ModelRouter/internal/lua/adapters/anthropic.lua

local adapter = {}
adapter.name = "anthropic"
adapter.version = "3.0.0"
adapter.endpoint = "/v1/messages"
adapter.headers = {
    ["anthropic-version"] = "2023-06-01",
}

-- ============================================================
-- helpers
-- ============================================================

local ROLE_WHITELIST = {
    system = true, user = true, assistant = true, tool = true,
}

local function collect_blocks(content)
    if type(content) == "string" then
        if content == "" then return {} end
        return { { type = "text", text = content } }
    end
    if type(content) ~= "table" then return {} end
    local blocks = {}
    for _, p in ipairs(content) do
        if type(p) == "string" then
            table.insert(blocks, { type = "text", text = p })
        elseif type(p) == "table" then
            if p.type == "text" then
                table.insert(blocks, { type = "text", text = p.text or "" })
            elseif p.type == "image_url" and type(p.image_url) == "table" and p.image_url.url then
                local url = p.image_url.url
                local mt, b64 = string.match(url, "^data:([^,]+);base64,(.+)$")
                if b64 then
                    table.insert(blocks, { type = "image", source = { type = "base64", media_type = mt or "image/png", data = b64 } })
                else
                    table.insert(blocks, { type = "image", source = { type = "url", url = url } })
                end
            else
                -- Unknown content type: preserve for forward compatibility
                table.insert(blocks, p)
            end
        end
    end
    return blocks
end

local function text_of(content)
    if type(content) == "string" then return content end
    local t = ""
    for _, p in ipairs(content or {}) do
        if type(p) == "string" then
            t = t .. p
        elseif type(p) == "table" and p.type == "text" and p.text then
            t = t .. p.text
        end
    end
    return t
end

--- Append blocks to the last user message (merge) or create a new one.
local function append_user(msgs, blocks)
    if #msgs > 0 and msgs[#msgs].role == "user" then
        for _, b in ipairs(blocks) do
            table.insert(msgs[#msgs].content, b)
        end
    else
        table.insert(msgs, { role = "user", content = blocks })
    end
end

-- ============================================================
-- transform_request  (OpenAI → Anthropic)
-- ============================================================

function adapter.transform_request(raw_body)
    local ok, req = pcall(json.decode, raw_body)
    if not ok or type(req) ~= "table" then return raw_body end

    local msgs = {}
    local system = ""
    local pending_tool = {}  -- accumulated tool_result blocks

    for _, m in ipairs(req.messages or {}) do
        local role = m.role or "user"

        if role == "system" then
            system = system .. text_of(m.content) .. "\n"

        elseif role == "assistant" then
            -- Build content array: text/image blocks from content + tool_use from tool_calls
            local blocks = collect_blocks(m.content)

            if type(m.tool_calls) == "table" then
                for _, tc in ipairs(m.tool_calls) do
                    if type(tc) == "table" then
                        local fn = tc["function"] or {}
                        local args = fn.arguments
                        local input = {}
                        if type(args) == "string" and args ~= "" then
                            local ok2, parsed = pcall(json.decode, args)
                            if ok2 and type(parsed) == "table" then input = parsed end
                        elseif type(args) == "table" then
                            input = args
                        end
                        table.insert(blocks, {
                            type  = "tool_use",
                            id    = tc.id or "",
                            name  = fn.name or "",
                            input = input,
                        })
                    end
                end
            end

            table.insert(msgs, { role = "assistant", content = blocks })

        elseif role == "tool" then
            -- Accumulate consecutive tool results; will be flushed as one
            -- user message with tool_result content blocks.
            table.insert(pending_tool, {
                type        = "tool_result",
                tool_use_id = m.tool_call_id or "",
                content     = text_of(m.content),
            })

        else
            -- user / any other role: flush pending tool results first
            if #pending_tool > 0 then
                append_user(msgs, pending_tool)
                pending_tool = {}
            end
            append_user(msgs, collect_blocks(m.content))
        end
    end

    -- Flush any trailing tool results
    if #pending_tool > 0 then
        append_user(msgs, pending_tool)
    end

    -- ── Build Anthropic request ──────────────────────────────
    local anthropic_req = {
        model      = req.model or "claude-sonnet-4-20250514",
        max_tokens = req.max_tokens or 4096,
        messages   = msgs,
        stream     = req.stream or false,
    }

