car-inference 0.14.0

//! Protocol abstraction — unified interface for remote model API providers.
//!
//! Each provider (OpenAI, Anthropic, Google) implements `ProtocolHandler`,
//! which owns message formatting, request building, response parsing, and
//! streaming. RemoteBackend becomes a thin dispatcher.

use crate::stream::StreamEvent;
use crate::tasks::generate::{ContentBlock, Message, ResponseFormat, ToolCall};
use crate::InferenceError;
use serde_json::Value;
use std::collections::HashMap;

/// Unified request parameters for any protocol.
#[derive(Debug, Clone)]
pub struct ApiRequest {
    pub model: String,
    pub messages: Vec<Value>,
    pub system: Option<String>,
    pub temperature: f64,
    pub max_tokens: usize,
    pub tools: Option<Vec<Value>>,
    pub tool_choice: Option<String>,
    pub parallel_tool_calls: Option<bool>,
    pub stream: bool,
    pub budget_tokens: usize,
    /// When true, mark system prompt and tool definitions for Anthropic prompt caching.
    /// Enables cache reuse across parent/child agent calls sharing the same prefix.
    pub cache_control: bool,
    /// Optional schema constraint on the response. Each handler maps
    /// this to its provider's native field — see `ResponseFormat`.
    pub response_format: Option<ResponseFormat>,
}

/// Unified response from any protocol.
#[derive(Debug, Clone)]
pub struct ApiResponse {
    pub text: String,
    pub tool_calls: Vec<ToolCall>,
    /// Token usage statistics parsed from the API response.
    pub usage: Option<crate::TokenUsage>,
}

/// Protocol handler trait — each provider implements this.
pub trait ProtocolHandler: Send + Sync {
    /// URL path for the API endpoint (appended to the base URL).
    fn endpoint_path(&self) -> &str;

    /// Build HTTP headers for authentication.
    fn auth_headers(&self, api_key: &str) -> Vec<(String, String)>;

    /// Build the JSON request body from unified ApiRequest.
    fn build_request_body(&self, req: &ApiRequest) -> Value;

    /// Parse a complete (non-streaming) response body into ApiResponse.
    fn parse_response(&self, body: &str) -> Result<ApiResponse, InferenceError>;

    /// Parse a single SSE line into StreamEvents (for streaming responses).
    /// Returns multiple events when a single SSE chunk contains multiple tool calls.
    fn parse_stream_event(&self, event_type: &str, data: &str) -> Vec<StreamEvent>;

    /// Convert conversation Messages to this protocol's message format.
    fn build_messages(
        &self,
        messages: &[Message],
        prompt: &str,
        context: Option<&str>,
        images: Option<&[ContentBlock]>,
    ) -> (Vec<Value>, Option<String>);

    /// Convert tool definitions to this protocol's format.
    fn build_tools(&self, tools: &[Value]) -> Vec<Value>;

    /// Whether this protocol supports streaming.
    fn supports_streaming(&self) -> bool {
        true
    }

    /// Whether this protocol supports extended thinking.
    fn supports_thinking(&self) -> bool {
        false
    }

    /// Whether this protocol accepts video content blocks natively on
    /// the generation endpoint. Defaults to `false`. When a request
    /// carries a video block and the selected protocol returns
    /// `false`, `RemoteBackend::execute_request` rejects the call
    /// with [`InferenceError::UnsupportedMode`] rather than silently
    /// downgrading the video to a text placeholder.
    fn supports_video(&self) -> bool {
        false
    }

    /// Whether this protocol accepts audio content blocks natively.
    /// Same contract as [`supports_video`] — false means the remote
    /// backend pre-check returns `UnsupportedMode` instead of
    /// silently stringifying the audio reference.
    fn supports_audio(&self) -> bool {
        false
    }

    /// Backend identifier used in `UnsupportedMode` error messages.
    /// Handlers override to something recognizable (e.g. "openai",
    /// "anthropic-messages-v1").
    fn protocol_name(&self) -> &'static str {
        "remote"
    }
}

// ---------------------------------------------------------------------------
// OpenAI-compatible handler (also covers OpenAI Responses API for chat)
// ---------------------------------------------------------------------------

pub struct OpenAiHandler;

impl ProtocolHandler for OpenAiHandler {
    fn endpoint_path(&self) -> &str {
        "/v1/chat/completions"
    }

    fn auth_headers(&self, api_key: &str) -> Vec<(String, String)> {
        vec![
            ("Authorization".into(), format!("Bearer {}", api_key)),
            ("Content-Type".into(), "application/json".into()),
        ]
    }

    fn build_request_body(&self, req: &ApiRequest) -> Value {
        let quirks = openai_quirks(&req.model);
        let mut body = serde_json::json!({
            "model": req.model,
            "messages": req.messages,
        });
        if quirks.uses_max_completion_tokens {
            body["max_completion_tokens"] = serde_json::json!(req.max_tokens);
        } else {
            body["max_tokens"] = serde_json::json!(req.max_tokens);
        }
        if req.temperature >= 0.0 && !quirks.rejects_temperature {
            body["temperature"] = serde_json::json!(req.temperature);
        }
        if let Some(ref tools) = req.tools {
            body["tools"] = serde_json::json!(tools);
            body["tool_choice"] = serde_json::json!(req.tool_choice.as_deref().unwrap_or("auto"));
            if let Some(parallel_tool_calls) = req.parallel_tool_calls {
                body["parallel_tool_calls"] = serde_json::json!(parallel_tool_calls);
            }
        }
        // Structured output. OpenAI's strict mode enforces the schema on
        // the response; the looser `json_object` form just demands valid
        // JSON. Schema name defaults to "response" — within the 64-char
        // alphanumeric/-_ constraint that OpenAI imposes on this field.
        match &req.response_format {
            Some(ResponseFormat::JsonSchema {
                schema,
                strict,
                name,
            }) => {
                body["response_format"] = serde_json::json!({
                    "type": "json_schema",
                    "json_schema": {
                        "name": name.as_deref().unwrap_or("response"),
                        "schema": schema,
                        "strict": strict,
                    },
                });
            }
            Some(ResponseFormat::JsonObject) => {
                body["response_format"] = serde_json::json!({ "type": "json_object" });
            }
            None => {}
        }
        if req.stream {
            body["stream"] = serde_json::json!(true);
            // `stream_options.include_usage` makes OpenAI emit a final
            // choiceless chunk with real prompt/completion token counts.
            // Providers that don't recognize the flag ignore it, so
            // always-on for the streaming path is safe and matches
            // pre-refactor behavior in `RemoteBackend::generate_stream`.
            body["stream_options"] = serde_json::json!({ "include_usage": true });
        }
        body
    }

    fn parse_response(&self, body: &str) -> Result<ApiResponse, InferenceError> {
        let parsed: Value = serde_json::from_str(body)
            .map_err(|e| InferenceError::InferenceFailed(format!("parse response: {e}")))?;

        let choice = parsed
            .get("choices")
            .and_then(|c| c.as_array())
            .and_then(|a| a.first())
            .ok_or_else(|| InferenceError::InferenceFailed("empty response".into()))?;

        let message = choice
            .get("message")
            .ok_or_else(|| InferenceError::InferenceFailed("no message in choice".into()))?;

        let text = message
            .get("content")
            .and_then(|c| c.as_str())
            .unwrap_or("")
            .to_string();

        let mut tool_calls = Vec::new();
        if let Some(tcs) = message.get("tool_calls").and_then(|t| t.as_array()) {
            for tc in tcs {
                if let Some(func) = tc.get("function") {
                    let id = tc.get("id").and_then(|i| i.as_str()).map(|s| s.to_string());
                    let name = func
                        .get("name")
                        .and_then(|n| n.as_str())
                        .unwrap_or("")
                        .to_string();
                    let args_str = func
                        .get("arguments")
                        .and_then(|a| a.as_str())
                        .unwrap_or("{}");
                    let arguments: HashMap<String, Value> =
                        serde_json::from_str(args_str).unwrap_or_default();
                    tool_calls.push(ToolCall {
                        id,
                        name,
                        arguments,
                    });
                }
            }
        }

