feat(gateway): pass through upstream finish_reason end-to-end

The gateway hardcoded "stop" on every terminating stream chunk, so
tool-call rounds reported finish_reason=stop and length caps were
invisible to clients. UnifiedChunk now carries finish_reason; adapters
emit it (with empty-string finish reasons like sensenova treated as
non-terminal), standardSSEChunk passes it through for un-adapted
upstreams, [DONE] no longer emits a duplicate reason-less done chunk,
and both streaming paths emit the real reason with "stop" as fallback.

Also vendor sensenova/agentrouter adapters into the repo: they were
WebUI-only uploads and a deploy sync silently removed them while live
AUTO-chain slots still referenced them.

internal/gateway/chat.go

@@ -656,6 +656,7 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 		return
 	}
 	var lastUsage *types.TokenUsage
+	lastFinish := ""
 	for ck := range chunks {
 		if ck.Usage != nil {
 			lastUsage = mergeUsage(lastUsage, ck.Usage)
@@ -672,8 +673,12 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 		}
 		choice := ChunkChoice{Index: 0, Delta: delta}
 		if ck.Done {
-			stop := "stop"
+			fin := ck.FinishReason
-			choice.FinishReason = &stop
+			if fin == "" {
+				fin = "stop"
+			}
+			lastFinish = fin
+			choice.FinishReason = &fin
 		}
 		chunk.Choices = []ChunkChoice{choice}
 		rec.Compl += int64(len(ck.Content)+len(ck.ReasoningContent)+len(ck.ToolCalls)) / 3
@@ -681,10 +686,13 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 			return
 		}
 	}
-	stop := "stop"
+	finalFinish := lastFinish
+	if finalFinish == "" {
+		finalFinish = "stop"
+	}
 	send(ChatChunk{
 		ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &stop}},
+		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &finalFinish}},
 	})
 	// Final usage chunk (OpenAI standard: empty choices + usage before [DONE]).
 	// Prefer the upstream's exact usage if the stream carried it; fall back to
@@ -826,6 +834,7 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 		return
 	}
 	var lastUsage *types.TokenUsage
+	lastFinish := ""
 	for ck := range chunks {
 		if ck.Usage != nil {
 			lastUsage = mergeUsage(lastUsage, ck.Usage)
@@ -842,8 +851,12 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 		}
 		choice := ChunkChoice{Index: 0, Delta: delta}
 		if ck.Done {
-			stop := "stop"
+			fin := ck.FinishReason
-			choice.FinishReason = &stop
+			if fin == "" {
+				fin = "stop"
+			}
+			lastFinish = fin
+			choice.FinishReason = &fin
 		}
 		chunk.Choices = []ChunkChoice{choice}
 		rec.Compl += int64(len(ck.Content)+len(ck.ReasoningContent)+len(ck.ToolCalls)) / 3
@@ -851,10 +864,13 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 			return
 		}
 	}
-	stop := "stop"
+	finalFinish := lastFinish
+	if finalFinish == "" {
+		finalFinish = "stop"
+	}
 	send(ChatChunk{
 		ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
-		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &stop}},
+		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &finalFinish}},
 	})
 	// Final usage chunk (OpenAI standard: empty choices + usage before [DONE]).
 	// Prefer the upstream's exact usage if the stream carried it; fall back to

internal/lua/adapters/agentrouter.lua

@@ -0,0 +1,111 @@
+local adapter = {}
+
+adapter.name = "agentrouter"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+-- AgentRouter 的 WAF 会按客户端指纹白名单校验，只放行官方客户端。
+-- 这里通过 build_headers 注入官方客户端 User-Agent 以通过校验。
+-- 默认 QwenCode 指纹（实测可通）；可在 source.meta.user_agent 覆盖成 Claude Code 指纹。
+
+local default_ua = "QwenCode/0.2.0 (linux; x64)"
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok then return raw_body end
+    req.disable_thinking = nil
+    req.extra_body = nil
+    if req.messages then
+        for _, msg in ipairs(req.messages) do
+            msg.reasoning_content = nil
+        end
+    end
+    return json.encode(req)
+end
+
+-- 注入 Authorization + 官方客户端 User-Agent
+function adapter.build_headers(meta)
+    local ua = default_ua
+    local meta_ua = (meta.source or {}).meta and (meta.source).meta.user_agent
+    if meta_ua and meta_ua ~= "" then
+        ua = meta_ua
+    end
+    return {
+        ["Content-Type"] = "application/json",
+        ["Authorization"] = "Bearer " .. meta.api_key,
+        ["User-Agent"] = ua,
+    }
+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
+    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 = 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
+
+    if not chunk.choices or #chunk.choices == 0 then 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 delta.reasoning_content then
+        unified.reasoning_content = delta.reasoning_content
+    end
+    if delta.tool_calls then
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+return adapter

