vinez/headroom
1
0
Fork 0
mirror of https://github.com/headroomlabs-ai/headroom.git synced 2026-08-10 14:27:00 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix(proxy): skip OpenAI tool_search deferral for Codex client (#2729)

Browse source 
## Description

OpenAI Responses tool_search deferral injects `defer_loading` and a
`tool_search` tool for eligible gpt-5.4+ requests. When the model later
calls a deferred tool, the function_call item carries a `namespace`
field. Codex CLI round-trip structs drop unknown fields, so the next
request omits `namespace` and OpenAI returns 400, killing the session
mid-run. Proxy logs for these Codex turns show no tool savings, so the
injection breaks Codex without benefit.

This skips OpenAI tool_search deferral when the classified client is
Codex, leaving other clients unchanged.

Closes #2726

## Type of Change

- [x] Bug fix (non-breaking change that fixes an issue)
- [ ] New feature (non-breaking change that adds functionality)
- [ ] Breaking change (fix or feature that would cause existing
functionality to change)
- [ ] Documentation update
- [ ] Performance improvement
- [ ] Code refactoring (no functional changes)

## Changes Made

- Add `openai_tool_search_client_supported` and a Codex-only unsupported
client set
- Pass optional `client` into `inject_tool_search_deferral_openai` and
no-op for Codex
- Plumb `client` through Responses compression (HTTP, WebSocket,
passthrough) with legacy-signature retries
- Add regression tests for Codex skip and non-Codex still injects

## Testing

- [x] Unit tests pass (`pytest`)
- [x] Linting passes (`ruff check .`)
- [ ] Type checking passes (`mypy headroom`)
- [x] New tests added for new functionality
- [ ] Manual testing performed

### Test Output

```text
$ .venv/bin/python -m pytest tests/test_openai_tool_search_deferral.py -q -o addopts=
34 passed, 1 warning in 0.27s

$ .venv/bin/python -m ruff check headroom/proxy/helpers.py headroom/proxy/handlers/openai.py tests/test_openai_tool_search_deferral.py
All checks passed!

$ .venv/bin/python -m ruff format --check headroom/proxy/helpers.py headroom/proxy/handlers/openai.py tests/test_openai_tool_search_deferral.py
3 files already formatted
```

## Real Behavior Proof

- Environment: Linux x86_64, Python 3.14.5 in repo .venv, shallow
checkout of headroomlabs-ai/headroom main at 01df245 plus this branch
- Exact command / steps: `.venv/bin/python -m pytest
tests/test_openai_tool_search_deferral.py -q -o addopts=`;
`.venv/bin/python -m ruff check headroom/proxy/helpers.py
headroom/proxy/handlers/openai.py
tests/test_openai_tool_search_deferral.py`; `.venv/bin/python -m ruff
format --check headroom/proxy/helpers.py
headroom/proxy/handlers/openai.py
tests/test_openai_tool_search_deferral.py`
- Observed result: 34 targeted tests passed, including Codex client
identity no-op and non-Codex still injecting tool_search; ruff check and
format check clean on touched files
- Not tested: live `codex exec` multi-turn session through headroom
proxy with >=12 tools; full monorepo `make ci-precheck`; mypy

## Review Readiness

- [x] I have performed a self-review
- [x] This PR is ready for human review

## Checklist

- [x] My code follows the project's style guidelines
- [x] I have performed a self-review of my code
- [ ] I have commented my code, particularly in hard-to-understand areas
- [ ] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective or that my
feature works
- [x] New and existing unit tests pass locally with my changes
- [x] I did **not** edit `CHANGELOG.md` - it is generated by
release-please from my Conventional Commit PR title (a CI guard enforces
this)

## Screenshots (if applicable)

N/A

## Additional Notes

- Scoped to OpenAI Responses tool_search deferral client gating only;
Anthropic tool_search path unchanged
- Related open work for OpenCode (#2696) is separate; this PR only
excludes Codex
- mypy not run on this VPS for this change
This commit is contained in:
AxelRay 2026-08-03 18:41:21 +07:00 • committed by GitHub
parent 01df245252
commit 56b3e4c1b1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 98 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

