link-assistant-router 0.91.0

//! Native Gemini ↔ `OpenAI` Chat Completions translation.
//!
//! The Gemini CLI speaks only `generateContent`/`streamGenerateContent`. Its
//! requests must therefore reach Codex, Claude and Qwen subscriptions through
//! the router's `OpenAI` Chat Completions path, which every other namespace
//! already shares. This module owns that translation in both directions so the
//! HTTP handler in [`crate::gemini`] stays a thin router.
//!
//! The translation is intentionally lossless for the parts Gemini CLI actually
//! uses: system instruction, multi-turn text, client function declarations,
//! function calls, function results, generation config and usage metadata.

use serde_json::{Map, Value, json};

/// Gemini part-role for assistant turns.
const MODEL_ROLE: &str = "model";

/// Translate a native Gemini `GenerateContentRequest` into an `OpenAI` Chat
/// Completions body for `model`.
#[must_use]
pub fn gemini_request_to_chat(model: &str, request: &Value) -> Value {
    let mut messages: Vec<Value> = Vec::new();

    if let Some(text) = system_instruction_text(request)
        && !text.is_empty()
    {
        messages.push(json!({ "role": "system", "content": text }));
    }

    // Gemini identifies a tool result only by function name, while OpenAI
    // requires the id of the call it answers. Track the calls we emitted so a
    // later `functionResponse` can be paired with the newest matching id.
    let mut pending_calls: Vec<(String, String)> = Vec::new();
    let mut call_counter = 0_usize;

    for content in request
        .get("contents")
        .and_then(Value::as_array)
        .map_or(&[][..], Vec::as_slice)
    {
        let role = content
            .get("role")
            .and_then(Value::as_str)
            .unwrap_or("user");
        let parts = content
            .get("parts")
            .and_then(Value::as_array)
            .map_or(&[][..], Vec::as_slice);

        let mut text = String::new();
        let mut tool_calls: Vec<Value> = Vec::new();

        for part in parts {
            if let Some(chunk) = part.get("text").and_then(Value::as_str) {
                text.push_str(chunk);
            } else if let Some(call) = part.get("functionCall") {
                let name = call
                    .get("name")
                    .and_then(Value::as_str)
                    .unwrap_or_default()
                    .to_string();
                let id = call
                    .get("id")
                    .and_then(Value::as_str)
                    .map_or_else(|| format!("call_{call_counter}"), ToString::to_string);
                call_counter += 1;
                let arguments = call
                    .get("args")
                    .map_or_else(|| "{}".to_string(), ToString::to_string);
                pending_calls.push((name.clone(), id.clone()));
                tool_calls.push(json!({
                    "id": id,
                    "type": "function",
                    "function": { "name": name, "arguments": arguments },
                }));
            } else if let Some(response) = part.get("functionResponse") {
                let name = response
                    .get("name")
                    .and_then(Value::as_str)
                    .unwrap_or_default();
                let id = take_call_id(&mut pending_calls, name)
                    .unwrap_or_else(|| format!("call_{name}"));
                let content = response
                    .get("response")
                    .map_or_else(String::new, ToString::to_string);
                messages.push(json!({
                    "role": "tool",
                    "tool_call_id": id,
                    "content": content,
                }));
            }
        }

        if role == MODEL_ROLE {
            if !text.is_empty() || !tool_calls.is_empty() {
                let mut message = Map::new();
                message.insert("role".into(), json!("assistant"));
                message.insert("content".into(), json!(text));
                if !tool_calls.is_empty() {
                    message.insert("tool_calls".into(), Value::Array(tool_calls));
                }
                messages.push(Value::Object(message));
            }
        } else if !text.is_empty() {
            messages.push(json!({ "role": "user", "content": text }));
        }
    }

    let mut chat = json!({ "model": model, "messages": messages });

    if let Some(config) = request.get("generationConfig") {
        for (gemini_key, openai_key) in [
            ("maxOutputTokens", "max_tokens"),
            ("temperature", "temperature"),
            ("topP", "top_p"),
            ("stopSequences", "stop"),
        ] {
            if let Some(value) = config.get(gemini_key) {
                chat[openai_key] = value.clone();
            }
        }
    }

    let tools = translate_tools(request.get("tools"));
    if !tools.is_empty() {
        chat["tools"] = Value::Array(tools);
    }
    if let Some(choice) = translate_tool_choice(request.get("toolConfig")) {
        chat["tool_choice"] = choice;
    }
    chat
}

