fix(skills): use command -v instead of which for codex detection

[mvanhorn]

Summary

Replace which codex with command -v codex in the Codex availability gate used by /review, /ship, /office-hours, /plan-ceo-review, /plan-eng-review, /plan-design-review, /plan-devex-review, /design-consultation, /design-review, and /codex itself.

Why this matters

which is not POSIX. Its behavior varies across platforms:

• Debian and Ubuntu ship /usr/bin/which from debianutils, which scans PATH
• Some systems alias which to a bash function that reads the hash table
• macOS ships a small csh-compatible script
• Busybox does not ship which at all

In non-interactive shells spawned by tool runners, PATH hydration and hash state can diverge in ways that cause which codex to miss a binary that command -v codex and type codex resolve correctly. The reporter on #1193 hit this on Ubuntu/Pop!_OS with @openai/codex@0.125.0 installed under nvm:

$ which codex 2>&1
/bin/bash: line 1: codex: command not found

$ command -v codex
/home/btallman/.nvm/versions/node/v24.12.0/bin/codex

When the availability probe misses, /review silently falls back to "Claude adversarial only" without telling the user. The reporter approved a PR that way.

command -v is POSIX-specified and uses the same resolution the shell itself uses to run a command, so a positive result means the binary is runnable.

Changes

Source files edited directly:

• codex/SKILL.md.tmpl line 45 (one callsite in the template used by /codex)
• scripts/resolvers/review.ts lines 269, 428, 557 (three callsites consumed by /review and /ship)
• scripts/resolvers/design.ts lines 13, 515, 691 (three callsites consumed by design/review skills)
• test/skill-validation.test.ts line 1320 (assertion updated to match new literal)
• test/skill-e2e-plan.test.ts line 779 (assertion updated to match new literal)

Regenerated via bun run gen:skill-docs --host all:

• codex/SKILL.md, review/SKILL.md, ship/SKILL.md, office-hours/SKILL.md, plan-ceo-review/SKILL.md, plan-eng-review/SKILL.md, plan-design-review/SKILL.md, plan-devex-review/SKILL.md, design-consultation/SKILL.md, design-review/SKILL.md
• test/fixtures/golden/claude-ship-SKILL.md and test/fixtures/golden/factory-ship-SKILL.md refreshed from regenerated ship output (the codex-ship golden was already byte-identical to the regenerated .agents/ output)

gstack already uses command -v codex in autoplan/SKILL.md.tmpl:250 and test/helpers/providers/gpt.ts:20. This aligns the rest of the codebase with that existing convention.

Not in scope

• pair-agent/SKILL.md.tmpl:131 and pair-agent/SKILL.md:1177 use the same which ngrok 2>/dev/null shape. The reporter's issue is about codex detection, so ngrok is left alone. Happy to open a follow-up PR for it.
• test/fixtures/golden-ship-claude.md (top-level, not under golden/) still contains which codex but is not referenced by any test and was last touched in 2026-04 via a community PR. Treated as an orphan artifact.

Testing

bun run gen:skill-docs --dry-run --host all     # exit 0, all files fresh
bun run skill:check                              # 43/43 skills OK, all hosts fresh
bun test test/host-config.test.ts -t "golden"    # 3/3 golden regression tests pass
bun test test/skill-validation.test.ts -t "codex/SKILL.md"  # 12/12 pass

Three unrelated tests fail on both upstream/main and this branch (no compiled binaries > 2MB, read_secret_to_env > rejects invalid var names, Opus 4.7 overlay > Fan out). Verified by running them on upstream/main HEAD before branching.

Fixes #1193