switchyard_translation/codecs/anthropic/

buffered.rs

// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

//! Buffered codec for Anthropic Messages request and response JSON.

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

use crate::codecs::common::{
    ANTHROPIC_REQUEST_KEY, is_anthropic_request, is_known_role_name, provider_extensions,
    text_from_blocks,
};
use crate::codecs::openai_chat::{decode_file_source, decode_image_source};
use crate::codecs::structured_output::decode_openai_schema_enforcement;
use crate::codecs::{
    DecodedRequest, DecodedResponse, EncodedRequest, EncodedResponse, FormatCodec,
};
use crate::diagnostic::TranslationDiagnostic;
use crate::error::{Result, TranslationError};
use crate::format::{FormatId, WireFormat};
use crate::llm::{
    AggLlmResponse, ContentBlock, FileSource, ImageSource, InstructionBlock, LlmRequest,
    MediaSource, Message, OutputParams, ProviderExtensions, ReasoningParams, ResponseOutput, Role,
    SamplingParams, StopReason, Tool, ToolCall, ToolChoice, ToolDefinition, ToolResult, Usage,
};
use crate::policy::{DeterministicIdPolicy, TranslationPolicy};
use crate::util::{
    capture_request_preservation, capture_response_preservation, desanitize_anthropic_tool_use_id,
    embed_preservation, exact_preserved_request, exact_preserved_response,
    sanitize_anthropic_tool_use_id,
};
use crate::util::{
    json_string, push_lossy, reject_responses_native_input_item, stable_id, string_value,
    validate_request_capabilities,
};
use base64::{Engine as _, engine::general_purpose::STANDARD};

/// Format codec for Anthropic Messages payloads.
pub struct AnthropicMessagesCodec;

impl FormatCodec for AnthropicMessagesCodec {
    fn format(&self) -> FormatId {
        WireFormat::AnthropicMessages.into()
    }

    fn decode_request(&self, body: &Value, policy: &TranslationPolicy) -> Result<DecodedRequest> {
        let body = crate::util::object(body, "$")?;
        if body
            .get("fallback_credit_token")
            .is_some_and(|value| !value.is_null())
            && body.get("fallbacks").is_some_and(|value| !value.is_null())
        {
            return Err(TranslationError::InvalidValue {
                path: "$.fallback_credit_token".to_string(),
                message: "fallback_credit_token cannot be combined with fallbacks".to_string(),
            });
        }
        let mut diagnostics = Vec::new();
        let max_output_tokens = body
            .get("max_tokens")
            .map(|value| {
                value
                    .as_u64()
                    .ok_or_else(|| TranslationError::InvalidValue {
                        path: "$.max_tokens".to_string(),
                        message: "expected a non-negative integer".to_string(),
                    })
            })
            .transpose()?;
        let response_format = decode_anthropic_output_format(body, &mut diagnostics, policy)?;
        let mut request = LlmRequest {
            model: body
                .get("model")
                .and_then(Value::as_str)
                .filter(|model| !model.is_empty())
                .map(ToOwned::to_owned),
            output: OutputParams {
                max_output_tokens,
                is_schema_enforced: response_format.as_ref().map(|_| true),
                response_format,
            },
            sampling: SamplingParams {
                temperature: body.get("temperature").and_then(Value::as_f64),
                top_p: body.get("top_p").and_then(Value::as_f64),
                top_k: body.get("top_k").and_then(Value::as_i64),
            },
            reasoning: ReasoningParams {
                effort: body
                    .get("output_config")
                    .and_then(Value::as_object)
                    .and_then(|object| object.get("effort"))
                    .and_then(Value::as_str)
                    .map(ToOwned::to_owned),
                raw: body.get("thinking").cloned(),
            },
            stream: body.get("stream").and_then(Value::as_bool).unwrap_or(false),
            preservation: capture_request_preservation(
                WireFormat::AnthropicMessages,
                &Value::Object(body.clone()),
                policy,
            ),
            ..LlmRequest::default()
        };
        if let Some(system) = body.get("system")
            && let Some(content) = decode_anthropic_system(system)?
        {
            request.instructions.push(InstructionBlock {
                role: Role::System,
                content,
            });
        }
        if let Some(messages) = body.get("messages") {
            let messages = messages
                .as_array()
                .ok_or_else(|| TranslationError::InvalidType {
                    path: "$.messages".to_string(),
                    expected: "array",
                })?;
            let mut generated_id = 0;
            for (index, message) in messages.iter().enumerate() {
                let Some(message) = message.as_object() else {
                    push_lossy(
                        &mut diagnostics,
                        policy,
                        format!("Anthropic message at index {index} is not an object"),
                    )?;
                    continue;
                };
                // Request decoding enforces the provider contract: an unknown
                // role is rejected rather than coerced to `user`. Anthropic
                // Messages only defines `user`/`assistant`, but other known
                // role names stay lenient (mapped to `user`) to preserve
                // historical cross-format behaviour.
                let role = match message.get("role").and_then(Value::as_str) {
                    Some("assistant") => Role::Assistant,
                    None => Role::User,
                    Some(other) if is_known_role_name(other) => Role::User,
                    Some(other) => {
                        return Err(TranslationError::unsupported_role(
                            format!("$.messages[{index}].role"),
                            other,
                        ));
                    }
                };
                generated_id += 1;
                let content = decode_anthropic_content(
                    message
                        .get("content")
                        .unwrap_or(&Value::String(String::new())),
                    role,
                    generated_id,
                    &mut diagnostics,
                    policy,
                )?;
                request.messages.push(Message::new(role, content));
            }
        }
        request.tools = decode_anthropic_tools(body.get("tools"));
        request.tool_choice = body.get("tool_choice").map(decode_anthropic_tool_choice);
        request.extensions.fields = provider_extensions(
            body,
            &[
                "model",
                "messages",
                "system",
                "tools",
                "tool_choice",
                "max_tokens",
                "temperature",
                "top_p",
                "top_k",
                "thinking",
                "output_config",
                "output_format",
                "stream",
                // OpenAI-only identity fields must not leak through Anthropic decoding.
                "safety_identifier",
            ],
        );
        if let Some(is_disabled) = body
            .get("tool_choice")
            .and_then(|choice| choice.get("disable_parallel_tool_use"))
            .and_then(Value::as_bool)
        {
            request
                .extensions
                .fields
                .insert("parallel_tool_calls".to_string(), Value::Bool(!is_disabled));
        }
        request
            .extensions
            .fields
            .insert(ANTHROPIC_REQUEST_KEY.to_string(), Value::Bool(true));