/// Remove and return the newest pending call id emitted for `name`.
fn take_call_id(pending: &mut Vec<(String, String)>, name: &str) -> Option<String> {
    let index = pending.iter().rposition(|(call, _)| call == name)?;
    Some(pending.remove(index).1)
}

fn system_instruction_text(request: &Value) -> Option<String> {
    let instruction = request
        .get("systemInstruction")
        .or_else(|| request.get("system_instruction"))?;
    let mut text = String::new();
    if let Some(parts) = instruction.get("parts").and_then(Value::as_array) {
        for part in parts {
            if let Some(chunk) = part.get("text").and_then(Value::as_str) {
                text.push_str(chunk);
            }
        }
    } else if let Some(chunk) = instruction.as_str() {
        text.push_str(chunk);
    }
    Some(text)
}

/// Translate Gemini tool declarations into `OpenAI` tool definitions.
///
/// Client tools arrive as `functionDeclarations`; Google's server-side search
/// tool arrives as an empty `googleSearch`/`google_search_retrieval` object and
/// maps onto the router's provider-agnostic `web_search` server tool, which
/// [`crate::capabilities`] gates per provider.
fn translate_tools(tools: Option<&Value>) -> Vec<Value> {
    let mut translated = Vec::new();
    for entry in tools
        .and_then(Value::as_array)
        .map_or(&[][..], Vec::as_slice)
    {
        if let Some(declarations) = entry
            .get("functionDeclarations")
            .or_else(|| entry.get("function_declarations"))
            .and_then(Value::as_array)
        {
            for declaration in declarations {
                let mut function = Map::new();
                for key in ["name", "description"] {
                    if let Some(value) = declaration.get(key) {
                        function.insert(key.into(), value.clone());
                    }
                }
                if let Some(parameters) = declaration
                    .get("parameters")
                    .or_else(|| declaration.get("parametersJsonSchema"))
                {
                    function.insert("parameters".into(), parameters.clone());
                }
                translated.push(json!({ "type": "function", "function": function }));
            }
        }
        if entry.get("googleSearch").is_some()
            || entry.get("google_search").is_some()
            || entry.get("googleSearchRetrieval").is_some()
            || entry.get("google_search_retrieval").is_some()
        {
            translated.push(json!({ "type": "web_search" }));
        }
    }
    translated
}

fn translate_tool_choice(config: Option<&Value>) -> Option<Value> {
    let config = config?;
    let mode = config
        .get("functionCallingConfig")
        .or_else(|| config.get("function_calling_config"))?
        .get("mode")
        .and_then(Value::as_str)?;
    match mode.to_ascii_uppercase().as_str() {
        "ANY" => Some(json!("required")),
        "NONE" => Some(json!("none")),
        "AUTO" => Some(json!("auto")),
        _ => None,
    }
}

/// Translate an `OpenAI` Chat Completion into a native Gemini
/// `GenerateContentResponse`.
#[must_use]
pub fn chat_to_gemini_response(chat: &Value, model: &str) -> Value {
    let choice = chat
        .get("choices")
        .and_then(Value::as_array)
        .and_then(|choices| choices.first());
    let message = choice.and_then(|choice| choice.get("message"));

