codex:rescue skill init consumes ~20% of session usage budget before actual work

Bug Report

Plugin Version

• @openai/codex-plugin-cc v1.0.1

Environment

• Claude Code CLI (Opus 4.6, 1M context)
• macOS (Darwin 25.3.0, Apple Silicon)

Description

When a code-reviewer agent calls Skill("codex:rescue") during a multi-pass code review, the initialization phase alone consumes approximately 20% of the Claude Code session usage budget before any actual review work begins.

In our case, the Codex call then hit an API rate limit — resulting in zero review output for ~20% usage cost.

Steps to Reproduce

1. Start a Claude Code session with Opus model
2. Spawn a code-reviewer agent within a team
3. Code-reviewer calls Skill("codex:rescue") with a list of 6 changed files and review instructions
4. Observe usage consumption during the init/setup phase

Expected Behavior

Plugin initialization should be lightweight. The bulk of usage should come from the actual review task execution, not the setup phase.

Actual Behavior

• Init phase consumed ~20% of session usage budget
• After init, the actual Codex API call hit a rate limit
• Net result: 20% usage burned, zero review output
• The plugin appears to load excessive context or make redundant API calls during setup

Impact

This makes codex:rescue impractical for routine code reviews — the init cost alone is prohibitive, especially when the actual review may fail due to rate limits after that cost is already paid.

Workaround

Skipping Skill("codex:rescue") entirely and relying on self-review or spawning a separate review agent instead.

Begin at the Skill("codex:rescue") initialization and setup path used by the code-reviewer agent, and trace what context loading or API calls occur before the review request. Reproduce with six changed files and the stated review instructions; done means initialization is lightweight and the review does not spend the reported budget on setup or redundant calls before execution.