        Ok(DecodedRequest {
            request,
            diagnostics,
        })
    }

    fn encode_request(
        &self,
        request: &LlmRequest,
        policy: &TranslationPolicy,
    ) -> Result<EncodedRequest> {
        if let Some(body) =
            exact_preserved_request(&request.preservation, WireFormat::AnthropicMessages, policy)
        {
            return Ok(EncodedRequest {
                body,
                diagnostics: Vec::new(),
            });
        }
        let mut diagnostics = Vec::new();
        validate_request_capabilities(request, &mut diagnostics, policy)?;
        let allowed = crate::codecs::common::allowed_function_tools(request)?;
        let (tools, tool_choice) = allowed.as_ref().map_or(
            (request.tools.as_slice(), request.tool_choice.as_ref()),
            |(tools, choice)| (tools.as_slice(), Some(choice)),
        );
        let mut body = Map::new();
        if let Some(model) = &request.model {
            body.insert("model".to_string(), Value::String(model.clone()));
        }
        let system_text = request
            .instructions
            .iter()
            .flat_map(|instruction| instruction.content.iter())
            .chain(
                request
                    .messages
                    .iter()
                    .filter(|message| matches!(message.role, Role::System | Role::Developer))
                    .flat_map(|message| message.content.iter()),
            )
            .filter_map(|block| match block {
                ContentBlock::Text { text } | ContentBlock::Refusal { text } => Some(text.as_str()),
                _ => None,
            })
            .collect::<Vec<_>>()
            .join("\n\n");
        if !system_text.is_empty() {
            body.insert("system".to_string(), Value::String(system_text));
        }
        // Anthropic `system` is a text field, so other instruction content is dropped.
        let instructions = request
            .instructions
            .iter()
            .enumerate()
            .map(|(index, block)| (block.role, &block.content, "instruction", index));
        let inline = request
            .messages
            .iter()
            .enumerate()
            .filter(|(_, message)| matches!(message.role, Role::System | Role::Developer))
            .map(|(index, message)| (message.role, &message.content, "message", index));
        for (role, content, kind, index) in instructions.chain(inline) {
            for block in content {
                if !matches!(
                    block,
                    ContentBlock::Text { .. } | ContentBlock::Refusal { .. }
                ) {
                    push_lossy(
                        &mut diagnostics,
                        policy,
                        format!(
                            "Anthropic system prompt accepts only text; dropped non-text content from {role} {kind} {index}"
                        ),
                    )?;
                }
            }
        }

        body.insert(
            "messages".to_string(),
            Value::Array(encode_anthropic_messages(
                &request.messages,
                &mut diagnostics,
                policy,
            )?),
        );

        if !tools.is_empty() {
            body.insert("tools".to_string(), encode_anthropic_tools(tools)?);
        }
        let parallel_tool_calls = request
            .extensions
            .fields
            .get("parallel_tool_calls")
            .and_then(Value::as_bool);
        if tool_choice.is_some() || (parallel_tool_calls.is_some() && !tools.is_empty()) {
            let mut choice = encode_anthropic_tool_choice(tool_choice.unwrap_or(&ToolChoice::Auto));
            if let Some(is_enabled) = parallel_tool_calls
                && let Some(object) = choice.as_object_mut()
                && object.get("type").and_then(Value::as_str) != Some("none")
            {
                object.insert(
                    "disable_parallel_tool_use".to_string(),
                    Value::Bool(!is_enabled),
                );
            }
            body.insert("tool_choice".to_string(), choice);
        }
        if is_anthropic_request(request) {
            for field in [
                "inference_geo",
                "service_tier",
                "stop_sequences",
                "metadata",
                "cache_control",
                "container",
                "speed",
                "diagnostics",
                "fallback_credit_token",
            ] {
                if let Some(value) = request.extensions.fields.get(field) {
                    body.insert(field.to_string(), value.clone());
                }
            }
        }
        if let Some(stop_sequences) =
            anthropic_stop_sequences_from_extensions(&request.extensions.fields)
        {
            body.entry("stop_sequences").or_insert(stop_sequences);
        }
        if let Some(max_tokens) = request.output.max_output_tokens {
            body.insert("max_tokens".to_string(), json!(max_tokens));
        } else {
            body.insert("max_tokens".to_string(), json!(64_000));
        }
        if let Some(value) = request.sampling.temperature {
            body.insert("temperature".to_string(), json!(value));
        }
        if let Some(value) = request.sampling.top_p {
            body.insert("top_p".to_string(), json!(value));
        }
        if let Some(value) = request.sampling.top_k {
            body.insert("top_k".to_string(), json!(value));
        }
        if request.stream {
            body.insert("stream".to_string(), Value::Bool(true));
        }
        // Native thinking controls the mode independently of output effort.
        // Other codecs store their own provider's reasoning object in `raw`.
        if is_anthropic_request(request)
            && let Some(thinking) = &request.reasoning.raw
        {
            body.insert("thinking".to_string(), thinking.clone());
        }
        if let Some(effort) = &request.reasoning.effort {
            body.entry("thinking".to_string())
                .or_insert_with(|| json!({"type": "adaptive"}));
            body.insert("output_config".to_string(), json!({"effort": effort}));
        }
        if let Some(response_format) = &request.output.response_format
            && let Some(format) = encode_anthropic_output_format(
                response_format,
                request.output.is_schema_enforced,
                &mut diagnostics,
                policy,
            )?
        {
            let output_config = body
                .entry("output_config".to_string())
                .or_insert_with(|| Value::Object(Map::new()));
            let Some(output_config) = output_config.as_object_mut() else {
                return Err(TranslationError::InvalidType {
                    path: "$.output_config".to_string(),
                    expected: "object",
                });
            };
            output_config.insert("format".to_string(), format);
        } else if request.output.response_format.is_none()
            && request.output.is_schema_enforced == Some(true)
        {
            push_lossy(
                &mut diagnostics,
                policy,
                "Structured-output schema enforcement requires a response format with a schema",
            )?;
        }

