feat(qoder): sign chat in-process instead of through workers

AGENTS.md

@@ -45,14 +45,14 @@ live-account acceptance, and release status separate.
 - After console UI changes, run `cd frontend && npm run sync` so `internal/webui/static` matches `frontend/dist`. Do not commit a stale hashed JS/CSS pair.
 - User-facing changes add one bilingual file in `changelog/unreleased/` (`### English` / `### 中文`). Do not put upcoming notes in `CHANGELOG.md` and do not add `## Unreleased` there. Do not freeze or unfreeze changelog sections by hand. Do not reuse a fragment filename until that tag's archive PR has merged.
 - When `main` is checked out in another worktree, merge PRs with `gh api` / GitHub; do not `git checkout main` here.
-- Keep architecture: auth / endpoint / executor / translate, plus `internal/store` (SQLite), `internal/control` (console facade: accounts, keys, settings, catalog, login, import), `internal/runtime` (Manager lifecycle), `internal/providers/qoder` (Qoder HOME/CLI/worker protocol and Adapter), `internal/gateway` (public protocol HTTP), `internal/console` (operator HTTP), `internal/server` (routes/middleware/webui), and `internal/app` (process assembly). `internal/api` is a test-only compatibility facade (`api.New` → `app.New`); do not add business there. Account entities stay in `accounts`; Pool/Item/RouteQuery/Classify and request Prepare live in `executor`. Display catalog cache lives in `control.Catalog`; catalog aggregation, identity filter, and settings decoration live in `control`, not `app`. Public `/v1/chat/completions`, `/v1/messages`, `/v1/responses`, and `/v1/models` live in `gateway`; console `/api/*` lives in `console`; `internal/server` registers both. Console HTTP decodes and maps errors; persist/apply for system settings and console-key rotation live in `control.System` / `control.KeyRotation`. Update job/maintenance lives in `internal/update.Coordinator`. SQLite lives in `internal/store`; process tables live in runtime. Runtime constructs the one Pool and injects it into executor. Qoder stays `child_process`; the registered Adapter omits Prober. Runtime catalog may use `adapter.Models`; quota/login/chat still use worker HTTP. `cmd/server` constructs `app.New`. `accounts` must not import `runtime` or `executor`. Executor Prepare must not take `*http.Request` or import store. Gateway, console, and server must not import store or runtime Manager. App/server/gateway/console must not import `internal/api`. Provider packages must not receive `http.ResponseWriter` or import executor taxonomy; OAuth loopback HTTP stays in `auth.ServeLoopback`. Adapter error classification and cooldown math live in `executor`; gateway only formats the result.
+- Keep architecture: auth / endpoint / executor / translate, plus `internal/store` (SQLite), `internal/control` (console facade: accounts, keys, settings, catalog, login, import), `internal/runtime` (Manager lifecycle), `internal/providers/qoder` (Qoder HOME/CLI/worker protocol and Adapter), `internal/gateway` (public protocol HTTP), `internal/console` (operator HTTP), `internal/server` (routes/middleware/webui), and `internal/app` (process assembly). `internal/api` is a test-only compatibility facade (`api.New` → `app.New`); do not add business there. Account entities stay in `accounts`; Pool/Item/RouteQuery/Classify and request Prepare live in `executor`. Display catalog cache lives in `control.Catalog`; catalog aggregation, identity filter, and settings decoration live in `control`, not `app`. Public `/v1/chat/completions`, `/v1/messages`, `/v1/responses`, and `/v1/models` live in `gateway`; console `/api/*` lives in `console`; `internal/server` registers both. Console HTTP decodes and maps errors; persist/apply for system settings and console-key rotation live in `control.System` / `control.KeyRotation`. Update job/maintenance lives in `internal/update.Coordinator`. SQLite lives in `internal/store`; process tables live in runtime. Runtime constructs the one Pool and injects it into executor. Qoder login, catalog, and quota stay on a per-account child process because the CLI credential exchange still lives there; chat does not. The registered Adapter omits Prober and sends chat through the in-process COSY client. `cmd/server` constructs `app.New`. `accounts` must not import `runtime` or `executor`. Executor Prepare must not take `*http.Request` or import store. Gateway, console, and server must not import store or runtime Manager. App/server/gateway/console must not import `internal/api`. Provider packages must not receive `http.ResponseWriter` or import executor taxonomy; OAuth loopback HTTP stays in `auth.ServeLoopback`. Adapter error classification and cooldown math live in `executor`; gateway only formats the result.
 - Prefer direct HTTP/SSE to Qoder cloud APIs
 - Pin qodercli / qoderclicn hooks in `worker/src/compat.mjs`; fail loudly on mismatch. Qoder CN is `provider=qoder` + `region=cn`, not a new family
 - Reasoning levels are catalog-driven: map client values through `internal/providers/reasoning.go` (`none`/`low`/`medium`/`high`/`xhigh`/`max`), clamp anything the model does not allow back to an allowed level, and treat the console value as a default only (it never locks a call or caps a higher client value)
 - Console UI: React + Tailwind v4 + **HeroUI only** for components
 - Follow `docs/DESIGN.md` (taste v1 adapted for this console)
 - Keep iterating Qoder login, usage, and account routing. Keep scheduling focused on personal-account routing rather than commercial gateway features
-- Qoder multi-account = one worker process per HOME; do not share WASM context. WorkBuddy / Trae / Devin use in-process adapters, not one child process per account
+- Qoder login still uses one worker process per HOME; do not share that WASM context. Qoder chat is an in-process COSY client and must not import the CLI bundle per request or per account. WorkBuddy / Trae / Devin use in-process adapters, not one child process per account
 - Schema changes go in a new numbered SQLite migration entry in `internal/store/migrations.go`. Never rewrite shipped SQL
 