        // Parse usage statistics
        let usage = parsed.get("usage").and_then(|u| {
            Some(crate::TokenUsage {
                prompt_tokens: u.get("prompt_tokens").and_then(|v| v.as_u64()).unwrap_or(0),
                completion_tokens: u
                    .get("completion_tokens")
                    .and_then(|v| v.as_u64())
                    .unwrap_or(0),
                total_tokens: u.get("total_tokens").and_then(|v| v.as_u64()).unwrap_or(0),
                context_window: 0, // filled in by caller from model schema
            })
        });

        Ok(ApiResponse {
            text,
            tool_calls,
            usage,
        })
    }

    fn parse_stream_event(&self, _event_type: &str, data: &str) -> Vec<StreamEvent> {
        crate::stream::parse_openai_sse_line(&format!("data: {}", data))
    }

    fn build_messages(
        &self,
        messages: &[Message],
        prompt: &str,
        context: Option<&str>,
        images: Option<&[ContentBlock]>,
    ) -> (Vec<Value>, Option<String>) {
        if !messages.is_empty() {
            let mut result = Vec::new();
            if let Some(ctx) = context {
                result.push(serde_json::json!({"role": "system", "content": ctx}));
            }
            for msg in messages {
                match msg {
                    Message::System { content } => {
                        result.push(serde_json::json!({"role": "system", "content": content}));
                    }
                    Message::User { content } => {
                        result.push(serde_json::json!({"role": "user", "content": content}));
                    }
                    Message::UserMultimodal { content } => {
                        let blocks: Vec<Value> = content
                            .iter()
                            .map(|block| match block {
                                ContentBlock::Text { text } => {
                                    serde_json::json!({"type": "text", "text": text})
                                }
                                ContentBlock::ImageBase64 { data, media_type } => {
                                    serde_json::json!({
                                        "type": "image_url",
                                        "image_url": {
                                            "url": format!("data:{};base64,{}", media_type, data),
                                        }
                                    })
                                }
                                ContentBlock::ImageUrl { url, detail } => {
                                    serde_json::json!({
                                        "type": "image_url",
                                        "image_url": {
                                            "url": url,
                                            "detail": detail,
                                        }
                                    })
                                }
                                // Video / audio on OpenAI-compatible:
                                // the API has no native input for
                                // these, and `RemoteBackend` pre-
                                // rejects them via
                                // `supports_video()` / `supports_audio()`.
                                // The prior implementation produced a
                                // text placeholder here (`[video: …]`)
                                // which silently corrupted behavior
                                // for any caller that bypassed the
                                // pre-check. Panic loudly instead;
                                // tests assert `UnsupportedMode`
                                // instead of exercising this arm.
                                ContentBlock::VideoPath { .. }
                                | ContentBlock::VideoUrl { .. }
                                | ContentBlock::VideoBase64 { .. }
                                | ContentBlock::AudioPath { .. }
                                | ContentBlock::AudioUrl { .. }
                                | ContentBlock::AudioBase64 { .. } => {
                                    unreachable!(
                                        "video/audio ContentBlock reached OpenAI \
                                         build_messages — should have been rejected \
                                         by RemoteBackend::execute_request"
                                    )
                                }
                            })
                            .collect();
                        result.push(serde_json::json!({"role": "user", "content": blocks}));
                    }
                    Message::Assistant {
                        content,
                        tool_calls,
                    } => {
                        if tool_calls.is_empty() {
                            result
                                .push(serde_json::json!({"role": "assistant", "content": content}));
                        } else {
                            let tc: Vec<Value> = tool_calls.iter().enumerate().map(|(i, tc)| {
                                let id = tc.id.clone().unwrap_or_else(|| format!("call_{}", i));
                                serde_json::json!({
                                    "id": id,
                                    "type": "function",
                                    "function": {
                                        "name": tc.name,
                                        "arguments": serde_json::to_string(&tc.arguments).unwrap_or_default(),
                                    }
                                })
                            }).collect();
                            let mut msg =
                                serde_json::json!({"role": "assistant", "tool_calls": tc});
                            if !content.is_empty() {
                                msg["content"] = serde_json::json!(content);
                            }
                            result.push(msg);
                        }
                    }
                    Message::ToolResult {
                        tool_use_id,
                        content,
                    } => {
                        result.push(serde_json::json!({
                            "role": "tool",
                            "tool_call_id": tool_use_id,
                            "content": content,
                        }));
                    }
                    // OpenAI Chat Completions has no native concept
                    // of provider output items; they're a Responses
                    // API construct. Drop them silently — the items
                    // remain in the caller's stored history and the
                    // chat-completions request still has the matching
                    // assistant tool_use blocks for context.
                    Message::ProviderOutputItems { .. } => continue,
                }
            }
            (result, None) // system already in messages
        } else {
            let mut msgs = Vec::new();
            if let Some(ctx) = context {
                msgs.push(serde_json::json!({"role": "system", "content": ctx}));
            }
            if let Some(images) = images.filter(|images| !images.is_empty()) {
                let mut blocks = vec![serde_json::json!({"type": "text", "text": prompt})];
                for image in images {
                    let block = match image {
                        ContentBlock::Text { text } => {
                            serde_json::json!({"type": "text", "text": text})
                        }
                        ContentBlock::ImageBase64 { data, media_type } => {
                            serde_json::json!({
                                "type": "image_url",
                                "image_url": {
                                    "url": format!("data:{};base64,{}", media_type, data),
                                }
                            })
                        }
                        ContentBlock::ImageUrl { url, detail } => {
                            serde_json::json!({
                                "type": "image_url",
                                "image_url": {
                                    "url": url,
                                    "detail": detail,
                                }
                            })
                        }
                        ContentBlock::VideoPath { .. }
                        | ContentBlock::VideoUrl { .. }
                        | ContentBlock::VideoBase64 { .. }
                        | ContentBlock::AudioPath { .. }
                        | ContentBlock::AudioUrl { .. }
                        | ContentBlock::AudioBase64 { .. } => {
                            unreachable!(
                                "video/audio ContentBlock reached OpenAI build_messages \
                                 — should have been rejected by RemoteBackend::execute_request"
                            )
                        }
                    };
                    blocks.push(block);
                }
                msgs.push(serde_json::json!({"role": "user", "content": blocks}));
            } else {
                msgs.push(serde_json::json!({"role": "user", "content": prompt}));
            }
            (msgs, None)
        }
    }

    fn build_tools(&self, tools: &[Value]) -> Vec<Value> {
        tools
            .iter()
            .map(|t| {
                if t.get("type").is_some() {
                    t.clone()
                } else {
                    serde_json::json!({"type": "function", "function": t})
                }
            })
            .collect()
    }

    fn protocol_name(&self) -> &'static str {
        "openai"
    }
}

/// Per-model OpenAI API quirks. Centralized so the next quirk to land
/// (reasoning_effort, verbosity, ...) gets one match in one place
/// instead of a new boolean and a new `starts_with` chain at every
/// call site.
struct OpenAiQuirks {
    /// Newer models reject `max_tokens` and require `max_completion_tokens`.
    uses_max_completion_tokens: bool,
    /// o-series models reject any non-default `temperature` value.
    rejects_temperature: bool,
}

fn openai_quirks(model: &str) -> OpenAiQuirks {
    let m = model.to_lowercase();
    let is_o_series = m.starts_with("o1") || m.starts_with("o3") || m.starts_with("o4");
    OpenAiQuirks {
        uses_max_completion_tokens: is_o_series
            || m.starts_with("gpt-5")
            || m.starts_with("gpt-4.1"),
        rejects_temperature: is_o_series,
    }
}