        let body = embed_preservation(Value::Object(body), &request.preservation, policy);
        Ok(EncodedRequest { body, diagnostics })
    }

    fn decode_response(
        &self,
        body: &Value,
        _policy: &TranslationPolicy,
    ) -> Result<DecodedResponse> {
        let body = crate::util::object(body, "$")?;
        let mut content = Vec::new();
        if let Some(blocks) = body.get("content").and_then(Value::as_array) {
            for (index, block) in blocks.iter().enumerate() {
                if let Some(block) = block.as_object() {
                    content.extend(decode_anthropic_content_block(
                        block,
                        Role::Assistant,
                        index + 1,
                        &mut Vec::new(),
                        &TranslationPolicy::default(),
                    )?);
                }
            }
        }
        if content.is_empty() {
            content.push(ContentBlock::Text {
                text: String::new(),
            });
        }
        let response = AggLlmResponse {
            id: body
                .get("id")
                .and_then(Value::as_str)
                .map(ToOwned::to_owned),
            model: body
                .get("model")
                .and_then(Value::as_str)
                .map(ToOwned::to_owned),
            outputs: vec![ResponseOutput {
                url_citations: Vec::new(),
                role: Role::Assistant,
                content,
                stop_reason: Some(map_anthropic_stop_reason(
                    body.get("stop_reason").and_then(Value::as_str),
                )),
            }],
            usage: decode_anthropic_usage(body.get("usage")),
            extensions: ProviderExtensions {
                fields: provider_extensions(
                    body,
                    &[
                        "id",
                        "type",
                        "role",
                        "model",
                        "content",
                        "stop_reason",
                        "usage",
                    ],
                ),
            },
            preservation: capture_response_preservation(
                WireFormat::AnthropicMessages,
                &Value::Object(body.clone()),
                _policy,
            ),
        };
        Ok(DecodedResponse {
            response,
            diagnostics: Vec::new(),
        })
    }

    fn encode_response(
        &self,
        response: &AggLlmResponse,
        _policy: &TranslationPolicy,
    ) -> Result<EncodedResponse> {
        super::super::responses::validate_response_output(response, WireFormat::AnthropicMessages)?;
        if let Some(body) = exact_preserved_response(
            &response.preservation,
            WireFormat::AnthropicMessages,
            _policy,
        ) {
            return Ok(EncodedResponse {
                body,
                diagnostics: Vec::new(),
            });
        }
        let output = response.first_output();
        let content = output
            .map(|output| encode_anthropic_content(&output.content))
            .transpose()?
            .unwrap_or_else(|| vec![json!({"type": "text", "text": ""})]);
        let normalized_stop_reason = output.and_then(|output| output.stop_reason);
        let body = json!({
            "id": response.id.clone().unwrap_or_else(|| "msg_switchyard".to_string()),
            "type": "message",
            "role": "assistant",
            "model": response.model.clone().unwrap_or_else(|| "unknown".to_string()),
            "content": content,
            "stop_reason": normalized_stop_reason
                .map(anthropic_stop_reason)
                .unwrap_or("end_turn"),
            "stop_sequence": Value::Null,
            "stop_details": normalized_stop_reason
                .map(|reason| {
                    anthropic_stop_details(reason, response.extensions.fields.get("stop_details"))
                })
                .unwrap_or(Value::Null),
            "usage": encode_anthropic_usage(&response.usage),
        });
        Ok(EncodedResponse {
            body: embed_preservation(body, &response.preservation, _policy),
            diagnostics: Vec::new(),
        })
    }
}

// Reads the current `output_config.format`, or the beta `output_format` it replaced,
// into the neutral OpenAI-shaped response format.
fn decode_anthropic_output_format(
    body: &Map<String, Value>,
    diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Option<Value>> {
    let Some(format) = body
        .get("output_config")
        .and_then(Value::as_object)
        .and_then(|config| config.get("format"))
        .or_else(|| body.get("output_format"))
    else {
        return Ok(None);
    };
    let Some(format) = format.as_object() else {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output format is not an object; the requested format was dropped",
        )?;
        return Ok(None);
    };
    if format.get("type").and_then(Value::as_str) != Some("json_schema") {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output maps only a json_schema format; the requested format was dropped",
        )?;
        return Ok(None);
    }
    // A non-object schema would reach the upstream as a malformed `json_schema.schema`.
    let Some(schema) = format.get("schema").filter(|schema| schema.is_object()) else {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output requires format.schema to be an object; the requested format was dropped",
        )?;
        return Ok(None);
    };
    Ok(Some(json!({
        "type": "json_schema",
        "json_schema": {
            // Anthropic identifies the schema by position; the neutral shape needs a name.
            "name": "response",
            "schema": schema.clone(),
        },
    })))
}

