Codex 支持 agents.toml config_overrides（exec -c 透传）

SETUP.md

@@ -491,7 +491,21 @@ If OpenCode's configured default model is disabled, the first prompt fails with
 
 Bridge drives Desktop-bundled or PATH `codex` through `codex exec --json`. It does not open ChatGPT.exe. Discovery order: `CODEX_CLI_PATH` (env or `$CODEX_HOME/config.toml` / `~/.codex/config.toml`), then `%LOCALAPPDATA%\\OpenAI\\Codex\\bin\\<hash>\\codex.exe` (newest mtime among binaries that pass `exec --help` capability checks), then PATH `codex`.
 
-Login stays in `$CODEX_HOME` when set, otherwise `~/.codex`. `--ignore-user-config` skips Desktop `config.toml` (MCP / Computer Use), and Bridge adds `-c cli_auth_credentials_store="auto"` so ChatGPT login still resolves from keyring or `auth.json`. Project `.codex/config.toml`, `AGENTS.md`, and execpolicy rules can still apply. Bridge inherits `CODEX_HOME` so a custom home used by Desktop is visible to the worker.
+Login stays in `$CODEX_HOME` when set, otherwise `~/.codex`. `--ignore-user-config` skips Desktop `config.toml` (MCP / Computer Use), and Bridge adds `-c cli_auth_credentials_store="auto"` so ChatGPT login still resolves from keyring or `auth.json`. Project `.codex/config.toml` and project hooks/rules normally do not load, because Codex reads project trust from the user config this flag skips; `AGENTS.md` and rules under `$CODEX_HOME/rules` still apply. Bridge inherits `CODEX_HOME` so a custom home used by Desktop is visible to the worker.
+
+`[agents.codex] config_overrides` adds raw `codex exec -c <item>` pairs — for settings `--ignore-user-config` would skip, e.g. a custom model provider. The block below merges field-wise into the bundled codex entry, so only `config_overrides` is needed:
+
+```toml
+[agents.codex]
+config_overrides = [
+  'model_provider="my-gateway"',
+  'model_providers.my-gateway.name="My Gateway"',
+  'model_providers.my-gateway.base_url="https://example.com/v1"',
+  'model_providers.my-gateway.env_key="MY_GATEWAY_API_KEY"',
+]
+```
+
+Keep secrets out of overrides: a key can only be referenced via `env_key` (command-line arguments are visible in the process list). Since Codex clears the MCP child environment, a variable named by `env_key` must also be listed in the host's `env_vars` / `[env.inherit]` like any other worker credential. Task-level `model`/`effort` are appended after overrides, so they keep priority.
 
 Default approval is `--approve-for-me` (auto review + workspace-write). `--yolo` is only used when `[agents.codex] session_meta = { yolo = true }`. Prompt is UTF-8 stdin (`-`), never an argv string.
 

src/agent_bridge/adapters/codex.py

@@ -62,6 +62,7 @@ def _build_cmd(self, session: Session, task: Task, env: dict[str, str] | None =
             effort=task.effort or session.effort,
             resume_id=session.native_session_id,
             yolo=yolo_requested(self.agent.session_meta),
+            config_overrides=self.agent.config_overrides,
         )
 
     async def _drain_stderr(self, proc: asyncio.subprocess.Process, session_id: str) -> str:

src/agent_bridge/codex_exec.py

@@ -221,6 +221,7 @@ def build_codex_exec_argv(
     effort: str | None = None,
     resume_id: str | None = None,
     yolo: bool = False,
+    config_overrides: Sequence[str] = (),
 ) -> list[str]:
     cmd = [
         *command,
@@ -241,6 +242,9 @@ def build_codex_exec_argv(
         "-C",
         cwd,
     ]
+    for item in config_overrides:
+        # Earlier than the task's -m / effort -c so per-dispatch values win.
+        cmd += ["-c", item]
     if model:
         cmd += ["-m", model]
     mapped = codex_effort(effort)

src/agent_bridge/config.py

@@ -126,6 +126,10 @@ class AgentConfig(BaseModel):
     # privilege), "allow_always" persists the grant inside the worker,
     # "deny" cancels every request.
     permission_policy: str = "allow_once"
+    # Codex only: each item is passed as `codex exec -c <item>` (raw TOML
+    # key=value). No key validation, no filtering, user config.toml is
+    # never read.
+    config_overrides: list[str] = Field(default_factory=list)
 
