fix(adapters): stop dropping non-streaming tool calls (agent loops died on turn 2)

Four adapters handled tool_calls in transform_stream_chunk but lost them in
transform_response, so any NON-streaming tool-using conversation broke on its
second request: the client received finish_reason:"tool_calls" with no
tool_calls payload, replayed an assistant message whose function
name/arguments were empty, and the upstream rejected the next turn with

    400 invalid tool_call function, function/name/arguments cannot be empty

The production audit trail shows 46 such failures on sensenova alone.

- sensenova.lua: forward message.tool_calls, decoding the arguments JSON string
  into an object as the unified shape expects.
- gemini.lua: collect functionCall parts from candidates[].content.parts. Also
  correct finish_reason, since Gemini reports "STOP" even when it emitted a
  function call and clients keyed on it treat that as a finished answer.
- ollama.lua: the field was initialized to an empty table and never filled;
  fill it and likewise correct done_reason "stop" -> "tool_calls".

trae is a different failure with the same symptom: trae-local-api's OpenAI
endpoint (/v1/chat/completions, src/server.js:353) never reads the request's
`tools` array — only its Anthropic endpoint does — so the relayed model is never
told the tool schema and instead PRINTS a <tool_call>{...}</tool_call> block into
content, leaving message.tool_calls null and finish_reason "stop". An OpenAI
client sees an ordinary completion and its agent loop ends mid-conversation.
trae.lua now recovers the structured call from that text, strips the block from
user-visible content, and corrects finish_reason. Both tag spellings
(<tool_call>/<toolcall>, the latter is what the same codebase's Anthropic prompt
asks for) and all three argument key names (arguments/params/input) are accepted.
This is a defensive fallback: fixing the upstream shim to honour `tools` remains
the real fix, since the model still guesses parameter names.

Tests: TestNonStreamToolCallsPreserved covers all ten OpenAI-shaped adapters,
TestGeminiNonStreamToolCalls and TestOllamaNonStreamToolCalls cover their native
shapes, TestTraeTextToolCallRecovery covers both tag spellings, prose around the
block, and asserts a plain text answer never gains tool_calls.

Verified end-to-end against mock upstreams reproducing each shape: a full
two-round agent loop (tool call -> tool result -> final answer) now completes for
both the structured and the text-emitted variants.

internal/lua/adapters/gemini.lua

@@ -81,16 +81,42 @@ function adapter.transform_response(raw_body)
 
     if resp.candidates and #resp.candidates > 0 then
         local cand = resp.candidates[1]
+        local tools = {}
         if cand.content and cand.content.parts then
             for _, part in ipairs(cand.content.parts) do
                 if part.text then
                     unified.content = unified.content .. part.text
+                elseif part.functionCall then
+                    -- Non-streaming tool calls used to be dropped here while
+                    -- transform_stream_chunk handled them, so a non-streaming
+                    -- agent turn looked like a plain text answer and the tool
+                    -- loop died. Gemini's args are already an object.
+                    local args = part.functionCall.args
+                    if type(args) == "string" then
+                        local aok, decoded = pcall(json.decode, args)
+                        args = aok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    table.insert(tools, {
+                        id = part.functionCall.id or ("call_" .. #tools),
+                        type = "function",
+                        name = part.functionCall.name or "",
+                        arguments = args
+                    })
                 end
             end
         end
         if cand.finishReason then
             unified.finish_reason = cand.finishReason
         end
+        if #tools > 0 then
+            unified.tool_calls = tools
+            -- Gemini reports finishReason "STOP" even when it emitted a
+            -- functionCall; clients keyed on finish_reason would treat that as
+            -- a completed answer and never run the tool.
+            unified.finish_reason = "tool_calls"
+        end
     end
 
     return json.encode(unified)

internal/lua/adapters/ollama.lua

@@ -58,13 +58,39 @@ function adapter.transform_response(raw_body)
     local unified = {
         content = "",
         finish_reason = resp.done_reason or "",
-        tool_calls = {},
         -- key must be token_usage to match Go's UnifiedResponse json tag
         token_usage = { prompt = p, completion = c, total = p + c }
     }
 