    -- ── tools ────────────────────────────────────────────────
    local has_tools = false
    if req.tools and type(req.tools) == "table" then
        local tools = {}
        for _, t in ipairs(req.tools) do
            if type(t) == "table" and t.type == "function"
                and type(t["function"]) == "table" then
                local fn = t["function"]
                table.insert(tools, {
                    name         = fn.name or "",
                    description  = fn.description or "",
                    input_schema = fn.parameters or { type = "object", properties = {} },
                })
            end
        end
        if #tools > 0 then
            anthropic_req.tools = tools
            has_tools = true
        end
    end

    -- ── tool_choice ──────────────────────────────────────────
    -- OpenAI → Anthropic mapping:
    --   "auto"     → {type:"auto"}
    --   "none"     → remove tools entirely (Anthropic has no "none")
    --   "required" → {type:"any"}
    --   {type:"function", function:{name:"X"}} → {type:"tool", name:"X"}
    if has_tools and req.tool_choice then
        local tc = req.tool_choice
        if type(tc) == "string" then
            if tc == "auto" then
                anthropic_req.tool_choice = { type = "auto" }
            elseif tc == "none" then
                anthropic_req.tools = nil
                anthropic_req.tool_choice = nil
            elseif tc == "required" then
                anthropic_req.tool_choice = { type = "any" }
            end
        elseif type(tc) == "table" and tc.type == "function" then
            local fn = tc["function"] or {}
            anthropic_req.tool_choice = { type = "tool", name = fn.name or "" }
        end
    end

    -- ── thinking (opt-in via extra_body) ──────────────────────
    -- Default: OFF. Client sends extra_body.thinking to enable.
    -- Example: {"extra_body": {"thinking": {"type": "enabled", "budget_tokens": 4096}}}
    if type(req.extra_body) == "table" and type(req.extra_body.thinking) == "table" then
        anthropic_req.thinking = req.extra_body.thinking
    end

    -- ── system ───────────────────────────────────────────────
    if system ~= "" then
        -- Strip trailing newline from accumulation
        anthropic_req.system = string.match(system, "^(.-)\n*$")
    end

    return json.encode(anthropic_req)
end

-- ============================================================
-- transform_response  (Anthropic → OpenAI, non-streaming)
-- ============================================================

function adapter.transform_response(raw_body)
    local ok, resp = pcall(json.decode, raw_body)
    if not ok or resp == nil then return raw_body end

    local unified = {
        content = "",
        finish_reason = "",
        token_usage = { prompt = 0, completion = 0, total = 0 },
    }

    if resp.usage then
        unified.token_usage.prompt    = resp.usage.input_tokens  or 0
        unified.token_usage.completion = resp.usage.output_tokens or 0
        unified.token_usage.total     = (resp.usage.input_tokens or 0) + (resp.usage.output_tokens or 0)
        -- Emit details whenever Anthropic reports the field, even at 0, so a
        -- reported cache miss stays distinguishable from "not reported".
        if resp.usage.cache_read_input_tokens ~= nil then
            unified.token_usage.prompt_tokens_details = {
                cached_tokens = resp.usage.cache_read_input_tokens
            }
        end
    end

    if resp.content and #resp.content > 0 then
        local tcs = {}
        for _, block in ipairs(resp.content) do
            if block.type == "text" then
                unified.content = unified.content .. (block.text or "")
            elseif block.type == "thinking" and block.thinking then
                unified.reasoning_content = (unified.reasoning_content or "") .. block.thinking
            elseif block.type == "tool_use" then
                table.insert(tcs, {
                    id        = block.id or "",
                    type      = "function",
                    name      = block.name or "",
                    arguments = block.input or {},
                })
            end
        end
        if #tcs > 0 then unified.tool_calls = tcs end
    end

    unified.finish_reason = resp.stop_reason or ""
    if unified.finish_reason == "tool_use" then
        unified.finish_reason = "tool_calls"
    elseif unified.finish_reason == "max_tokens" then
        unified.finish_reason = "length"
    elseif unified.finish_reason == "end_turn"
        or unified.finish_reason == "stop_sequence" then
        unified.finish_reason = "stop"
    end

    return json.encode(unified)
end

-- ============================================================
-- transform_stream_chunk  (Anthropic SSE → OpenAI SSE delta)
-- ============================================================

function adapter.transform_stream_chunk(raw_chunk)
    local ok, chunk = pcall(json.decode, raw_chunk)
    if not ok then return "" end

