potato_type/prompt/

interface.rs

use crate::anthropic::v1::request::{AnthropicSettings, MessageParam as AnthropicMessage};
use crate::error::TypeError;
use crate::google::v1::generate::request::{GeminiContent, GeminiSettings};
use crate::openai::v1::chat::request::ChatMessage as OpenAIChatMessage;
use crate::openai::v1::chat::settings::OpenAIChatSettings;
use crate::prompt::builder::{to_provider_request, ProviderRequest};
use crate::prompt::settings::ModelSettings;
use crate::prompt::types::parse_response_to_json;
use crate::prompt::types::ResponseType;
use crate::prompt::types::Role;
use crate::prompt::{AnthropicMessageList, GeminiContentList, MessageNum, OpenAIMessageList};
use crate::tools::AgentToolDefinition;
use crate::traits::MessageFactory;
use crate::SettingsType;
use crate::{Provider, SaveName};
use potato_util::utils::extract_string_value;
use potato_util::PyHelperFuncs;
use potatohead_macro::try_extract_message;
use pyo3::prelude::*;
use pyo3::types::{PyDict, PyList, PyString, PyTuple};
use pythonize::pythonize;
use serde::{Deserialize, Deserializer, Serialize};
use serde_json::Value;
use std::collections::BTreeSet;
use std::path::{Path, PathBuf};

/// Deserializes `messages` from either a single string or a list of strings.
/// This allows YAML block scalars (`|`) to be used for multi-line prompts.
fn deserialize_string_or_vec<'de, D>(deserializer: D) -> Result<Vec<String>, D::Error>
where
    D: Deserializer<'de>,
{
    #[derive(Deserialize)]
    #[serde(untagged)]
    enum StringOrVec {
        Single(String),
        List(Vec<String>),
    }

    match StringOrVec::deserialize(deserializer)? {
        StringOrVec::Single(s) => Ok(vec![s]),
        StringOrVec::List(v) => Ok(v),
    }
}

/// Generic prompt configuration structure for user-friendly YAML/JSON format.
///
/// Example:
/// ```yaml
/// model: gemini-1.5-pro
/// provider: Google         # optional; falls back to POTATO_HEAD_DEFAULT_PROVIDER
/// messages:
///   - "Hello ${variable1}"
/// ```
/// `messages` also accepts a single block-scalar string (YAML `|`).
#[derive(Debug, Deserialize)]
pub struct GenericPromptConfig {
    model: String,
    #[serde(default)]
    provider: Option<String>,
    #[serde(deserialize_with = "deserialize_string_or_vec")]
    messages: Vec<String>,
    #[serde(default)]
    system_instructions: Option<Vec<String>>,
    #[serde(default)]
    settings: Option<Value>,
    response_format: Option<Value>,
}

const PROMPT_FILE_EXTENSIONS: [&str; 3] = ["yaml", "yml", "json"];

fn push_attempted_path(attempted_paths: &mut Vec<PathBuf>, path: PathBuf) {
    if !attempted_paths.iter().any(|existing| existing == &path) {
        attempted_paths.push(path);
    }
}

fn format_candidate_paths(paths: &[PathBuf]) -> String {
    paths
        .iter()
        .map(|path| path.display().to_string())
        .collect::<Vec<_>>()
        .join(", ")
}

fn create_message_for_provider(
    content: String,
    provider: &Provider,
    role: &str,
) -> Result<MessageNum, TypeError> {
    match provider {
        Provider::OpenAI => {
            OpenAIChatMessage::from_text(content, role).map(MessageNum::OpenAIMessageV1)
        }
        Provider::Anthropic => {
            AnthropicMessage::from_text(content, role).map(MessageNum::AnthropicMessageV1)
        }
        Provider::Gemini | Provider::Google | Provider::Vertex | Provider::GoogleAdk => {
            GeminiContent::from_text(content, role).map(MessageNum::GeminiContentV1)
        }
        _ => Err(TypeError::Error(format!(
            "Unsupported provider for message creation: {:?}",
            provider
        ))),
    }
}

fn parse_single_message(
    message: &Bound<'_, PyAny>,
    provider: &Provider,
    default_role: &str,
) -> Result<MessageNum, TypeError> {
    // String conversion (most common case)
    if message.is_instance_of::<PyString>() {
        let text = message.extract::<String>()?;
        return create_message_for_provider(text, provider, default_role);
    }

    // Try each message type using macro
    try_extract_message!(
        message,
        OpenAIChatMessage => MessageNum::OpenAIMessageV1,
        AnthropicMessage => MessageNum::AnthropicMessageV1,
        GeminiContent => MessageNum::GeminiContentV1,
    );

    Err(TypeError::InvalidMessageTypeInList(
        message.get_type().name()?.to_string(),
    ))
}

fn parse_messages(
    messages: &Bound<'_, PyAny>,
    provider: &Provider,
    default_role: &str,
) -> Result<Vec<MessageNum>, TypeError> {
    // Single message
    let mut messages =
        if !messages.is_instance_of::<PyList>() && !messages.is_instance_of::<PyTuple>() {
            vec![parse_single_message(messages, provider, default_role)?]
        } else {
            // List/tuple of messages
            messages
                .try_iter()?
                .map(|item| {
                    let item = item?;
                    parse_single_message(&item, provider, default_role)
                })
                .collect::<Result<Vec<_>, _>>()?
        };

    // Convert Anthropic system messages to TextBlockParam format
    // optimize this later - maybe
    if provider == &Provider::Anthropic
        && (default_role == Role::System.as_str()
            || default_role == Role::Assistant.as_str()
            || default_role == Role::Developer.as_str())
    {
        for msg in messages.iter_mut() {
            msg.anthropic_message_to_system_message()?;
        }
    }

    Ok(messages)
}

fn get_system_role(provider: &Provider) -> &'static str {
    match provider {
        Provider::OpenAI => Role::Developer.into(),
        Provider::Gemini | Provider::Vertex | Provider::Google | Provider::GoogleAdk => {
            Role::Model.into()
        }
        Provider::Anthropic => Role::System.into(),
        _ => Role::System.into(),
    }
}

/// Create a single system `MessageNum` for the given provider from a plain Rust string.
/// Useful for `AgentBuilder::system_prompt()` which runs in pure-Rust (no Python context).
pub fn create_system_message_for_provider(
    content: String,
    provider: &Provider,
) -> Result<MessageNum, TypeError> {
    let role = get_system_role(provider);
    let mut msg = create_message_for_provider(content, provider, role)?;
    // Anthropic system messages must be TextBlockParam
    if provider == &Provider::Anthropic {
        msg.anthropic_message_to_system_message()?;
    }
    Ok(msg)
}

/// Helper for extracting system instructions from optional parameter
pub fn extract_system_instructions(
    system_instruction: Option<&Bound<'_, PyAny>>,
    provider: &Provider,
) -> Result<Option<Vec<MessageNum>>, TypeError> {
    let system_instructions = if let Some(sys_inst) = system_instruction {
        Some(parse_messages(
            sys_inst,
            provider,
            get_system_role(provider),
        )?)
    } else {
        None
    };

    Ok(system_instructions)
}

#[pyclass(from_py_object)]
#[derive(Debug, Serialize, Clone, PartialEq)]
pub struct Prompt {
    pub request: ProviderRequest,

    #[pyo3(get)]
    pub model: String,

    #[pyo3(get)]
    pub provider: Provider,

    pub version: String,

    #[pyo3(get)]
    #[serde(default)]
    pub parameters: Vec<String>,

    #[pyo3(get)]
    #[serde(default)]
    pub media_parameters: Vec<String>,

    #[serde(default)]
    pub response_type: ResponseType,
}

/// ModelSettings variant based on the type of settings provided.
fn extract_model_settings(model_settings: &Bound<'_, PyAny>) -> Result<ModelSettings, TypeError> {
    let settings_type = model_settings
        .call_method0("settings_type")?
        .extract::<SettingsType>()?;

    match settings_type {
        SettingsType::OpenAIChat => model_settings
            .extract::<OpenAIChatSettings>()
            .map(ModelSettings::OpenAIChat),
        SettingsType::GoogleChat => model_settings
            .extract::<GeminiSettings>()
            .map(ModelSettings::GoogleChat),
        SettingsType::Anthropic => model_settings
            .extract::<AnthropicSettings>()
            .map(ModelSettings::AnthropicChat),
        SettingsType::ModelSettings => model_settings.extract::<ModelSettings>(),
    }
    .map_err(Into::into)
}

#[derive(Debug, Deserialize)]
#[serde(untagged)]
enum PromptFormat {
    Generic(GenericPromptConfig),
    Full(Box<PromptInternal>),
}

