[Done] New features

[YiweiShen]

/codex --fetch https://github.com/openai/codex

--no-pr Are there any new features we can work on?

[github-actions]

Here are concrete, high‑impact feature ideas tailored to this Codez Action (Codex-on-GitHub-Actions), grouped for clarity. I’ve prioritized items you can ship incrementally.

Safety & Control

• Label/approval gate: require a maintainer label (e.g., ai-approved) before applying edits or opening PRs.
• Dry‑run mode: analyze and propose a patch as a comment without touching the repo.
• Directory/file allowlist: restrict edits to specified globs (e.g., src/, docs/).
• Concurrency guard: auto-cancel older in‑flight runs on the same issue/PR to avoid racing updates.

User Experience

• Review mode: instead of applying patches, post GitHub review comments on the PR (inline suggestions).
• Custom PR metadata: inputs to set PR labels, assignees, reviewers, and draft status.
• Better progress UX: optional live log thread, and a final “summary with links” block (diff, PR, artifacts).
• Comment templates: configurable output templates for results/failures.

Capabilities & Context

• Tests integration: optional run tests and include failing test output in context (beyond fix-build logs).
• Repo search/context: include matches from GitHub Code Search or ripgrep for keywords referenced in the prompt.
• MCP support passthrough: inputs to inject MCP server config and materialize ~/.codex/config.toml before invoking Codex.
• URL fetch policy: per-domain allowlist/denylist and max-size limits when using --fetch.

Providers & Models

• Azure/OpenRouter compatibility: inputs for API variant, version, model mapping; robust error hints for misconfig.
• Model routing: choose model by task type (e.g., use cheaper model for planning, stronger for patching).

Performance & Reliability

• Clone caching: use actions/cache keyed by repo+ref to speed repeated runs.
• Retry & backoff: standardized retry for GitHub/OpenAI calls, with rate-limit handling.
• Time-sliced execution: optional phased runs (plan → apply → test) with separate timeouts to avoid hitting the global cap.

Security & Privacy

• Expanded masking: redact common secret patterns (AWS keys, tokens) beyond just the provided tokens.
• Allowlist actors: optional explicit list of users allowed to trigger, even if they have write (defense-in-depth).

Automation & Planning

• Plan-and-execute mode: first produce a checklist, then apply step-by-step with comment updates; halt on user feedback.
• Issue planner: enhance --create-issues by supporting milestones, labels, and cross-linking to the parent thread.

Quick wins we can implement next

• Dry‑run mode (no PR): add a --dry-run flag that posts unified diffs and commit message preview only.
• Label/approval gate: new input approve-label; if missing, only comment results.
• Custom PR metadata: inputs pr-labels, pr-assignees, pr-reviewers, pr-draft to shape PRs.

Want me to start with dry‑run mode + approval gate and PR metadata? I can implement those behind inputs, update docs, and wire into runAction promptly.