/// Maps the neutral OpenAI-shaped JSON schema to Anthropic's output format.
fn encode_anthropic_output_format(
    response_format: &Value,
    enforcement: Option<bool>,
    diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Option<Value>> {
    let Some(json_schema) = response_format
        .as_object()
        .filter(|format| format.get("type").and_then(Value::as_str) == Some("json_schema"))
        .and_then(|format| format.get("json_schema"))
        .and_then(Value::as_object)
    else {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output requires an OpenAI JSON schema response format",
        )?;
        return Ok(None);
    };
    let Some(schema) = json_schema.get("schema") else {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output requires json_schema.schema",
        )?;
        return Ok(None);
    };

    if enforcement.or_else(|| decode_openai_schema_enforcement(Some(response_format)))
        == Some(false)
    {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output enforces the schema; advisory schema enforcement was strengthened",
        )?;
    }
    let mut schema = schema.clone();
    if strip_anthropic_unsupported_constraints(&mut schema) {
        push_lossy(
            diagnostics,
            policy,
            "Anthropic structured output dropped unsupported JSON Schema constraints",
        )?;
    }
    Ok(Some(json!({"type": "json_schema", "schema": schema})))
}

/// Removes constraints unsupported by Anthropic's structured-output grammar.
fn strip_anthropic_unsupported_constraints(value: &mut Value) -> bool {
    match value {
        Value::Object(object) => {
            let mut removed = false;
            for key in ["minimum", "maximum", "minLength", "maxLength"] {
                removed |= object.remove(key).is_some();
            }
            for value in object.values_mut() {
                removed |= strip_anthropic_unsupported_constraints(value);
            }
            removed
        }
        Value::Array(values) => {
            let mut removed = false;
            for value in values {
                removed |= strip_anthropic_unsupported_constraints(value);
            }
            removed
        }
        _ => false,
    }
}

// Decodes Anthropic's `system` field into instruction blocks.
fn decode_anthropic_system(value: &Value) -> Result<Option<Vec<ContentBlock>>> {
    match value {
        Value::String(text) if !text.is_empty() => {
            Ok(Some(vec![ContentBlock::Text { text: text.clone() }]))
        }
        Value::String(_) | Value::Null => Ok(None),
        Value::Array(blocks) => {
            let mut content = Vec::new();
            for block in blocks {
                if let Some(block) = block.as_object()
                    && block.get("type").and_then(Value::as_str) == Some("text")
                {
                    let text = block
                        .get("text")
                        .and_then(Value::as_str)
                        .unwrap_or_default()
                        .to_string();
                    content.push(ContentBlock::Text { text });
                }
            }
            Ok((!content.is_empty()).then_some(content))
        }
        _ => Err(TranslationError::InvalidType {
            path: "$.system".to_string(),
            expected: "string or array of text blocks",
        }),
    }
}

// Decodes Anthropic message content into normalized content blocks.
fn decode_anthropic_content(
    value: &Value,
    role: Role,
    generated_counter: usize,
    diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Vec<ContentBlock>> {
    match value {
        Value::String(text) => Ok(vec![ContentBlock::Text { text: text.clone() }]),
        Value::Null => Ok(vec![ContentBlock::Text {
            text: String::new(),
        }]),
        Value::Array(blocks) => {
            let mut content = Vec::new();
            for (index, block) in blocks.iter().enumerate() {
                let Some(block) = block.as_object() else {
                    push_lossy(
                        diagnostics,
                        policy,
                        format!("Anthropic content block {index} is not an object"),
                    )?;
                    continue;
                };
                content.extend(decode_anthropic_content_block(
                    block,
                    role,
                    generated_counter + index,
                    diagnostics,
                    policy,
                )?);
            }
            if content.is_empty() {
                content.push(ContentBlock::Text {
                    text: String::new(),
                });
            }
            Ok(content)
        }
        other => Ok(vec![ContentBlock::Text {
            text: string_value(other).unwrap_or_default(),
        }]),
    }
}

// Decodes one Anthropic content block into one or more IR blocks.
fn decode_anthropic_content_block(
    block: &Map<String, Value>,
    _role: Role,
    generated_counter: usize,
    _diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Vec<ContentBlock>> {
    Ok(match block.get("type").and_then(Value::as_str) {
        Some("text") => vec![ContentBlock::Text {
            text: block
                .get("text")
                .and_then(Value::as_str)
                .unwrap_or_default()
                .to_string(),
        }],
        Some("thinking") => vec![ContentBlock::Reasoning {
            text: block
                .get("thinking")
                .and_then(Value::as_str)
                .unwrap_or_default()
                .to_string(),
            signature: block
                .get("signature")
                .and_then(Value::as_str)
                .filter(|signature| !signature.is_empty())
                .map(ToOwned::to_owned),
            details: Vec::new(),
        }],
        Some("tool_use") => vec![ContentBlock::ToolCall(ToolCall {
            id: block
                .get("id")
                .and_then(Value::as_str)
                .filter(|id| !id.is_empty())
                .map(desanitize_anthropic_tool_use_id)
                .unwrap_or_else(|| match &policy.deterministic_ids {
                    DeterministicIdPolicy::GenerateStable { prefix } => {
                        stable_id(prefix, generated_counter)
                    }
                    DeterministicIdPolicy::Preserve => String::new(),
                }),
            name: block
                .get("name")
                .and_then(Value::as_str)
                .unwrap_or_default()
                .to_string(),
            arguments: block.get("input").cloned().unwrap_or_else(|| json!({})),
        })],
        Some("tool_result") => vec![ContentBlock::ToolResult(ToolResult {
            tool_call_id: block
                .get("tool_use_id")
                .and_then(Value::as_str)
                .map(desanitize_anthropic_tool_use_id)
                .unwrap_or_default(),
            content: decode_tool_result_content(block.get("content").unwrap_or(&Value::Null)),
            is_error: block.get("is_error").and_then(Value::as_bool),
        })],
        Some("image") => vec![ContentBlock::Image {
            source: ImageSource::Raw(Value::Object(block.clone())),
        }],
        Some("input_image") | Some("image_url") => decode_image_source(block)
            .map(|source| vec![ContentBlock::Image { source }])
            .unwrap_or_default(),
        Some("input_file") | Some("file") => vec![ContentBlock::File {
            source: decode_file_source(block),
        }],
        Some("document") => vec![ContentBlock::File {
            source: decode_anthropic_file_source(block),
        }],
        _ => vec![ContentBlock::Unknown {
            provider: WireFormat::AnthropicMessages.into(),
            raw: Value::Object(block.clone()),
        }],
    })
}

