test(openai): set Gin mode once per package

[warelik]

Summary

Port the accepted CPA gin test-mode race fix to CPAPlus. Sets gin.TestMode once in a package-level TestMain and removes 59 per-test gin.SetMode(gin.TestMode) calls across 9 test files in sdk/api/handlers/openai.

Why (race)

The per-test gin.SetMode(gin.TestMode) calls race under parallel tests: parallel tests write the package-global gin mode (gin package mode var) while other parallel tests read it through gin.New / gin.CreateTestContext. The result is a data race flagged by go test -race.

Reproduced pre-fix (unmodified base, this branch's base 4823235a)

go test -race -count=2 -run 'TestPrepareCodexMultiAgentV2ToolsAtResponsesBoundary$|TestResponsesPreparesCodexMultiAgentV2ToolsForHTTPAndSSE|TestPrepareCodexMultiAgentV2ToolsAtResponsesBoundarySkipsOtherClients' ./sdk/api/handlers/openai/

WARNING: DATA RACE
Write at 0x... by goroutine 20:
  gin.SetMode() gin@v1.10.1/mode.go:72
  openai.TestPrepareCodexMultiAgentV2ToolsAtResponsesBoundarySkipsOtherClients() openai_responses_multi_agent_test.go:186
Previous write at 0x... by goroutine 18:
  gin.SetMode() gin@v1.10.1/mode.go:72
  openai.TestPrepareCodexMultiAgentV2ToolsAtResponsesBoundary() openai_responses_multi_agent_test.go:26
--- FAIL: <3 parallel tests> race detected during execution of test

Change

• Add sdk/api/handlers/openai/gin_testmode_test.go — package openai, TestMain that calls gin.SetMode(gin.TestMode) once before os.Exit(m.Run()).
• Remove all per-test gin.SetMode(gin.TestMode) calls (write-write race source) from:
  • gitlab_duo_handler_test.go (2)
  • openai_images_handlers_test.go (1)
  • openai_responses_compact_test.go (3)
  • openai_responses_handlers_stream_error_test.go (5)
  • openai_responses_handlers_stream_test.go (2)
  • openai_responses_multi_agent_test.go (3)
  • openai_responses_signature_test.go (2)
  • openai_responses_websocket_test.go (35)
  • openai_videos_handlers_test.go (6)
• All 3 t.Parallel() and all gin imports preserved (still used by gin.New/gin.CreateTestContext).

Stat: 10 files changed, 18 insertions(+), 59 deletions(-).

Verification (all green)

• Post-fix focused race -count=3 (the 3 parallel multi-agent tests): ok, no DATA RACE.
• Full package race -count=1: ok.
• Package normal: ok.
• Full go test ./...: exit 0.
• go build ./...: exit 0.
• go vet ./sdk/api/handlers/openai/: exit 0.
• gofmt clean; git diff --check clean.
• Invariant: exactly 1 TestMain, exactly 1 real gin.SetMode(gin.TestMode) (in TestMain), zero per-test calls.

Functionally isolated from #175 (author-failover-recovery) — different branch (fix/gin-testmode-race), test-infrastructure only, no production code.

Mirror

CPA: router-for-me/CLIProxyAPI#4948 (identical changes)

[warelik]

Evidence — race fixed, plus CI build check classification

Change (accepted CPA port, functionally isolated to openai tests)

• New sdk/api/handlers/openai/gin_testmode_test.go: package openai, TestMain calls gin.SetMode(gin.TestMode) once.
• Removed 59 per-test gin.SetMode(gin.TestMode) calls across 9 test files (all in sdk/api/handlers/openai/).
• Stat: 10 files changed, 18 insertions(+), 59 deletions(-). Zero production/non-openai changes.

Pre-fix race (reproduced on unmodified base 4823235a)

go test -race -count=2 on the 3 parallel multi-agent tests → WARNING: DATA RACE
(write-write on gin.SetMode, mode.go:72, at multi_agent_test.go:26/134/186); all 3 FAIL.

Post-fix verification (all green)

• Focused race -count=3 (the 3 parallel tests): ok, no DATA RACE.
• Full openai package race -count=1: ok. Package normal: ok.
• Full go test ./...: exit 0. go build ./...: exit 0. go vet ./sdk/api/handlers/openai/: exit 0.
• gofmt clean; git diff --check clean.
• Invariant: exactly 1 TestMain, exactly 1 real gin.SetMode(gin.TestMode) (in TestMain), zero per-test calls, 3 t.Parallel unchanged.

Build CHECK failure is unrelated (pre-existing flake)

build run 31718293427 failed at go test ./... on:
internal/runtime/executor/websocket_session_target_test.go:286: second request error = websocket: close 1006 (abnormal closure): unexpected EOF
(subtest TestWebsocketRetryBindFailureClearsActiveSessionState/Codex_nonstream).

• This PR touches only sdk/api/handlers/openai/; internal/runtime/executor is unchanged (git diff --name-only HEAD~1 HEAD).
• That test uses raw websocket/httptest (never gin), so no shared global/gin-mode coupling, and my change lives in a different test binary package.
• Standalone repro does NOT reproduce:
  • PR head -count=1: ok
  • PR head -count=5: ok (0.753s)
  • Clean base 4823235a (disposable worktree): ok (0.823s)
• The test forces a past write deadline (SetWriteDeadline(now.Add(-1s))) to induce a write failure, then races active-session recovery — timing-sensitive under full parallel load.

Recommended for maintainer: re-run the build check (or confirm as known flake). No code/test change was made to mask it.

[warelik]

@codex review

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.