 ## Don't

changelog/unreleased/qoder-checkin-machine-headers.md

@@ -0,0 +1,7 @@
+### English
+
+- Qoder CN check-in no longer sends a partial machine identity. Without the official runtime-info identity, those headers hid the claimable credit campaign.
+
+### 中文
+
+- Qoder 国内版签到在没有官方 runtime-info 机器身份时不再发送半套机器头。之前这些头会让可领取的积分活动从列表里消失。

changelog/unreleased/qoder-inprocess-chat.md

@@ -0,0 +1,7 @@
+### English
+
+- Qoder chat now signs requests in the Go process and calls the gateway directly. Per-account Node workers stay for login, model catalog, and quota, and are no longer on the chat path.
+
+### 中文
+
+- Qoder 聊天改为在 Go 进程内签名并直连网关。每个账号的 Node worker 仍负责登录、模型目录和配额，不再参与聊天请求。

docs/ARCHITECTURE_SUMMARY.md

@@ -20,7 +20,7 @@ CLI2API 是一个面向个人部署的 Go + Node 网关：
 
 - 对外提供 OpenAI Chat Completions、Anthropic Messages、OpenAI Responses 和 Models 兼容接口。
 - 管理多个上游账号，并按 provider、region、model、pin、会话粘性和冷却状态选号。
-- Qoder 使用每账号隔离的 Node child process；WorkBuddy、Trae CN Work 和 Devin 使用 Go 进程内 adapter。
+- Qoder 登录、目录和配额仍使用每账号隔离的 Node child process；Qoder 聊天在 Go 进程内签名直连网关。WorkBuddy、Trae CN Work 和 Devin 使用 Go 进程内 adapter。
 - 控制台负责账号、密钥、设置、目录、日志、登录、导入和更新操作。
 - 这是个人网关，不实现计费、Redis 槽位、多租户商业网关或公开暴露的管理端口。
 
@@ -71,8 +71,9 @@ cmd/server
 ## Provider 边界
 
 - Qoder Global 和 Qoder CN 是同一个 `provider=qoder`，通过 `region` 区分，不创建新的 provider family。
-- Qoder 每个账号使用独立 HOME 和独立 child process；不得为每个请求启动完整 CLI agent。
+- Qoder 登录仍使用独立 HOME 和独立 child process，不共享 WASM context；聊天不加载 CLI bundle。不得为每个请求启动完整 CLI agent。
 - Qoder 的 CLI / worker 兼容性版本固定在 `worker/src/compat.mjs`，不兼容时应明确失败。
+- Qoder 聊天改为在 Go 进程内使用 COSY 客户端签名直连网关（[`internal/providers/qoder/cosy.go`](../internal/providers/qoder/cosy.go)）。COSY 协议版本固定在 [`COSYVersion = "1.1.32"`](cosy.go:23)，支持中国区（`gateway.qoder.com.cn`）和全球区（`api1.qoder.sh`）双端点。
 - WorkBuddy、Trae CN Work、Devin 使用进程内 adapter，不复制 Qoder worker 生命周期。
 - provider 负责上游事实映射；executor 负责是否切号、冷却多久和是否 failover。
 - Provider 能力、模型目录和 reasoning level 必须以实际 catalog 声明为准，不凭空增加模型能力。