internal/lua/adapters/anthropic.lua

@@ -124,12 +124,23 @@ function adapter.transform_stream_chunk(raw_chunk)
                 uses = { prompt = p, completion = c, total = p + c }
             end
         end
-        local done = (chunk.delta and chunk.delta.stop_reason ~= nil)
+        local finish = nil
+        if chunk.delta and chunk.delta.stop_reason ~= nil then
+            -- Anthropic stop_reason -> OpenAI finish_reason
+            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
         if uses ~= nil then
             -- completion is final here; prompt is merged from message_start
-            return json.encode({ content = "", done = done, usage = uses })
+            return json.encode({ content = "", done = (finish ~= nil), finish_reason = finish, usage = uses })
         end
-        return json.encode({ content = "", done = done })
+        return json.encode({ content = "", done = (finish ~= nil), finish_reason = finish })
     end
     if chunk.type == "content_block_start" and chunk.content_block
         and chunk.content_block.type == "tool_use" then

internal/lua/adapters/deepseek.lua

@@ -106,10 +106,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/gemini.lua

@@ -110,7 +110,19 @@ function adapter.transform_stream_chunk(raw_chunk)
         return ""
     end
     local cand = chunk.candidates[1]
-    local unified = { content = "", done = (cand.finishReason ~= nil) }
+    -- Gemini finishReason -> OpenAI finish_reason
+    local finish = nil
+    if cand.finishReason ~= nil then
+        if cand.finishReason == "MAX_TOKENS" then
+            finish = "length"
+        elseif cand.finishReason == "SAFETY" or cand.finishReason == "RECITATION"
+            or cand.finishReason == "BLOCKLIST" then
+            finish = "content_filter"
+        else
+            finish = "stop"
+        end
+    end
+    local unified = { content = "", done = (finish ~= nil), finish_reason = finish }
     local reasoning = ""
     local tools = {}
     if cand.content and cand.content.parts then

internal/lua/adapters/github.lua

@@ -87,10 +87,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/groq.lua

@@ -86,10 +86,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/kimicode.lua

@@ -119,10 +119,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/mistral.lua

@@ -86,10 +86,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/ollama.lua

@@ -85,16 +85,27 @@ function adapter.transform_stream_chunk(raw_chunk)
         end
     end
 
+    -- Ollama done_reason -> OpenAI finish_reason ("length" 透传，其余归一 stop)
+    local finish = nil
+    if chunk.done then
+        if chunk.done_reason == "length" then
+            finish = "length"
+        else
+            finish = "stop"
+        end
+    end
+
     if not chunk.message then
         if uses ~= nil then
-            return json.encode({ content = "", done = true, usage = uses })
+            return json.encode({ content = "", done = true, finish_reason = finish, usage = uses })
         end
-        return ""
+        return json.encode({ content = "", done = true, finish_reason = finish })
     end
 
     local unified = {
         content = chunk.message.content or "",
-        done = chunk.done or false
+        done = chunk.done or false,
+        finish_reason = finish
     }
     if uses ~= nil then
         unified.usage = uses

internal/lua/adapters/openai.lua

@@ -90,10 +90,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/opencode.lua

@@ -159,10 +159,15 @@ function adapter.transform_stream_chunk(raw_chunk)
     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 = (fr ~= nil)
+        done = (finish ~= nil)
     }
+    if finish then
+        unified.finish_reason = finish
+    end
     if uses ~= nil then
         unified.usage = uses
     end

internal/lua/adapters/sensenova.lua

@@ -0,0 +1,108 @@
+-- sensenova API format adapter
+-- streaming: delta only has reasoning_content, no content field
+-- non-streaming: has both content and reasoning_content
+
+local adapter = {}
+adapter.name = "sensenova"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+-- Same as openai - 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
+    req.disable_thinking = nil
+    req.extra_body = nil
+    if req.messages then
+        for _, msg in ipairs(req.messages) do
+            msg.reasoning_content = nil
+        end
+    end
+    return json.encode(req)
+end
+
+-- Same as openai - extract content from response
+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
+    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
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+-- Sensenova-specific stream handling
+-- Upstream puts content in reasoning_content only (no content field)
+-- Also sends finish_reason="" (empty string) on every chunk
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok then return "" end
+
+    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,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+
+    local delta = chunk.choices[1].delta or {}
+
+    -- Sensenova: delta has reasoning_content but no content
+    local content = delta.content or ""
+    if content == "" and delta.reasoning_content then
+        content = delta.reasoning_content
+    end
+
+    -- Sensenova: finish_reason is "" on every chunk, "stop" on last
+    local fr = chunk.choices[1].finish_reason
+    local done = (fr == "stop" or fr == "length")
+
+    local unified = {
+        content = content,
+        done = done,
+    }
+    if chunk.choices[1].finish_reason and chunk.choices[1].finish_reason ~= "" then
+        unified.finish_reason = chunk.choices[1].finish_reason
+    end
+    if delta.tool_calls then
+        unified.tool_calls = delta.tool_calls
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+
+    return json.encode(unified)
+end
+
+return adapter

