feat(launch): let a codex wrapper sit behind cue's profile picker

[NagyVikt]

Problem

A wrapper like oh-my-codex could only be reached by shadowing codex in the shell — a fish function, an alias. That wins the name outright, so cue launch codex never ran and the profile picker never appeared. claude got a picker; codex got nothing, with the wrapper silently passing whatever cue current had already resolved.

Change

$CUE_REAL_CODEX tells cue launch codex to exec the wrapper in place of the codex binary, so cue resolves the profile and materializes the runtime first and the wrapper inherits both. Unlike $CUE_REAL_CLAUDE, which the launch path deliberately ignores, this one is honored there — the point is dispatch, not discovery.

The override is held to the same bar as a PATH candidate (real, executable file), so a stale value falls back to the PATH walk instead of failing as an unexplained exit 127.

The loop this had to avoid

Wrappers spawn codex by bare name. With cue's shim dir still on PATH that spawn lands right back in cue launch — cue → wrapper → cue → wrapper, something MAX_LAUNCH_DEPTH can only abort with an error, never fix. stripShimDirFromPath drops the dir for the child's process tree and the override is unset alongside it.

Trade-off worth knowing: this takes cue's claude shim out of that process tree too.

Bundled fix

CUE_SMART_SUBSET folded the entire passthrough argv into the skill classifier, flags included. codex --madmax measured as:

smart-subset: 4/21 skills kept — "--madmax" likely refers to a cue profile

That gutted the profile the user had just picked. passthroughPrompt folds only the prose. Latent for claude --resume too, so this changes behavior for all flag-bearing launches, not just codex. Both remaining imprecisions (a flag's value folds in; a prompt opening with - drops out) fail toward loading more skills, never fewer.

Verification

All three exec sites (CUE_BYPASS fast path, help passthrough, full launch), against a real wrapper:

WRAPPER-ARGV: --madmax
WRAPPER-CUE_PROFILE: core
WRAPPER-CODEX_HOME: ~/.config/cue/runtime/core/codex
WRAPPER-CUE_REAL_CODEX: <unset>          ← no re-dispatch
WRAPPER-shim-on-PATH: 0                  ← loop broken
WRAPPER-bare-codex-resolves-to: ~/.local/bin/codex   ← real binary

Subset fix both directions: --madmax alone → 23 skills (full profile); -p "fix the medusa checkout bug" → still classifies, now on the prose.

Check	Result
Typecheck	Pass
Lint	Pass (same 6 warnings as base, none in touched files)
Tests	3177 pass / 10 fail — failing set byte-identical to origin/main, measured in a separate base worktree
New tests	+17

The 10 failures are MCP-catalog artifacts of the test worktree and are present on an untouched origin/main checkout too.

Review

Self-reviewed: 0 CRITICAL, 0 HIGH, 2 MEDIUM (both fixed before this PR — dead findRealCodexBin removed, exec guard added), 4 LOW documented in code comments.

🤖 Generated with Claude Code