Improve Codex autoreview workflow

[DanShaders]

No description provided.

[github-actions]

Findings

• .github/workflows/codex-review.yml:51 – The workflow now references openai/codex-action@v1 instead of the previously pinned SHA. Because this job runs in a pull_request_target context with repo write permissions and secrets, using a floating tag reintroduces a supply-chain risk: the action owner could retag v1, or the repo could be compromised, and the attacker would immediately execute arbitrary code with your token. Please re-pin the exact commit SHA to keep the workflow safe.
• .github/workflows/codex-review.yml:60 – The conditional prompt line uses ${{ condition && 'text' }} without a fallback. In GitHub expressions false && '...' yields the literal boolean false, so for every PR not authored by DanShaders the prompt will contain a stray false line, which degrades the instruction quality sent to Codex. Use a construct that evaluates to an empty string when the condition is false (condition && 'text' || '') or wrap it in an if block.

[github-actions]

Findings

• High – .github/workflows/codex-review.yml:38-52 now checks out refs/pull/.../merge under pull_request_target, so the runner pulls untrusted PR code while secrets (OpenAI key, PR write token) are available and persist-credentials defaults to true. The prior flow deliberately stayed on the trusted base commit to avoid executing PR content. This change reintroduces a secret-exfiltration/RCE vector if a malicious or compromised allowed account opens a PR. Please revert to checking out the trusted base ref (or disable credentials) when handling untrusted PR code.