Parent epic: #273
Decision
Codex evaluation has been moved out of glitchwerks/github-actions to a new scratch repo:
https://github.com/glitchwerks/github-actions-codex (private)
Rationale
glitchwerks/github-actions has accumulated three phases of Claude-specific scaffolding — the 696-line review prompt in pr-review/action.yml, the runtime/ overlay tree, the digest-pinned claude-*.yml reusable workflows, the spec/plan corpus under docs/superpowers/. Evaluating Codex inside that scaffolding biases the answer: any Codex test either ports Claude conventions onto Codex or fights them, neither of which reveals whether Codex is good on its own terms.
The scratch repo is an unbiased evaluation surface — Codex App installed, no pre-existing assumptions about review shape, no Claude prompts to translate.
Scope of the scratch repo
Per session decision: scratch sandbox, not a production-candidate fork. Goal of exploration is to answer "do I want to commit to Codex as the next platform?" — not to port github-actions's surface.
If exploration concludes positively → return here, restart sub-issues #A–#Q with Codex-native authoring informed by the exploration.
If exploration concludes negatively → reopen spec PR's decision (#278's dual-surface choice) with alternative fallback (stay on Anthropic via API key, alternative vendor, drop AI review).
Hard deadline still real
Anthropic OAuth EOL ~2026-06-20. Exploration must produce a decision well before that date, or a fallback plan must be in place by then. Tracked separately (no specific sub-issue yet).
State of existing codex-pivot work
| Item |
State |
Action |
| Spec PR #278 (merged) |
Lives on main as the locked design from the original decision |
Leave; will revisit if exploration outcome diverges from spec |
| Sub-issues #A–#Q (#279–#295) |
All open |
Deferral comments added; left open as a record of the planned phase-1 surface |
| Draft PR #296 (canonical AGENTS.md) |
Open, draft |
Closed as superseded — AGENTS.md will be authored in the scratch repo if Codex proves out |
pr-review/action.yml and Claude workflows |
Live in production |
Continue running unchanged until OAuth deadline forces a commit |
What this issue tracks
The pivot itself plus the exploration's outcome. Close when exploration concludes (positive or negative) with a comment summarizing the finding and the next direction.
🤖 Generated by Claude Code on behalf of @cbeaulieu-gt
Parent epic: #273
Decision
Codex evaluation has been moved out of
glitchwerks/github-actionsto a new scratch repo:https://github.com/glitchwerks/github-actions-codex (private)
Rationale
glitchwerks/github-actionshas accumulated three phases of Claude-specific scaffolding — the 696-line review prompt inpr-review/action.yml, theruntime/overlay tree, the digest-pinnedclaude-*.ymlreusable workflows, the spec/plan corpus underdocs/superpowers/. Evaluating Codex inside that scaffolding biases the answer: any Codex test either ports Claude conventions onto Codex or fights them, neither of which reveals whether Codex is good on its own terms.The scratch repo is an unbiased evaluation surface — Codex App installed, no pre-existing assumptions about review shape, no Claude prompts to translate.
Scope of the scratch repo
Per session decision: scratch sandbox, not a production-candidate fork. Goal of exploration is to answer "do I want to commit to Codex as the next platform?" — not to port
github-actions's surface.If exploration concludes positively → return here, restart sub-issues #A–#Q with Codex-native authoring informed by the exploration.
If exploration concludes negatively → reopen spec PR's decision (#278's dual-surface choice) with alternative fallback (stay on Anthropic via API key, alternative vendor, drop AI review).
Hard deadline still real
Anthropic OAuth EOL ~2026-06-20. Exploration must produce a decision well before that date, or a fallback plan must be in place by then. Tracked separately (no specific sub-issue yet).
State of existing codex-pivot work
mainas the locked design from the original decisionpr-review/action.ymland Claude workflowsWhat this issue tracks
The pivot itself plus the exploration's outcome. Close when exploration concludes (positive or negative) with a comment summarizing the finding and the next direction.
🤖 Generated by Claude Code on behalf of @cbeaulieu-gt