#[derive(Debug, Deserialize)]
struct PromptInternal {
    request: Value,
    model: String,
    provider: Provider,
    version: String,
    #[serde(default)]
    parameters: Vec<String>,
    #[serde(default)]
    media_parameters: Vec<String>,
    #[serde(default)]
    response_type: ResponseType,
}

impl<'de> Deserialize<'de> for Prompt {
    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
    where
        D: serde::Deserializer<'de>,
    {
        let format = PromptFormat::deserialize(deserializer)?;

        match format {
            PromptFormat::Generic(config) => Self::from_generic_config(config)
                .map_err(|e| serde::de::Error::custom(e.to_string())),
            PromptFormat::Full(internal) => Ok(Prompt {
                request: provider_request_from_value(internal.provider.clone(), internal.request)
                    .map_err(|e| serde::de::Error::custom(e.to_string()))?,
                model: internal.model,
                provider: internal.provider,
                version: internal.version,
                parameters: internal.parameters,
                media_parameters: internal.media_parameters,
                response_type: internal.response_type,
            }),
        }
    }
}

fn provider_request_from_value(
    provider: Provider,
    value: Value,
) -> Result<ProviderRequest, TypeError> {
    match provider {
        Provider::OpenAI => Ok(ProviderRequest::OpenAIV1(serde_json::from_value(value)?)),
        Provider::Anthropic => Ok(ProviderRequest::AnthropicV1(serde_json::from_value(value)?)),
        Provider::Gemini | Provider::Google | Provider::Vertex | Provider::GoogleAdk => {
            Ok(ProviderRequest::GeminiV1(serde_json::from_value(value)?))
        }
        Provider::Undefined => Err(TypeError::UnsupportedProviderForRequestCreation),
    }
}

#[pymethods]
impl Prompt {
    /// Creates a new Prompt object.
    /// Main parsing logic is as follows:
    /// 1. Extract model settings if provided, otherwise use provider default settings.
    /// 2. Message and system instructions are expected to be a variant of MessageNum (OpenAIChatMessage, AnthropicMessage or GeminiContent).
    /// 3. On instantiation, message will be check if is_instance_of pystring. If pystring, provider will be used to map to appropriate message Text type
    /// 4. If message is a pylist, each item will be checked for is_instance_of pystring or MessageNum variant and converted accordingly.
    /// 5. If message is a single MessageNum variant, it will be extracted and wrapped in a vec.
    /// 6. After messages are parsed, a full provider request struct will by built using to_provider_request function.
    /// # Arguments:
    /// * `message`: A single message or list of messages representing user input.
    /// * `model`: The model identifier to use for the prompt.
    /// * `provider`: Optional. Defaults to env var POTATO_HEAD_DEFAULT_PROVIDER if unset.
    /// * `system_instruction`: Optional system instruction message or list of messages.
    /// * `model_settings`: Optional model settings to use for the prompt.
    /// * `output_type`: Optional output type to enforce structured output.
    #[new]
    #[pyo3(signature = (messages, model, provider=None, system_instructions=None, model_settings=None, output_type=None))]
    pub fn new(
        py: Python<'_>,
        messages: &Bound<'_, PyAny>,
        model: &str,
        provider: Option<&Bound<'_, PyAny>>,
        system_instructions: Option<&Bound<'_, PyAny>>,
        model_settings: Option<&Bound<'_, PyAny>>,
        output_type: Option<&Bound<'_, PyAny>>, // can be a pydantic model or one of Opsml's predefined outputs
    ) -> Result<Self, TypeError> {
        // 1. get model settings if provided
        let model_settings = model_settings
            .as_ref()
            .map(|s| extract_model_settings(s))
            .transpose()?;

        // 2. extract provider
        let provider = Provider::resolve_from_py(provider)?;

        // 3. Parse user messages with "user" role
        // We'll use this to figure out the type of request struct to create
        let messages = parse_messages(messages, &provider, Role::User.into())?;
        let system_instructions = if let Some(sys_inst) = system_instructions {
            parse_messages(sys_inst, &provider, get_system_role(&provider))?
        } else {
            vec![]
        };

        // 4.  validate response_json_schema
        let (response_type, response_json_schema) = match output_type {
            Some(output_type) => {
                // check if output_type is a pydantic model and extract the model json schema
                parse_response_to_json(py, output_type)?
            }
            None => (ResponseType::Null, None),
        };

        Self::new_rs(
            messages,
            model,
            provider,
            system_instructions,
            model_settings,
            response_json_schema,
            response_type,
        )
    }

    #[getter]
    pub fn model_settings<'py>(&self, py: Python<'py>) -> Result<Bound<'py, PyAny>, TypeError> {
        self.request.model_settings(py)
    }

    #[getter]
    pub fn model_identifier(&self) -> String {
        format!("{}:{}", self.provider.as_str(), self.model)
    }

    #[pyo3(signature = (path = None))]
    pub fn save_prompt(&self, path: Option<PathBuf>) -> Result<PathBuf, TypeError> {
        let save_path = path.unwrap_or_else(|| PathBuf::from(SaveName::Prompt));
        PyHelperFuncs::save_to_json(self, &save_path)?;
        Ok(save_path.with_extension("json"))
    }

    #[staticmethod]
    pub fn from_path(path: PathBuf) -> Result<Self, TypeError> {
        Self::load_prompt_from_path(path.as_path(), None)
    }

    #[staticmethod]
    pub fn model_validate_json(json_string: String) -> Result<Self, TypeError> {
        let json_value: Value = serde_json::from_str(&json_string)?;
        let model: Self = serde_json::from_value(json_value)?;

        Ok(model)
    }

    pub fn model_dump_json(&self) -> String {
        serde_json::to_string(self).unwrap()
    }

    pub fn __str__(&self) -> String {
        PyHelperFuncs::__str__(self)
    }

    #[getter]
    /// Returns all messages as Python objects, including system instructions, user messages, and assistant messages.
    pub fn all_messages<'py>(&self, py: Python<'py>) -> Result<Bound<'py, PyList>, TypeError> {
        self.request.get_all_py_messages(py)
    }

    #[getter]
    /// Returns User messages as Python objects. This means, system instructions are excluded.
    pub fn messages<'py>(&self, py: Python<'py>) -> Result<Bound<'py, PyList>, TypeError> {
        self.request.get_py_messages(py)
    }

    #[getter]
    /// Returns the last User message as a Python object. This means, system instructions are excluded.
    pub fn message<'py>(&self, py: Python<'py>) -> Result<Bound<'py, PyAny>, TypeError> {
        self.request.get_py_message(py)
    }

    #[getter]
    /// Returns the messages as OpenAI ChatMessage Python objects
    /// This is a helper that provide strict typing when working with OpenAI prompts
    pub fn openai_messages(&self) -> Result<OpenAIMessageList, TypeError> {
        if self.provider != Provider::OpenAI {
            return Err(TypeError::Error(
                "Prompt provider is not OpenAI".to_string(),
            ));
        }
        let messages = self
            .request
            .messages()
            .iter()
            .filter(|msg| msg.is_user_message())
            .filter_map(|msg| match msg {
                MessageNum::OpenAIMessageV1(m) => Some(m.clone()),
                _ => None,
            })
            .collect::<Vec<_>>();
        Ok(OpenAIMessageList { messages })
    }

    #[getter]
    /// Returns the last message as an OpenAI ChatMessage Python object
    /// This is a helper that provide strict typing when working with OpenAI prompts
    pub fn openai_message(&self) -> Result<OpenAIChatMessage, TypeError> {
        if self.provider != Provider::OpenAI {
            return Err(TypeError::Error(
                "Prompt provider is not OpenAI".to_string(),
            ));
        }
        self.request.get_openai_message()
    }

    #[getter]
    /// Returns the messages as Google GeminiContent Python objects
    /// This is a helper that provide strict typing when working with Google/Gemini/Vertex prompts
    pub fn gemini_messages(&self) -> Result<GeminiContentList, TypeError> {
        if !self.is_google_provider() {
            return Err(TypeError::Error(
                "Prompt provider is not Google, Gemini, or Vertex".to_string(),
            ));
        }
        let messages = self
            .request
            .messages()
            .iter()
            .filter(|msg| msg.is_user_message())
            .filter_map(|msg| match msg {
                MessageNum::GeminiContentV1(m) => Some(m.clone()),
                _ => None,
            })
            .collect::<Vec<_>>();

        Ok(GeminiContentList { messages })
    }

    #[getter]
    /// Returns the last message as a Google GeminiContent Python object
    /// This is a helper that provide strict typing when working with Google/Gemini/Vertex prompts
    pub fn gemini_message(&self) -> Result<GeminiContent, TypeError> {
        if !self.is_google_provider() {
            return Err(TypeError::Error(
                "Prompt provider is not Google, Gemini, or Vertex".to_string(),
            ));
        }
        self.request.get_gemini_message()
    }