// ---------------------------------------------------------------------------
// Anthropic handler
// ---------------------------------------------------------------------------

pub struct AnthropicHandler;

impl ProtocolHandler for AnthropicHandler {
    fn endpoint_path(&self) -> &str {
        "/v1/messages"
    }

    fn auth_headers(&self, api_key: &str) -> Vec<(String, String)> {
        vec![
            ("x-api-key".into(), api_key.to_string()),
            ("anthropic-version".into(), "2023-06-01".into()),
            ("anthropic-beta".into(), "prompt-caching-2024-07-31".into()),
            ("Content-Type".into(), "application/json".into()),
        ]
    }

    fn build_request_body(&self, req: &ApiRequest) -> Value {
        let mut body = serde_json::json!({
            "model": req.model,
            "max_tokens": req.max_tokens,
            "messages": req.messages,
        });

        if req.budget_tokens > 0 {
            body["thinking"] = serde_json::json!({
                "type": "enabled",
                "budget_tokens": req.budget_tokens,
            });
            // Anthropic requires temperature unset when thinking is enabled
        } else if req.temperature >= 0.0 {
            body["temperature"] = serde_json::json!(req.temperature);
        }

        if let Some(ref system) = req.system {
            if req.cache_control {
                // Use structured system prompt with cache_control for prompt caching
                body["system"] = serde_json::json!([{
                    "type": "text",
                    "text": system,
                    "cache_control": {"type": "ephemeral"}
                }]);
            } else {
                body["system"] = Value::String(system.clone());
            }
        }

        if let Some(ref tools) = req.tools {
            body["tools"] = Value::Array(tools.clone());
            if req.cache_control && !tools.is_empty() {
                // Mark last tool definition for cache breakpoint
                if let Some(arr) = body["tools"].as_array_mut() {
                    if let Some(last) = arr.last_mut() {
                        if let Some(obj) = last.as_object_mut() {
                            obj.insert(
                                "cache_control".to_string(),
                                serde_json::json!({"type": "ephemeral"}),
                            );
                        }
                    }
                }
            }
            body["tool_choice"] = serde_json::json!({"type": "auto"});
        }

        // Anthropic has no native response_format field as of early 2026.
        // Schema-validated output is achieved via tool_use coercion
        // (define a tool whose `input_schema` is your schema, set
        // `tool_choice: required`). Surfacing that here would conflict
        // with caller-supplied tools, so we log and ignore — the caller
        // can keep using the tool-coercion idiom that's worked all along.
        if req.response_format.is_some() {
            tracing::warn!(
                "response_format set on Anthropic request — Anthropic has no native \
                 JSON-schema field; the request will run unconstrained. Use tool_use \
                 with tool_choice=required to enforce a schema on Claude.",
            );
        }

        if req.stream {
            body["stream"] = serde_json::json!(true);
        }

        body
    }

    fn parse_response(&self, body: &str) -> Result<ApiResponse, InferenceError> {
        let parsed: Value = serde_json::from_str(body)
            .map_err(|e| InferenceError::InferenceFailed(format!("parse response: {e}")))?;

        let mut text = String::new();
        let mut tool_calls = Vec::new();
        let mut thinking_text = String::new();

        if let Some(content) = parsed.get("content").and_then(|c| c.as_array()) {
            for block in content {
                match block.get("type").and_then(|t| t.as_str()) {
                    Some("text") => {
                        if let Some(t) = block.get("text").and_then(|t| t.as_str()) {
                            text.push_str(t);
                        }
                    }
                    Some("thinking") => {
                        // Extended thinking block — capture as fallback if no text blocks exist
                        if let Some(t) = block.get("thinking").and_then(|t| t.as_str()) {
                            tracing::debug!(thinking_len = t.len(), "extended thinking block");
                            thinking_text.push_str(t);
                        }
                    }
                    Some("tool_use") => {
                        if let (Some(name), Some(input)) = (
                            block.get("name").and_then(|n| n.as_str()),
                            block.get("input"),
                        ) {
                            let id = block
                                .get("id")
                                .and_then(|i| i.as_str())
                                .map(|s| s.to_string());
                            let arguments: HashMap<String, Value> =
                                serde_json::from_value(input.clone()).unwrap_or_default();
                            tool_calls.push(ToolCall {
                                id,
                                name: name.to_string(),
                                arguments,
                            });
                        }
                    }
                    _ => {}
                }
            }
        }

        // Fallback: if the model produced only thinking blocks with no text output,
        // use the thinking content. This prevents empty responses when extended
        // thinking is auto-enabled and the model reasons entirely in thinking blocks.
        if text.is_empty() && !thinking_text.is_empty() && tool_calls.is_empty() {
            tracing::warn!(
                thinking_len = thinking_text.len(),
                "response had only thinking blocks, using thinking content as text fallback"
            );
            text = thinking_text;
        }

        // Parse usage statistics
        let usage = parsed.get("usage").and_then(|u| {
            Some(crate::TokenUsage {
                prompt_tokens: u.get("input_tokens").and_then(|v| v.as_u64()).unwrap_or(0),
                completion_tokens: u.get("output_tokens").and_then(|v| v.as_u64()).unwrap_or(0),
                total_tokens: u.get("input_tokens").and_then(|v| v.as_u64()).unwrap_or(0)
                    + u.get("output_tokens").and_then(|v| v.as_u64()).unwrap_or(0),
                context_window: 0, // filled in by caller from model schema
            })
        });

        Ok(ApiResponse {
            text,
            tool_calls,
            usage,
        })
    }

    fn parse_stream_event(&self, event_type: &str, data: &str) -> Vec<StreamEvent> {
        crate::stream::parse_anthropic_sse_line(event_type, data)
    }

    fn build_messages(
        &self,
        messages: &[Message],
        prompt: &str,
        context: Option<&str>,
        images: Option<&[ContentBlock]>,
    ) -> (Vec<Value>, Option<String>) {
        // Anthropic's `/v1/messages` takes `system` as a top-level
        // field — pushing `{"role": "system", ...}` into the messages
        // array returns a 400. Fold any `Message::System` into the
        // same slot that `context` uses; when both are present, the
        // caller-supplied context wins (it's the deliberate system
        // prompt built by the runtime) and we append the incoming
        // system message on a new line so nothing is silently dropped.
        let mut system = context.map(|c| c.to_string());
        if !messages.is_empty() {
            for msg in messages {
                if let Message::System { content } = msg {
                    system = Some(match system {
                        Some(existing) if !existing.is_empty() => {
                            format!("{existing}\n\n{content}")
                        }
                        _ => content.clone(),
                    });
                }
            }
        }