internal/app/app.go

@@ -120,6 +120,13 @@ func New(cfg config.Config) *App {
 	providerReg.Register(zhipu.NewClient(store).Adapter())
 	qoderClient := qoder.NewClient()
 	qoderClient.Bind(manager.AccountURL, manager.ProxyAPIKey)
+	qoderClient.SetDirect(qoder.NewDirect(store, func(ctx context.Context, accountID string) (string, error) {
+		account, err := store.Get(ctx, accountID)
+		if err != nil {
+			return "", err
+		}
+		return account.ProviderRegion, nil
+	}))
 	providerReg.Register(qoderClient.Adapter())
 	manager.SetProviders(providerReg)
 	manager.SetWorkBuddy(workbuddyClient)

internal/app/architecture_test.go

@@ -29,9 +29,6 @@ var importAllowlist = map[string]map[string]string{
 	modulePath + "/internal/api": {
 		modulePath + "/internal/app": "S15: api is the remaining test-only compatibility facade (api.New → app.New).",
 	},
-	modulePath + "/internal/executor": {
-		modulePath + "/internal/providers/qoder": "S09 leftover: Qoder chat still uses worker HTTP via qoder.NewChatRequest until production chat switches to Adapter.",
-	},
 	modulePath + "/internal/runtime": {
 		modulePath + "/internal/providers/qoder": "S08/S09 leftover: Qoder child spawn, HOME, catalog, and quota still call providers/qoder until remaining capabilities go through Adapter.",
 	},

internal/app/regression_test.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
-	"io"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -16,6 +15,7 @@ import (
 	"github.com/caigee-cmd/cli2api/internal/app"
 	"github.com/caigee-cmd/cli2api/internal/config"
 	"github.com/caigee-cmd/cli2api/internal/executor"
+	"github.com/caigee-cmd/cli2api/internal/providers/qoder"
 	"github.com/caigee-cmd/cli2api/internal/update"
 )
 
@@ -79,36 +79,39 @@ func TestConsoleKeyRotationUpdatesExistingHandlerAndWorkerRequests(t *testing.T)
 		}
 		currentKey = result.Secret
 	}
-	// A fake ready worker uses the runtime's current key, without spawning a CLI.
+	// Chat no longer enters the login worker. The gateway sees a COSY bearer,
+	// never the console key that rotation just changed.
 	var calls atomic.Int32
 	worker := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/chat/completions" {
-			http.NotFound(w, r)
-			return
-		}
 		calls.Add(1)
-		if r.Header.Get("Authorization") != "Bearer "+currentKey {
-			t.Error("worker received stale key")
+		if strings.Contains(r.Header.Get("Authorization"), currentKey) {
+			t.Error("upstream received the console key")
 			w.WriteHeader(401)
 			return
 		}
-		var req struct {
-			Stream bool `json:"stream"`
-		}
-		_ = json.NewDecoder(r.Body).Decode(&req)
-		if req.Stream {
-			w.Header().Set("Content-Type", "text/event-stream")
-			fmt.Fprint(w, "data: {\"id\":\"test\",\"model\":\"glm-5.2\",\"choices\":[{\"delta\":{\"content\":\"OK\"}}]}\n\ndata: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}]}\n\ndata: [DONE]\n\n")
+		if !strings.HasPrefix(r.Header.Get("Authorization"), "Bearer COSY.") {
+			t.Errorf("authorization = %q", r.Header.Get("Authorization"))
+			w.WriteHeader(401)
 			return
 		}
-		io.WriteString(w, `{"model":"glm-5.2","choices":[{"message":{"content":"OK"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`)
+		w.Header().Set("Content-Type", "text/event-stream")
+		fmt.Fprint(w, "data: {\"body\":{\"choices\":[{\"delta\":{\"content\":\"OK\"},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":1,\"completion_tokens\":1}}}\n\n")
 	}))
 	defer worker.Close()
 	account, err := a.Manager.Store().Create(context.Background(), accounts.CreateAccount{Name: "fake-worker", Enabled: false})
 	if err != nil {
 		t.Fatal(err)
 	}
-	a.Pool.Upsert(executor.Item{ID: account.ID, Provider: "qoder", Runtime: "child_process", URL: worker.URL, Models: []string{"glm-5.2"}})
+	if err := a.Manager.Store().SaveCredential(context.Background(), account.ID, "oauth", accounts.NativeCredential{
+		UserBlob:  []byte(`{"uid":"u-rotation","access_token":"dt-rotation"}`),
+		MachineID: "0123456789abcdef",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	original := qoder.ChatEndpointHook
+	qoder.ChatEndpointHook = func(string) string { return worker.URL + "/algo/api/v2/service/pro/sse/agent_chat_generation?Encode=1" }
+	t.Cleanup(func() { qoder.ChatEndpointHook = original })
+	a.Pool.Upsert(executor.Item{ID: account.ID, Provider: "qoder", Runtime: "child_process", URL: "http://127.0.0.1:1", Models: []string{"glm-5.2"}})
 	// Catalog probing is tested separately; keep the fake worker deterministic.
 	a.Gateway.Catalogs = nil
 	for _, path := range []string{"/v1/chat/completions", "/api/chat", "/v1/messages", "/v1/responses"} {

