fix: route codex payloads around the SDK's GIL-holding request transform (#93650)

[kshitijk4poor]

Fixes #93650. Salvage of #93681 (@kchernev).

Summary

Routes the bulk payload fields (input, tools) through extra_body instead of the typed SDK kwargs, bypassing the OpenAI SDK's maybe_transform walk that can wedge for hours holding the GIL on large Codex Responses API calls.

Problem

A codex_responses streaming call can freeze the entire agent for hours before a single byte leaves the process. The OpenAI SDK's responses.create re-walks the whole request body against the ResponseCreateParams union graph (openai/_utils/_transform.py::maybe_transform) client-side, holding the GIL. In the incident documented in #93650 (openai 2.24.0, CPython 3.11.16), one such walk over a ~1.4 MB / 1,093-item conversation wedged for 12.5 CPU-hours at ~99% of a core.

Core-dump forensics showed the worker in maybe_transform → isinstance(obj, Mapping) → typing.__subclasscheck__, with 9 other threads in take_gil futex waits — including the TTFB/stale watchdog poller, which never got to run any of its detectors. No in-process watchdog can rescue this: the GIL-holding walk starves them, and even when they run their remedy is socket-level (_close_request_client_once) — a pre-network hang has no socket to kill.

Fix

Hermes assembles Codex payloads from JSON round-trips — they are already wire format, so the transform walk is pure overhead plus this failure mode. The SDK merges extra_body into the JSON body after the transform (_base_client._build_request, shallow merge), a mechanism this codebase already relies on for prompt_cache_retention. This PR routes the bulk payload fields (input, tools) through extra_body, skipping maybe_transform for them entirely while producing an identical request body.

Safety rails:

• Plain-JSON guard — a field is only moved if every node is a plain JSON type (dict/list/str/int/float/bool/None with str keys). Non-JSON types keep the typed SDK path.
• Caller precedence — an explicit caller-provided extra_body entry wins over a moved field.
• Escape hatch — HERMES_CODEX_SDK_TRANSFORM=1 restores the pre-fix behavior.

Applied at both live responses.create call sites: the primary stream path (codex_runtime._open_codex_stream) and the auxiliary adapter (auxiliary_client._CodexCompletionsAdapter).

Changes

• agent/codex_runtime.py: +74 lines — _is_plain_json_data, _bypass_sdk_request_transform, wired into _open_codex_stream
• agent/auxiliary_client.py: +7/-1 — wired into _CodexCompletionsAdapter.create
• tests/run_agent/test_codex_sdk_transform_bypass.py: +153 lines — 10 new tests
• tests/agent/test_auxiliary_client.py: +14/-1 — fake-client capture updated for both wire shapes
• tests/run_agent/test_native_compaction.py: +5/-1 — fake-client capture updated
• tests/run_agent/test_run_agent_codex_responses.py: +6/-1 — retention guard updated for new wire shape
• contributors/emails/: AUTHOR_MAP entry for @kchernev

Validation

	Before	After
Targeted tests (codex + aux + compaction)	—	245 passed
New bypass tests	—	10 passed
E2E (real imports, _bypass_sdk_request_transform)	—	6/6 passed
Ruff check	—	clean

Competitor research

• LiteLLM: bypasses the SDK entirely (builds raw httpx requests), so maybe_transform never runs
• Codex CLI (Rust): builds JSON requests directly, structurally immune
• LangChain: uses chat.completions.create (smaller typedict graph, not the massive ResponseCreateParams union)
• Claude Code: routes through Anthropic's native API, not the OpenAI SDK

No other major harness uses the OpenAI Python SDK's responses.create() with payloads as large as Hermes does (558K tokens, 1.4MB body). Hermes is uniquely exposed.