    #[getter]
    pub fn anthropic_messages(&self) -> Result<AnthropicMessageList, TypeError> {
        if self.provider != Provider::Anthropic {
            return Err(TypeError::Error(
                "Prompt provider is not Anthropic".to_string(),
            ));
        }
        let messages = self
            .request
            .messages()
            .iter()
            .filter(|msg| msg.is_user_message())
            .filter_map(|msg| match msg {
                MessageNum::AnthropicMessageV1(m) => Some(m.clone()),
                _ => None,
            })
            .collect::<Vec<_>>();

        Ok(AnthropicMessageList { messages })
    }

    #[getter]
    /// Returns the last message as an Anthropic MessageParam Python object
    pub fn anthropic_message(&self) -> Result<AnthropicMessage, TypeError> {
        if self.provider != Provider::Anthropic {
            return Err(TypeError::Error(
                "Prompt provider is not Anthropic".to_string(),
            ));
        }
        self.request.get_anthropic_message()
    }

    #[getter]
    pub fn system_instructions<'py>(
        &self,
        py: Python<'py>,
    ) -> Result<Bound<'py, PyList>, TypeError> {
        self.request.get_py_system_instructions(py)
    }

    /// Binds a variable in the prompt to a value. This will return a new Prompt with the variable bound to the value.
    /// This will iterate over all user messages and bind the variable in each message.
    /// # Arguments:
    /// * `name`: The name of the variable to bind.
    /// * `value`: The value to bind the variable to.
    /// # Returns:
    /// * `Result<Self, PromptError>`: Returns a new Prompt with the variable bound to the value.
    #[pyo3(signature = (name=None, value=None, **kwargs))]
    pub fn bind(
        &self,
        name: Option<&str>,
        value: Option<&Bound<'_, PyAny>>,
        kwargs: Option<&Bound<'_, PyDict>>,
    ) -> Result<Self, TypeError> {
        let mut new_prompt = self.clone();

        if let (Some(name), Some(value)) = (name, value) {
            let var_value = extract_string_value(value)?;
            for message in new_prompt.request.messages_mut() {
                message.bind_mut(name, &var_value)?;
            }
        }

        if let Some(kwargs) = kwargs {
            for (key, val) in kwargs.iter() {
                let var_name = key.extract::<String>()?;
                let var_value = extract_string_value(&val)?;

                for message in new_prompt.request.messages_mut() {
                    message.bind_mut(&var_name, &var_value)?;
                }
            }
        }

        if name.is_none() && kwargs.is_none_or(|k| k.is_empty()) {
            return Err(TypeError::Error(
                "Must provide either (name, value) or keyword arguments for binding".to_string(),
            ));
        }

        Ok(new_prompt)
    }

    /// Binds a variable in the prompt to a value. This will mutate the current Prompt and bind the variable in each user message.
    /// # Arguments:
    /// * `name`: The name of the variable to bind.
    /// * `value`: The value to bind the variable to.
    /// # Returns:
    /// * `Result<(), PromptError>`: Returns Ok(()) on success or an error if the binding fails.
    #[pyo3(signature = (name=None, value=None, **kwargs))]
    pub fn bind_mut(
        &mut self,
        name: Option<&str>,
        value: Option<&Bound<'_, PyAny>>,
        kwargs: Option<&Bound<'_, PyDict>>,
    ) -> Result<(), TypeError> {
        if let (Some(name), Some(value)) = (name, value) {
            let var_value = extract_string_value(value)?;
            for message in self.request.messages_mut() {
                message.bind_mut(name, &var_value)?;
            }
        }

        if let Some(kwargs) = kwargs {
            for (key, val) in kwargs.iter() {
                let var_name = key.extract::<String>()?;
                let var_value = extract_string_value(&val)?;

                for message in self.request.messages_mut() {
                    message.bind_mut(&var_name, &var_value)?;
                }
            }
        }

        if name.is_none() && kwargs.is_none_or(|k| k.is_empty()) {
            return Err(TypeError::Error(
                "Must provide either (name, value) or keyword arguments for binding".to_string(),
            ));
        }

        Ok(())
    }

    pub fn bind_media(
        &self,
        name: &str,
        media: &crate::prompt::media::MediaRef,
    ) -> Result<Self, TypeError> {
        let mut new_prompt = self.clone();
        new_prompt.bind_media_mut(name, media)?;
        Ok(new_prompt)
    }

    pub fn bind_media_mut(
        &mut self,
        name: &str,
        media: &crate::prompt::media::MediaRef,
    ) -> Result<(), TypeError> {
        let token = format!("${{media:{name}}}");
        let mut found = false;
        for message in self.request.messages_mut() {
            if message.bind_media_mut(&token, media, &self.provider)? {
                found = true;
            }
        }
        if !found {
            return Err(TypeError::MediaPlaceholderNotFound {
                name: name.to_string(),
            });
        }
        Ok(())
    }

    #[getter]
    pub fn response_json_schema_pretty(&self) -> Option<String> {
        Some(PyHelperFuncs::__str__(
            self.request.response_json_schema().as_ref()?,
        ))
    }

    #[getter]
    #[pyo3(name = "response_json_schema")]
    pub fn response_json_schema_py(&self) -> Option<String> {
        Some(self.request.response_json_schema().as_ref()?.to_string())
    }

    pub fn model_dump<'py>(&self, py: Python<'py>) -> Result<Bound<'py, PyAny>, TypeError> {
        let request = &self.request.to_json()?;
        Ok(pythonize(py, request)?)
    }
}

impl Prompt {
    fn read_prompt_file(path: &Path) -> Result<Self, TypeError> {
        let content = std::fs::read_to_string(path)?;

        let extension = path
            .extension()
            .and_then(|ext| ext.to_str())
            .ok_or_else(|| TypeError::Error(format!("Invalid file path: {:?}", path)))?;

        let mut prompt: Prompt = match extension.to_lowercase().as_str() {
            "json" => serde_json::from_str(&content)?,
            "yaml" | "yml" => serde_yaml::from_str(&content)?,
            _ => {
                return Err(TypeError::Error(format!(
                    "Unsupported file extension '{}'. Expected .json, .yaml, or .yml",
                    extension
                )))
            }
        };

        if prompt.parameters.is_empty() {
            let system_instructions: Vec<MessageNum> = prompt
                .request
                .system_instructions()
                .iter()
                .map(|msg| (*msg).clone())
                .collect();
            let parameters =
                Self::extract_variables(prompt.request.messages(), &system_instructions);
            prompt.parameters = parameters;
        }

        if prompt.media_parameters.is_empty() {
            let system_instructions: Vec<MessageNum> = prompt
                .request
                .system_instructions()
                .iter()
                .map(|msg| (*msg).clone())
                .collect();
            let media_parameters =
                Self::extract_media_variables(prompt.request.messages(), &system_instructions);
            prompt.media_parameters = media_parameters;
        }

        Ok(prompt)
    }

    fn resolve_prompt_candidate(
        requested_path: &Path,
        candidate_path: PathBuf,
        attempted_paths: &mut Vec<PathBuf>,
    ) -> Result<Option<PathBuf>, TypeError> {
        push_attempted_path(attempted_paths, candidate_path.clone());

        if candidate_path.is_file() {
            return Ok(Some(candidate_path));
        }

        if requested_path.extension().is_some() {
            return Ok(None);
        }

        let mut matches = Vec::new();

        for extension in PROMPT_FILE_EXTENSIONS {
            let extension_candidate = candidate_path.with_extension(extension);
            push_attempted_path(attempted_paths, extension_candidate.clone());

            if extension_candidate.is_file() {
                matches.push(extension_candidate);
            }
        }

        match matches.len() {
            0 => Ok(None),
            1 => Ok(matches.into_iter().next()),
            _ => Err(TypeError::AmbiguousPromptPath {
                requested_path: requested_path.display().to_string(),
                candidate_paths: format_candidate_paths(&matches),
            }),
        }
    }

    fn resolve_prompt_path(path: &Path, base_dir: Option<&Path>) -> Result<PathBuf, TypeError> {
        let mut attempted_paths = Vec::new();

        if path.is_absolute() {
            return Self::resolve_prompt_candidate(path, path.to_path_buf(), &mut attempted_paths)?
                .ok_or_else(|| TypeError::PromptPathNotFound {
                    requested_path: path.display().to_string(),
                    attempted_paths: format_candidate_paths(&attempted_paths),
                });
        }

        let mut candidate_roots = Vec::new();
        if let Some(base_dir) = base_dir {
            candidate_roots.push(base_dir.to_path_buf());
        }

        let current_dir = std::env::current_dir()?;
        if !candidate_roots.iter().any(|root| root == &current_dir) {
            candidate_roots.push(current_dir);
        }

        for root in candidate_roots {
            let candidate_path = root.join(path);
            if let Some(resolved_path) =
                Self::resolve_prompt_candidate(path, candidate_path, &mut attempted_paths)?
            {
                return Ok(resolved_path);
            }
        }