// Preserves supported Anthropic tool-result blocks in the neutral IR.
fn decode_tool_result_content(value: &Value) -> Vec<ContentBlock> {
    match value {
        Value::String(text) => vec![ContentBlock::Text { text: text.clone() }],
        Value::Array(blocks) => {
            let mut content = Vec::new();
            for block in blocks {
                if let Some(block) = block.as_object() {
                    match block.get("type").and_then(Value::as_str) {
                        Some("text") => content.push(ContentBlock::Text {
                            text: block
                                .get("text")
                                .and_then(Value::as_str)
                                .unwrap_or_default()
                                .to_string(),
                        }),
                        Some("image") => content.push(ContentBlock::Image {
                            source: ImageSource::Raw(Value::Object(block.clone())),
                        }),
                        Some("document") => content.push(ContentBlock::File {
                            source: decode_anthropic_file_source(block),
                        }),
                        _ => content.push(ContentBlock::Unknown {
                            provider: WireFormat::AnthropicMessages.into(),
                            raw: Value::Object(block.clone()),
                        }),
                    }
                }
            }
            content
        }
        Value::Null => vec![ContentBlock::Text {
            text: String::new(),
        }],
        other => vec![ContentBlock::Text {
            text: json_string(other),
        }],
    }
}

// Keeps Anthropic document fields together for same-format re-encoding.
fn decode_anthropic_file_source(block: &Map<String, Value>) -> FileSource {
    FileSource::Raw(Value::Object(block.clone()))
}

// Decodes Anthropic tool definitions into normalized tool definitions.
fn decode_anthropic_tools(value: Option<&Value>) -> Vec<Tool> {
    value
        .and_then(Value::as_array)
        .into_iter()
        .flatten()
        .filter_map(Value::as_object)
        .filter_map(|tool| {
            if tool
                .get("type")
                .and_then(Value::as_str)
                .is_some_and(|kind| kind != "custom")
            {
                return Some(Tool::Provider {
                    provider: WireFormat::AnthropicMessages.into(),
                    raw: Value::Object(tool.clone()),
                });
            }
            let name = tool.get("name").and_then(Value::as_str)?.to_string();
            (!name.is_empty()).then(|| {
                Tool::Function(ToolDefinition {
                    name,
                    description: tool
                        .get("description")
                        .and_then(Value::as_str)
                        .map(ToOwned::to_owned),
                    parameters: tool
                        .get("input_schema")
                        .cloned()
                        .unwrap_or_else(|| json!({})),
                    strict: tool.get("strict").and_then(Value::as_bool),
                })
            })
        })
        .collect()
}

// Decodes Anthropic tool-choice values into normalized policy.
fn decode_anthropic_tool_choice(value: &Value) -> ToolChoice {
    match value {
        Value::String(text) if text == "auto" => ToolChoice::Auto,
        Value::String(text) if text == "any" => ToolChoice::Required,
        Value::String(text) if text == "none" => ToolChoice::None,
        Value::Object(object) => match object.get("type").and_then(Value::as_str) {
            Some("auto") => ToolChoice::Auto,
            Some("any") => ToolChoice::Required,
            Some("none") => ToolChoice::None,
            Some("tool") => object
                .get("name")
                .and_then(Value::as_str)
                .map(|name| ToolChoice::Tool {
                    name: name.to_string(),
                })
                .unwrap_or_else(|| ToolChoice::Raw(value.clone())),
            _ => ToolChoice::Raw(value.clone()),
        },
        _ => ToolChoice::Raw(value.clone()),
    }
}

// Encodes one normalized message into Anthropic message JSON.
fn encode_anthropic_message(
    message: &Message,
    diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Value> {
    let role = match message.role {
        Role::Assistant => "assistant",
        Role::User | Role::Tool | Role::System | Role::Developer => "user",
    };
    let content = encode_anthropic_content_with_policy(&message.content, diagnostics, policy)?;
    let simple_text = content.len() == 1
        && content
            .first()
            .and_then(Value::as_object)
            .and_then(|object| object.get("type"))
            .and_then(Value::as_str)
            == Some("text");
    let content = if simple_text {
        content
            .first()
            .and_then(Value::as_object)
            .and_then(|object| object.get("text"))
            .cloned()
            .unwrap_or_else(|| Value::String(String::new()))
    } else {
        Value::Array(content)
    };
    Ok(json!({"role": role, "content": content}))
}