        if !messages.is_empty() {
            let mut result = Vec::new();
            for msg in messages {
                match msg {
                    // Already folded into the system field above — skip.
                    Message::System { .. } => continue,
                    Message::User { content } => {
                        result.push(serde_json::json!({"role": "user", "content": content}));
                    }
                    Message::UserMultimodal { content } => {
                        let blocks: Vec<Value> = content
                            .iter()
                            .map(|block| {
                                match block {
                                    ContentBlock::Text { text } => {
                                        serde_json::json!({"type": "text", "text": text})
                                    }
                                    ContentBlock::ImageBase64 { data, media_type } => {
                                        serde_json::json!({
                                            "type": "image",
                                            "source": {
                                                "type": "base64",
                                                "media_type": media_type,
                                                "data": data,
                                            }
                                        })
                                    }
                                    ContentBlock::ImageUrl { url, .. } => {
                                        // Anthropic uses source.url format
                                        serde_json::json!({
                                            "type": "image",
                                            "source": {
                                                "type": "url",
                                                "url": url,
                                            }
                                        })
                                    }
                                    // Anthropic's /v1/messages has no
                                    // native video or audio path.
                                    // RemoteBackend pre-rejects these;
                                    // panic here so direct callers
                                    // can't accidentally ship a silent
                                    // `[video: …]` placeholder.
                                    ContentBlock::VideoPath { .. }
                                    | ContentBlock::VideoUrl { .. }
                                    | ContentBlock::VideoBase64 { .. }
                                    | ContentBlock::AudioPath { .. }
                                    | ContentBlock::AudioUrl { .. }
                                    | ContentBlock::AudioBase64 { .. } => {
                                        unreachable!(
                                            "video/audio ContentBlock reached Anthropic \
                                             build_messages — should have been rejected \
                                             by RemoteBackend::execute_request"
                                        )
                                    }
                                }
                            })
                            .collect();
                        result.push(serde_json::json!({"role": "user", "content": blocks}));
                    }
                    Message::Assistant {
                        content,
                        tool_calls,
                    } => {
                        let mut blocks: Vec<Value> = Vec::new();
                        if !content.is_empty() {
                            blocks.push(serde_json::json!({"type": "text", "text": content}));
                        }
                        for (i, tc) in tool_calls.iter().enumerate() {
                            let id = tc.id.clone().unwrap_or_else(|| format!("toolu_{}", i));
                            blocks.push(serde_json::json!({
                                "type": "tool_use",
                                "id": id,
                                "name": tc.name,
                                "input": tc.arguments,
                            }));
                        }
                        if blocks.is_empty() {
                            blocks.push(serde_json::json!({"type": "text", "text": ""}));
                        }
                        result.push(serde_json::json!({"role": "assistant", "content": blocks}));
                    }
                    Message::ToolResult {
                        tool_use_id,
                        content,
                    } => {
                        result.push(serde_json::json!({
                            "role": "user",
                            "content": [{
                                "type": "tool_result",
                                "tool_use_id": tool_use_id,
                                "content": content,
                            }]
                        }));
                    }
                    // Provider output items are an OpenAI Responses
                    // API construct; Anthropic's `/v1/messages` has
                    // no equivalent. Drop them — the caller's stored
                    // history still has the matching tool_use blocks.
                    Message::ProviderOutputItems { .. } => continue,
                }
            }
            (result, system)
        } else {
            let content = if let Some(images) = images.filter(|images| !images.is_empty()) {
                let mut blocks = vec![serde_json::json!({"type": "text", "text": prompt})];
                for image in images {
                    let block = match image {
                        ContentBlock::Text { text } => {
                            serde_json::json!({"type": "text", "text": text})
                        }
                        ContentBlock::ImageBase64 { data, media_type } => {
                            serde_json::json!({
                                "type": "image",
                                "source": {
                                    "type": "base64",
                                    "media_type": media_type,
                                    "data": data,
                                }
                            })
                        }
                        ContentBlock::ImageUrl { url, .. } => {
                            serde_json::json!({
                                "type": "image",
                                "source": {
                                    "type": "url",
                                    "url": url,
                                }
                            })
                        }
                        ContentBlock::VideoPath { .. }
                        | ContentBlock::VideoUrl { .. }
                        | ContentBlock::VideoBase64 { .. }
                        | ContentBlock::AudioPath { .. }
                        | ContentBlock::AudioUrl { .. }
                        | ContentBlock::AudioBase64 { .. } => {
                            unreachable!(
                                "video/audio ContentBlock reached Anthropic build_messages \
                                 — should have been rejected by RemoteBackend::execute_request"
                            )
                        }
                    };
                    blocks.push(block);
                }
                Value::Array(blocks)
            } else {
                Value::String(prompt.to_string())
            };
            let msgs = vec![serde_json::json!({"role": "user", "content": content})];
            (msgs, system)
        }
    }

    fn build_tools(&self, tools: &[Value]) -> Vec<Value> {
        tools.iter().filter_map(|t| {
            let func = t.get("function").unwrap_or(t);
            Some(serde_json::json!({
                "name": func.get("name")?,
                "description": func.get("description").and_then(|d| d.as_str()).unwrap_or(""),
                "input_schema": func.get("parameters").cloned().unwrap_or(serde_json::json!({"type": "object"})),
            }))
        }).collect()
    }

    fn supports_thinking(&self) -> bool {
        true
    }

    fn protocol_name(&self) -> &'static str {
        "anthropic"
    }
}

// ---------------------------------------------------------------------------
// Google (Gemini) handler
// ---------------------------------------------------------------------------

pub struct GoogleHandler;

impl ProtocolHandler for GoogleHandler {
    fn endpoint_path(&self) -> &str {
        ""
    } // URL is fully custom

    fn auth_headers(&self, _api_key: &str) -> Vec<(String, String)> {
        // Google uses query param auth, not headers
        vec![("Content-Type".into(), "application/json".into())]
    }

    fn build_request_body(&self, req: &ApiRequest) -> Value {
        let mut body = serde_json::json!({
            "contents": req.messages,
        });

        if let Some(ref system) = req.system {
            body["systemInstruction"] = serde_json::json!({
                "parts": [{"text": system}],
            });
        }

        let mut generation_config = serde_json::json!({
            "maxOutputTokens": req.max_tokens,
        });
        if req.temperature >= 0.0 {
            generation_config["temperature"] = serde_json::json!(req.temperature);
        }
        // Gemini exposes structured output via generationConfig.responseMimeType
        // and (optionally) responseSchema. The schema dialect is a subset
        // of JSON Schema — basic types/enums/required/object-properties.
        // We pass schemas through as-is and let the upstream validator
        // surface dialect mismatches; this keeps the field portable
        // across Gemini model versions.
        match &req.response_format {
            Some(ResponseFormat::JsonSchema { schema, .. }) => {
                generation_config["responseMimeType"] = serde_json::json!("application/json");
                generation_config["responseSchema"] = schema.clone();
            }
            Some(ResponseFormat::JsonObject) => {
                generation_config["responseMimeType"] = serde_json::json!("application/json");
            }
            None => {}
        }
        body["generationConfig"] = generation_config;

        if let Some(ref tools) = req.tools {
            body["tools"] = serde_json::json!([{
                "functionDeclarations": tools,
            }]);
            body["toolConfig"] = serde_json::json!({
                "functionCallingConfig": {
                    "mode": "AUTO",
                }
            });
        }

        body
    }

    fn parse_response(&self, body: &str) -> Result<ApiResponse, InferenceError> {
        let parsed: Value = serde_json::from_str(body)
            .map_err(|e| InferenceError::InferenceFailed(format!("parse response: {e}")))?;

        let parts = parsed
            .get("candidates")
            .and_then(|c| c.as_array())
            .and_then(|a| a.first())
            .and_then(|c| c.get("content"))
            .and_then(|c| c.get("parts"))
            .and_then(|p| p.as_array())
            .cloned()
            .unwrap_or_default();

        let mut text_chunks = Vec::new();
        let mut tool_calls = Vec::new();
        for part in parts {
            if let Some(text) = part.get("text").and_then(|t| t.as_str()) {
                text_chunks.push(text.to_string());
            }
            if let Some(function_call) = part
                .get("functionCall")
                .or_else(|| part.get("function_call"))
            {
                let name = function_call
                    .get("name")
                    .and_then(|n| n.as_str())
                    .unwrap_or_default()
                    .to_string();
                let arguments = function_call
                    .get("args")
                    .or_else(|| function_call.get("arguments"))
                    .and_then(|args| args.as_object())
                    .map(|map| {
                        map.iter()
                            .map(|(k, v)| (k.clone(), v.clone()))
                            .collect::<HashMap<_, _>>()
                    })
                    .unwrap_or_default();
                if !name.is_empty() {
                    tool_calls.push(ToolCall {
                        id: None,
                        name,
                        arguments,
                    });
                }
            }
        }

        let usage = parsed.get("usageMetadata").map(|usage| crate::TokenUsage {
            prompt_tokens: usage
                .get("promptTokenCount")
                .and_then(|v| v.as_u64())
                .unwrap_or(0),
            completion_tokens: usage
                .get("candidatesTokenCount")
                .and_then(|v| v.as_u64())
                .unwrap_or(0),
            total_tokens: usage
                .get("totalTokenCount")
                .and_then(|v| v.as_u64())
                .unwrap_or(0),
            context_window: 0,
        });