    let mut parts: Vec<Value> = Vec::new();
    if let Some(text) = message
        .and_then(|message| message.get("content"))
        .and_then(Value::as_str)
        && !text.is_empty()
    {
        parts.push(json!({ "text": text }));
    }
    for call in message
        .and_then(|message| message.get("tool_calls"))
        .and_then(Value::as_array)
        .map_or(&[][..], Vec::as_slice)
    {
        let function = call.get("function");
        let name = function
            .and_then(|function| function.get("name"))
            .and_then(Value::as_str)
            .unwrap_or_default();
        let args = function
            .and_then(|function| function.get("arguments"))
            .and_then(Value::as_str)
            .and_then(|arguments| serde_json::from_str::<Value>(arguments).ok())
            .unwrap_or_else(|| json!({}));
        let mut call_part = json!({ "name": name, "args": args });
        if let Some(id) = call.get("id").and_then(Value::as_str) {
            call_part["id"] = json!(id);
        }
        parts.push(json!({ "functionCall": call_part }));
    }

    let finish_reason = choice
        .and_then(|choice| choice.get("finish_reason"))
        .and_then(Value::as_str)
        .map_or("STOP", map_finish_reason);

    let usage = chat.get("usage");
    let prompt_tokens = usage
        .and_then(|usage| usage.get("prompt_tokens"))
        .and_then(Value::as_u64)
        .unwrap_or(0);
    let completion_tokens = usage
        .and_then(|usage| usage.get("completion_tokens"))
        .and_then(Value::as_u64)
        .unwrap_or(0);

    json!({
        "candidates": [{
            "index": 0,
            "content": { "role": MODEL_ROLE, "parts": parts },
            "finishReason": finish_reason,
        }],
        "usageMetadata": {
            "promptTokenCount": prompt_tokens,
            "candidatesTokenCount": completion_tokens,
            "totalTokenCount": prompt_tokens + completion_tokens,
        },
        "modelVersion": model,
    })
}

fn map_finish_reason(openai: &str) -> &'static str {
    match openai {
        "length" => "MAX_TOKENS",
        "content_filter" => "SAFETY",
        _ => "STOP",
    }
}

/// Translate an `OpenAI`-shaped error body into Gemini's error envelope.
#[must_use]
pub fn openai_error_to_gemini(status: u16, body: &Value) -> Value {
    let message = body
        .pointer("/error/message")
        .and_then(Value::as_str)
        .unwrap_or("upstream request failed");
    let status_text = match status {
        400 => "INVALID_ARGUMENT",
        401 | 403 => "PERMISSION_DENIED",
        404 => "NOT_FOUND",
        429 => "RESOURCE_EXHAUSTED",
        503 => "UNAVAILABLE",
        _ => "INTERNAL",
    };
    json!({
        "error": {
            "code": status,
            "message": message,
            "status": status_text,
        }
    })
}

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

    #[test]
    fn translates_system_instruction_and_multi_turn_text() {
        let request = json!({
            "systemInstruction": {"parts": [{"text": "be terse"}]},
            "contents": [
                {"role": "user", "parts": [{"text": "hi"}]},
                {"role": "model", "parts": [{"text": "hello"}]},
                {"role": "user", "parts": [{"text": "more"}]}
            ],
            "generationConfig": {"maxOutputTokens": 64, "temperature": 0.25}
        });
        let chat = gemini_request_to_chat("gpt-5.4-mini", &request);
        assert_eq!(chat["model"], "gpt-5.4-mini");
        let messages = chat["messages"].as_array().unwrap();
        assert_eq!(messages.len(), 4);
        assert_eq!(messages[0]["role"], "system");
        assert_eq!(messages[0]["content"], "be terse");
        assert_eq!(messages[2]["role"], "assistant");
        assert_eq!(chat["max_tokens"], 64);
        assert_eq!(chat["temperature"], 0.25);
    }

    #[test]
    fn translates_function_declarations_and_tool_mode() {
        let request = json!({
            "contents": [{"role": "user", "parts": [{"text": "weather?"}]}],
            "tools": [{"functionDeclarations": [{
                "name": "get_weather",
                "description": "look up weather",
                "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}
            }]}, {"googleSearch": {}}],
            "toolConfig": {"functionCallingConfig": {"mode": "ANY"}}
        });
        let chat = gemini_request_to_chat("claude-opus-4-7", &request);
        let tools = chat["tools"].as_array().unwrap();
        assert_eq!(tools[0]["function"]["name"], "get_weather");
        assert_eq!(tools[0]["function"]["parameters"]["type"], "object");
        assert_eq!(tools[1]["type"], "web_search");
        assert_eq!(chat["tool_choice"], "required");
    }