     if resp.message then
         unified.content = resp.message.content or ""
+        -- Non-streaming tool calls were previously dropped: the field was
+        -- initialized to an empty table and never filled, while
+        -- transform_stream_chunk handled them. A non-streaming agent turn thus
+        -- looked like a plain answer and the tool loop stopped.
+        if type(resp.message.tool_calls) == "table" and #resp.message.tool_calls > 0 then
+            local tcs = {}
+            for _, tc in ipairs(resp.message.tool_calls) do
+                local fn = tc["function"] or {}
+                -- Ollama sends arguments as an object already
+                local args = fn.arguments
+                if type(args) == "string" then
+                    local aok, decoded = pcall(json.decode, args)
+                    args = aok and decoded or {}
+                elseif type(args) ~= "table" then
+                    args = {}
+                end
+                table.insert(tcs, {
+                    id = tc.id or ("call_" .. #tcs),
+                    type = tc.type or "function",
+                    name = fn.name or "",
+                    arguments = args
+                })
+            end
+            unified.tool_calls = tcs
+            -- Ollama reports done_reason "stop" alongside tool calls
+            unified.finish_reason = "tool_calls"
+        end
     end
 
     return json.encode(unified)

internal/lua/adapters/sensenova.lua

@@ -61,6 +61,34 @@ function adapter.transform_response(raw_body)
                 -- sensenova-6.8-flash-lite 用 reasoning 字段而不是 reasoning_content
                 unified.reasoning_content = ch.message.reasoning
             end
+            -- Tool calls MUST be forwarded. Dropping them while keeping
+            -- finish_reason="tool_calls" makes the client replay an assistant
+            -- message whose function name/arguments are empty, and sensenova
+            -- then rejects the next turn with
+            --   400 invalid tool_call function, function/name/arguments cannot be empty
+            -- i.e. a tool-using conversation dies on its second request.
+            if type(ch.message.tool_calls) == "table" and #ch.message.tool_calls > 0 then
+                local tcs = {}
+                for _, tc in ipairs(ch.message.tool_calls) do
+                    local fn = tc["function"] or {}
+                    -- arguments arrives as a JSON *string* on the wire; the
+                    -- unified shape expects a decoded object.
+                    local args = fn.arguments
+                    if type(args) == "string" then
+                        local args_ok, decoded = pcall(json.decode, args)
+                        args = args_ok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = fn.name,
+                        arguments = args
+                    })
+                end
+                unified.tool_calls = tcs
+            end
         end
         unified.finish_reason = ch.finish_reason or ""
     end

internal/lua/adapters/trae.lua

@@ -25,6 +25,78 @@ function adapter.transform_request(raw_body)
     return json.encode(req)
 end
 