        Ok(ApiResponse {
            text: text_chunks.join("\n"),
            tool_calls,
            usage,
        })
    }

    fn parse_stream_event(&self, _event_type: &str, _data: &str) -> Vec<StreamEvent> {
        Vec::new() // Google streaming not implemented
    }

    fn build_messages(
        &self,
        messages: &[Message],
        prompt: &str,
        context: Option<&str>,
        images: Option<&[ContentBlock]>,
    ) -> (Vec<Value>, Option<String>) {
        // Fold any Message::System entries into Gemini's
        // `systemInstruction` slot (returned via the second tuple
        // element). Multiple System messages are joined — the caller-
        // supplied `context` wins when both are set.
        let mut system_instruction: Option<String> = context.map(|c| c.to_string());
        if !messages.is_empty() {
            for msg in messages {
                if let Message::System { content } = msg {
                    system_instruction = Some(match system_instruction {
                        Some(existing) if !existing.is_empty() => {
                            format!("{existing}\n\n{content}")
                        }
                        _ => content.clone(),
                    });
                }
            }
        }

        if !messages.is_empty() {
            let contents = messages
                .iter()
                // System messages are already folded into
                // system_instruction — drop them from contents[].
                // Provider output items are an OpenAI Responses
                // API construct with no Gemini equivalent — also
                // drop, the caller's stored history still has the
                // matching tool_use blocks.
                .filter(|msg| {
                    !matches!(
                        msg,
                        Message::System { .. } | Message::ProviderOutputItems { .. }
                    )
                })
                .map(|msg| match msg {
                    // Unreachable after the filter, but the compiler
                    // still needs the arms for exhaustiveness.
                    Message::System { .. } => unreachable!("System filtered above"),
                    Message::ProviderOutputItems { .. } => {
                        unreachable!("ProviderOutputItems filtered above")
                    }
                    Message::User { content } => serde_json::json!({
                        "role": "user",
                        "parts": [{"text": content}],
                    }),
                    Message::UserMultimodal { content } => serde_json::json!({
                        "role": "user",
                        "parts": content.iter().map(google_part_from_block).collect::<Vec<_>>(),
                    }),
                    Message::Assistant {
                        content,
                        tool_calls,
                    } => {
                        let mut parts = Vec::new();
                        if !content.is_empty() {
                            parts.push(serde_json::json!({"text": content}));
                        }
                        for tool_call in tool_calls {
                            parts.push(serde_json::json!({
                                "functionCall": {
                                    "name": tool_call.name,
                                    "args": tool_call.arguments,
                                }
                            }));
                        }
                        serde_json::json!({
                            "role": "model",
                            "parts": parts,
                        })
                    }
                    Message::ToolResult {
                        tool_use_id,
                        content,
                    } => serde_json::json!({
                        "role": "tool",
                        "parts": [{
                            "functionResponse": {
                                "name": tool_use_id,
                                "response": {"content": content},
                            }
                        }],
                    }),
                })
                .collect();
            (contents, system_instruction)
        } else {
            let mut parts = vec![serde_json::json!({"text": prompt})];
            if let Some(images) = images.filter(|images| !images.is_empty()) {
                parts.extend(images.iter().map(google_part_from_block));
            }
            (
                vec![serde_json::json!({
                    "role": "user",
                    "parts": parts,
                })],
                system_instruction,
            )
        }
    }

    fn build_tools(&self, tools: &[Value]) -> Vec<Value> {
        tools
            .iter()
            .filter_map(|t| {
                let func = t.get("function").unwrap_or(t);
                Some(serde_json::json!({
                    "name": func.get("name")?,
                    "description": func.get("description").and_then(|d| d.as_str()).unwrap_or(""),
                    "parameters": func.get("parameters").cloned().unwrap_or(serde_json::json!({"type": "object"})),
                }))
            })
            .collect()
    }

    fn supports_streaming(&self) -> bool {
        false
    }

    fn supports_video(&self) -> bool {
        // Google Gemini's generateContent API accepts video parts via
        // fileData (URL) or inlineData (base64) — the only mainstream
        // multimodal provider that does, as of this writeup.
        true
    }

    fn supports_audio(&self) -> bool {
        // Gemini accepts audio natively via fileData/inlineData on
        // the same contents[] array as text/image/video.
        true
    }

    fn protocol_name(&self) -> &'static str {
        "google-gemini"
    }
}

fn google_part_from_block(block: &ContentBlock) -> Value {
    match block {
        ContentBlock::Text { text } => serde_json::json!({"text": text}),
        ContentBlock::ImageBase64 { data, media_type } => serde_json::json!({
            "inlineData": {
                "mimeType": media_type,
                "data": data,
            }
        }),
        ContentBlock::ImageUrl { url, .. } => serde_json::json!({
            "fileData": {
                "mimeType": infer_mime_type_from_url(url),
                "fileUri": url,
            }
        }),
        // Google's Gemini API does accept video fileData/inlineData, so
        // emit the native representation when we have a concrete source.
        // Base64 inline video gets the raw `data` pass-through, keyed to
        // the caller's media_type.
        ContentBlock::VideoPath { path, .. } => serde_json::json!({
            "fileData": {
                "mimeType": "video/mp4",
                "fileUri": format!("file://{path}"),
            }
        }),
        ContentBlock::VideoUrl { url, .. } => serde_json::json!({
            "fileData": {
                "mimeType": "video/mp4",
                "fileUri": url,
            }
        }),
        ContentBlock::VideoBase64 {
            data, media_type, ..
        } => serde_json::json!({
            "inlineData": {
                "mimeType": media_type,
                "data": data,
            }
        }),
        // Gemini accepts audio as fileData or inlineData, same shape
        // as video. Default mime type is wav; callers who supply an
        // explicit media_type via AudioBase64 get exact pass-through.
        ContentBlock::AudioPath { path, .. } => serde_json::json!({
            "fileData": {
                "mimeType": "audio/wav",
                "fileUri": format!("file://{path}"),
            }
        }),
        ContentBlock::AudioUrl { url, .. } => serde_json::json!({
            "fileData": {
                "mimeType": "audio/wav",
                "fileUri": url,
            }
        }),
        ContentBlock::AudioBase64 {
            data, media_type, ..
        } => serde_json::json!({
            "inlineData": {
                "mimeType": media_type,
                "data": data,
            }
        }),
    }
}

fn infer_mime_type_from_url(url: &str) -> &'static str {
    let lower = url.to_ascii_lowercase();
    if lower.ends_with(".png") {
        "image/png"
    } else if lower.ends_with(".webp") {
        "image/webp"
    } else if lower.ends_with(".heic") {
        "image/heic"
    } else if lower.ends_with(".heif") {
        "image/heif"
    } else {
        "image/jpeg"
    }
}

// ---------------------------------------------------------------------------
// Factory
// ---------------------------------------------------------------------------

/// Get the appropriate protocol handler for an API protocol.
pub fn handler_for(protocol: crate::schema::ApiProtocol) -> Box<dyn ProtocolHandler> {
    match protocol {
        crate::schema::ApiProtocol::OpenAiCompat | crate::schema::ApiProtocol::OpenAiResponses => {
            Box::new(OpenAiHandler)
        }
        crate::schema::ApiProtocol::Anthropic => Box::new(AnthropicHandler),
        crate::schema::ApiProtocol::Google => Box::new(GoogleHandler),
        crate::schema::ApiProtocol::AzureOpenAi => Box::new(AzureOpenAiHandler),
    }
}

// ---------------------------------------------------------------------------
// Azure OpenAI handler — same request format as OpenAI but different auth/URL
// ---------------------------------------------------------------------------

pub struct AzureOpenAiHandler;

impl ProtocolHandler for AzureOpenAiHandler {
    fn endpoint_path(&self) -> &str {
        // Azure uses deployment-based URLs; the actual path is built in remote.rs
        "/openai/deployments"
    }

    fn auth_headers(&self, api_key: &str) -> Vec<(String, String)> {
        vec![
            ("api-key".into(), api_key.to_string()),
            ("Content-Type".into(), "application/json".into()),
        ]
    }

    fn build_request_body(&self, req: &ApiRequest) -> Value {
        // Same as OpenAI format
        OpenAiHandler.build_request_body(req)
    }

    fn parse_response(&self, body: &str) -> Result<ApiResponse, crate::InferenceError> {
        OpenAiHandler.parse_response(body)
    }

    fn parse_stream_event(&self, event_type: &str, data: &str) -> Vec<crate::stream::StreamEvent> {
        OpenAiHandler.parse_stream_event(event_type, data)
    }

    fn build_messages(
        &self,
        messages: &[Message],
        prompt: &str,
        context: Option<&str>,
        images: Option<&[ContentBlock]>,
    ) -> (Vec<Value>, Option<String>) {
        OpenAiHandler.build_messages(messages, prompt, context, images)
    }

    fn build_tools(&self, tools: &[Value]) -> Vec<Value> {
        OpenAiHandler.build_tools(tools)
    }

    fn protocol_name(&self) -> &'static str {
        "azure-openai"
    }
}

// ---------------------------------------------------------------------------
// Google URL builder (special case — uses query param auth)
// ---------------------------------------------------------------------------

