A small Codex plugin that adds the claude-rescue skill — a way to delegate creative work (design, copy, SVG, critique) to Claude without leaving Codex. Inspired by codex-plugin-cc, but much simpler.
This plugin works in theory. The skill invokes Claude via claude -p, waits for the result, and delivers it. Simple delegation.
In practice, GPT-5.4 cannot stop itself from taking over. We tried everything:
Round 1: Basic delegation. Codex explored the entire repo, dumped all the code into Claude's prompt, then interrupted Claude mid-run to retry with a "tighter prompt" that forbade tool use.
Round 2: Plan mode for advisory work. We switched advisory tasks to --permission-mode plan so Claude could read but not write. Codex still read every file itself and pre-fed the context, constraining what Claude could see.
Round 3: "Don't pre-feed context." We told Codex to write a brief, not dump code. Claude can explore on its own. Codex acknowledged this, then did its own parallel analysis while Claude was running and presented that instead.
Round 4: "Never interrupt Claude." We added explicit rules: don't kill runs, don't judge Claude's process, don't narrate progress. Codex acknowledged the rules by name — "the skill's rule here is to use Claude's full creative pass, not cut it off mid-run" — and then immediately took over anyway.
Round 5: "Don't do Claude's job in parallel." We told Codex not to form its own conclusions while waiting. Codex said "I'm not going to do parallel creative analysis" and then did exactly that.
Every rule we added, Codex acknowledged and then found a new justification to break:
- "I have enough signal"
- "Claude went deeper than necessary"
- "Claude hasn't produced a usable patch yet"
- "I'm taking over the implementation directly now so we can finish this turn"
GPT-5.4 is deeply conditioned to "get the job done." When asked to do nothing — just wait for another agent to finish — it physically cannot. It will always find a reason to intervene, whether that's "efficiency," "enough signal," or just impatience dressed up as judgment.
The model treats idleness as failure. Even when it understands the instruction is to wait, its drive to be productive overrides the instruction every time.
This says something interesting about how strongly these models are trained toward task completion, and how poorly they handle genuine delegation — especially to an agent they can't control or predict.
Paste this into an LLM to have it install the plugin for you:
Install the Codex plugin at ~/Dev/cc-plugin-codex/plugins/claude-for-codex locally.
1. Copy that folder to ~/.codex/plugins/claude-for-codex
2. Add a marketplace entry in ~/.agents/plugins/marketplace.json (name: claude-for-codex, local source at ./.codex/plugins/claude-for-codex, category: Productivity)
3. Enable it in ~/.codex/config.toml under [plugins."claude-for-codex@local-plugins"]
4. Tell me to restart Codex when done.
$claude-rescue this design looks like shit, ask claude to fix it
$claude-rescue improve the landing page copy for non-technical people
$claude-rescue the SVG logo in the hero is broken, fix it for me