Add continuity and inference execution tests, implement state management

- Introduced `continuity_e2e_test.go` to validate end-to-end continuity scenarios, ensuring proper state handling across tenant boundaries.
- Added `continuity_test.go` for unit tests on continuity state persistence, corruption handling, and opaque state token management.
- Implemented `inference_execution.go` to manage inference execution, including state binding and continuity logic.
- Created `tenant.go` to manage tenant context within requests, ensuring proper isolation and handling of tenant namespaces.
- Enhanced state token management in responses, ensuring secure handling of opaque tokens and previous response IDs.

Author: Futureppo

.env.example

@@ -11,16 +11,17 @@ GROK_RETRY_BASE_DELAY=200ms
 GROK_RATE_LIMIT_COOLDOWN=1m
 # 上游账号或免费模型额度耗尽后的默认冷却时间
 GROK_QUOTA_COOLDOWN=24h
-# 会话与上游账号亲和关系在内存中的有效时间
+# 会话与上游账号亲和关系的有效时间；hard binding 会哈希化持久保存，soft binding 仅在内存中
 GROK_AFFINITY_TTL=1h
-# 内存中最多保存的会话亲和关系数量
+# 最多保存的会话亲和关系数量
 GROK_AFFINITY_MAX_ENTRIES=100000
 
 
 # 可选。留空（默认）表示不启用访问控制。设置后，受保护的 API 和上游透传接口
 # 必须携带 `Authorization: Bearer <key>`、`x-api-key: <key>` 或
 # `api-key: <key>`。OpenAI SDK 客户端会自动将 `api_key` 参数作为 Bearer 令牌发送
-# 支持配置多个 Key，可为每个客户端分配独立 Key，无需重启服务
+# 支持配置多个 Key，可为每个客户端分配独立 Key。
+# 每个 Key 还会派生独立的 tenant namespace；不配置 Key 时所有调用方共享 public tenant。
 
 GROK_API_KEYS=key1,key2,key3
 
@@ -30,16 +31,17 @@ GROK_API_KEYS=key1,key2,key3
 GROK_ADMIN_KEY=
 
 
-# 每个账号对应一个 OAuth JSON 凭证文件，并直接放入此目录，支持直接导入CPA格式的凭证
-# 文件支持热加载，刷新后的令牌会写回。
+# 直接放入 Grok CLI auth.json 或兼容凭证，支持旧单凭证、tokens 包装和多 scope 文件。
+# 文件支持热加载；刷新时只在文件锁内原子更新目标 scope。state v2 与哈希化 hard affinity
+# 也保存在此目录，因此目录必须可写。
 GROK_AUTHS_DIR=./auths
 # 扫描凭证目录并热加载新增、更新或删除文件的时间间隔
 GROK_AUTHS_RELOAD_INTERVAL=30s
 # 同时执行 OAuth 令牌刷新的最大并发数
 GROK_AUTH_REFRESH_CONCURRENCY=4
 # 每个账号允许同时发往上游的最大请求数。超出的请求会等待其他账号或已释放的容量，
-GROK_ACCOUNT_MAX_INFLIGHT=8
-# 按此时间间隔刷新每个账号持久化保存的上游模型目录。
+GROK_ACCOUNT_MAX_INFLIGHT=16
+# 按此时间间隔刷新 state v2 中每个账号的结构化上游模型目录与 ETag。
 GROK_MODELS_REFRESH_INTERVAL=6h
 
 
@@ -69,14 +71,20 @@ GROK_MODELS_REFRESH_INTERVAL=6h
 # Grok CLI 连接的地址
 GROK_CHAT_PROXY_BASE_URL=https://cli-chat-proxy.grok.com
 GROK_CHAT_PROXY_VERSION=v1
+# API Key 账号固定使用运维配置的 xAI API origin，远端模型字段不能覆盖
+GROK_XAI_API_BASE_URL=https://api.x.ai
 # 流式传输默认使用 identity，避免 gzip 层缓冲较小的 SSE 事件
 # 仅在需要兼容性回退时设置为 gzip
 GROK_STREAM_COMPRESSION=identity
 # 如果上游返回 HTTP 426，并要求更高的最低版本，请更新此值
-GROK_CLIENT_VERSION=0.2.93
+GROK_CLIENT_VERSION=0.2.102
 GROK_CLIENT_SURFACE=tui
+# 默认使用非交互的 headless 客户端模式；可选 interactive
+GROK_CLIENT_MODE=headless
 GROK_CLIENT_IDENTIFIER=grok-shell
 GROK_TOKEN_AUTH=xai-grok-cli
+# 可选的托管部署标识
+# GROK_DEPLOYMENT_ID=
 # x-grok-* 请求头配置
 GROK_CLIENT_NAME=grok-shell
 # 仅可在受控调试环境中关闭 TLS 证书验证

.github/workflows/ci.yml

@@ -23,16 +23,16 @@ jobs:
           cache: false
       - name: Check formatting
         shell: bash
-        run: test -z "$(gofmt -l .)"
+        run: test -z "$(gofmt -l $(git ls-files '*.go'))"
       - name: Test
-        run: go test ./...
+        run: go test -count=1 ./...
       - name: Race detector
         if: runner.os == 'Linux'
-        run: go test -race ./...
+        run: go test -race -count=1 ./...
       - name: Vet
         run: go vet ./...
       - name: Build
-        run: go build ./cmd/grok2api
+        run: go build -trimpath ./cmd/grok2api
 
   docker:
     needs: test

CONTRIBUTING.md

@@ -7,13 +7,28 @@ Thanks for contributing to `grokcli2api-go`.
 Requirements: Go 1.23 or newer.
 
 ```bash
-go test ./...
+test -z "$(gofmt -l $(git ls-files '*.go'))"
+go test -count=1 ./...
+go test -race -count=1 ./... # Linux
 go vet ./...
-go build ./cmd/grok2api
+go build -trimpath ./cmd/grok2api
+docker build .
 ```
 
 Run `gofmt` on changed Go files. Keep the production implementation on the Go standard library unless a dependency is clearly justified.
 
+## Live smoke
+
+The real inference smoke is disabled by default. Run it only after every offline gate above has passed, including Linux race testing and the Docker build:
+
+```bash
+GROK_LIVE_SMOKE=1 \
+GROK_LIVE_SMOKE_OFFLINE_GATES=passed \
+go test -count=1 -run '^TestLiveInferenceSmoke$' ./internal/grok
+```
+
+It reads `auths/live-01.json` by default; set `GROK_LIVE_SMOKE_AUTH_FILE` to use another file. The source must resolve to exactly one logical credential so initialization performs one account catalog discovery. The smoke gives the service only a temporary `0600` copy with refresh tokens, ID tokens, email fields, and stale legacy model lists removed. It refuses access tokens with less than ten minutes remaining and fails if the source credential hash or Git worktree changes. Response bodies, tokens, and account identifiers are never printed.
+
 ## Pull requests
 
 - Keep changes focused and describe protocol compatibility implications.

Dockerfile

@@ -3,7 +3,7 @@ WORKDIR /src
 COPY go.mod ./
 COPY cmd ./cmd
 COPY internal ./internal
-RUN CGO_ENABLED=0 go test ./... && \
+RUN CGO_ENABLED=0 go test -count=1 ./... && \
     CGO_ENABLED=0 go build -trimpath -ldflags="-s -w" -o /out/grok2api ./cmd/grok2api
 
 FROM alpine:3.22