/// Build the full Google API URL with model and key.
pub fn google_url(endpoint: &str, model: &str, api_key: &str) -> String {
    let base = endpoint.trim_end_matches('/');
    format!(
        "{}/v1beta/models/{}:generateContent?key={}",
        base, model, api_key
    )
}

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

    #[test]
    fn openai_single_turn_messages() {
        let handler = OpenAiHandler;
        let (msgs, system) = handler.build_messages(&[], "Hello", Some("Be helpful"), None);
        assert_eq!(msgs.len(), 2);
        assert_eq!(msgs[0]["role"], "system");
        assert_eq!(msgs[1]["content"], "Hello");
        assert!(system.is_none()); // system is in messages
    }

    #[test]
    fn openai_multi_turn_messages() {
        let handler = OpenAiHandler;
        let messages = vec![
            Message::User {
                content: "Hi".into(),
            },
            Message::Assistant {
                content: "Hello!".into(),
                tool_calls: vec![],
            },
            Message::User {
                content: "Search for X".into(),
            },
        ];
        let (msgs, _) = handler.build_messages(&messages, "", None, None);
        assert_eq!(msgs.len(), 3);
        assert_eq!(msgs[2]["content"], "Search for X");
    }

    #[test]
    fn openai_tool_call_messages() {
        let handler = OpenAiHandler;
        let tc = ToolCall {
            id: None,
            name: "search".into(),
            arguments: [("q".into(), Value::String("rust".into()))].into(),
        };
        let messages = vec![
            Message::User {
                content: "Search".into(),
            },
            Message::Assistant {
                content: String::new(),
                tool_calls: vec![tc],
            },
            Message::ToolResult {
                tool_use_id: "call_0".into(),
                content: "found it".into(),
            },
        ];
        let (msgs, _) = handler.build_messages(&messages, "", None, None);
        assert_eq!(msgs.len(), 3);
        assert!(msgs[1].get("tool_calls").is_some());
        assert_eq!(msgs[2]["role"], "tool");
    }

    #[test]
    fn anthropic_system_separate() {
        let handler = AnthropicHandler;
        let (msgs, system) = handler.build_messages(&[], "Hello", Some("Be helpful"), None);
        assert_eq!(msgs.len(), 1); // only user message
        assert_eq!(system, Some("Be helpful".into()));
    }

    #[test]
    fn anthropic_tool_use_format() {
        let handler = AnthropicHandler;
        let tc = ToolCall {
            id: None,
            name: "search".into(),
            arguments: [("q".into(), Value::String("test".into()))].into(),
        };
        let messages = vec![
            Message::User {
                content: "Search".into(),
            },
            Message::Assistant {
                content: String::new(),
                tool_calls: vec![tc],
            },
            Message::ToolResult {
                tool_use_id: "toolu_0".into(),
                content: "result".into(),
            },
        ];
        let (msgs, _) = handler.build_messages(&messages, "", None, None);
        assert_eq!(msgs.len(), 3);
        let assistant_content = msgs[1].get("content").unwrap().as_array().unwrap();
        assert_eq!(assistant_content[0]["type"], "tool_use");
        let user_content = msgs[2].get("content").unwrap().as_array().unwrap();
        assert_eq!(user_content[0]["type"], "tool_result");
    }

    #[test]
    fn anthropic_thinking_in_request() {
        let handler = AnthropicHandler;
        let req = ApiRequest {
            model: "claude".into(),
            messages: vec![serde_json::json!({"role": "user", "content": "plan"})],
            system: None,
            temperature: 0.7,
            max_tokens: 4096,
            tools: None,
            tool_choice: None,
            parallel_tool_calls: None,
            stream: false,
            budget_tokens: 8000,
            cache_control: false,
            response_format: None,
        };
        let body = handler.build_request_body(&req);
        assert!(body.get("thinking").is_some());
        assert_eq!(body["thinking"]["budget_tokens"], 8000);
        assert!(body.get("temperature").is_none()); // removed when thinking enabled
    }

    fn empty_request(model: &str) -> ApiRequest {
        ApiRequest {
            model: model.into(),
            messages: vec![serde_json::json!({"role": "user", "content": "hi"})],
            system: None,
            temperature: 0.7,
            max_tokens: 256,
            tools: None,
            tool_choice: None,
            parallel_tool_calls: None,
            stream: false,
            budget_tokens: 0,
            cache_control: false,
            response_format: None,
        }
    }

    #[test]
    fn openai_streaming_request_carries_stream_options_include_usage() {
        // Streaming refactor (#125): build_request_body is now the
        // single source of truth for request shape. When stream=true
        // it must add `stream_options.include_usage = true` so the
        // streaming tail-chunk carries real prompt/completion token
        // counts. Without this, RemoteBackend::generate_stream loses
        // the usage signal that #79's TokenUsage parser depends on.
        let handler = OpenAiHandler;
        let mut req = empty_request("gpt-5");
        req.stream = true;
        let body = handler.build_request_body(&req);
        assert_eq!(body["stream"], serde_json::json!(true));
        assert_eq!(
            body["stream_options"]["include_usage"],
            serde_json::json!(true),
            "streaming bodies must include `stream_options.include_usage` so usage flows back"
        );
    }

    #[test]
    fn openai_non_streaming_request_omits_stream_options() {
        // Inverse — non-streaming requests must not carry stream
        // fields at all (some compatible providers reject the
        // unknown `stream_options` outside a streaming call).
        let handler = OpenAiHandler;
        let req = empty_request("gpt-5");
        let body = handler.build_request_body(&req);
        assert!(body.get("stream").is_none());
        assert!(body.get("stream_options").is_none());
    }

    #[test]
    fn anthropic_streaming_request_marks_stream_true() {
        // Sanity check that Anthropic's body also gets stream=true
        // through the same build_request_body abstraction. Anthropic
        // doesn't have a `stream_options` field — it emits usage in
        // its `message_start` and `message_delta` SSE frames already.
        let handler = AnthropicHandler;
        let mut req = empty_request("claude-opus-4-7");
        req.stream = true;
        let body = handler.build_request_body(&req);
        assert_eq!(body["stream"], serde_json::json!(true));
        assert!(
            body.get("stream_options").is_none(),
            "Anthropic has no stream_options field; usage flows via SSE frames"
        );
    }

    #[test]
    fn openai_emits_strict_json_schema_response_format() {
        let mut req = empty_request("gpt-5");
        req.response_format = Some(ResponseFormat::JsonSchema {
            schema: serde_json::json!({
                "type": "object",
                "properties": {"answer": {"type": "string"}},
                "required": ["answer"]
            }),
            strict: true,
            name: Some("answer_schema".into()),
        });
        let body = OpenAiHandler.build_request_body(&req);
        let rf = body.get("response_format").expect("response_format set");
        assert_eq!(rf["type"], "json_schema");
        assert_eq!(rf["json_schema"]["name"], "answer_schema");
        assert_eq!(rf["json_schema"]["strict"], true);
        assert_eq!(rf["json_schema"]["schema"]["required"][0], "answer");
    }

    #[test]
    fn openai_emits_json_object_when_no_schema() {
        let mut req = empty_request("gpt-4o");
        req.response_format = Some(ResponseFormat::JsonObject);
        let body = OpenAiHandler.build_request_body(&req);
        assert_eq!(body["response_format"]["type"], "json_object");
        // No json_schema sub-object on the looser variant.
        assert!(body["response_format"].get("json_schema").is_none());
    }

    #[test]
    fn openai_omits_response_format_when_none() {
        let req = empty_request("gpt-4o");
        let body = OpenAiHandler.build_request_body(&req);
        assert!(body.get("response_format").is_none());
    }