// Encodes messages while grouping adjacent tool-result-only messages correctly.
fn encode_anthropic_messages(
    messages: &[Message],
    diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Vec<Value>> {
    let mut encoded = Vec::new();
    let mut index = 0;

    while let Some(message) = messages.get(index) {
        // Inline instructions are emitted in the top-level system field.
        if matches!(message.role, Role::System | Role::Developer) {
            index += 1;
            continue;
        }
        if !message_is_tool_result_only(message) {
            encoded.push(encode_anthropic_message(message, diagnostics, policy)?);
            index += 1;
            continue;
        }

        let mut content = Vec::new();
        while let Some(tool_message) = messages.get(index) {
            if matches!(tool_message.role, Role::System | Role::Developer) {
                index += 1;
                continue;
            }
            if !message_is_tool_result_only(tool_message) {
                break;
            }
            content.extend(encode_anthropic_content_with_policy(
                &tool_message.content,
                diagnostics,
                policy,
            )?);
            index += 1;
        }
        encoded.push(json!({"role": "user", "content": content}));
    }

    Ok(encoded)
}

// Maps preserved OpenAI-style stop extensions to Anthropic stop sequences.
fn anthropic_stop_sequences_from_extensions(extensions: &Map<String, Value>) -> Option<Value> {
    match extensions.get("stop") {
        Some(Value::String(stop)) => Some(json!([stop])),
        Some(Value::Array(stops)) => Some(Value::Array(stops.clone())),
        _ => None,
    }
}

// Checks whether a message contains only tool-result blocks.
fn message_is_tool_result_only(message: &Message) -> bool {
    (message.role == Role::Tool || message.role == Role::User)
        && !message.content.is_empty()
        && message
            .content
            .iter()
            .all(|block| matches!(block, ContentBlock::ToolResult(_)))
}

// Encodes content while applying lossy-conversion policy to unknown blocks.
fn encode_anthropic_content_with_policy(
    content: &[ContentBlock],
    diagnostics: &mut Vec<TranslationDiagnostic>,
    policy: &TranslationPolicy,
) -> Result<Vec<Value>> {
    let mut blocks = Vec::new();
    for block in content {
        crate::codecs::openai_media::validate_media(block, WireFormat::AnthropicMessages)?;
        match block {
            ContentBlock::Unknown { provider, raw } => {
                reject_responses_native_input_item(provider, raw, WireFormat::AnthropicMessages)?;
                push_lossy(
                    diagnostics,
                    policy,
                    "unknown content block encoded as text for Anthropic",
                )?;
                blocks.push(json!({"type": "text", "text": json_string(raw)}));
            }
            other => blocks.extend(encode_one_anthropic_block(other)?),
        }
    }
    if blocks.is_empty() {
        blocks.push(json!({"type": "text", "text": ""}));
    }
    Ok(blocks)
}

fn encode_anthropic_file(source: &FileSource) -> Result<Value> {
    match source {
        FileSource::FileData { data, filename } => {
            let (media_type, data) = split_base64_data_uri(data)
                .unwrap_or_else(|| (document_media_type(filename.as_deref()), data.as_str()));
            let source = match media_type {
                "text/plain" => {
                    let bytes =
                        STANDARD
                            .decode(data)
                            .map_err(|_| TranslationError::InvalidValue {
                                path: "file_data".into(),
                                message: "invalid base64 text document".into(),
                            })?;
                    let text =
                        String::from_utf8(bytes).map_err(|_| TranslationError::InvalidValue {
                            path: "file_data".into(),
                            message: "text document must be UTF-8".into(),
                        })?;
                    json!({"type": "text", "media_type": media_type, "data": text})
                }
                "application/pdf" => {
                    json!({"type": "base64", "media_type": media_type, "data": data})
                }
                _ => {
                    return Err(TranslationError::LossyConversion(
                        "Anthropic requires PDF or plain-text documents".into(),
                    ));
                }
            };
            let mut block = json!({"type": "document", "source": source});
            if let Some(filename) = filename {
                block["title"] = filename.clone().into();
            }
            Ok(block)
        }
        FileSource::Raw(raw) if raw.get("type").and_then(Value::as_str) == Some("input_file") => {
            let url = raw.get("file_url").and_then(Value::as_str).ok_or_else(|| {
                TranslationError::LossyConversion("unsupported Anthropic document source".into())
            })?;
            let mut block = json!({"type": "document", "source": {"type": "url", "url": url}});
            if let Some(filename) = raw.get("filename") {
                block["title"] = filename.clone();
            }
            Ok(block)
        }
        FileSource::Raw(raw) => Ok(raw.clone()),
        FileSource::FileId(file_id) => Ok(json!({
            "type": "document", "source": {"type": "file", "file_id": file_id}
        })),
    }
}

// Encodes content without producing diagnostics for response paths.
fn encode_anthropic_content(content: &[ContentBlock]) -> Result<Vec<Value>> {
    let mut blocks = Vec::new();
    for block in content {
        crate::codecs::openai_media::validate_media(block, WireFormat::AnthropicMessages)?;
        blocks.extend(encode_one_anthropic_response_block(block)?);
    }
    if blocks.is_empty() {
        blocks.push(json!({"type": "text", "text": ""}));
    }
    Ok(blocks)
}

// Encodes response content, where synthetic reasoning may be shown to clients.
fn encode_one_anthropic_response_block(block: &ContentBlock) -> Result<Vec<Value>> {
    match block {
        ContentBlock::Reasoning {
            text,
            signature: None,
            ..
        } => Ok(vec![json!({
            "type": "thinking",
            "thinking": text,
            "signature": "",
        })]),
        other => encode_one_anthropic_block(other),
    }
}