internal/lua/vm_test.go

@@ -422,3 +422,35 @@ func TestOpenCodeAdapterKeepsWhitelistedRolesAndDropsMultimodal(t *testing.T) {
 		t.Fatalf("unexpected roles: %s", out)
 	}
 }
+
+func TestAdaptersPassFinishReason(t *testing.T) {
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+	// OpenAI-style chunk with a real finish reason must surface finish_reason
+	// and done=true; an empty-string finish_reason (sensenova sends "" on
+	// every chunk) must NOT terminate the stream.
+	chunk := `{"choices":[{"index":0,"finish_reason":"tool_calls","delta":{"content":""}}]}`
+	empty := `{"choices":[{"index":0,"finish_reason":"","delta":{"content":"x"}}]}`
+	for _, name := range []string{"openai", "deepseek", "github", "groq", "kimicode", "mistral", "opencode"} {
+		out, err := vm.Transform(name, "transform_stream_chunk", chunk)
+		if err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+		if !strings.Contains(out, `"finish_reason":"tool_calls"`) {
+			t.Fatalf("%s: finish_reason lost: %s", name, out)
+		}
+		if !strings.Contains(out, `"done":true`) {
+			t.Fatalf("%s: done not set on real finish: %s", name, out)
+		}
+		out, err = vm.Transform(name, "transform_stream_chunk", empty)
+		if err != nil {
+			t.Fatalf("%s: %v", name, err)
+		}
+		if strings.Contains(out, `"done":true`) {
+			t.Fatalf("%s: empty finish_reason must not end stream: %s", name, out)
+		}
+	}
+}

internal/provider/provider.go

@@ -803,6 +803,7 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 		scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
 		var chunks int
 		var doneSeen bool
+		var doneSent bool
 		for scanner.Scan() {
 			line := strings.TrimSpace(scanner.Text())
 			if line == "" || !strings.HasPrefix(line, "data:") {
@@ -814,9 +815,15 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 			}
 			if data == "[DONE]" {
 				doneSeen = true
-				select {
+				// adapters that already emitted their terminating done chunk
-				case ch <- types.UnifiedChunk{Done: true}:
+				// (with the real finish reason) must not get a second,
-				case <-ctx.Done():
+				// reason-less done from [DONE] — it would override the true
+				// finish_reason downstream.
+				if !doneSent {
+					select {
+					case ch <- types.UnifiedChunk{Done: true}:
+					case <-ctx.Done():
+					}
 				}
 				continue
 			}
@@ -834,6 +841,9 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 			if err := json.Unmarshal([]byte(unified), &ck); err != nil {
 				continue
 			}
+			if ck.Done {
+				doneSent = true
+			}
 			chunks++
 			select {
 			case ch <- ck:
@@ -995,6 +1005,15 @@ func standardSSEChunk(data string) string {
 			Total:      pickFirst(pu.TotalTokens, pu.Total),
 		}
 	}
+	finish := ""
+	done := false
+	if len(raw.Choices) > 0 && raw.Choices[0].FinishReason != nil {
+		// empty-string finish reasons (sensenova sends "" on every chunk)
+		// are not a finish signal
+		if fr := *raw.Choices[0].FinishReason; fr != "" {
+			finish, done = fr, true
+		}
+	}
 	out, _ := json.Marshal(types.UnifiedChunk{
 		Content: func() string {
 			if len(raw.Choices) > 0 {
@@ -1002,8 +1021,9 @@ func standardSSEChunk(data string) string {
 			}
 			return ""
 		}(),
-		Done:  len(raw.Choices) > 0 && raw.Choices[0].FinishReason != nil,
+		Done:         done,
-		Usage: usage,
+		FinishReason: finish,
+		Usage:        usage,
 	})
 	return string(out)
 }

internal/provider/provider_test.go

@@ -8,6 +8,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"path/filepath"
+	"strings"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -472,3 +473,14 @@ func TestImageAutoUsesImageModel(t *testing.T) {
 		t.Fatalf("AUTO image must use the image-kind model, got %q", gotModel)
 	}
 }
+
+func TestStandardSSEChunkFinishReason(t *testing.T) {
+	out := standardSSEChunk(`{"choices":[{"index":0,"finish_reason":"tool_calls","delta":{}}]}`)
+	if !strings.Contains(out, `"done":true`) || !strings.Contains(out, `"finish_reason":"tool_calls"`) {
+		t.Fatalf("real finish reason must pass through: %s", out)
+	}
+	out = standardSSEChunk(`{"choices":[{"index":0,"finish_reason":"","delta":{"content":"hi"}}]}`)
+	if strings.Contains(out, `"Done":true`) {
+		t.Fatalf("empty-string finish reason is not a finish signal: %s", out)
+	}
+}

internal/types/types.go

@@ -136,6 +136,10 @@ type ImageGenResponse struct {
 type UnifiedChunk struct {
 	Content          string          `json:"content"`
 	Done             bool            `json:"done"`
+	// FinishReason carries the upstream finish/stop reason ("tool_calls",
+	// "length", ...) when the adapter provides it; the gateway emits it on
+	// the terminating chunk instead of the default "stop".
+	FinishReason     string          `json:"finish_reason,omitempty"`
 	ToolCalls        json.RawMessage `json:"tool_calls,omitempty"`
 	ReasoningContent string          `json:"reasoning_content,omitempty"`
 	// Usage carries the upstream token usage when the stream chunk provides