    -- ── message_start: initial usage ─────────────────────────
    if chunk.type == "message_start" then
        if chunk.message and type(chunk.message.usage) == "table" then
            local u = chunk.message.usage
            local p = u.input_tokens or 0
            local c = u.output_tokens or 0
            if p > 0 or c > 0 then
                local uses = { prompt = p, completion = c, total = p + c }
                if u.cache_read_input_tokens ~= nil then
                    uses.prompt_tokens_details = {
                        cached_tokens = u.cache_read_input_tokens
                    }
                end
                return json.encode({ usage = uses, done = false })
            end
        end
        return ""
    end

    -- ── message_delta: stop_reason + final usage ─────────────
    if chunk.type == "message_delta" then
        local finish = nil
        if chunk.delta and chunk.delta.stop_reason ~= nil then
            local sr = chunk.delta.stop_reason
            if sr == "max_tokens" then
                finish = "length"
            elseif sr == "tool_use" then
                finish = "tool_calls"
            else
                finish = "stop"
            end
        end
        local uses = nil
        if type(chunk.usage) == "table" then
            local u = chunk.usage
            local p = u.input_tokens or 0
            local c = u.output_tokens or 0
            if p > 0 or c > 0 then
                uses = { prompt = p, completion = c, total = p + c }
            end
        end
        if uses ~= nil then
            return json.encode({ content = "", done = (finish ~= nil), finish_reason = finish, usage = uses })
        end
        return json.encode({ content = "", done = (finish ~= nil), finish_reason = finish })
    end

    -- ── content_block_start: begin text / thinking / tool_use ─
    if chunk.type == "content_block_start" and chunk.content_block then
        local cb = chunk.content_block
        if cb.type == "tool_use" then
            -- Pass Anthropic content_block index through; Go-side rewrites
            -- to sequential OpenAI tool_call ordinal for multi-tool streams.
            return json.encode({
                content = "", done = false,
                tool_calls = { {
                    index = chunk.index or 0,
                    id    = cb.id or "",
                    type  = "function",
                    ["function"] = { name = cb.name or "", arguments = "" },
                } },
            })
        end
        return ""  -- text / thinking block start: no OpenAI equivalent
    end

    -- ── content_block_delta: incremental content ──────────────
    if chunk.type == "content_block_delta" and chunk.delta then
        local d = chunk.delta
        if d.type == "input_json_delta" then
            -- Tool call argument fragment; Go-side rewrites index.
            return json.encode({
                content = "", done = false,
                tool_calls = { {
                    index = chunk.index or 0,
                    id    = "",
                    type  = "function",
                    ["function"] = { name = "", arguments = d.partial_json or "" },
                } },
            })
        end
        if d.type == "thinking_delta" and d.thinking then
            return json.encode({ content = "", done = false, reasoning_content = d.thinking })
        end
        if d.type == "text_delta" and d.text then
            return json.encode({ content = d.text, done = false })
        end
    end

    -- ── content_block_stop / message_stop ─────────────────────
    if chunk.type == "message_stop" then
        -- The message_delta event already emitted the true finish_reason.
        -- Do NOT emit done=true here: an empty finish_reason would
        -- overwrite the real one (tool_calls) in the Go gateway's
        -- lastFinish tracker, causing the final SSE chunk to say
        -- finish_reason=stop instead of tool_calls.
        return ""
    end

    return ""
end

-- ============================================================
-- transform_error  (Anthropic error → human-readable string)
-- ============================================================

function adapter.transform_error(status, body)
    local ok, resp = pcall(json.decode, body)
    if not ok or type(resp) ~= "table" then return nil end

    -- Anthropic envelope: {type:"error", error:{type, message}}
    if resp.type == "error" and type(resp.error) == "table"
        and type(resp.error.message) == "string" then
        return resp.error.message
    end

    -- Flat envelope: {error: {message: "..."}}
    if type(resp.error) == "table" and type(resp.error.message) == "string" then
        return resp.error.message
    end

    return nil
end

return adapter