ModelRouter/internal/lua/adapters/openai.lua

local adapter = {}

adapter.name = "openai"
adapter.version = "2.0.0"
adapter.endpoint = "/chat/completions"
adapter.headers = {}

-- Claude-behind-OpenAI upstreams (tabitoken, 扇贝, …) convert /chat/completions
-- to the Anthropic Messages API internally, so they inherit Anthropic's
-- tool id rule: ^[a-zA-Z0-9_-]{1,64}$, enforced by rejecting the WHOLE request
-- with "Invalid tool use format" / "tool_use.id: String should match pattern".
-- Plain OpenAI has no such rule, so an OpenAI-compatible model can mint an id
-- like "bash:0" (observed from moonshotai/kimi-k3). In a fan-out router that id
-- is replayed to every other source, so one such id kills every Claude slot at
-- once and an AUTO request falls through all tiers.
--
-- safe_tool_id is pure and deterministic, so a tool_calls entry and its
-- matching tool_call_id are rewritten identically within one request. A
-- rewritten id keeps an 8-hex digest of the ORIGINAL id, without which two
-- distinct ids could collapse into one ("a:b" and "a_b") and become an
-- unpaired/duplicate tool call. Already-legal ids are returned untouched, so
-- well-behaved traffic is byte-identical to before.
-- (anthropic.lua carries the same helper; Lua adapters have no shared prelude.)
local TOOL_ID_MAX = 64

local function safe_tool_id(id)
    if type(id) ~= "string" or id == "" then return id end
    local clean = string.gsub(id, "[^A-Za-z0-9_-]", "_")
    if clean == id and #clean <= TOOL_ID_MAX then
        return clean
    end
    local digest = string.sub(sha256_hex(id), 1, 8)
    local keep = TOOL_ID_MAX - #digest - 1
    if #clean > keep then clean = string.sub(clean, 1, keep) end
    return clean .. "_" .. digest
end

-- OpenAI /chat/completions format (pass-through, strip provider-specific fields)
function adapter.transform_request(raw_body)
    local ok, req = pcall(json.decode, raw_body)
    if not ok then return raw_body end
    -- The gateway's private control fields are GONE from the body (v2): the
    -- neutral `thinking` field carries that intent and each adapter maps it in
    -- its own protocol terms. These upstreams have no thinking switch at all,
    -- so the field is simply dropped here — which is what the nil-ing below
    -- used to do, except the field being dropped is no longer the kernel's
    -- private one.
    req.thinking = nil
    req.extra_body = nil
    if req.messages then
        for _, msg in ipairs(req.messages) do
            msg.reasoning_content = nil
            if msg.tool_call_id ~= nil then
                msg.tool_call_id = safe_tool_id(msg.tool_call_id)
            end
            if type(msg.tool_calls) == "table" then
                for _, tc in ipairs(msg.tool_calls) do
                    if type(tc) == "table" and tc.id ~= nil then
                        tc.id = safe_tool_id(tc.id)
                    end
                end
            end
        end
    end
    return json.encode(req)
end

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 type(resp.usage) == "table" then
        unified.token_usage.prompt = resp.usage.prompt_tokens or 0
        unified.token_usage.completion = resp.usage.completion_tokens or 0
        unified.token_usage.total = resp.usage.total_tokens or 0
        -- Cache passthrough: OpenAI v2 prompt_tokens_details.cached_tokens
        -- and DeepSeek-legacy prompt_cache_hit/miss_tokens. dsh reads
        -- prompt_tokens_details.cached_tokens (falls back to the legacy
        -- standalone field), so both shapes reach clients.
        local hit = 0
        if type(resp.usage.prompt_tokens_details) == "table" and resp.usage.prompt_tokens_details.cached_tokens ~= nil then
            hit = resp.usage.prompt_tokens_details.cached_tokens
            unified.token_usage.prompt_tokens_details = { cached_tokens = hit }
        end
        if (resp.usage.prompt_cache_hit_tokens or 0) > 0 then
            unified.token_usage.prompt_cache_hit_tokens = resp.usage.prompt_cache_hit_tokens
            unified.token_usage.prompt_cache_miss_tokens = resp.usage.prompt_cache_miss_tokens or 0
            if hit == 0 then
                unified.token_usage.prompt_tokens_details = { cached_tokens = resp.usage.prompt_cache_hit_tokens }
            end
        end
    end