81
headroom/proxy/handlers/openai.py
View file

@ -2209,6 +2209,7 @@ class OpenAIHandlerMixin:
model: str,
request_id: str,
timing: dict[str, float] | None = None,
client: str | None = None,
) -> tuple[dict[str, Any], bool, int, list[str], str | None, int, int, int]:
"""Compress an OpenAI Responses payload through the shared router.
@ -2346,7 +2347,9 @@ class OpenAIHandlerMixin:
# one — hence a transform tag but no tokens_saved claim.
from headroom.proxy.helpers import inject_tool_search_deferral_openai
_deferred_tools = inject_tool_search_deferral_openai(working.get("tools"), model)
_deferred_tools = inject_tool_search_deferral_openai(
working.get("tools"), model, client=client
)
if _deferred_tools is not working.get("tools"):
if working is payload:
working = copy.deepcopy(payload)
@ -2548,6 +2551,7 @@ class OpenAIHandlerMixin:
model: str,
request_id: str,
timeout: float = COMPRESSION_TIMEOUT_SECONDS,
client: str | None = None,
) -> tuple[dict[str, Any], bool, int, list[str], str | None, int, int, int, dict[str, float]]:
timing: dict[str, float] = {}
@ -2562,21 +2566,32 @@ class OpenAIHandlerMixin:
shape_labels, shape_mutated = _shape_openai_responses_payload(
payload, model=model, request_id=request_id
)
try:
result = self._compress_openai_responses_payload(
payload,
model=model,
request_id=request_id,
timing=timing,
)
except TypeError as exc:
if "unexpected keyword argument 'timing'" not in str(exc):
raise
result = self._compress_openai_responses_payload(
payload,
model=model,
request_id=request_id,
)
compression_kwargs: dict[str, Any] = {
"model": model,
"request_id": request_id,
"timing": timing,
"client": client,
}
while True:
try:
result = self._compress_openai_responses_payload(
payload,
**compression_kwargs,
)
break
except TypeError as exc:
unsupported_kwarg = next(
(
name
for name in ("client", "timing")
if f"unexpected keyword argument '{name}'" in str(exc)
and name in compression_kwargs
),
None,
)
if unsupported_kwarg is None:
raise
compression_kwargs.pop(unsupported_kwarg)
if shape_labels:
# Carry the shaper labels on the transforms channel so the
# outcome funnel feeds the output-savings ledger
@ -4917,6 +4932,7 @@ class OpenAIHandlerMixin:
body,
model=model,
request_id=request_id,
client=client,
)
attempted_input_tokens = int(_attempted_tokens)
if _transforms:
@ -6559,6 +6575,7 @@ class OpenAIHandlerMixin:
timeout=_codex_ws_compression_timeout_seconds()
if client == "codex"
else COMPRESSION_TIMEOUT_SECONDS,
client=client,
)
for _timing_name, _timing_ms in _ws_compression_timing.items():
_record_ws_compression_timing(_timing_name, _timing_ms)
@ -6904,6 +6921,7 @@ class OpenAIHandlerMixin:
timeout=_codex_ws_compression_timeout_seconds()
if client == "codex"
else COMPRESSION_TIMEOUT_SECONDS,
client=client,
)
for _timing_name, _timing_ms in frame_compression_timing.items():
_record_ws_compression_timing(_timing_name, _timing_ms)
@ -8650,7 +8668,9 @@ class OpenAIHandlerMixin:
},
)
async def _maybe_compress_passthrough_responses(self, body: bytes) -> bytes:
async def _maybe_compress_passthrough_responses(
self, body: bytes, *, client: str | None = None
) -> bytes:
"""Compress an OpenAI Responses-shaped passthrough body, fail-open.
Reuses the native `/v1/responses` compression path so custom
@ -8669,15 +8689,22 @@ class OpenAIHandlerMixin:
model = str(payload.get("model") or "passthrough")
request_id = await self._next_request_id()
try:
(
compressed_payload,
modified,
*_rest,
) = await self._compress_openai_responses_payload_in_executor(
payload,
model=model,
request_id=request_id,
)
try:
result = await self._compress_openai_responses_payload_in_executor(
payload,
model=model,
request_id=request_id,
client=client,
)
except TypeError as exc:
if "unexpected keyword argument 'client'" not in str(exc):
raise
result = await self._compress_openai_responses_payload_in_executor(
payload,
model=model,
request_id=request_id,
)
compressed_payload, modified, *_rest = result
except Exception as exc: # noqa: BLE001 — fail-open on any compressor error
logger.warning(
"[%s] passthrough Responses compression failed, forwarding verbatim: %s",
@ -8792,7 +8819,7 @@ class OpenAIHandlerMixin:
and path.rstrip("/").endswith("/responses")
and body
):
compressed = await self._maybe_compress_passthrough_responses(body)
compressed = await self._maybe_compress_passthrough_responses(body, client=client)
if compressed != body:
body = compressed
# Body size changed — let httpx recompute Content-Length.

