejects_rag_store_before_network_io() {
        let model = GeminiModel::new("test-key", "gemini-3.7-flash")
            .expect("construct studio model")
            .with_vertex_rag_store(test_store());

        let error = model
            .validate_request_contract(&LlmRequest::new("gemini-3.7-flash", Vec::new()))
            .expect_err("studio backend must reject the rag store locally");
        assert_eq!(error.code, "model.gemini.rag_store_requires_vertex");
        assert_eq!(error.category, adk_core::ErrorCategory::Unsupported);
        assert_eq!(error.details.provider.as_deref(), Some("gemini"));
    }

    #[test]
    fn no_rag_store_means_no_gate() {
        let model = GeminiModel::new("test-key", "gemini-3.7-flash").expect("construct model");
        assert_eq!(model.vertex_rag_store(), None);
        model
            .validate_request_contract(&LlmRequest::new("gemini-3.7-flash", Vec::new()))
            .expect("no store configured, nothing to reject");
    }
}

#[cfg(all(test, feature = "gemini-vertex"))]
mod vertex_transport_tests {
    use super::*;

    #[test]
    fn vertex_transport_failures_are_retryable() {
        let error = adk_gemini::ClientError::Io {
            source: std::io::Error::other(
                "the transport reports an error: client error (SendRequest): http2 error",
            ),
        };
        let mapped = gemini_error_to_adk(&error);
        assert_eq!(
            (mapped.category, mapped.code, mapped.is_retryable()),
            (ErrorCategory::Unavailable, "model.gemini.unavailable", true)
        );

        let denied =
            adk_gemini::ClientError::Io { source: std::io::Error::other("permission denied") };
        assert_eq!(gemini_error_to_adk(&denied).category, ErrorCategory::Internal);
    }
}