    if type(resp.choices) == "table" and #resp.choices > 0 then
        local ch = resp.choices[1]
        if type(ch.message) == "table" then
            unified.content = ch.message.content or ""
            if ch.message.reasoning_content then
                unified.reasoning_content = ch.message.reasoning_content
            end
            if type(ch.message.tool_calls) == "table" then
                local tcs = {}
                for _, tc in ipairs(ch.message.tool_calls) do
                    local args_ok, args = pcall(json.decode, tc["function"].arguments)
                    if not args_ok then args = {} end
                    table.insert(tcs, {
                        id = safe_tool_id(tc.id),
                        type = tc.type or "function",
                        name = tc["function"].name,
                        arguments = args
                    })
                end
                unified.tool_calls = tcs
            end
        end
        unified.finish_reason = ch.finish_reason or ""
    end

    return json.encode(unified)
end

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

    -- OpenAI-style streams may attach usage to a chunk with empty choices
    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
    local uses = nil
    if type(chunk.usage) == "table" then
        uses = {
            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
            total = chunk.usage.total_tokens or chunk.usage.total or 0,
        }
        if type(chunk.usage.prompt_tokens_details) == "table" and chunk.usage.prompt_tokens_details.cached_tokens ~= nil then
            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_tokens_details.cached_tokens }
        elseif (chunk.usage.prompt_cache_hit_tokens or 0) > 0 then
            uses.prompt_cache_hit_tokens = chunk.usage.prompt_cache_hit_tokens
            uses.prompt_cache_miss_tokens = chunk.usage.prompt_cache_miss_tokens or 0
            uses.prompt_tokens_details = { cached_tokens = chunk.usage.prompt_cache_hit_tokens }
        end
    end

    if not chunk.choices or #chunk.choices == 0 then
        if uses ~= nil then
            -- usage-only chunk is not a content/finish signal; the gateway
            -- emits its own terminal stop chunk and merges this usage.
            return json.encode({ usage = uses, done = false })
        end
        return ""
    end
    local delta = chunk.choices[1].delta or {}
    local fr = chunk.choices[1].finish_reason

    local finish = (type(fr) == "string" and fr ~= "") and fr or nil

    local unified = {
        content = delta.content or "",
        done = (finish ~= nil)
    }
    if finish then
        unified.finish_reason = finish
    end
    if uses ~= nil then
        unified.usage = uses
    end
    if delta.reasoning_content then
        unified.reasoning_content = delta.reasoning_content
    end
    if delta.tool_calls then
        -- Sanitize on the way OUT too: an id this upstream happily minted (it
        -- does not validate them) becomes a landmine once the client replays
        -- it to a Claude upstream. Only the first fragment of a streamed call
        -- carries an id; later argument fragments have none and are untouched.
        for _, tc in ipairs(delta.tool_calls) do
            if type(tc) == "table" and tc.id ~= nil then
                tc.id = safe_tool_id(tc.id)
            end
        end
        unified.tool_calls = delta.tool_calls
    end
    return json.encode(unified)
end

-- 错误收敛：标准 OpenAI 信封 {error:{message,...}}
function adapter.transform_error(status, body)
    -- Non-JSON body (nginx HTML error pages, plain text): extract a short
    -- human-readable reason instead of letting the raw body reach the log.
    local ok, resp = pcall(json.decode, body)
    if not ok or type(resp) ~= "table" then
        -- HTML error page: pull the <title> text (e.g. "413 Request Entity Too Large")
        local title = string.match(body or "", "<title>(.-)</title>")
        if title and title ~= "" then return title end
        -- Bare text: first non-empty line, capped
        local line = string.match(body or "", "^%s*([^\r\n]+)")
        if line and line ~= "" and not string.match(line, "^<") then
            return string.sub(line, 1, 200)
        end
        return nil
    end

    -- Standard OpenAI envelope: {error:{message,...}} or {error:"..."}
    local e = resp.error
    if type(e) == "table" and type(e.message) == "string" then
        return e.message
    end
    if type(e) == "string" then return e end

    -- Flat envelope used by many OpenAI-compatible gateways:
    --   {"code":20012,"message":"Model does not exist..."}
    --   {"code":"INVALID_API_KEY","message":"Invalid API key"}
    if type(resp.message) == "string" and resp.message ~= "" then
        if resp.code ~= nil then
            return tostring(resp.code) .. ": " .. resp.message
        end
        return resp.message
    end

    -- Some gateways use {detail:"..."} (FastAPI style)
    if type(resp.detail) == "string" and resp.detail ~= "" then
        return resp.detail
    end

    return nil
end

return adapter