        Err(TypeError::PromptPathNotFound {
            requested_path: path.display().to_string(),
            attempted_paths: format_candidate_paths(&attempted_paths),
        })
    }

    fn load_prompt_from_path(path: &Path, base_dir: Option<&Path>) -> Result<Self, TypeError> {
        let resolved_path = Self::resolve_prompt_path(path, base_dir)?;
        Self::read_prompt_file(&resolved_path)
    }

    pub fn from_path_with_base(
        path: impl AsRef<Path>,
        base_dir: impl AsRef<Path>,
    ) -> Result<Self, TypeError> {
        Self::load_prompt_from_path(path.as_ref(), Some(base_dir.as_ref()))
    }

    /// Converts a generic prompt configuration to a Prompt instance.
    /// This handles the user-friendly YAML/JSON format parsing.
    pub fn from_generic_config(config: GenericPromptConfig) -> Result<Self, TypeError> {
        // Validate that messages is not empty
        if config.messages.is_empty() {
            return Err(TypeError::Error(
                "Prompt has no messages. Generic prompt format requires at least one message."
                    .to_string(),
            ));
        }

        // Parse provider string to Provider enum
        let provider = Provider::resolve(config.provider.as_deref())?;

        // Convert message strings to MessageNum based on provider
        let messages: Vec<MessageNum> = config
            .messages
            .into_iter()
            .map(|msg| create_message_for_provider(msg, &provider, Role::User.as_str()))
            .collect::<Result<Vec<_>, _>>()?;

        // Convert system instructions if present
        let system_instructions = if let Some(sys_inst) = config.system_instructions {
            sys_inst
                .into_iter()
                .map(|msg| create_message_for_provider(msg, &provider, get_system_role(&provider)))
                .collect::<Result<Vec<_>, _>>()?
        } else {
            Vec::new()
        };

        // Convert settings to ModelSettings based on provider
        let model_settings = if let Some(settings) = config.settings {
            Some(Self::settings_from_value(settings, &provider)?)
        } else {
            None
        };

        // Create the prompt using new_rs
        Self::new_rs(
            messages,
            &config.model,
            provider,
            system_instructions,
            model_settings,
            config.response_format,
            ResponseType::Null,
        )
    }

    /// Converts a JSON Value to ModelSettings based on the provider.
    /// This handles the provider-specific settings structure from generic configs.
    fn settings_from_value(value: Value, provider: &Provider) -> Result<ModelSettings, TypeError> {
        match provider {
            Provider::OpenAI => {
                let settings: OpenAIChatSettings = serde_json::from_value(value)?;
                Ok(ModelSettings::OpenAIChat(settings))
            }
            Provider::Anthropic => {
                let settings: AnthropicSettings = serde_json::from_value(value)?;
                Ok(ModelSettings::AnthropicChat(settings))
            }
            Provider::Gemini | Provider::Google | Provider::Vertex | Provider::GoogleAdk => {
                let settings: GeminiSettings = serde_json::from_value(value)?;
                Ok(ModelSettings::GoogleChat(settings))
            }
            _ => Err(TypeError::Error(format!(
                "Settings not supported for provider: {:?}",
                provider
            ))),
        }
    }

    pub fn response_json_schema(&self) -> Option<&Value> {
        self.request.response_json_schema()
    }

    pub fn new_rs(
        mut messages: Vec<MessageNum>,
        model: &str,
        provider: Provider,
        mut system_instructions: Vec<MessageNum>,
        model_settings: Option<ModelSettings>,
        response_json_schema: Option<Value>,
        response_type: ResponseType,
    ) -> Result<Self, TypeError> {
        let model = model.to_string();
        // get version from crate
        let version = potato_util::version();
        // If model_settings is not provided, set model and provider to undefined if missing
        let model_settings = match model_settings {
            Some(settings) => {
                // validates if provider and settings are compatible
                settings.validate_provider(&provider)?;
                settings
            }
            None => ModelSettings::provider_default_settings(&provider),
        };

        if system_instructions
            .iter()
            .any(|msg| !msg.extract_media_variables().is_empty())
        {
            return Err(TypeError::MediaInSystemMessage);
        }

        for msg in messages.iter_mut() {
            msg.split_media_placeholders()?;
        }
        for msg in system_instructions.iter_mut() {
            msg.split_media_placeholders()?;
        }

        // extract named parameters in prompt
        let parameters = Self::extract_variables(&messages, &system_instructions);
        let media_parameters = Self::extract_media_variables(&messages, &system_instructions);

        // Build the provider request
        let request = to_provider_request(
            messages,
            system_instructions,
            model.clone(),
            model_settings,
            response_json_schema,
        )?;

        Ok(Self {
            request,
            version,
            parameters,
            media_parameters,
            response_type,
            model,
            provider,
        })
    }

    fn is_google_provider(&self) -> bool {
        matches!(
            self.provider,
            Provider::Google | Provider::Gemini | Provider::Vertex | Provider::GoogleAdk
        )
    }

    pub fn add_tools(&mut self, tools: Vec<AgentToolDefinition>) -> Result<(), TypeError> {
        self.request.add_tools(tools)
    }

    pub fn extract_variables(
        messages: &[MessageNum],
        system_instructions: &[MessageNum],
    ) -> Vec<String> {
        let mut variables = BTreeSet::new();

        // Extract from system instructions
        for msg in system_instructions {
            variables.extend(msg.extract_variables());
        }

        // Extract from user messages
        for msg in messages {
            variables.extend(msg.extract_variables());
        }

        variables.into_iter().collect()
    }

    pub fn extract_media_variables(
        messages: &[MessageNum],
        system_instructions: &[MessageNum],
    ) -> Vec<String> {
        let mut variables = BTreeSet::new();

        for msg in system_instructions {
            variables.extend(msg.extract_media_variables());
        }

        for msg in messages {
            variables.extend(msg.extract_media_variables());
        }

        variables.into_iter().collect()
    }

    pub fn model_dump_value(&self) -> Value {
        // Convert the Prompt to a JSON Value
        serde_json::to_value(self).unwrap_or(Value::Null)
    }

    pub fn to_request_json(&self) -> Result<Value, TypeError> {
        // Convert the Prompt to a JSON Value
        let json_value = serde_json::to_value(self)?;

        Ok(json_value)
    }

    pub fn set_response_json_schema(
        &mut self,
        response_json_schema: Option<Value>,
        response_type: ResponseType,
    ) {
        self.request.set_response_json_schema(response_json_schema);
        self.response_type = response_type;
    }
}

// tests
#[cfg(test)]
mod tests {
    use super::*;
    use crate::anthropic::v1::request::{
        Base64ImageSource, Base64PDFSource, ContentBlockParam, DocumentBlockParam, ImageBlockParam,
        MessageParam, PlainTextSource, TextBlockParam, UrlImageSource, UrlPDFSource,
    };
    use crate::google::{DataNum, GeminiContent, Part};
    use crate::openai::v1::chat::request::{
        ChatMessage as OpenAIChatMessage, ContentPart, FileContentPart, ImageContentPart,
        TextContentPart,
    };
    use crate::prompt::types::Score;
    use crate::StructuredOutput;
    use std::fs;
    use std::path::{Path, PathBuf};

    fn create_temp_prompt_dir() -> PathBuf {
        let dir = std::env::temp_dir().join(format!(
            "potatohead-prompt-tests-{}",
            potato_util::create_uuid7()
        ));
        fs::create_dir_all(&dir).unwrap();
        dir
    }

    fn write_generic_prompt(path: &Path, provider: &str, model: &str, message: &str) {
        let content =
            format!("model: {model}\nprovider: {provider}\nmessages:\n  - \"{message}\"\n");
        fs::write(path, content).unwrap();
    }

    fn create_openai_chat_message() -> OpenAIChatMessage {
        let text_part = TextContentPart::new("What company is this logo from?".to_string());
        let text_content_part = ContentPart::Text(text_part);
        OpenAIChatMessage {
            role: "user".to_string(),
            content: vec![text_content_part],
            name: None,
        }
    }

    fn create_system_openai_chat_message() -> OpenAIChatMessage {
        let text_part = TextContentPart::new("system_prompt".to_string());
        let text_content_part = ContentPart::Text(text_part);
        OpenAIChatMessage {
            role: "developer".to_string(),
            content: vec![text_content_part],
            name: None,
        }
    }

    fn create_openai_image_message() -> OpenAIChatMessage {
        let image_part = ImageContentPart::new("https://iili.io/3Hs4FMg.png".to_string(), None);
        let image_content_part = ContentPart::ImageUrl(image_part);
        OpenAIChatMessage {
            role: "user".to_string(),
            content: vec![image_content_part],
            name: None,
        }
    }

    fn create_openai_file_message() -> OpenAIChatMessage {
        let file_part = FileContentPart::new(
            Some("filedata".to_string()),
            Some("fileid".to_string()),
            Some("filename".to_string()),
        );
        let file_content_part = ContentPart::FileContent(file_part);
        OpenAIChatMessage {
            role: "user".to_string(),
            content: vec![file_content_part],
            name: None,
        }
    }