// Encodes a single normalized content block into Anthropic block JSON.
fn encode_one_anthropic_block(block: &ContentBlock) -> Result<Vec<Value>> {
    Ok(match block {
        ContentBlock::Text { text } | ContentBlock::Refusal { text } => {
            vec![json!({"type": "text", "text": text})]
        }
        ContentBlock::Reasoning {
            text,
            signature: Some(signature),
            ..
        } if !signature.is_empty() => vec![json!({
            "type": "thinking",
            "thinking": text,
            "signature": signature,
        })],
        ContentBlock::Reasoning { .. } => Vec::new(),
        ContentBlock::ToolCall(call) => vec![json!({
            "type": "tool_use",
            "id": sanitize_anthropic_tool_use_id(&call.id),
            "name": call.name,
            "input": anthropic_tool_input(&call.arguments),
        })],
        ContentBlock::ToolResult(result) => {
            let content = if result.content.iter().all(|block| {
                matches!(
                    block,
                    ContentBlock::Text { .. } | ContentBlock::Refusal { .. }
                )
            }) {
                Value::String(text_from_blocks(&result.content, " "))
            } else {
                let mut content = Vec::new();
                for block in &result.content {
                    content.extend(encode_one_anthropic_tool_result_block(block)?);
                }
                Value::Array(content)
            };
            let mut item = json!({
                "type": "tool_result",
                "tool_use_id": sanitize_anthropic_tool_use_id(&result.tool_call_id),
                "content": content,
            });
            if let Some(is_error) = result.is_error {
                item["is_error"] = Value::Bool(is_error);
            }
            vec![item]
        }
        ContentBlock::Image { source } => vec![match source {
            ImageSource::Url { url, .. } => match split_base64_data_uri(url) {
                Some((media_type, data)) => json!({
                    "type": "image",
                    "source": {"type": "base64", "media_type": media_type, "data": data},
                }),
                None => json!({"type": "image", "source": {"type": "url", "url": url}}),
            },
            ImageSource::Base64 { media_type, data } => json!({
                "type": "image",
                "source": {
                    "type": "base64",
                    "media_type": media_type.clone().unwrap_or_else(|| "image/png".to_string()),
                    "data": data,
                },
            }),
            ImageSource::Raw(raw) => raw.clone(),
        }],
        ContentBlock::File { source } => vec![encode_anthropic_file(source)?],
        ContentBlock::Audio { source } => vec![match source {
            MediaSource::Url { url, media_type } => {
                json!({"type": "audio", "source": {"type": "url", "url": url, "media_type": media_type}})
            }
            MediaSource::Base64 { media_type, data } => json!({
                "type": "audio",
                "source": {
                    "type": "base64",
                    "media_type": media_type.clone().unwrap_or_else(|| "audio/mpeg".to_string()),
                    "data": data,
                },
            }),
            MediaSource::Raw(raw) => raw.clone(),
        }],
        ContentBlock::Video { source } => vec![match source {
            MediaSource::Url { url, media_type } => {
                json!({"type": "video", "source": {"type": "url", "url": url, "media_type": media_type}})
            }
            MediaSource::Base64 { media_type, data } => json!({
                "type": "video",
                "source": {
                    "type": "base64",
                    "media_type": media_type.clone().unwrap_or_else(|| "video/mp4".to_string()),
                    "data": data,
                },
            }),
            MediaSource::Raw(raw) => raw.clone(),
        }],
        ContentBlock::Mcp(_) => {
            return Err(TranslationError::LossyConversion(
                "MCP history encoding is not implemented".into(),
            ));
        }
        ContentBlock::Unknown { raw, .. } => vec![raw.clone()],
    })
}

// Encodes only provider-safe block shapes inside Anthropic tool results.
fn encode_one_anthropic_tool_result_block(block: &ContentBlock) -> Result<Vec<Value>> {
    match block {
        ContentBlock::Text { .. }
        | ContentBlock::Refusal { .. }
        | ContentBlock::Image { .. }
        | ContentBlock::File { .. } => encode_one_anthropic_block(block),
        ContentBlock::Unknown { provider, raw }
            if provider.as_str() == WireFormat::AnthropicMessages.as_str() =>
        {
            Ok(vec![raw.clone()])
        }
        ContentBlock::Unknown { raw, .. } => {
            Ok(vec![json!({"type": "text", "text": json_string(raw)})])
        }
        ContentBlock::Mcp(_) => Err(TranslationError::LossyConversion(
            "MCP history encoding is not implemented".into(),
        )),
        ContentBlock::Reasoning { .. }
        | ContentBlock::Audio { .. }
        | ContentBlock::Video { .. }
        | ContentBlock::ToolCall(_)
        | ContentBlock::ToolResult(_) => Ok(Vec::new()),
    }
}

// Anthropic accepts PDF and plain-text documents; a file with no media type is a PDF
// unless its name says otherwise.
fn document_media_type(filename: Option<&str>) -> &'static str {
    match filename
        .and_then(|name| name.rsplit_once('.'))
        .map(|(_, ext)| ext)
    {
        Some(ext) if ext.eq_ignore_ascii_case("txt") || ext.eq_ignore_ascii_case("md") => {
            "text/plain"
        }
        _ => "application/pdf",
    }
}

// Splits `data:<media type>[;<parameter>...];base64,<payload>`, the form OpenAI-compatible
// clients use for inline images. A percent-encoded payload has to be decoded before it is
// base64, and a data URI with no media type leaves Anthropic's `media_type` with nothing to
// carry, so both keep the handling they had before and are relayed as-is.
fn split_base64_data_uri(url: &str) -> Option<(&str, &str)> {
    let (metadata, data) = url.strip_prefix("data:")?.split_once(',')?;
    let parameters = metadata.strip_suffix(";base64")?;
    let media_type = parameters
        .split_once(';')
        .map_or(parameters, |(media_type, _)| media_type);
    if media_type.is_empty() || data.contains('%') {
        return None;
    }
    Some((media_type, data))
}

// Anthropic requires `tool_use.input` to be object-shaped, while OpenAI and
// Responses commonly carry function-call arguments as JSON strings.
fn anthropic_tool_input(arguments: &Value) -> Value {
    match arguments {
        Value::Object(object) => Value::Object(object.clone()),
        Value::String(text) => serde_json::from_str::<Value>(text)
            .map_or_else(|_| json!({"raw": text}), ensure_anthropic_tool_input_object),
        Value::Null => json!({}),
        other => json!({"value": other}),
    }
}

