Seed CLAUDE_CODE_AUTO_MODE_SERVER=0 with claudeSetup

config.default.yaml

@@ -21,7 +21,8 @@ copilotBaseUrl: https://api.githubcopilot.com
 
 # When true, start updates ~/.claude/settings.json with ANTHROPIC_BASE_URL
 # and a dummy ANTHROPIC_AUTH_TOKEN so Claude Code can use this proxy.
-# Also seeds context/output budgets from model discovery when not already set.
+# Also seeds context/output budgets from model discovery, and
+# CLAUDE_CODE_AUTO_MODE_SERVER=0, when not already set.
 claudeSetup: true
 
 # Log verbosity:

src/lib/claude-settings.ts

@@ -134,6 +134,14 @@ export async function applyClaudeConfig(
     env.CLAUDE_CODE_MAX_OUTPUT_TOKENS = String(input.maxOutputTokens)
   }
 
+  // In auto mode Claude Code asks the server to check actions. In #162 those checks did not reach
+  // it through the relay, so it made its own classifier requests and printed a notice asking the
+  // gateway to support them. "0" skips the server check and the notice; the classifier requests
+  // stay as they were. Seeded only when absent, so a value the user set is kept.
+  if (env.CLAUDE_CODE_AUTO_MODE_SERVER === undefined) {
+    env.CLAUDE_CODE_AUTO_MODE_SERVER = "0"
+  }
+
   const next: Record<string, unknown> = {
     ...parsed,
     ...(nextModel !== undefined && { model: nextModel }),

tests/unit/claude-settings.test.ts

@@ -83,6 +83,38 @@ test("preserves explicit client context and output overrides", async () => {
   })
 })
 
+// #162: through the relay, auto mode's server checks did not reach Claude Code, and it showed a
+// notice. The writer seeds CLAUDE_CODE_AUTO_MODE_SERVER=0, but only when the key is absent.
+test("seeds CLAUDE_CODE_AUTO_MODE_SERVER=0 only when it is absent", async () => {
+  const input = (configPath: string) => ({
+    baseUrl: "http://127.0.0.1:4142",
+    configPath,
+    gptModel: "gpt-6-astra",
+    gptLimits: astraLimits,
+    maxOutputTokens: 128_000,
+  })
+
+  await withTemporarySettings(async (configPath) => {
+    await applyClaudeConfig(input(configPath))
+    const env = (await readSettings(configPath)).env as Record<string, unknown>
+
+    assert.equal(env.CLAUDE_CODE_AUTO_MODE_SERVER, "0")
+    assert.equal((await applyClaudeConfig(input(configPath))).changed, false)
+  })
+
+  for (const value of ["0", "1"]) {
+    await withTemporarySettings(async (configPath) => {
+      await fs.mkdir(path.dirname(configPath), { recursive: true })
+      await fs.writeFile(configPath, JSON.stringify({ env: { CLAUDE_CODE_AUTO_MODE_SERVER: value } }))
+
+      await applyClaudeConfig(input(configPath))
+      const env = (await readSettings(configPath)).env as Record<string, unknown>
+
+      assert.equal(env.CLAUDE_CODE_AUTO_MODE_SERVER, value)
+    })
+  }
+})
+
 // Why: a fresh managed Claude Code setup must actually select the configured GPT
 // default and expose its 1M client-side context identity without changing the
 // canonical model that Copilot receives.

