The codex-offered-design-review test was failing with error_api because
reading the full plan-design-review/SKILL.md (1331 lines, 77KB) bloated
the agent context to 142k tokens, exceeding API limits. All 3 retry
attempts failed consistently.
Fix: extract only the codex/outside-voice section (~180 lines) instead
of copying the full file. Follows the CLAUDE.md rule: "NEVER copy a
full SKILL.md file into an E2E test fixture." Applied to all 4 skills
in the test suite for consistency.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Garry Tan
30e6747814