// Preserve valid objects and wrap every other JSON shape in a dictionary so
// translated requests satisfy Anthropic's schema without discarding payloads.
fn ensure_anthropic_tool_input_object(arguments: Value) -> Value {
    match arguments {
        Value::Object(_) => arguments,
        Value::Null => json!({}),
        other => json!({"value": other}),
    }
}

// Encodes normalized tool definitions into Anthropic tool JSON.
fn encode_anthropic_tools(tools: &[Tool]) -> Result<Value> {
    Ok(Value::Array(
        tools
            .iter()
            .map(|tool| {
                if let Tool::Provider { provider, raw } = tool
                    && provider.as_str() == WireFormat::AnthropicMessages.as_str()
                {
                    return Ok(raw.clone());
                }
                let tool = tool.function().ok_or_else(|| {
                    TranslationError::LossyConversion(
                        "MCP and provider tool encoding is not implemented".into(),
                    )
                })?;
                let mut item = json!({
                    "name": tool.name,
                    "description": tool.description.clone().unwrap_or_default(),
                    "input_schema": tool.parameters,
                });
                if let Some(strict) = tool.strict {
                    item["strict"] = Value::Bool(strict);
                }
                Ok(item)
            })
            .collect::<Result<_>>()?,
    ))
}

// Encodes normalized tool choice into Anthropic tool-choice JSON.
fn encode_anthropic_tool_choice(choice: &ToolChoice) -> Value {
    match choice {
        ToolChoice::Auto => json!({"type": "auto"}),
        ToolChoice::Required => json!({"type": "any"}),
        ToolChoice::None => json!({"type": "none"}),
        ToolChoice::Tool { name } => json!({"type": "tool", "name": name}),
        ToolChoice::Raw(value) => value.clone(),
    }
}

// Normalizes Anthropic usage fields.
fn decode_anthropic_usage(value: Option<&Value>) -> Usage {
    let Some(value) = value.and_then(Value::as_object) else {
        return Usage::default();
    };
    let input_tokens = value.get("input_tokens").and_then(Value::as_u64);
    let cached_input_tokens = value.get("cache_read_input_tokens").and_then(Value::as_u64);
    let cache_creation_input_tokens = value
        .get("cache_creation_input_tokens")
        .and_then(Value::as_u64);
    let output_tokens = value.get("output_tokens").and_then(Value::as_u64);
    Usage {
        provider_cost: crate::codecs::common::provider_cost(value),
        input_tokens,
        cache: crate::codecs::common::cache_usage_details(
            value,
            cached_input_tokens,
            cache_creation_input_tokens,
        ),
        output_tokens,
        total_tokens: input_tokens.zip(output_tokens).map(|(input, output)| {
            input
                + cached_input_tokens.unwrap_or(0)
                + cache_creation_input_tokens.unwrap_or(0)
                + output
        }),
        reasoning_tokens: value.get("output_tokens_details").and_then(|details| {
            details
                .get("thinking_tokens")
                .and_then(Value::as_u64)
                .or_else(|| details.get("reasoning_tokens").and_then(Value::as_u64))
        }),
    }
}

// Encodes normalized usage into Anthropic usage JSON.
fn encode_anthropic_usage(usage: &Usage) -> Value {
    let mut value = json!({
        "input_tokens": usage.input_tokens.unwrap_or(0),
        "output_tokens": usage.output_tokens.unwrap_or(0),
    });
    if let Some(cached_tokens) = usage.cached_input_tokens() {
        value["cache_read_input_tokens"] = json!(cached_tokens);
    }
    if let Some(cache_creation_tokens) = usage.cache_creation_input_tokens() {
        value["cache_creation_input_tokens"] = json!(cache_creation_tokens);
    }
    if let Some(thinking_tokens) = usage.reasoning_tokens {
        value["output_tokens_details"] = json!({"thinking_tokens": thinking_tokens});
    }
    crate::codecs::common::encode_usage_metadata(value, usage)
}

// Maps Anthropic stop reasons to normalized stop reasons.
fn map_anthropic_stop_reason(reason: Option<&str>) -> StopReason {
    match reason {
        Some("max_tokens") => StopReason::MaxTokens,
        Some("tool_use") => StopReason::ToolUse,
        Some("refusal") => StopReason::ContentFilter,
        Some("end_turn") | None => StopReason::EndTurn,
        _ => StopReason::Unknown,
    }
}

// Maps normalized stop reasons back to Anthropic's vocabulary.
fn anthropic_stop_reason(reason: StopReason) -> &'static str {
    match reason {
        StopReason::MaxTokens => "max_tokens",
        StopReason::ToolUse => "tool_use",
        StopReason::ContentFilter => "refusal",
        StopReason::EndTurn | StopReason::Error | StopReason::Unknown => "end_turn",
    }
}

// Emits the metadata object Anthropic pairs with a `refusal` stop reason.
//
// An Anthropic source keeps its `stop_details` in provider extensions, so replay that
// object and preserve the named policy category and its explanation. Only a refusal
// synthesized from a provider that reports no category falls back to the null form,
// which Anthropic documents as the normal value for a refusal that maps to no named
// category.
fn anthropic_stop_details(reason: StopReason, source: Option<&Value>) -> Value {
    match reason {
        StopReason::ContentFilter => source
            .filter(|details| !details.is_null())
            .cloned()
            .unwrap_or_else(|| {
                json!({
                    "type": "refusal",
                    "category": Value::Null,
                    "explanation": Value::Null,
                })
            }),
        _ => Value::Null,
    }
}