Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

fix(codex): restore openai_base_url top-level injection for subscription routing

JerrettDavisd54c5b6a5840
diff --git a/headroom/cli/wrap.py b/headroom/cli/wrap.py
index 3686f2752..668fce5b0 100644
--- a/headroom/cli/wrap.py
+++ b/headroom/cli/wrap.py
@@ -542,10 +542,16 @@ def _prepare_wrap_rtk(verbose: bool = False, *, label: str | None = None) -> Pat
def _inject_codex_provider_config(port: int) -> None:
"""Inject a Headroom model provider into Codex's config.toml.
- Codex ignores OPENAI_BASE_URL for WebSocket transport unless a custom
- provider declares ``supports_websockets = true``. This writes a
- ``[model_providers.headroom]`` section that routes both HTTP and WS
- through the proxy, and sets ``model_provider = "headroom"``.
+ Two keys are written in the top-level block:
+
+ * ``model_provider = "headroom"`` — selects the custom provider for
+ API-key mode traffic.
+ * ``openai_base_url = "http://127.0.0.1:{port}/v1"`` — overrides the
+ built-in ``openai`` provider's base URL. This is the critical key for
+ **subscription (ChatGPT plan) users**: Codex detects subscription auth
+ and routes through the built-in ``openai`` provider regardless of
+ ``model_provider``, so without this override it bypasses the proxy and
+ hits ``https://chatgpt.com/backend-api/codex`` directly.
Safe to call multiple times — the injected block is fully replaced on
each call, so re-running with a different ``port`` updates the config.
@@ -563,7 +569,10 @@ def _inject_codex_provider_config(port: int) -> None:
# stripping them is unambiguous and never consumes user content that
# happens to sit between the two.
top_level_block = (
- f'{_CODEX_TOP_LEVEL_MARKER}\nmodel_provider = "headroom"\n{_CODEX_END_MARKER}\n'
+ f"{_CODEX_TOP_LEVEL_MARKER}\n"
+ f'model_provider = "headroom"\n'
+ f'openai_base_url = "http://127.0.0.1:{port}/v1"\n'
+ f"{_CODEX_END_MARKER}\n"
)
provider_section = (
f"{_CODEX_TOP_LEVEL_MARKER}\n"
diff --git a/tests/test_install/test_providers.py b/tests/test_install/test_providers.py
index ba1e4525f..2aae48d0f 100644
--- a/tests/test_install/test_providers.py
+++ b/tests/test_install/test_providers.py
@@ -508,14 +508,20 @@ def test_headroom_provider_block_never_sets_requires_openai_auth(
assert "[model_providers.headroom]" in content
-def test_inject_codex_provider_config_does_not_write_openai_base_url(
+def test_inject_codex_provider_config_writes_openai_base_url(
monkeypatch: pytest.MonkeyPatch, tmp_path: Path
) -> None:
- """_inject_codex_provider_config must NOT write a top-level openai_base_url key.
+ """_inject_codex_provider_config MUST write a top-level openai_base_url key.
- Bug 3 (#406): injecting openai_base_url at the top level of config.toml
- causes ChatGPT/subscription users to bypass headroom and ride the built-in
- openai provider directly. The field must be absent after injection.
+ Bug 3 (#406) has two halves:
+ 1. Strip ``requires_openai_auth`` from the headroom provider block (done in 3ca48d3).
+ 2. Inject ``openai_base_url`` at the top level so subscription (ChatGPT plan)
+ users are also routed through headroom.
+
+ Without the top-level ``openai_base_url`` override, Codex subscription mode
+ uses the built-in ``openai`` provider with ``chatgpt.com/backend-api/codex``
+ as the base URL, bypassing the proxy entirely. This key is the only way to
+ intercept subscription traffic.
"""
from headroom.cli import wrap as wrap_mod
@@ -529,10 +535,12 @@ def test_inject_codex_provider_config_does_not_write_openai_base_url(
assert config_file.exists(), "inject must create ~/.codex/config.toml"
content = config_file.read_text()
- # The top-level openai_base_url key must be absent. Its presence would
- # bypass headroom for subscription/ChatGPT users.
- assert "openai_base_url" not in content, (
- "openai_base_url must not appear in config.toml after injection; "
+ # The top-level openai_base_url key must be present at exactly this value.
+ # Any port drift (e.g. 9999) will break subscription routing and must cause
+ # this test to fail.
+ assert 'openai_base_url = "http://127.0.0.1:8787/v1"' in content, (
+ "openai_base_url must appear at the top level of config.toml after injection "
+ "so that Codex subscription (ChatGPT plan) users are routed through headroom; "
f"got:\n{content}"
)
# Sanity: the provider block is actually there.
@@ -543,3 +551,58 @@ def test_inject_codex_provider_config_does_not_write_openai_base_url(
"requires_openai_auth must be absent from the injected provider block; "
f"got:\n{content}"
)
+ # openai_base_url must be in the top-level block, NOT inside [model_providers.*]
+ lines = content.splitlines()
+ in_provider_section = False
+ for line in lines:
+ stripped = line.strip()
+ if stripped.startswith("["):
+ in_provider_section = stripped.startswith("[model_providers")
+ if in_provider_section and stripped.startswith("openai_base_url"):
+ raise AssertionError(
+ f"openai_base_url must be top-level, not inside a [model_providers.*] section; "
+ f"found it inside a provider section. Config:\n{content}"
+ )
+
+
+def test_unwrap_removes_top_level_openai_base_url(
+ monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+) -> None:
+ """After unwrap, the top-level openai_base_url key must be gone.
+
+ Verifies that ``_restore_codex_provider_config`` (unwrap) removes the
+ ``openai_base_url`` key so orphaned entries don't accumulate between
+ wrap/unwrap cycles.
+ """
+ from headroom.cli import wrap as wrap_mod
+
+ home = tmp_path
+ monkeypatch.setenv("HOME", str(home))
+ monkeypatch.setenv("USERPROFILE", str(home))
+
+ wrap_mod._inject_codex_provider_config(8787)
+ config_file = home / ".codex" / "config.toml"
+ assert 'openai_base_url = "http://127.0.0.1:8787/v1"' in config_file.read_text()
+
+ wrap_mod._restore_codex_provider_config()
+
+ # After unwrap the file should be gone (no prior content) or cleaned.
+ if config_file.exists():
+ content = config_file.read_text()
+ assert "openai_base_url" not in content, (
+ "openai_base_url must not remain in config.toml after unwrap; "
+ f"got:\n{content}"
+ )
+ # Also verify via _strip_codex_headroom_blocks directly — the orphan-cleanup
+ # path is exercised when there is no backup file (crash-recovery path).
+ orphan_content = (
+ 'model = "gpt-4o"\n'
+ 'openai_base_url = "http://127.0.0.1:8787/v1"\n'
+ 'model_provider = "headroom"\n'
+ )
+ stripped = wrap_mod._strip_codex_headroom_blocks(orphan_content)
+ assert "openai_base_url" not in stripped, (
+ "_strip_codex_headroom_blocks must remove orphaned openai_base_url lines; "
+ f"got:\n{stripped}"
+ )
+ assert 'model = "gpt-4o"' in stripped