    #[test]
    fn pairs_function_responses_with_the_call_they_answer() {
        let request = json!({
            "contents": [
                {"role": "user", "parts": [{"text": "weather?"}]},
                {"role": "model", "parts": [
                    {"functionCall": {"name": "get_weather", "args": {"city": "Lisbon"}}}
                ]},
                {"role": "user", "parts": [
                    {"functionResponse": {"name": "get_weather", "response": {"c": 21}}}
                ]}
            ]
        });
        let chat = gemini_request_to_chat("gpt-5.4-mini", &request);
        let messages = chat["messages"].as_array().unwrap();
        let call_id = messages[1]["tool_calls"][0]["id"].as_str().unwrap();
        assert_eq!(
            messages[1]["tool_calls"][0]["function"]["arguments"],
            "{\"city\":\"Lisbon\"}"
        );
        assert_eq!(messages[2]["role"], "tool");
        assert_eq!(messages[2]["tool_call_id"], call_id);
    }

    #[test]
    fn honours_a_client_supplied_function_call_id() {
        let request = json!({
            "contents": [
                {"role": "model", "parts": [
                    {"functionCall": {"id": "toolu_42", "name": "ls", "args": {}}}
                ]},
                {"role": "user", "parts": [
                    {"functionResponse": {"name": "ls", "response": {"files": []}}}
                ]}
            ]
        });
        let chat = gemini_request_to_chat("gpt-5.4-mini", &request);
        let messages = chat["messages"].as_array().unwrap();
        assert_eq!(messages[0]["tool_calls"][0]["id"], "toolu_42");
        assert_eq!(messages[1]["tool_call_id"], "toolu_42");
    }

    #[test]
    fn translates_chat_completion_text_and_usage() {
        let chat = json!({
            "choices": [{
                "index": 0,
                "message": {"role": "assistant", "content": "answer"},
                "finish_reason": "length"
            }],
            "usage": {"prompt_tokens": 3, "completion_tokens": 5}
        });
        let gemini = chat_to_gemini_response(&chat, "gpt-5.4-mini");
        assert_eq!(
            gemini["candidates"][0]["content"]["parts"][0]["text"],
            "answer"
        );
        assert_eq!(gemini["candidates"][0]["content"]["role"], "model");
        assert_eq!(gemini["candidates"][0]["finishReason"], "MAX_TOKENS");
        assert_eq!(gemini["usageMetadata"]["totalTokenCount"], 8);
        assert_eq!(gemini["modelVersion"], "gpt-5.4-mini");
    }

    #[test]
    fn translates_chat_tool_calls_into_function_call_parts() {
        let chat = json!({
            "choices": [{
                "message": {
                    "role": "assistant",
                    "content": "",
                    "tool_calls": [{
                        "id": "call_1",
                        "type": "function",
                        "function": {"name": "get_weather", "arguments": "{\"city\":\"Lisbon\"}"}
                    }]
                },
                "finish_reason": "tool_calls"
            }]
        });
        let gemini = chat_to_gemini_response(&chat, "claude-opus-4-7");
        let part = &gemini["candidates"][0]["content"]["parts"][0]["functionCall"];
        assert_eq!(part["name"], "get_weather");
        assert_eq!(part["args"]["city"], "Lisbon");
        assert_eq!(part["id"], "call_1");
        assert_eq!(gemini["candidates"][0]["finishReason"], "STOP");
    }

    #[test]
    fn maps_openai_errors_onto_the_gemini_envelope() {
        let error = openai_error_to_gemini(
            429,
            &json!({"error": {"message": "slow down", "type": "rate_limit_error"}}),
        );
        assert_eq!(error["error"]["code"], 429);
        assert_eq!(error["error"]["message"], "slow down");
        assert_eq!(error["error"]["status"], "RESOURCE_EXHAUSTED");
    }
}