    fn create_anthropic_text_message() -> MessageParam {
        let text_block =
            TextBlockParam::new_rs("What company is this logo from?".to_string(), None, None);
        MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Text(text_block),
            }],
        }
    }

    fn create_anthropic_system_message() -> MessageParam {
        let text_block = TextBlockParam::new_rs("system_prompt".to_string(), None, None);
        MessageParam {
            role: "assistant".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Text(text_block),
            }],
        }
    }

    fn create_anthropic_base64_image_message() -> MessageParam {
        let image_source =
            Base64ImageSource::new("image/png".to_string(), "base64data".to_string()).unwrap();
        let image_block = ImageBlockParam {
            source: crate::anthropic::v1::request::ImageSource::Base64(image_source),
            cache_control: None,
            r#type: "image".to_string(),
        };
        MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Image(image_block),
            }],
        }
    }

    fn create_anthropic_url_image_message() -> MessageParam {
        let image_source = UrlImageSource::new("https://iili.io/3Hs4FMg.png".to_string());
        let image_block = ImageBlockParam {
            source: crate::anthropic::v1::request::ImageSource::Url(image_source),
            cache_control: None,
            r#type: "image".to_string(),
        };
        MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Image(image_block),
            }],
        }
    }

    fn create_anthropic_base64_pdf_message() -> MessageParam {
        let pdf_source = Base64PDFSource::new("base64pdfdata".to_string()).unwrap();
        let document_block = DocumentBlockParam {
            source: crate::anthropic::v1::request::DocumentSource::Base64(pdf_source),
            cache_control: None,
            title: Some("test_document.pdf".to_string()),
            context: None,
            r#type: "document".to_string(),
            citations: None,
        };
        MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Document(document_block),
            }],
        }
    }

    fn create_anthropic_url_pdf_message() -> MessageParam {
        let pdf_source = UrlPDFSource::new("https://example.com/document.pdf".to_string());
        let document_block = DocumentBlockParam {
            source: crate::anthropic::v1::request::DocumentSource::Url(pdf_source),
            cache_control: None,
            title: Some("test_document.pdf".to_string()),
            context: None,
            r#type: "document".to_string(),
            citations: None,
        };
        MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Document(document_block),
            }],
        }
    }

    fn create_anthropic_plain_text_document_message() -> MessageParam {
        let text_source = PlainTextSource::new("Plain text document content".to_string());
        let document_block = DocumentBlockParam {
            source: crate::anthropic::v1::request::DocumentSource::Text(text_source),
            cache_control: None,
            title: Some("text_document.txt".to_string()),
            context: Some("Context for the document".to_string()),
            r#type: "document".to_string(),
            citations: None,
        };
        MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Document(document_block),
            }],
        }
    }

    #[test]
    fn test_task_list_add_and_get() {
        let text_part = TextContentPart::new("Test prompt. ${param1} ${param2}".to_string());
        let content_part = ContentPart::Text(text_part);
        let message = OpenAIChatMessage {
            role: "user".to_string(),
            content: vec![content_part],
            name: None,
        };

        let prompt = Prompt::new_rs(
            vec![MessageNum::OpenAIMessageV1(message)],
            "gpt-4o",
            Provider::OpenAI,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check if the prompt was created successfully
        assert_eq!(prompt.request.messages().len(), 1);

        // check prompt parameters
        assert!(prompt.parameters.len() == 2);

        // sort parameters to ensure order does not affect the test
        let mut parameters = prompt.parameters.clone();
        parameters.sort();

        assert_eq!(parameters[0], "param1");
        assert_eq!(parameters[1], "param2");

        // bind parameter
        let bound_msg = prompt.request.messages()[0]
            .bind("param1", "Value1")
            .unwrap();
        let bound_msg = bound_msg.bind("param2", "Value2").unwrap();

        // Check if the bound message contains the correct values
        match bound_msg.clone() {
            MessageNum::OpenAIMessageV1(msg) => {
                if let ContentPart::Text(text_part) = &msg.content[0] {
                    assert_eq!(text_part.text, "Test prompt. Value1 Value2");
                } else {
                    panic!("Expected TextContentPart");
                }
            }
            _ => panic!("Expected OpenAIMessageV1"),
        }
    }

    #[test]
    fn test_image_prompt() {
        let text_message = create_openai_chat_message();
        let image_message = create_openai_image_message();

        let system_text_part = TextContentPart::new("system_prompt".to_string());
        let system_text_content_part = ContentPart::Text(system_text_part);

        let system_text_message = OpenAIChatMessage {
            role: "assistant".to_string(),
            content: vec![system_text_content_part],
            name: None,
        };

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::OpenAIMessageV1(text_message),
                MessageNum::OpenAIMessageV1(image_message),
            ],
            "gpt-4o",
            Provider::OpenAI,
            vec![MessageNum::OpenAIMessageV1(system_text_message)],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check the first user message
        if let MessageNum::OpenAIMessageV1(msg) = &prompt.request.messages()[1] {
            if let ContentPart::Text(text_part) = &msg.content[0] {
                assert_eq!(text_part.text, "What company is this logo from?");
            } else {
                panic!("Expected TextContentPart for the first user message");
            }
        } else {
            panic!("Expected OpenAIMessageV1 for the first user message");
        }

        // Check the second user message (ImageUrl)
        if let MessageNum::OpenAIMessageV1(msg) = &prompt.request.messages()[2] {
            if let ContentPart::ImageUrl(image_url) = &msg.content[0] {
                assert_eq!(image_url.image_url.url, "https://iili.io/3Hs4FMg.png");
                assert_eq!(image_url.r#type, "image_url");
            } else {
                panic!("Expected ContentPart::Image for the second user message");
            }
        } else {
            panic!("Expected OpenAIMessageV1 for the second user message");
        }
    }

    #[test]
    fn test_document_prompt() {
        let text_message = create_openai_chat_message();
        let file_message = create_openai_file_message();
        let system_message = create_system_openai_chat_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::OpenAIMessageV1(text_message),
                MessageNum::OpenAIMessageV1(file_message),
            ],
            "gpt-4o",
            Provider::OpenAI,
            vec![MessageNum::OpenAIMessageV1(system_message)],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check the 2nd user message (file)
        if let MessageNum::OpenAIMessageV1(msg) = &prompt.request.messages()[2] {
            if let ContentPart::FileContent(file_content) = &msg.content[0] {
                assert_eq!(file_content.file.file_id.as_ref().unwrap(), "fileid");
                assert_eq!(file_content.file.filename.as_ref().unwrap(), "filename");
            } else {
                panic!("Expected ContentPart::FileContent for the second user message");
            }
        } else {
            panic!("Expected OpenAIMessageV1 for the first user message");
        }
    }

    #[test]
    fn test_response_format_score() {
        let text_message = create_openai_chat_message();
        let prompt = Prompt::new_rs(
            vec![MessageNum::OpenAIMessageV1(text_message)],
            "gpt-4o",
            Provider::OpenAI,
            vec![],
            None,
            Some(Score::get_structured_output_schema()),
            ResponseType::Null,
        )
        .unwrap();

        // Check if the response json schema is set correctly
        assert!(prompt.response_json_schema().is_some());
    }

    #[test]
    fn test_anthropic_text_message_binding() {
        let text_block =
            TextBlockParam::new_rs("Test prompt. ${param1} ${param2}".to_string(), None, None);
        let message = MessageParam {
            role: "user".to_string(),
            content: vec![ContentBlockParam {
                inner: crate::anthropic::v1::request::ContentBlock::Text(text_block),
            }],
        };

        let prompt = Prompt::new_rs(
            vec![MessageNum::AnthropicMessageV1(message)],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        assert_eq!(prompt.request.messages().len(), 1);
        assert_eq!(prompt.parameters.len(), 2);

        let mut parameters = prompt.parameters.clone();
        parameters.sort();
        assert_eq!(parameters[0], "param1");
        assert_eq!(parameters[1], "param2");

        // Test parameter binding
        let bound_msg = prompt.request.messages()[0]
            .bind("param1", "Value1")
            .unwrap();
        let bound_msg = bound_msg.bind("param2", "Value2").unwrap();

        match bound_msg {
            MessageNum::AnthropicMessageV1(msg) => {
                if let crate::anthropic::v1::request::ContentBlock::Text(text_block) =
                    &msg.content[0].inner
                {
                    assert_eq!(text_block.text, "Test prompt. Value1 Value2");
                } else {
                    panic!("Expected TextBlockParam");
                }
            }
            _ => panic!("Expected AnthropicMessageV1"),
        }
    }

