An open letter from a power user: Claude Code is losing the terminal to Codex

This is not a bug report. This is feedback from someone who doesn't want to leave.

I hold two Max subscriptions. I have been using Claude Code since its early days -- not casually, but as the backbone of a multi-workspace development operation spanning 14 domain hubs, 100+ skills, custom hooks, orchestrators, and FABRICATE pipelines. I am not a weekend hobbyist. Claude Code is my production environment.

I am writing this because over the last two weeks, I have been gradually shifting critical work to OpenAI Codex CLI with GPT-5.4. Not because I wanted to. Because Claude Code forced my hand.

---

What changed

1. The v2.1.78 permission regression is still unresolved.

A security hardening change introduced hardcoded protected-directory logic that overrides defaultMode: "dontAsk", broad allow-list wildcards (Bash(*), Edit(*), Write(*)), skipDangerousModePermissionPrompt: true, AND PreToolUse hook decisions returning permissionDecision: "allow". I had a carefully layered security config at four settings levels -- user, user-local, project, project-local -- all aligned. The binary ignores all of it.

I reported this (#37496, closed as duplicate), commented on #36168 with confirmation, tested the community-suggested hook workaround, and confirmed it doesn't work. The hook fires, logs the auto-approve, but the binary-level check runs as a separate gate afterward and prompts anyway.

There has been zero acknowledgment from Anthropic staff on any of the 30+ related issues. Not a "we're looking into it." Not a "this is intentional." Nothing.

2. Delivery completeness has declined.

Claude Code increasingly fails to complete tasks end-to-end. It starts strong, then trails off -- leaving partial implementations, skipping verification, or losing context mid-task. I've built autonomous TDD loops, self-healing retry patterns, and multi-agent verification pipelines specifically to compensate for this. That's not a workflow -- that's a workaround for an unreliable tool.

3. Instruction adherence has regressed.

Skills get ignored. Explicit CLAUDE.md directives get skipped. Structured output formats break mid-response. This aligns with issue #36093 ("severe instruction adherence and reasoning regression since ~March 16-17"), which also has no Anthropic response.

---

What Codex is doing right

I switched reluctantly. Here's what I found:

• Permission model works. Configure it once, it respects it. No binary overrides, no silent downgrades, no mid-session mode changes.
• Consistent delivery. GPT-5.4 completes tasks. Not partially. Not with trailing "I'll leave the rest for you." Completely.
• No productivity tax. I don't need to build elaborate hook systems to compensate for the tool fighting me.

---

What I'm asking for

1. Acknowledge the permission regression publicly. 30+ issues, zero staff response. That's not acceptable for a tool people pay $200/month for.
2. Fix it or revert it. If the protected-directory hardening can't respect user-configured bypass settings, roll it back to v2.1.77 behavior and ship a proper implementation later.
3. Prioritize terminal-first users. Channels, Remote Control, VS Code integrations, desktop apps -- these are commercially smart. But the terminal is where your power users live. If claude -p pipelines break because permissions can't be bypassed even with --dangerously-skip-permissions, you're losing the people who evangelize your product.
4. Quality over velocity. Seven releases in ten days. The permission regression shipped in v2.1.78 and persists through v2.1.81 -- four releases later. Slow down. Test more. Ship less.

---

The competitive reality

"OpenAI is shipping everything. Anthropic is perfecting one thing."

That was true six months ago. Today, OpenAI is perfecting -- and Anthropic is slipping. Codex CLI with GPT-5.4 delivers more consistently, respects its configuration, and doesn't fight its users. I don't say this to be inflammatory. I say it because I want Claude Code to be better, and right now it isn't.

I'm not leaving yet. But I'm using Codex more every day. And I suspect I'm not the only one.

---

Two Max subscriptions. 14 domain hubs. 100+ custom skills. Daily user since early access. Writing this because I care, not because I've given up.