    #[test]
    fn google_emits_response_mime_and_schema() {
        let mut req = empty_request("gemini-2.5-pro");
        req.response_format = Some(ResponseFormat::JsonSchema {
            schema: serde_json::json!({"type": "object"}),
            strict: false,
            name: None,
        });
        let body = GoogleHandler.build_request_body(&req);
        let cfg = body.get("generationConfig").expect("generationConfig");
        assert_eq!(cfg["responseMimeType"], "application/json");
        assert_eq!(cfg["responseSchema"]["type"], "object");
    }

    #[test]
    fn google_json_object_skips_schema() {
        let mut req = empty_request("gemini-2.5-pro");
        req.response_format = Some(ResponseFormat::JsonObject);
        let body = GoogleHandler.build_request_body(&req);
        let cfg = body.get("generationConfig").expect("generationConfig");
        assert_eq!(cfg["responseMimeType"], "application/json");
        assert!(cfg.get("responseSchema").is_none());
    }

    #[test]
    fn anthropic_does_not_emit_response_format_field() {
        // No native response_format field on Anthropic — request must
        // fly without it (handler logs a warning, documented in the
        // ResponseFormat doc comment).
        let mut req = empty_request("claude-opus-4-7");
        req.response_format = Some(ResponseFormat::JsonSchema {
            schema: serde_json::json!({"type": "object"}),
            strict: true,
            name: None,
        });
        let body = AnthropicHandler.build_request_body(&req);
        assert!(body.get("response_format").is_none());
        assert!(body.get("responseSchema").is_none());
    }

    #[test]
    fn openai_tools_wrapped() {
        let handler = OpenAiHandler;
        let tools = vec![serde_json::json!({"name": "search", "parameters": {}})];
        let built = handler.build_tools(&tools);
        assert_eq!(built[0]["type"], "function");
        assert!(built[0].get("function").is_some());
    }

    #[test]
    fn openai_request_preserves_required_tool_choice_and_parallel_tool_calls() {
        let handler = OpenAiHandler;
        let req = ApiRequest {
            model: "gpt-5.4-mini".into(),
            messages: vec![serde_json::json!({"role": "user", "content": "extract"})],
            system: None,
            temperature: 0.0,
            max_tokens: 1024,
            tools: Some(vec![serde_json::json!({
                "type": "function",
                "function": {
                    "name": "extract_action_items",
                    "parameters": {"type": "object", "additionalProperties": false}
                }
            })]),
            tool_choice: Some("required".into()),
            parallel_tool_calls: Some(false),
            stream: false,
            budget_tokens: 0,
            cache_control: false,
            response_format: None,
        };

        let body = handler.build_request_body(&req);

        assert_eq!(body["tool_choice"], "required");
        assert_eq!(body["parallel_tool_calls"], false);
        assert_eq!(body["tools"][0]["function"]["name"], "extract_action_items");
    }

    #[test]
    fn anthropic_tools_format() {
        let handler = AnthropicHandler;
        let tools = vec![
            serde_json::json!({"function": {"name": "search", "description": "Search", "parameters": {"type": "object"}}}),
        ];
        let built = handler.build_tools(&tools);
        assert_eq!(built[0]["name"], "search");
        assert!(built[0].get("input_schema").is_some());
    }

    #[test]
    fn google_no_streaming() {
        let handler = GoogleHandler;
        assert!(!handler.supports_streaming());
    }

    #[test]
    fn anthropic_supports_thinking() {
        let handler = AnthropicHandler;
        assert!(handler.supports_thinking());
    }

    #[test]
    fn handler_factory() {
        use crate::schema::ApiProtocol;
        let h = handler_for(ApiProtocol::Anthropic);
        assert!(h.supports_thinking());

        let h = handler_for(ApiProtocol::OpenAiCompat);
        assert!(!h.supports_thinking());
    }

    #[test]
    fn openai_parse_text_response() {
        let handler = OpenAiHandler;
        let body = r#"{"choices":[{"message":{"content":"Hello world"}}]}"#;
        let resp = handler.parse_response(body).unwrap();
        assert_eq!(resp.text, "Hello world");
        assert!(resp.tool_calls.is_empty());
    }

    #[test]
    fn openai_parse_usage() {
        let handler = OpenAiHandler;
        let body = r#"{"choices":[{"message":{"content":"Hi"}}],"usage":{"prompt_tokens":10,"completion_tokens":5,"total_tokens":15}}"#;
        let resp = handler.parse_response(body).unwrap();
        let usage = resp.usage.unwrap();
        assert_eq!(usage.prompt_tokens, 10);
        assert_eq!(usage.completion_tokens, 5);
        assert_eq!(usage.total_tokens, 15);
    }

    #[test]
    fn openai_parse_multiple_tool_calls() {
        let handler = OpenAiHandler;
        let body = r#"{"choices":[{"message":{"content":"","tool_calls":[{"function":{"name":"read_file","arguments":"{\"path\":\"a.rs\"}"}},{"function":{"name":"read_file","arguments":"{\"path\":\"b.rs\"}"}}]}}]}"#;
        let resp = handler.parse_response(body).unwrap();
        assert_eq!(resp.tool_calls.len(), 2);
        assert_eq!(resp.tool_calls[0].name, "read_file");
        assert_eq!(resp.tool_calls[1].name, "read_file");
    }

    #[test]
    fn anthropic_parse_tool_response() {
        let handler = AnthropicHandler;
        let body = r#"{"content":[{"type":"text","text":"Let me search"},{"type":"tool_use","name":"search","id":"t1","input":{"q":"rust"}}]}"#;
        let resp = handler.parse_response(body).unwrap();
        assert_eq!(resp.text, "Let me search");
        assert_eq!(resp.tool_calls.len(), 1);
        assert_eq!(resp.tool_calls[0].name, "search");
    }

    #[test]
    fn anthropic_parse_usage() {
        let handler = AnthropicHandler;
        let body = r#"{"content":[{"type":"text","text":"Hi"}],"usage":{"input_tokens":12,"output_tokens":3}}"#;
        let resp = handler.parse_response(body).unwrap();
        let usage = resp.usage.unwrap();
        assert_eq!(usage.prompt_tokens, 12);
        assert_eq!(usage.completion_tokens, 3);
        assert_eq!(usage.total_tokens, 15);
    }

    #[test]
    fn anthropic_parse_multiple_tool_calls() {
        let handler = AnthropicHandler;
        let body = r#"{"content":[{"type":"text","text":"I'll read both files"},{"type":"tool_use","name":"read","id":"t1","input":{"path":"a.rs"}},{"type":"tool_use","name":"read","id":"t2","input":{"path":"b.rs"}}]}"#;
        let resp = handler.parse_response(body).unwrap();
        assert_eq!(resp.text, "I'll read both files");
        assert_eq!(resp.tool_calls.len(), 2);
        assert_eq!(resp.tool_calls[0].name, "read");
        assert_eq!(resp.tool_calls[1].name, "read");
    }

    #[test]
    fn anthropic_cache_control_system_prompt() {
        let handler = AnthropicHandler;
        let req = ApiRequest {
            model: "claude".into(),
            messages: vec![serde_json::json!({"role": "user", "content": "hello"})],
            system: Some("You are helpful.".into()),
            temperature: 0.7,
            max_tokens: 1024,
            tools: None,
            tool_choice: None,
            parallel_tool_calls: None,
            stream: false,
            budget_tokens: 0,
            cache_control: true,
            response_format: None,
        };
        let body = handler.build_request_body(&req);
        // System should be an array with cache_control block
        let system = body.get("system").unwrap();
        assert!(system.is_array());
        let blocks = system.as_array().unwrap();
        assert_eq!(blocks.len(), 1);
        assert_eq!(blocks[0]["type"], "text");
        assert_eq!(blocks[0]["text"], "You are helpful.");
        assert_eq!(blocks[0]["cache_control"]["type"], "ephemeral");
    }

