strip :THINKING in e2ee_round_trip, drop per-model defaults

lua/claude_handler.lua

@@ -54,7 +54,7 @@ function _M.handle()
     if not is_streaming then
         -- Non-streaming
         local decrypted, round_err = e2ee.e2ee_round_trip(
-            api_key, model, oai_body, false, "/v1/chat/completions"
+            api_key, model, oai_body, false, "/v1/chat/completions", nil, oai_request
         )
 
         if not decrypted then
@@ -101,7 +101,8 @@ function _M.handle()
                         ngx.flush(true)
                     end
                 end
-            end)
+            end,
+            oai_request)
 
         if round_err then
             if round_err.raw then

lua/e2ee_handler.lua

@@ -14,9 +14,20 @@ local cjson = require("cjson.safe")
 local http = require("resty.http")
 
 local API_BASE = "https://api.chutes.ai"
+local payload_cjson
 
 local _M = {}
 
+local function get_payload_cjson()
+    if not payload_cjson then
+        payload_cjson = cjson.new()
+        if payload_cjson.decode_array_with_array_mt then
+            payload_cjson.decode_array_with_array_mt(true)
+        end
+    end
+    return payload_cjson
+end
+
 --- Extract API key from Authorization header or x-api-key
 function _M.get_api_key()
     local headers = ngx.req.get_headers()
@@ -107,9 +118,70 @@ end
 --- E2EE round-trip: encrypt request, send, decrypt response
 -- For non-streaming: returns (decrypted_json_string, nil) or (nil, {status=N, message=S})
 -- For streaming: calls on_chunk(line) for each decrypted SSE data line, on_chunk(nil) at end
-function _M.e2ee_round_trip(api_key, model, body_json, is_streaming, e2e_path, on_chunk)
+function _M.e2ee_round_trip(api_key, model, body_json, is_streaming, e2e_path, on_chunk, request_payload)
     local err
 
+    local thinking
+    local base_model = model:match("^(.-):THINKING$")
+    local changed = false
+
+    if base_model then
+        model = base_model
+        thinking = true
+    end
+
+    local h = ngx.var and ngx.var.http_x_enable_thinking
+    if type(h) == "string" then
+        h = h:lower()
+        if h == "true" then
+            thinking = true
+        elseif h == "false" then
+            thinking = false
+        end
+    end
+
+    local kwargs_hint = type(request_payload) == "table"
+        and type(request_payload.chat_template_kwargs) == "table"
+        and request_payload.chat_template_kwargs or nil
+    local normalize_kwargs = kwargs_hint
+        and ((kwargs_hint.thinking ~= nil and kwargs_hint.enable_thinking == nil)
+            or (kwargs_hint.enable_thinking ~= nil and kwargs_hint.thinking == nil))
+
+    if thinking ~= nil or normalize_kwargs then
+        local json = get_payload_cjson()
+        local payload = json.decode(body_json)
+        if payload then
+            if base_model then
+                payload.model = model
+                changed = true
+            end
+
+            local kwargs = type(payload.chat_template_kwargs) == "table" and payload.chat_template_kwargs or nil
+            if kwargs and getmetatable(kwargs) == json.array_mt then
+                kwargs = nil
+            end
+            if thinking ~= nil then
+                kwargs = kwargs or {}
+                kwargs.thinking = thinking
+                kwargs.enable_thinking = thinking
+                payload.chat_template_kwargs = kwargs
+                changed = true
+            elseif kwargs then
+                if kwargs.thinking ~= nil and kwargs.enable_thinking == nil then
+                    kwargs.enable_thinking = kwargs.thinking
+                    changed = true
+                elseif kwargs.enable_thinking ~= nil and kwargs.thinking == nil then
+                    kwargs.thinking = kwargs.enable_thinking
+                    changed = true
+                end
+            end
+
+            if changed then
+                body_json = json.encode(payload)
+            end
+        end
+    end
+
     -- Resolve model -> chute_id
     local chute_id
     chute_id, err = discovery.resolve_chute_id(model, api_key)
@@ -354,7 +426,9 @@ function _M.handle()
     end
 
     if not is_streaming then
-        local decrypted, round_err = _M.e2ee_round_trip(api_key, model, body, false, original_path)
+        local decrypted, round_err = _M.e2ee_round_trip(
+            api_key, model, body, false, original_path, nil, payload
+        )
         if not decrypted then
             if round_err.raw then
                 ngx.status = round_err.status
@@ -384,7 +458,8 @@ function _M.handle()
                         ngx.flush(true)
                     end
                 end
-            end)
+            end,
+            payload)
 
         if round_err then
             if round_err.raw then

lua/responses_handler.lua

@@ -54,7 +54,7 @@ function _M.handle()
     if not is_streaming then
         -- Non-streaming
         local decrypted, round_err = e2ee.e2ee_round_trip(
-            api_key, model, oai_body, false, "/v1/chat/completions"
+            api_key, model, oai_body, false, "/v1/chat/completions", nil, oai_request
         )
 
         if not decrypted then
@@ -101,7 +101,8 @@ function _M.handle()
                         ngx.flush(true)
                     end
                 end
-            end)
+            end,
+            oai_request)
 
         if round_err then
             if round_err.raw then