+-- parse_text_tool_calls extracts tool calls that an upstream emitted as PLAIN
+-- TEXT instead of using the OpenAI tool_calls field.
+--
+-- trae-local-api's OpenAI endpoint (/v1/chat/completions) does not read the
+-- request's `tools` array at all, so the relayed model is never told the tool
+-- schema; it falls back to printing
+--     <tool_call>
+--     {"name": "get_weather", "arguments": {"location": "北京"}}
+--     </tool_call>
+-- into message.content, leaves message.tool_calls null, and reports
+-- finish_reason="stop". A client following the OpenAI contract therefore never
+-- sees a tool call: the agent loop terminates unexpectedly mid-conversation
+-- (and a hand-written replay produces an empty function name next turn).
+--
+-- Both tag spellings are accepted: the same codebase's Anthropic endpoint
+-- instructs models to emit <toolcall>, and models mix the two. Key names vary
+-- too (arguments / params / input), so all are tried.
+--
+-- Returns (tool_calls_array_or_nil, content_with_blocks_removed).
+local function parse_text_tool_calls(content)
+    if type(content) ~= "string" or content == "" then return nil, content end
+    if not (content:find("<tool_call", 1, true) or content:find("<toolcall", 1, true)) then
+        return nil, content
+    end
+
+    local tcs = {}
+    local idx = 0
+
+    local function collect(pattern)
+        for payload in content:gmatch(pattern) do
+            local ok, obj = pcall(json.decode, payload)
+            if ok and type(obj) == "table" then
+                -- some models wrap it as {"function":{"name":..,"arguments":..}}
+                local fn = obj["function"]
+                local name = obj.name or (type(fn) == "table" and fn.name) or nil
+                if name then
+                    local args = obj.arguments or obj.params or obj.input
+                    if args == nil and type(fn) == "table" then
+                        args = fn.arguments or fn.params
+                    end
+                    if type(args) == "string" then
+                        local aok, decoded = pcall(json.decode, args)
+                        args = aok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
+                    idx = idx + 1
+                    table.insert(tcs, {
+                        id = obj.id or ("call_text_" .. idx),
+                        type = "function",
+                        name = name,
+                        arguments = args
+                    })
+                end
+            end
+        end
+    end
+
+    -- `<tag ...>` allows attributes; %s* handles the "</tool_call >" spacing
+    -- that trae-local-api's own prompt example uses.
+    collect("<tool_call[^>]*>%s*(.-)%s*</tool_call%s*>")
+    collect("<toolcall[^>]*>%s*(.-)%s*</toolcall%s*>")
+
+    if #tcs == 0 then return nil, content end
+
+    -- drop the blocks from user-visible content; keep any surrounding prose
+    local stripped = content:gsub("<tool_call[^>]*>%s*.-%s*</tool_call%s*>", "")
+    stripped = stripped:gsub("<toolcall[^>]*>%s*.-%s*</toolcall%s*>", "")
+    stripped = stripped:gsub("^%s+", ""):gsub("%s+$", "")
+    return tcs, stripped
+end
+
 function adapter.transform_response(raw_body)
     -- trae-local-api 偶发在非流式请求中返回 SSE 格式数据，
     -- 表现为多个 data: {...} 行或混合了 reasoning_chunk 等。
@@ -71,15 +143,21 @@ function adapter.transform_response(raw_body)
             if ch.message.reasoning_content then
                 unified.reasoning_content = ch.message.reasoning_content
             end
-            if type(ch.message.tool_calls) == "table" then
+            if type(ch.message.tool_calls) == "table" and #ch.message.tool_calls > 0 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
+                    local fn = tc["function"] or {}
+                    local args = fn.arguments
+                    if type(args) == "string" then
+                        local args_ok, decoded = pcall(json.decode, args)
+                        args = args_ok and decoded or {}
+                    elseif type(args) ~= "table" then
+                        args = {}
+                    end
                     table.insert(tcs, {
                         id = tc.id,
                         type = tc.type or "function",
-                        name = tc["function"].name,
+                        name = fn.name,
                         arguments = args
                     })
                 end
@@ -87,6 +165,22 @@ function adapter.transform_response(raw_body)
             end
         end
         unified.finish_reason = ch.finish_reason or ""
+
+        -- Fallback: trae-local-api's OpenAI endpoint drops the request's
+        -- `tools` array entirely, so the relayed model is never told the tool
+        -- schema and instead PRINTS a <tool_call>{...}</tool_call> block into
+        -- content, leaving message.tool_calls null and finish_reason="stop".
+        -- A client following the OpenAI contract then sees a normal completion
+        -- and its agent loop terminates mid-conversation. Recover the
+        -- structured call so the loop can continue.
+        if unified.tool_calls == nil then
+            local recovered, cleaned = parse_text_tool_calls(unified.content)
+            if recovered then
+                unified.tool_calls = recovered
+                unified.content = cleaned
+                unified.finish_reason = "tool_calls"
+            end
+        end
     end
 
     return json.encode(unified)