    #[test]
    fn anthropic_cache_control_disabled() {
        let handler = AnthropicHandler;
        let req = ApiRequest {
            model: "claude".into(),
            messages: vec![serde_json::json!({"role": "user", "content": "hello"})],
            system: Some("You are helpful.".into()),
            temperature: 0.7,
            max_tokens: 1024,
            tools: None,
            tool_choice: None,
            parallel_tool_calls: None,
            stream: false,
            budget_tokens: 0,
            cache_control: false,
            response_format: None,
        };
        let body = handler.build_request_body(&req);
        // System should be a plain string when cache_control is false
        assert!(body.get("system").unwrap().is_string());
    }

    #[test]
    fn anthropic_cache_control_tools() {
        let handler = AnthropicHandler;
        let tools = vec![
            serde_json::json!({"name": "search", "description": "Search", "input_schema": {"type": "object"}}),
            serde_json::json!({"name": "read", "description": "Read file", "input_schema": {"type": "object"}}),
        ];
        let req = ApiRequest {
            model: "claude".into(),
            messages: vec![serde_json::json!({"role": "user", "content": "hello"})],
            system: None,
            temperature: 0.7,
            max_tokens: 1024,
            tools: Some(tools),
            tool_choice: None,
            parallel_tool_calls: None,
            stream: false,
            budget_tokens: 0,
            cache_control: true,
            response_format: None,
        };
        let body = handler.build_request_body(&req);
        let tools_arr = body["tools"].as_array().unwrap();
        // Only the last tool should have cache_control
        assert!(tools_arr[0].get("cache_control").is_none());
        assert_eq!(tools_arr[1]["cache_control"]["type"], "ephemeral");
    }

    #[test]
    fn anthropic_beta_header_included() {
        let handler = AnthropicHandler;
        let headers = handler.auth_headers("test-key");
        let beta = headers.iter().find(|(k, _)| k == "anthropic-beta");
        assert!(beta.is_some());
        assert_eq!(beta.unwrap().1, "prompt-caching-2024-07-31");
    }

    #[test]
    fn google_parse_response() {
        let handler = GoogleHandler;
        let body = r#"{"candidates":[{"content":{"parts":[{"text":"Hello from Gemini"}]}}]}"#;
        let resp = handler.parse_response(body).unwrap();
        assert_eq!(resp.text, "Hello from Gemini");
    }

    #[test]
    fn google_tools_format() {
        let handler = GoogleHandler;
        let tools = vec![serde_json::json!({
            "function": {
                "name": "search",
                "description": "Search docs",
                "parameters": {"type": "object"}
            }
        })];
        let built = handler.build_tools(&tools);
        assert_eq!(built[0]["name"], "search");
        assert!(built[0].get("parameters").is_some());
    }

    #[test]
    fn google_builds_multimodal_messages() {
        let handler = GoogleHandler;
        let messages = vec![Message::UserMultimodal {
            content: vec![
                ContentBlock::Text {
                    text: "Describe this image.".to_string(),
                },
                ContentBlock::ImageUrl {
                    url: "https://example.com/cat.jpg".to_string(),
                    detail: "auto".to_string(),
                },
            ],
        }];
        let (msgs, system) = handler.build_messages(&messages, "", Some("Be concise"), None);
        assert_eq!(msgs.len(), 1);
        assert_eq!(msgs[0]["role"], "user");
        let parts = msgs[0]["parts"].as_array().unwrap();
        assert_eq!(parts[0]["text"], "Describe this image.");
        assert!(parts[1].get("fileData").is_some());
        assert_eq!(system, Some("Be concise".to_string()));
    }

    #[test]
    fn google_request_body_includes_tools_and_system() {
        let handler = GoogleHandler;
        let req = ApiRequest {
            model: "gemini-2.5-flash".into(),
            messages: vec![serde_json::json!({
                "role": "user",
                "parts": [{"text": "Find the file and summarize it."}],
            })],
            system: Some("Use tools when needed.".into()),
            temperature: 0.2,
            max_tokens: 512,
            tools: Some(vec![serde_json::json!({
                "name": "search",
                "description": "Search files",
                "parameters": {"type": "object"}
            })]),
            tool_choice: None,
            parallel_tool_calls: None,
            stream: false,
            budget_tokens: 0,
            cache_control: false,
            response_format: None,
        };
        let body = handler.build_request_body(&req);
        assert!(body.get("systemInstruction").is_some());
        assert!(body.get("tools").is_some());
        assert_eq!(body["toolConfig"]["functionCallingConfig"]["mode"], "AUTO");
        assert_eq!(body["generationConfig"]["maxOutputTokens"], 512);
    }

    #[test]
    fn google_parse_multiple_tool_calls_and_usage() {
        let handler = GoogleHandler;
        let body = r#"{
            "candidates":[{"content":{"parts":[
                {"text":"Let me do that."},
                {"functionCall":{"name":"search","args":{"q":"rust"}}},
                {"functionCall":{"name":"read_file","args":{"path":"src/lib.rs"}}}
            ]}}],
            "usageMetadata":{"promptTokenCount":10,"candidatesTokenCount":4,"totalTokenCount":14}
        }"#;
        let resp = handler.parse_response(body).unwrap();
        assert_eq!(resp.text, "Let me do that.");
        assert_eq!(resp.tool_calls.len(), 2);
        assert_eq!(resp.tool_calls[0].name, "search");
        assert_eq!(resp.tool_calls[1].name, "read_file");
        let usage = resp.usage.unwrap();
        assert_eq!(usage.prompt_tokens, 10);
        assert_eq!(usage.completion_tokens, 4);
        assert_eq!(usage.total_tokens, 14);
    }

    #[test]
    fn openai_vision_message() {
        let handler = OpenAiHandler;
        let messages = vec![Message::UserMultimodal {
            content: vec![
                ContentBlock::Text {
                    text: "What is in this image?".to_string(),
                },
                ContentBlock::ImageUrl {
                    url: "https://example.com/cat.jpg".to_string(),
                    detail: "auto".to_string(),
                },
            ],
        }];
        let (msgs, _) = handler.build_messages(&messages, "", None, None);
        assert_eq!(msgs.len(), 1);
        let content = msgs[0]["content"].as_array().unwrap();
        assert_eq!(content.len(), 2);
        assert_eq!(content[0]["type"], "text");
        assert_eq!(content[1]["type"], "image_url");
    }

    #[test]
    fn anthropic_vision_message() {
        let handler = AnthropicHandler;
        let messages = vec![Message::UserMultimodal {
            content: vec![
                ContentBlock::Text {
                    text: "Describe this.".to_string(),
                },
                ContentBlock::ImageBase64 {
                    data: "iVBOR...".to_string(),
                    media_type: "image/png".to_string(),
                },
            ],
        }];
        let (msgs, _) = handler.build_messages(&messages, "", None, None);
        assert_eq!(msgs.len(), 1);
        let content = msgs[0]["content"].as_array().unwrap();
        assert_eq!(content[0]["type"], "text");
        assert_eq!(content[1]["type"], "image");
        assert_eq!(content[1]["source"]["type"], "base64");
    }

    #[test]
    fn openai_single_turn_images() {
        let handler = OpenAiHandler;
        let images = vec![ContentBlock::ImageUrl {
            url: "https://example.com/cat.jpg".to_string(),
            detail: "high".to_string(),
        }];
        let (msgs, _) = handler.build_messages(&[], "Describe this image", None, Some(&images));
        let content = msgs[0]["content"].as_array().unwrap();
        assert_eq!(content.len(), 2);
        assert_eq!(content[0]["type"], "text");
        assert_eq!(content[1]["type"], "image_url");
    }

    #[test]
    fn anthropic_single_turn_images() {
        let handler = AnthropicHandler;
        let images = vec![ContentBlock::ImageBase64 {
            data: "iVBOR...".to_string(),
            media_type: "image/png".to_string(),
        }];
        let (msgs, _) = handler.build_messages(&[], "Describe this image", None, Some(&images));
        let content = msgs[0]["content"].as_array().unwrap();
        assert_eq!(content.len(), 2);
        assert_eq!(content[0]["type"], "text");
        assert_eq!(content[1]["type"], "image");
        assert_eq!(content[1]["source"]["type"], "base64");
    }
}