internal/executor/chat.go

@@ -1,6 +1,7 @@
 package executor
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -14,7 +15,6 @@ import (
 
 	"github.com/caigee-cmd/cli2api/internal/accounts"
 	"github.com/caigee-cmd/cli2api/internal/providers"
-	"github.com/caigee-cmd/cli2api/internal/providers/qoder"
 	"github.com/caigee-cmd/cli2api/internal/translate"
 )
 
@@ -486,6 +486,17 @@ func keyGrantedSingleRegion(providerFilter string, allowed []string) (string, bo
 	return regions[0], true
 }
 
+// qoderDirect reports whether this Qoder account chats through the registered
+// adapter instead of its login worker. Tests that stub a worker URL and do not
+// register the adapter stay on the worker path.
+func (e ChatExecutor) qoderDirect(item Item) bool {
+	if itemProvider(item) != "qoder" || e.Providers == nil {
+		return false
+	}
+	adapter, ok := e.Providers.Get("qoder")
+	return ok && adapter.Chat != nil
+}
+
 func isInProcessItem(item Item) bool {
 	if item.Runtime == string(providers.RuntimeInProcess) {
 		return true
@@ -604,6 +615,34 @@ func (e ChatExecutor) recordAttempt(ctx context.Context, attempt accounts.Reques
 	e.OnAttempt(attempt)
 }
 
+// buildWorkerChatPayload is the OpenAI-shaped body a non-Qoder worker stub
+// still accepts. Qoder production chat does not use it.
+func buildWorkerChatPayload(req translate.ChatRequest, stream bool) map[string]any {
+	payload := map[string]any{
+		"model":    req.Model,
+		"messages": req.Messages,
+		"stream":   stream,
+	}
+	if len(req.MaxCompletionTokens) > 0 {
+		payload["max_tokens"] = json.RawMessage(req.MaxCompletionTokens)
+	} else if len(req.MaxTokens) > 0 {
+		payload["max_tokens"] = json.RawMessage(req.MaxTokens)
+	}
+	if len(req.Temperature) > 0 {
+		payload["temperature"] = json.RawMessage(req.Temperature)
+	}
+	if len(req.Tools) > 0 {
+		payload["tools"] = json.RawMessage(req.Tools)
+	}
+	if len(req.ToolChoice) > 0 {
+		payload["tool_choice"] = json.RawMessage(req.ToolChoice)
+	}
+	if len(req.ReasoningEffort) > 0 {
+		payload["reasoning_effort"] = json.RawMessage(req.ReasoningEffort)
+	}
+	return payload
+}
+
 func (e ChatExecutor) newWorkerRequest(ctx context.Context, item Item, payload []byte, prefer string) (*http.Request, error) {
 	account := prefer
 	if account == "" {
@@ -613,7 +652,21 @@ func (e ChatExecutor) newWorkerRequest(ctx context.Context, item Item, payload [
 	if e.WorkerKeySource != nil {
 		key = e.WorkerKeySource()
 	}
-	return qoder.NewChatRequest(ctx, item.URL, account, RequestIDFromContext(ctx), key, payload)
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, strings.TrimRight(item.URL, "/")+"/v1/chat/completions", bytes.NewReader(payload))
+	if err != nil {
+		return nil, err
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	if key != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+key)
+	}
+	if account != "" {
+		httpReq.Header.Set("X-Qoder-Account", account)
+	}
+	if requestID := RequestIDFromContext(ctx); requestID != "" {
+		httpReq.Header.Set("X-Request-Id", requestID)
+	}
+	return httpReq, nil
 }
 