    #[test]
    fn test_anthropic_url_image_prompt() {
        let text_message = create_anthropic_text_message();
        let image_message = create_anthropic_url_image_message();
        let system_message = create_anthropic_system_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::AnthropicMessageV1(text_message),
                MessageNum::AnthropicMessageV1(image_message),
            ],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![MessageNum::AnthropicMessageV1(system_message)],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check first message (text)
        if let MessageNum::AnthropicMessageV1(msg) = &prompt.request.messages()[0] {
            if let crate::anthropic::v1::request::ContentBlock::Text(text_block) =
                &msg.content[0].inner
            {
                assert_eq!(text_block.text, "What company is this logo from?");
            } else {
                panic!("Expected TextBlock for first message");
            }
        } else {
            panic!("Expected AnthropicMessageV1");
        }

        // Check second message (image URL)
        if let MessageNum::AnthropicMessageV1(msg) = &prompt.request.messages()[1] {
            if let crate::anthropic::v1::request::ContentBlock::Image(image_block) =
                &msg.content[0].inner
            {
                match &image_block.source {
                    crate::anthropic::v1::request::ImageSource::Url(url_source) => {
                        assert_eq!(url_source.url, "https://iili.io/3Hs4FMg.png");
                        assert_eq!(url_source.r#type, "url");
                    }
                    _ => panic!("Expected URL image source"),
                }
                assert_eq!(image_block.r#type, "image");
            } else {
                panic!("Expected ImageBlock for second message");
            }
        } else {
            panic!("Expected AnthropicMessageV1");
        }
    }

    #[test]
    fn test_anthropic_base64_image_prompt() {
        let text_message = create_anthropic_text_message();
        let image_message = create_anthropic_base64_image_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::AnthropicMessageV1(text_message),
                MessageNum::AnthropicMessageV1(image_message),
            ],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check second message (base64 image)
        if let MessageNum::AnthropicMessageV1(msg) = &prompt.request.messages()[1] {
            if let crate::anthropic::v1::request::ContentBlock::Image(image_block) =
                &msg.content[0].inner
            {
                match &image_block.source {
                    crate::anthropic::v1::request::ImageSource::Base64(base64_source) => {
                        assert_eq!(base64_source.media_type, "image/png");
                        assert_eq!(base64_source.data, "base64data");
                        assert_eq!(base64_source.r#type, "base64");
                    }
                    _ => panic!("Expected Base64 image source"),
                }
            } else {
                panic!("Expected ImageBlock");
            }
        } else {
            panic!("Expected AnthropicMessageV1");
        }
    }

    // Test: Anthropic PDF document (base64)
    #[test]
    fn test_anthropic_base64_pdf_document_prompt() {
        let text_message = create_anthropic_text_message();
        let pdf_message = create_anthropic_base64_pdf_message();
        let system_message = create_anthropic_system_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::AnthropicMessageV1(text_message),
                MessageNum::AnthropicMessageV1(pdf_message),
            ],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![MessageNum::AnthropicMessageV1(system_message)],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check second message (PDF document)
        if let MessageNum::AnthropicMessageV1(msg) = &prompt.request.messages()[1] {
            if let crate::anthropic::v1::request::ContentBlock::Document(document_block) =
                &msg.content[0].inner
            {
                match &document_block.source {
                    crate::anthropic::v1::request::DocumentSource::Base64(pdf_source) => {
                        assert_eq!(pdf_source.media_type, "application/pdf");
                        assert_eq!(pdf_source.data, "base64pdfdata");
                        assert_eq!(pdf_source.r#type, "base64");
                    }
                    _ => panic!("Expected Base64 PDF source"),
                }
                assert_eq!(document_block.r#type, "document");
                assert_eq!(document_block.title.as_ref().unwrap(), "test_document.pdf");
            } else {
                panic!("Expected DocumentBlock");
            }
        } else {
            panic!("Expected AnthropicMessageV1");
        }
    }

    // Test: Anthropic URL PDF document
    #[test]
    fn test_anthropic_url_pdf_document_prompt() {
        let text_message = create_anthropic_text_message();
        let pdf_message = create_anthropic_url_pdf_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::AnthropicMessageV1(text_message),
                MessageNum::AnthropicMessageV1(pdf_message),
            ],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check second message (URL PDF)
        if let MessageNum::AnthropicMessageV1(msg) = &prompt.request.messages()[1] {
            if let crate::anthropic::v1::request::ContentBlock::Document(document_block) =
                &msg.content[0].inner
            {
                match &document_block.source {
                    crate::anthropic::v1::request::DocumentSource::Url(url_source) => {
                        assert_eq!(url_source.url, "https://example.com/document.pdf");
                        assert_eq!(url_source.r#type, "url");
                    }
                    _ => panic!("Expected URL PDF source"),
                }
            } else {
                panic!("Expected DocumentBlock");
            }
        } else {
            panic!("Expected AnthropicMessageV1");
        }
    }

    // Test: Anthropic plain text document
    #[test]
    fn test_anthropic_plain_text_document_prompt() {
        let text_message = create_anthropic_text_message();
        let text_doc_message = create_anthropic_plain_text_document_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::AnthropicMessageV1(text_message),
                MessageNum::AnthropicMessageV1(text_doc_message),
            ],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        // Check second message (plain text document)
        if let MessageNum::AnthropicMessageV1(msg) = &prompt.request.messages()[1] {
            if let crate::anthropic::v1::request::ContentBlock::Document(document_block) =
                &msg.content[0].inner
            {
                match &document_block.source {
                    crate::anthropic::v1::request::DocumentSource::Text(text_source) => {
                        assert_eq!(text_source.media_type, "text/plain");
                        assert_eq!(text_source.data, "Plain text document content");
                        assert_eq!(text_source.r#type, "text");
                    }
                    _ => panic!("Expected Text document source"),
                }
                assert_eq!(
                    document_block.context.as_ref().unwrap(),
                    "Context for the document"
                );
            } else {
                panic!("Expected DocumentBlock");
            }
        } else {
            panic!("Expected AnthropicMessageV1");
        }
    }

    // Test: Mixed Anthropic content (text + multiple documents)
    #[test]
    fn test_anthropic_mixed_content_prompt() {
        let text_message = create_anthropic_text_message();
        let pdf_message = create_anthropic_base64_pdf_message();
        let text_doc_message = create_anthropic_plain_text_document_message();
        let system_message = create_anthropic_system_message();

        let prompt = Prompt::new_rs(
            vec![
                MessageNum::AnthropicMessageV1(text_message),
                MessageNum::AnthropicMessageV1(pdf_message),
                MessageNum::AnthropicMessageV1(text_doc_message),
            ],
            "claude-3-5-sonnet-20241022",
            Provider::Anthropic,
            vec![MessageNum::AnthropicMessageV1(system_message)],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        assert_eq!(prompt.request.messages().len(), 3);
        assert_eq!(prompt.request.system_instructions().len(), 1);
        assert_eq!(prompt.provider, Provider::Anthropic);
        assert_eq!(prompt.model, "claude-3-5-sonnet-20241022");
    }

    // gemini test
    #[test]
    fn test_gemini_chat_message() {
        let text = Part::from_text("Test prompt. ${param1} ${param2}".to_string());
        let message = GeminiContent {
            role: "user".to_string(),
            parts: vec![text],
        };

        let prompt = Prompt::new_rs(
            vec![MessageNum::GeminiContentV1(message)],
            "gemini-1.5-pro",
            Provider::Google,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        assert_eq!(prompt.request.messages().len(), 1);
        assert_eq!(prompt.parameters.len(), 2);

        let mut parameters = prompt.parameters.clone();
        parameters.sort();
        assert_eq!(parameters[0], "param1");
        assert_eq!(parameters[1], "param2");

        // Test parameter binding
        let bound_msg = prompt.request.messages()[0]
            .bind("param1", "Value1")
            .unwrap();
        let bound_msg = bound_msg.bind("param2", "Value2").unwrap();

        match bound_msg {
            MessageNum::GeminiContentV1(msg) => {
                if let DataNum::Text(text_part) = &msg.parts[0].data {
                    assert_eq!(text_part, "Test prompt. Value1 Value2");
                } else {
                    panic!("Expected Text Part");
                }
            }
            _ => panic!("Expected GeminiContentV1"),
        }
    }