17
headroom/proxy/helpers.py
View file

@ -2421,6 +2421,7 @@ def inject_tool_search_deferral(
_OPENAI_TOOL_SEARCH_TYPE = "tool_search"
_OPENAI_TOOL_SEARCH_MIN_TOOLS = 12
_OPENAI_TOOL_SEARCH_RESIDENT_NAMES = frozenset({"terminal"})
_OPENAI_TOOL_SEARCH_UNSUPPORTED_CLIENTS = frozenset({"codex"})
# gpt-5.4 is the first model with Responses tool_search (OpenAI docs). Version-
# gated by default; overridable per deployment via a regex in
# HEADROOM_OPENAI_TOOL_SEARCH_MODELS (matched against the model name) so new
@ -2451,24 +2452,34 @@ def _model_supports_openai_tool_search(model: str | None) -> bool:
return (major, minor) >= _OPENAI_TOOL_SEARCH_MIN_VERSION
def openai_tool_search_client_supported(client: str | None) -> bool:
"""Return whether OpenAI tool search deferral is safe for this client."""
normalized = client.strip().lower() if client else ""
return normalized not in _OPENAI_TOOL_SEARCH_UNSUPPORTED_CLIENTS
def inject_tool_search_deferral_openai(
tools: Any,
model: str | None,
*,
client: str | None = None,
core_tools: frozenset[str] = _TOOL_SEARCH_CORE_TOOLS,
) -> Any:
"""Return a new Responses ``tools`` list with non-core function/MCP tools
deferred + a ``{"type": "tool_search"}`` tool injected, or the original list
unchanged when injection doesn't apply.
No-op when: the model doesn't support tool search (gpt-5.4+ only), ``tools``
No-op for Codex, whose round-trip structs drop deferred-call namespaces. Also
no-op when: the model doesn't support tool search (gpt-5.4+ only), ``tools``
is not a list, there are fewer than ``_OPENAI_TOOL_SEARCH_MIN_TOOLS``, a
tool_search tool is already present (client already defers), or nothing would
be deferred. Core coding tools and hosted/typed tools (web_search,
file_search, code_interpreter, computer, …) stay resident and unchanged, so
routine edit/read/run loops never pay a search round-trip and the request
file_search, code_interpreter, computer, ...) stay resident and unchanged,
so routine edit/read/run loops never pay a search round-trip and the request
stays valid; the injected search tool is itself resident.
"""
if not openai_tool_search_client_supported(client):
return tools
if not _model_supports_openai_tool_search(model):
return tools
if not isinstance(tools, list) or len(tools) < _OPENAI_TOOL_SEARCH_MIN_TOOLS:

30
tests/test_openai_tool_search_deferral.py
View file

@ -15,6 +15,7 @@ import pytest
from headroom.proxy.helpers import (
_model_supports_openai_tool_search,
inject_tool_search_deferral_openai,
openai_tool_search_client_supported,
)
@ -57,6 +58,35 @@ def test_env_override_wins_then_falls_back(monkeypatch):
# --- deferral behavior -------------------------------------------------------
@pytest.mark.parametrize(
("client", "supported"),
[(None, True), ("codex", False), (" CODEX ", False), ("opencode", True), ("claude", True)],
)
def test_client_supported(client, supported):
assert openai_tool_search_client_supported(client) is supported
def test_codex_client_does_not_inject():
tools = _tools()
out = inject_tool_search_deferral_openai(tools, "gpt-5.5", client="codex")
assert out is tools
assert all(tool.get("type") != "tool_search" for tool in out)
assert all("defer_loading" not in tool for tool in out)
@pytest.mark.parametrize("client", [None, "opencode"])
def test_supported_clients_still_inject(client):
tools = _tools()
out = inject_tool_search_deferral_openai(tools, "gpt-5.5", client=client)
assert out is not tools
assert out[0] == {"type": "tool_search"}
assert any(tool.get("defer_loading") is True for tool in out)
def test_defers_non_core_and_injects_search_tool():
tools = _tools()
out = inject_tool_search_deferral_openai(tools, "gpt-5.5")