refactor: rename gpt-5-4-prompting skill to codex-prompting

[tonyyunyang]

Summary

• Rename the gpt-5-4-prompting skill to codex-prompting so the label doesn't go stale as GPT iterates. The skill is really about prompting Codex (the runtime), not a specific GPT generation, and Codex already routes to whatever model the user has configured.
• Drop "GPT-5.4" mentions from the skill body and reference files. Update cross-references in agents/codex-rescue.md and skills/codex-cli-runtime/SKILL.md.
• Replace example gpt-5.4-mini strings in the README and the agent's model-passthrough rule with a <your-model> placeholder, since the actual model selection flows through ~/.codex/config.toml and --model.
• Update assertions in tests/commands.test.mjs to match.

Net-zero diff: 8 files, 21 insertions, 21 deletions. Folder rename detected by git (history preserved).

Test plan

• node --test tests/commands.test.mjs — all 8 assertions pass on the new branch
• No remaining gpt-5-4 / GPT-5.4 / gpt-5.4-mini strings in user-facing .md/.mjs/.json/.toml (verified via grep)
• Pre-existing failures in runtime.test.mjs and state.test.mjs confirmed unrelated — same 4 failures appear on baseline main when this branch's changes are stashed

Notes

• tests/fake-codex-fixture.mjs still has gpt-5.4 as a mock-server fallback. It's internal test infrastructure (not user-facing, not asserted on); left untouched to keep the diff minimal.
• The skill name has to be static — Claude Code resolves it from frontmatter at load time and can't reflect runtime config dynamically. Making it model-agnostic is the closest practical equivalent to "tied to whatever model the user has configured."

[12zuev]

Endorsing the version-agnostic approach over a gpt-5-5-prompting rename.

GPT releases are now on a ~6-week cadence (5.4 → 5.5 was 6 weeks), so a version-pinned skill name becomes a continuous-rename treadmill — anyone running the previous plugin tag sees stale labels until the next release, and downstream users who patch their local cache to match the current model lose the fix on the next claude plugin update.

PR #301 took the gpt-5-5-prompting path but was self-closed before review, so the version-pinned approach hasn't actually been weighed against this one. codex-prompting lands at the right level of abstraction: the skill encodes Codex-runtime prompting patterns, which are stable across GPT generations, while model selection already lives in ~/.codex/config.toml and --model. Net-zero diff with tests green is a low-risk landing for maintainers.

[tonyyunyang]

Endorsing the version-agnostic approach over a gpt-5-5-prompting rename.

GPT releases are now on a ~6-week cadence (5.4 → 5.5 was 6 weeks), so a version-pinned skill name becomes a continuous-rename treadmill — anyone running the previous plugin tag sees stale labels until the next release, and downstream users who patch their local cache to match the current model lose the fix on the next claude plugin update.

PR #301 took the gpt-5-5-prompting path but was self-closed before review, so the version-pinned approach hasn't actually been weighed against this one. codex-prompting lands at the right level of abstraction: the skill encodes Codex-runtime prompting patterns, which are stable across GPT generations, while model selection already lives in ~/.codex/config.toml and --model. Net-zero diff with tests green is a low-risk landing for maintainers.

Thanks for the endorsement.