    #[test]
    fn test_from_path_with_base_resolves_missing_extension() {
        let temp_dir = create_temp_prompt_dir();
        let prompt_path = temp_dir.join("prompt.yaml");
        write_generic_prompt(&prompt_path, "openai", "gpt-4o", "Hello ${name}");

        let prompt = Prompt::from_path_with_base("prompt", &temp_dir).unwrap();

        assert_eq!(prompt.model, "gpt-4o");
        assert_eq!(prompt.provider, Provider::OpenAI);
        assert_eq!(prompt.parameters, vec!["name".to_string()]);

        fs::remove_dir_all(temp_dir).unwrap();
    }

    #[test]
    fn test_from_path_reports_ambiguous_matches() {
        let temp_dir = create_temp_prompt_dir();
        write_generic_prompt(&temp_dir.join("prompt.yaml"), "openai", "gpt-4o", "Hello");
        fs::write(
            temp_dir.join("prompt.json"),
            r#"{"model":"gpt-4o","provider":"openai","messages":["Hello"]}"#,
        )
        .unwrap();

        let error = Prompt::from_path_with_base("prompt", &temp_dir).unwrap_err();

        match error {
            TypeError::AmbiguousPromptPath {
                requested_path,
                candidate_paths,
            } => {
                assert_eq!(requested_path, "prompt");
                assert!(candidate_paths.contains("prompt.yaml"));
                assert!(candidate_paths.contains("prompt.json"));
            }
            other => panic!("expected AmbiguousPromptPath, got {other:?}"),
        }

        fs::remove_dir_all(temp_dir).unwrap();
    }

    #[test]
    fn test_from_path_reports_attempted_paths_when_missing() {
        let temp_dir = create_temp_prompt_dir();
        let missing_path = temp_dir.join("missing_prompt");

        let error = Prompt::from_path(missing_path.clone()).unwrap_err();

        match error {
            TypeError::PromptPathNotFound {
                requested_path,
                attempted_paths,
            } => {
                assert_eq!(requested_path, missing_path.display().to_string());
                assert!(attempted_paths.contains("missing_prompt"));
                assert!(attempted_paths.contains("missing_prompt.yaml"));
                assert!(attempted_paths.contains("missing_prompt.yml"));
                assert!(attempted_paths.contains("missing_prompt.json"));
            }
            other => panic!("expected PromptPathNotFound, got {other:?}"),
        }

        fs::remove_dir_all(temp_dir).unwrap();
    }

    #[test]
    fn test_save_prompt_returns_written_json_path() {
        let temp_dir = create_temp_prompt_dir();
        let text_part = TextContentPart::new("Hello ${name}".to_string());
        let content_part = ContentPart::Text(text_part);
        let message = OpenAIChatMessage {
            role: "user".to_string(),
            content: vec![content_part],
            name: None,
        };
        let prompt = Prompt::new_rs(
            vec![MessageNum::OpenAIMessageV1(message)],
            "gpt-4o",
            Provider::OpenAI,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();

        let saved_path = prompt
            .save_prompt(Some(temp_dir.join("saved_prompt")))
            .unwrap();
        let loaded_prompt = Prompt::from_path(saved_path.clone()).unwrap();

        assert_eq!(
            saved_path.extension().and_then(|ext| ext.to_str()),
            Some("json")
        );
        assert!(saved_path.is_file());
        assert_eq!(loaded_prompt.model, "gpt-4o");
        assert_eq!(loaded_prompt.provider, Provider::OpenAI);

        fs::remove_dir_all(temp_dir).unwrap();
    }
}

#[cfg(test)]
mod media_binding_tests {
    use super::*;
    use crate::anthropic::v1::request::{
        ContentBlock, ContentBlockParam, DocumentSource, ImageSource, MessageParam, TextBlockParam,
    };
    use crate::google::v1::generate::request::{DataNum, GeminiContent, Part};
    use crate::openai::v1::chat::request::{ChatMessage, ContentPart, TextContentPart};
    use crate::prompt::media::{MediaKind, MediaRef};
    use crate::prompt::types::MessageNum;
    use crate::Provider;
    use base64::Engine;

    fn make_user_anthropic(text: &str) -> MessageNum {
        MessageNum::AnthropicMessageV1(MessageParam {
            content: vec![ContentBlockParam {
                inner: ContentBlock::Text(TextBlockParam::new_rs(text.to_string(), None, None)),
            }],
            role: "user".to_string(),
        })
    }

    fn make_user_openai(text: &str) -> MessageNum {
        MessageNum::OpenAIMessageV1(ChatMessage {
            role: "user".to_string(),
            content: vec![ContentPart::Text(TextContentPart::new(text.to_string()))],
            name: None,
        })
    }

    fn make_user_gemini(text: &str) -> MessageNum {
        MessageNum::GeminiContentV1(GeminiContent {
            role: "user".to_string(),
            parts: vec![Part {
                data: DataNum::Text(text.to_string()),
                ..Default::default()
            }],
        })
    }

    fn make_system_openai(text: &str) -> MessageNum {
        MessageNum::OpenAIMessageV1(ChatMessage {
            role: "developer".to_string(),
            content: vec![ContentPart::Text(TextContentPart::new(text.to_string()))],
            name: None,
        })
    }

    fn make_system_gemini(text: &str) -> MessageNum {
        MessageNum::GeminiContentV1(GeminiContent {
            role: "model".to_string(),
            parts: vec![Part {
                data: DataNum::Text(text.to_string()),
                ..Default::default()
            }],
        })
    }

    fn build_prompt(provider: Provider, model: &str, msg: MessageNum) -> Prompt {
        Prompt::new_rs(
            vec![msg],
            model,
            provider,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap()
    }

    #[test]
    fn anthropic_splitter_isolates_token() {
        let p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("hello ${media:chart} world"),
        );
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        assert_eq!(blocks.len(), 3);
        let texts: Vec<String> = blocks
            .iter()
            .filter_map(|b| match &b.inner {
                ContentBlock::Text(t) => Some(t.text.clone()),
                _ => None,
            })
            .collect();
        assert_eq!(texts, vec!["hello ", "${media:chart}", " world"]);
    }

    #[test]
    fn openai_splitter_isolates_token() {
        let p = build_prompt(
            Provider::OpenAI,
            "gpt-4o",
            make_user_openai("a ${media:x} b"),
        );
        let parts = match &p.request.messages()[0] {
            MessageNum::OpenAIMessageV1(m) => &m.content,
            _ => panic!(),
        };
        assert_eq!(parts.len(), 3);
    }

    #[test]
    fn gemini_splitter_isolates_token() {
        let p = build_prompt(
            Provider::Gemini,
            "gemini-2.0-flash",
            make_user_gemini("foo ${media:y}"),
        );
        let parts = match &p.request.messages()[0] {
            MessageNum::GeminiContentV1(m) => &m.parts,
            _ => panic!(),
        };
        assert_eq!(parts.len(), 2);
    }

    #[test]
    fn splitter_handles_token_at_start_and_end() {
        let p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:a}${media:b}"),
        );
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        assert_eq!(blocks.len(), 2);
    }

    #[test]
    fn parameters_and_media_parameters_disjoint() {
        let p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${greeting} on ${media:chart} for ${media:doc}"),
        );
        assert_eq!(p.parameters, vec!["greeting"]);
        let mut media = p.media_parameters.clone();
        media.sort();
        assert_eq!(media, vec!["chart".to_string(), "doc".to_string()]);
    }

    #[test]
    fn duplicate_media_token_dedups() {
        let p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:x} foo ${media:x}"),
        );
        assert_eq!(p.media_parameters, vec!["x".to_string()]);
    }

    #[test]
    fn anthropic_image_url() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:chart}"),
        );
        p.bind_media_mut(
            "chart",
            &MediaRef::image_url("https://x/y.png".into(), None),
        )
        .unwrap();
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        match &blocks[0].inner {
            ContentBlock::Image(b) => match &b.source {
                ImageSource::Url(s) => assert_eq!(s.url, "https://x/y.png"),
                _ => panic!("expected url source"),
            },
            _ => panic!("expected image block"),
        }
    }

    #[test]
    fn anthropic_image_bytes_roundtrip_serde() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:chart}"),
        );
        p.bind_media_mut(
            "chart",
            &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"FAKEPNG"),
        )
        .unwrap();
        let json = serde_json::to_value(&p).unwrap();
        let restored: Prompt = serde_json::from_value(json).unwrap();
        assert_eq!(p, restored);
    }

    #[test]
    fn anthropic_document_base64() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:doc}"),
        );
        p.bind_media_mut(
            "doc",
            &MediaRef::from_bytes(MediaKind::Document, "application/pdf".into(), b"%PDF"),
        )
        .unwrap();
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        match &blocks[0].inner {
            ContentBlock::Document(b) => match &b.source {
                DocumentSource::Base64(s) => {
                    assert_eq!(s.media_type, "application/pdf");
                    assert_eq!(
                        s.data,
                        base64::engine::general_purpose::STANDARD.encode(b"%PDF")
                    );
                }
                _ => panic!(),
            },
            _ => panic!(),
        }
    }

    #[test]
    fn anthropic_document_url() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:doc}"),
        );
        p.bind_media_mut(
            "doc",
            &MediaRef::document_url("https://x/y.pdf".into(), None),
        )
        .unwrap();
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        assert!(matches!(
            &blocks[0].inner,
            ContentBlock::Document(b)
                if matches!(&b.source, DocumentSource::Url(s) if s.url == "https://x/y.pdf")
        ));
    }