wiki/EN-Configuration.md

@@ -665,6 +665,7 @@ ANTHROPIC_BASE_URL=http://127.0.0.1:4142
 ANTHROPIC_AUTH_TOKEN=<dummy local token>
 CLAUDE_CODE_MAX_CONTEXT_TOKENS=<discovered GPT context window>
 CLAUDE_CODE_MAX_OUTPUT_TOKENS=<largest discovered output budget of the model pair>
+CLAUDE_CODE_AUTO_MODE_SERVER=0
 ```
 
 into:
@@ -684,6 +685,17 @@ With `claudeSetup: false`, configure these client variables yourself.
 The settings writer uses the same snapshot/atomic replacement boundary; malformed
 or existing empty settings files are not overwritten, and unrelated values remain.
 
+`CLAUDE_CODE_AUTO_MODE_SERVER=0` is also written only when the key is absent. In
+auto mode, Claude Code asks the server to check actions such as shell commands.
+Through the relay those checks did not reach Claude Code on the `chat-completions`
+or `messages` route, so it made its own classifier requests and printed a notice
+that the session isn't eligible for auto mode's no-charge classifier requests.
+With `0`, Claude Code skips the server check and the notice; the classifier
+requests stay as they were. The value stays in `~/.claude/settings.json`: remove
+or change it when Claude Code connects without the relay, or through a gateway
+that supports the server checks. Claude Code documents the variable as temporary;
+see [auto mode classifier billing](https://code.claude.com/docs/en/auto-mode-classifier-billing).
+
 The local dummy token is **not network authentication**. Host/Origin checks and
 JSON content-type validation reduce browser-origin misuse, not access by an
 arbitrary network client. Keep the listener on loopback; see

wiki/EN-Logging-Troubleshooting.md

@@ -924,10 +924,26 @@ Expected values:
 
 - `ANTHROPIC_BASE_URL` points at `http://127.0.0.1:4142`
 - `ANTHROPIC_AUTH_TOKEN` exists; it is a dummy value for local relay use
+- `CLAUDE_CODE_AUTO_MODE_SERVER` is `0`, unless you set another value
 
 Changing `host` or `port` requires restarting the relay, because the listening
 socket cannot move during hot reload.
 
+## Claude Code says the session isn't eligible for auto mode
+
+Claude Code can print a notice like this, in the terminal or, with `claude -p`,
+on stderr:
+
+```text
+We're changing auto mode to no longer charge for classifier requests in Claude Code. However, this session isn't eligible because your requests go through 127.0.0.1:4142, which isn't compatible with this update. Nothing breaks: auto mode keeps working, and its classifier requests are billed as before.
+```
+
+It is expected through the relay: auto mode's server checks do not reach Claude
+Code, so it makes its own classifier requests. `CLAUDE_CODE_AUTO_MODE_SERVER=0` in
+Claude Code's environment stops the server check and the notice. With
+`claudeSetup: true`, `copilot-relay start` writes it when it is absent; see
+[Configuration](EN-Configuration.md).
+
 ## Safe log sharing
 
 **Never share raw captures wholesale.** Prompts, tool arguments/results, signed

wiki/ZH-Configuration.md

@@ -572,6 +572,7 @@ ANTHROPIC_BASE_URL=http://127.0.0.1:4142
 ANTHROPIC_AUTH_TOKEN=<dummy local token>
 CLAUDE_CODE_MAX_CONTEXT_TOKENS=<发现的 GPT context 窗口>
 CLAUDE_CODE_MAX_OUTPUT_TOKENS=<两个配置模型中最大的已公布输出预算>
+CLAUDE_CODE_AUTO_MODE_SERVER=0
 ```
 
 这里的 token 是本地 relay 占位值。真正访问 GitHub Copilot 用的是
@@ -584,6 +585,15 @@ CLAUDE_CODE_MAX_OUTPUT_TOKENS=<两个配置模型中最大的已公布输出预
 设置写入同样使用快照及原子替换；不会覆盖格式错误或已经存在的空设置文件，其他无关
 值也会保留。
 
+`CLAUDE_CODE_AUTO_MODE_SERVER=0` 同样只在缺少该键时写入。auto mode 下，Claude Code
+会请服务器检查 shell 命令等操作。经过 relay 时，这些检查在 `chat-completions` 和
+`messages` 路由上都没有到达 Claude Code，于是它自己发起 classifier 请求，并提示该
+会话不能享受 auto mode 不再对 classifier 请求收费的变更。设为 `0` 后，Claude Code
+不再请求服务器检查，也不再显示该提示；classifier 请求与原来相同。这个值会留在
+`~/.claude/settings.json` 里：Claude Code 不经过 relay 连接，或改用支持服务器检查的
+网关时，请删除或修改它。Claude Code 文档说明该变量是临时的，见
+[auto mode classifier billing](https://code.claude.com/docs/en/auto-mode-classifier-billing)。
+
 本地占位 token **不是网络认证**。Host/Origin 与 JSON content-type 检查减少的是浏览器
 来源滥用，不能阻止任意网络客户端。请保持 loopback 监听，见[架构](ZH-Architecture.md)。
 

wiki/ZH-Logging-Troubleshooting.md

@@ -816,9 +816,23 @@ cat ~/.claude/settings.json
 
 - `ANTHROPIC_BASE_URL` 指向 `http://127.0.0.1:4142`
 - `ANTHROPIC_AUTH_TOKEN` 存在；它是给本地中继用的占位值
+- `CLAUDE_CODE_AUTO_MODE_SERVER` 为 `0`，除非你设置了其他值
 
 改 `host` 或 `port` 需要重启中继，因为监听 socket 无法在热重载期间迁移。
 
+## Claude Code 提示会话不符合 auto mode 的条件
+
+Claude Code 可能在终端中，或用 `claude -p` 时在 stderr 上，打印类似这样的提示：
+
+```text
+We're changing auto mode to no longer charge for classifier requests in Claude Code. However, this session isn't eligible because your requests go through 127.0.0.1:4142, which isn't compatible with this update. Nothing breaks: auto mode keeps working, and its classifier requests are billed as before.
+```
+
+经过中继时这是预期行为：auto mode 的服务器检查到达不了 Claude Code，所以它自己发起
+classifier 请求。在 Claude Code 的环境中设置 `CLAUDE_CODE_AUTO_MODE_SERVER=0`，可停止
+服务器检查和这条提示。`claudeSetup: true` 时，`copilot-relay start` 会在缺少该值时
+写入它；见[配置](ZH-Configuration.md)。
+
 ## 安全分享日志
 
 **绝不要整份分享原始捕获。** 即使排除了认证 header，提示词、工具参数/结果、签名