fix(integration): install codex CLI + isolate codex auth from the deepseek judge

[Yiminnn]

Third follow-up to #806/#807 to get the full L3 review green. The first full run on #803 reached review-pack with plan ✓ + all 10 rollouts ✓ + deterministic grader ✓ — only codex failed.

Root causes

1. codex binary not found: 'codex' — the review-pack job never installed the codex CLI.
2. codex auth clobbered to DeepSeek — provider-select exports OPENAI_*=DeepSeek for the cheap Pass-1 judge; codex_review.py then wrote the codex auth.json from that key and ran codex exec against the DeepSeek base.

Fix

• Install @openai/codex@0.141.0 in both review-pack jobs (L2 advisory, L3 required), continue-on-error so a failed install still hits codex's fail-closed verdict.
• _codex_env in codex_review.py isolates the host codex exec: real OpenAI key via CODEX_API_KEY + DeepSeek OPENAI_BASE_URL dropped (default OpenAI endpoint). The Pass-1 deepseek judge keeps the original env. Workflows pass CODEX_API_KEY: secrets.OPENAI_API_KEY (L2 also drops its OPENAI_API_KEY override, which paired the real key with the deepseek base).

Test plan

• ruff + YAML valid; _codex_env unit test (real-key isolation + backward-compat) green
• Re-dispatch L3 on Preserve pi-acp model metadata through LiteLLM proxy #803 → review-pack runs codex against the real OpenAI endpoint → real verdict (not 'codex unavailable')

[greptile-apps]

Greptile Summary

This PR fixes two root causes that prevented codex from completing during L3 review-pack runs: the codex CLI was never installed, and the OPENAI_* environment variables were being clobbered by the DeepSeek judge config before codex could authenticate.

• Adds npm install -g @openai/codex@0.141.0 (with continue-on-error) to both L2 and L3 review-pack jobs so the binary is available when codex_review.py tries to invoke it.
• Introduces _codex_env() in codex_review.py to isolate the codex CLI's env: when CODEX_API_KEY is set it becomes OPENAI_API_KEY and OPENAI_BASE_URL is dropped, preventing the DeepSeek judge clobber from leaking into codex auth; write_codex_auth, has_codex_auth, and run_codex_verdict are all updated to use this isolated env.
• Both workflow files switch from OPENAI_API_KEY: secrets.OPENAI_API_KEY to CODEX_API_KEY: secrets.OPENAI_API_KEY on the codex step, leaving Pass-1 DeepSeek variables intact in the job env.

Confidence Score: 4/5

Safe to merge; the auth-isolation logic is correct and well-tested, and both workflow files apply the fix consistently.

The _codex_env function correctly overwrites OPENAI_API_KEY and drops OPENAI_BASE_URL before spawning codex, but it leaves CODEX_API_KEY in the subprocess environment. Today's codex CLI ignores that variable, but it is a non-standard name that a future codex release might interpret, making the subprocess env slightly leakier than the doc-comment implies. Everything else — auth writing, auth checking, the fail-closed paths, and both workflow files — looks correct and consistent.

_codex_env in .github/scripts/codex_review.py — the CODEX_API_KEY key is not stripped before the env dict is handed to the codex subprocess.

Important Files Changed

Filename	Overview
.github/scripts/codex_review.py	Adds _codex_env() to isolate codex CLI auth from the DeepSeek judge env; main() now uses codex_env for write_codex_auth, has_codex_auth, and run_codex_verdict
.github/workflows/integration-final-review.yml	Adds npm install -g @openai/codex@0.141.0 (continue-on-error) before the codex step; switches from no explicit CODEX_API_KEY to CODEX_API_KEY: secrets.OPENAI_API_KEY
.github/workflows/integration-scope.yml	Adds codex CLI install step; replaces OPENAI_API_KEY: secrets.OPENAI_API_KEY with CODEX_API_KEY: secrets.OPENAI_API_KEY so the Pass-1 judge keeps the DeepSeek env while codex gets the real OpenAI key
tests/test_codex_review.py	Adds two unit tests for _codex_env: one verifying the DeepSeek-isolation path and one for the backward-compatible no-CODEX_API_KEY path