 
 class EnvConfig(BaseModel):

src/agent_bridge/quota_endpoints.py

@@ -85,7 +85,7 @@ def _codex(cfg: AgentConfig, env: Mapping[str, str]) -> str | None:
     if any(arg.split("=", 1)[0] in ("--oss", "--local-provider") for arg in cfg.command):
         return CUSTOM_ENDPOINT
     override = _option(cfg.command, "-c", "--config")
-    if override is not None:
+    if override is not None or cfg.config_overrides:
         # CLI overrides are arbitrary TOML paths. Do not build a second Codex
         # configuration resolver merely to guess which account they select.
         return "quota lookup is unsupported with Codex CLI configuration overrides"

tests/test_codex_exec.py

@@ -87,6 +87,66 @@ def test_build_argv_resume_and_yolo():
     assert cmd[-3:] == ["resume", "thread-1", "-"]
 
 
+def test_build_argv_config_overrides_sit_before_task_flags():
+    overrides = [
+        'model_provider="my-gateway"',
+        'model_providers.my-gateway.base_url="https://example.com/v1"',
+    ]
+    cmd = build_codex_exec_argv(
+        ["codex.exe"],
+        cwd=r"E:\proj",
+        model="gpt-5.6-sol",
+        effort="high",
+        config_overrides=overrides,
+    )
+    assert "--ignore-user-config" in cmd
+    # Each override is its own `-c <item>` pair, in order, after `-C <cwd>`
+    # and before the task-level `-m`/effort so those still win.
+    first = cmd.index(overrides[0])
+    assert cmd[first - 1] == "-c"
+    assert cmd[cmd.index(overrides[1]) - 1] == "-c"
+    assert cmd.index("-C") < first < cmd.index(overrides[1]) < cmd.index("-m")
+    assert cmd.index(overrides[1]) < cmd.index('model_reasoning_effort="high"')
+    assert cmd[-1] == "-"
+
+
+def test_build_argv_empty_overrides_unchanged():
+    base = build_codex_exec_argv(
+        ["codex.exe"], cwd=r"E:\proj", model="gpt-5.6-sol", effort="high"
+    )
+    with_empty = build_codex_exec_argv(
+        ["codex.exe"],
+        cwd=r"E:\proj",
+        model="gpt-5.6-sol",
+        effort="high",
+        config_overrides=[],
+    )
+    assert with_empty == base
+
+
+def test_adapter_config_overrides_reach_argv(tmp_path: Path, monkeypatch):
+    monkeypatch.setattr(
+        "agent_bridge.adapters.codex.resolve_codex_command",
+        lambda *args, **kwargs: ["codex.exe"],
+    )
+    adapter = CodexAdapter(
+        AgentConfig(
+            name="codex",
+            protocol="codex",
+            command=["codex"],
+            config_overrides=['model_provider="my-gateway"'],
+        ),
+        tmp_path,
+    )
+    session = Session(session_id="s1", agent="codex", cwd=str(tmp_path))
+    cmd = adapter._build_cmd(
+        session,
+        Task(task_id="t1", session_id="s1", agent="codex", message="hi", cwd=str(tmp_path)),
+    )
+    assert 'model_provider="my-gateway"' in cmd
+    assert cmd[cmd.index('model_provider="my-gateway"') - 1] == "-c"
+
+
 def test_jsonl_takes_last_agent_message_and_requires_turn_completed():
     state = CodexTurnState()
     apply_codex_event(state, {"type": "thread.started", "thread_id": "t1"})

tests/test_quota_endpoints.py

@@ -103,6 +103,19 @@ async def test_fallback_endpoint_overrides_cannot_reuse_official_cache(name, fla
         await cache.close()
 
 
+def test_codex_agent_config_overrides_block_quota_lookup():
+    # config_overrides become `codex exec -c` pairs; like `-c` flags in
+    # command they can select a custom provider, so fail closed.
+    agent = AgentConfig(
+        name="codex",
+        protocol="codex",
+        command=["codex"],
+        config_overrides=['model_provider="my-gateway"'],
+    )
+    reason = quota_block_reason(agent, {})
+    assert reason is not None and "configuration overrides" in reason
+
+
 @pytest.mark.parametrize("fallback", [False, True])
 @pytest.mark.parametrize(("name", "folder", "flag", "text"), [
     ("codex", ".codex", "-pwork", '[profiles.work]\nmodel_provider="custom"'),