 func classifyWorkerErr(resp *http.Response, body string) Classified {
@@ -725,10 +778,6 @@ func (l routeLoop) pickFailure(err error) (int, string, string, error) {
 func (e ChatExecutor) ChatNonStream(ctx context.Context, req translate.ChatRequest, prefer, providerFilter string) (result ChatResult, returnErr error) {
 	loop := e.newRouteLoop(ctx, prefer, providerFilter, req)
 	defer func() { result.Routing = loop.routing.Source }()
-	payload, err := json.Marshal(qoder.BuildChatPayload(req, false))
-	if err != nil {
-		return ChatResult{}, err
-	}
 	for loop.index < loop.attempts {
 		item, i, err := loop.pickNext(e, req.Model)
 		if err != nil {
@@ -738,7 +787,7 @@ func (e ChatExecutor) ChatNonStream(ctx context.Context, req translate.ChatReque
 			}
 			return ChatResult{}, pickErr
 		}
-		if isInProcessItem(item) {
+		if isInProcessItem(item) || e.qoderDirect(item) {
 			result, classified, err := e.chatInProcessNonStreamAttempt(ctx, item, req, i)
 			if err == nil {
 				result.AttemptCount = i + 1
@@ -756,6 +805,10 @@ func (e ChatExecutor) ChatNonStream(ctx context.Context, req translate.ChatReque
 			}
 			return ChatResult{AttemptCount: i + 1, AccountID: item.ID, Provider: item.Provider}, err
 		}
+		payload, err := json.Marshal(buildWorkerChatPayload(req, false))
+		if err != nil {
+			return ChatResult{}, err
+		}
 		headerAccount := loop.headerAccount(item, i)
 		httpReq, err := e.newWorkerRequest(ctx, item, payload, headerAccount)
 		if err != nil {
@@ -1140,14 +1193,6 @@ func (e ChatExecutor) chatStreamProxy(ctx context.Context, req translate.ChatReq
 		loop.attempts = e.attemptsFor(loop.providerFilter, loop.regionFilter, req.Model, loop.allowed, loop.eligible)
 	}
 	defer func() { result.Routing = loop.routing.Source }()
-	var payload []byte
-	if !preferNativeResponses || native == nil {
-		var err error
-		payload, err = json.Marshal(qoder.BuildChatPayload(req, true))
-		if err != nil {
-			return StreamResult{}, err
-		}
-	}
 	startedAll := time.Now()
 	for loop.index < loop.attempts {
 		item, i, err := loop.pickNext(e, req.Model)
@@ -1158,7 +1203,7 @@ func (e ChatExecutor) chatStreamProxy(ctx context.Context, req translate.ChatReq
 			}
 			return StreamResult{}, pickErr
 		}
-		if isInProcessItem(item) {
+		if isInProcessItem(item) || (e.qoderDirect(item) && !(preferNativeResponses && native != nil)) {
 			result, classified, err := e.chatInProcessStreamAttempt(ctx, item, req, native, i, preferNativeResponses)
 
 			if err == nil {
@@ -1176,6 +1221,10 @@ func (e ChatExecutor) chatStreamProxy(ctx context.Context, req translate.ChatReq
 			}
 			return StreamResult{AttemptCount: i + 1, AccountID: item.ID, Provider: item.Provider}, err
 		}
+		payload, err := json.Marshal(buildWorkerChatPayload(req, true))
+		if err != nil {
+			return StreamResult{}, err
+		}
 		headerAccount := loop.headerAccount(item, i)
 		httpReq, err := e.newWorkerRequest(ctx, item, payload, headerAccount)
 		if err != nil {