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant WF as GitHub Actions Job
    participant PS as provider-select
    participant CR as codex_review.py
    participant DS as DeepSeek Judge (Pass 1)
    participant CX as codex exec (Pass 2)

    WF->>PS: run provider-select
    PS-->>WF: "writes OPENAI_API_KEY=ds-key,<br/>OPENAI_BASE_URL=deepseek to $GITHUB_ENV"

    WF->>WF: "inject CODEX_API_KEY=secrets.OPENAI_API_KEY (step env)"

    WF->>CR: "uv run codex_review.py (env includes DeepSeek OPENAI_*<br/>+ CODEX_API_KEY=real-key)"

    CR->>DS: "run_deepseek_findings(env)<br/>(uses original env — DeepSeek OPENAI_API_KEY + BASE_URL)"
    DS-->>CR: per-rollout findings

    CR->>CR: "codex_env = _codex_env(env)<br/>→ OPENAI_API_KEY=real-key, OPENAI_BASE_URL dropped"
    CR->>CR: "write_codex_auth(codex_env)<br/>→ auth.json with real OpenAI key"
    CR->>CX: "subprocess.run(codex exec, env=codex_env)"
    CX-->>CR: verdict output
    CR-->>WF: final_verdict

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant WF as GitHub Actions Job
    participant PS as provider-select
    participant CR as codex_review.py
    participant DS as DeepSeek Judge (Pass 1)
    participant CX as codex exec (Pass 2)

    WF->>PS: run provider-select
    PS-->>WF: "writes OPENAI_API_KEY=ds-key,<br/>OPENAI_BASE_URL=deepseek to $GITHUB_ENV"

    WF->>WF: "inject CODEX_API_KEY=secrets.OPENAI_API_KEY (step env)"

    WF->>CR: "uv run codex_review.py (env includes DeepSeek OPENAI_*<br/>+ CODEX_API_KEY=real-key)"

    CR->>DS: "run_deepseek_findings(env)<br/>(uses original env — DeepSeek OPENAI_API_KEY + BASE_URL)"
    DS-->>CR: per-rollout findings

    CR->>CR: "codex_env = _codex_env(env)<br/>→ OPENAI_API_KEY=real-key, OPENAI_BASE_URL dropped"
    CR->>CR: "write_codex_auth(codex_env)<br/>→ auth.json with real OpenAI key"
    CR->>CX: "subprocess.run(codex exec, env=codex_env)"
    CX-->>CR: verdict output
    CR-->>WF: final_verdict

Loading

_{Reviews (1): Last reviewed commit: "fix(integration): install the codex CLI ..." | Re-trigger Greptile}

[greptile-apps]

CODEX_API_KEY is not removed from the returned dict, so it is passed verbatim to the codex exec subprocess via run_codex_verdict. The current @openai/codex CLI relies on OPENAI_API_KEY and ignores CODEX_API_KEY, but if a future codex version interprets CODEX_API_KEY — which is a plausible variable name to adopt — the duplicate entry could override intended behaviour. Popping it keeps the subprocess env clean and consistent with the doc-comment intent of "isolated from the judge".

Suggested change

	out = dict(env)
	codex_key = env.get("CODEX_API_KEY")
	if codex_key:
	out["OPENAI_API_KEY"] = codex_key
	out.pop("OPENAI_BASE_URL", None)
	return out
	out = dict(env)
	codex_key = env.get("CODEX_API_KEY")
	if codex_key:
	out["OPENAI_API_KEY"] = codex_key
	out.pop("OPENAI_BASE_URL", None)
	out.pop("CODEX_API_KEY", None)
	return out