    #[test]
    fn openai_image_bytes_becomes_data_url() {
        let mut p = build_prompt(
            Provider::OpenAI,
            "gpt-4o",
            make_user_openai("${media:chart}"),
        );
        p.bind_media_mut(
            "chart",
            &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"FAKE"),
        )
        .unwrap();
        let parts = match &p.request.messages()[0] {
            MessageNum::OpenAIMessageV1(m) => &m.content,
            _ => panic!(),
        };
        match &parts[0] {
            ContentPart::ImageUrl(p) => {
                assert!(p.image_url.url.starts_with("data:image/png;base64,"));
            }
            _ => panic!("expected image_url"),
        }
    }

    #[test]
    fn openai_image_url_passthrough() {
        let mut p = build_prompt(Provider::OpenAI, "gpt-4o", make_user_openai("${media:c}"));
        p.bind_media_mut("c", &MediaRef::image_url("https://x/y.png".into(), None))
            .unwrap();
        let parts = match &p.request.messages()[0] {
            MessageNum::OpenAIMessageV1(m) => &m.content,
            _ => panic!(),
        };
        assert!(matches!(
            &parts[0],
            ContentPart::ImageUrl(p) if p.image_url.url == "https://x/y.png"
        ));
    }

    #[test]
    fn openai_document_url_rejected() {
        let mut p = build_prompt(Provider::OpenAI, "gpt-4o", make_user_openai("${media:doc}"));
        let err = p
            .bind_media_mut(
                "doc",
                &MediaRef::document_url("https://x/y.pdf".into(), None),
            )
            .unwrap_err();
        assert!(matches!(err, TypeError::UnsupportedMediaForProvider { .. }));
    }

    #[test]
    fn openai_document_bytes_becomes_file_content() {
        let mut p = build_prompt(Provider::OpenAI, "gpt-4o", make_user_openai("${media:doc}"));
        p.bind_media_mut(
            "doc",
            &MediaRef::from_bytes(MediaKind::Document, "application/pdf".into(), b"%PDF"),
        )
        .unwrap();
        let parts = match &p.request.messages()[0] {
            MessageNum::OpenAIMessageV1(m) => &m.content,
            _ => panic!(),
        };
        match &parts[0] {
            ContentPart::FileContent(p) => {
                assert_eq!(p.r#type, "file");
                assert!(p
                    .file
                    .file_data
                    .as_deref()
                    .unwrap()
                    .starts_with("data:application/pdf;base64,"));
            }
            _ => panic!("expected file content"),
        }
    }

    #[test]
    fn gemini_inline_data_from_bytes() {
        let mut p = build_prompt(
            Provider::Gemini,
            "gemini-2.0-flash",
            make_user_gemini("${media:chart}"),
        );
        p.bind_media_mut(
            "chart",
            &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"FAKE"),
        )
        .unwrap();
        let parts = match &p.request.messages()[0] {
            MessageNum::GeminiContentV1(m) => &m.parts,
            _ => panic!(),
        };
        match &parts[0].data {
            DataNum::InlineData(b) => assert_eq!(b.mime_type, "image/png"),
            _ => panic!(),
        }
    }

    #[test]
    fn gemini_https_url_rejected() {
        let mut p = build_prompt(
            Provider::Gemini,
            "gemini-2.0-flash",
            make_user_gemini("${media:c}"),
        );
        let err = p
            .bind_media_mut(
                "c",
                &MediaRef::image_url("https://x/y.png".into(), Some("image/png".into())),
            )
            .unwrap_err();
        assert!(matches!(err, TypeError::UnsupportedMediaForProvider { .. }));
    }

    #[test]
    fn gemini_gs_url_accepted() {
        let mut p = build_prompt(
            Provider::Gemini,
            "gemini-2.0-flash",
            make_user_gemini("${media:c}"),
        );
        p.bind_media_mut(
            "c",
            &MediaRef::image_url("gs://b/c.png".into(), Some("image/png".into())),
        )
        .unwrap();
        let parts = match &p.request.messages()[0] {
            MessageNum::GeminiContentV1(m) => &m.parts,
            _ => panic!(),
        };
        match &parts[0].data {
            DataNum::FileData(f) => {
                assert_eq!(f.file_uri, "gs://b/c.png");
                assert_eq!(f.mime_type, "image/png");
            }
            _ => panic!(),
        }
    }

    #[test]
    fn gemini_url_without_mime_rejected() {
        let mut p = build_prompt(
            Provider::Gemini,
            "gemini-2.0-flash",
            make_user_gemini("${media:c}"),
        );
        let err = p
            .bind_media_mut("c", &MediaRef::image_url("gs://b/c.png".into(), None))
            .unwrap_err();
        assert!(matches!(err, TypeError::InvalidMediaType(_)));
    }

    #[test]
    fn missing_placeholder_errors() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:foo}"),
        );
        let err = p
            .bind_media_mut(
                "bar",
                &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"X"),
            )
            .unwrap_err();
        assert!(matches!(err, TypeError::MediaPlaceholderNotFound { .. }));
    }

    #[test]
    fn media_in_system_message_rejected_at_construction() {
        let sys = MessageNum::AnthropicSystemMessageV1(TextBlockParam::new_rs(
            "system ${media:x}".to_string(),
            None,
            None,
        ));
        let result = Prompt::new_rs(
            vec![make_user_anthropic("hi")],
            "claude-sonnet-4-5",
            Provider::Anthropic,
            vec![sys],
            None,
            None,
            ResponseType::Null,
        );
        assert!(matches!(result, Err(TypeError::MediaInSystemMessage)));
    }

    #[test]
    fn media_in_openai_system_message_rejected_at_construction() {
        let result = Prompt::new_rs(
            vec![make_user_openai("hi")],
            "gpt-4o",
            Provider::OpenAI,
            vec![make_system_openai("system ${media:x}")],
            None,
            None,
            ResponseType::Null,
        );
        assert!(matches!(result, Err(TypeError::MediaInSystemMessage)));
    }

    #[test]
    fn media_in_gemini_system_message_rejected_at_construction() {
        let result = Prompt::new_rs(
            vec![make_user_gemini("hi")],
            "gemini-2.0-flash",
            Provider::Gemini,
            vec![make_system_gemini("system ${media:x}")],
            None,
            None,
            ResponseType::Null,
        );
        assert!(matches!(result, Err(TypeError::MediaInSystemMessage)));
    }

    #[test]
    fn bind_does_not_touch_media_token() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${name} and ${media:name}"),
        );
        for m in p.request.messages_mut() {
            m.bind_mut("name", "Steven").unwrap();
        }
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        match &blocks.last().unwrap().inner {
            ContentBlock::Text(t) => assert_eq!(t.text, "${media:name}"),
            _ => panic!(),
        }
    }

    #[test]
    fn bind_then_bind_media_independent() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${greeting} ${media:img}"),
        );
        for m in p.request.messages_mut() {
            m.bind_mut("greeting", "Hello").unwrap();
        }
        p.bind_media_mut(
            "img",
            &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"X"),
        )
        .unwrap();
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        let mut saw_text = false;
        let mut saw_image = false;
        for b in blocks {
            match &b.inner {
                ContentBlock::Text(t) if t.text.contains("Hello") => saw_text = true,
                ContentBlock::Image(_) => saw_image = true,
                _ => {}
            }
        }
        assert!(saw_text && saw_image);
    }

    #[test]
    fn multiple_media_bindings_in_sequence() {
        let mut p = build_prompt(
            Provider::Anthropic,
            "claude-sonnet-4-5",
            make_user_anthropic("${media:a} ${media:b}"),
        );
        p.bind_media_mut(
            "a",
            &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"A"),
        )
        .unwrap();
        p.bind_media_mut(
            "b",
            &MediaRef::from_bytes(MediaKind::Document, "application/pdf".into(), b"%PDF"),
        )
        .unwrap();
        let blocks = match &p.request.messages()[0] {
            MessageNum::AnthropicMessageV1(m) => &m.content,
            _ => panic!(),
        };
        let mut saw_image = false;
        let mut saw_doc = false;
        for b in blocks {
            match &b.inner {
                ContentBlock::Image(_) => saw_image = true,
                ContentBlock::Document(_) => saw_doc = true,
                _ => {}
            }
        }
        assert!(saw_image && saw_doc);
    }

    #[test]
    fn placeholder_replaced_across_multiple_messages() {
        let mut p = Prompt::new_rs(
            vec![
                make_user_anthropic("${media:x}"),
                make_user_anthropic("again ${media:x}"),
            ],
            "claude-sonnet-4-5",
            Provider::Anthropic,
            vec![],
            None,
            None,
            ResponseType::Null,
        )
        .unwrap();
        p.bind_media_mut(
            "x",
            &MediaRef::from_bytes(MediaKind::Image, "image/png".into(), b"X"),
        )
        .unwrap();
        for msg in p.request.messages() {
            let has_image = match msg {
                MessageNum::AnthropicMessageV1(m) => m
                    .content
                    .iter()
                    .any(|b| matches!(b.inner, ContentBlock::Image(_))),
                _ => false,
            };
            assert!(has_image);
        }
    }
}