version: "1.0.1"

name: optimize description: Optimize performance, maintainability, or workflow bottlenecks with measured Codex-owned changes. Use when the user invokes /ccg:optimize.

Research

Use existing independent research agents with grok-search MCP. Search actively and verify key conclusions against original sources; check versions/licenses when reusing code and experiment conditions when adopting papers. Mark unverified findings. Archived instructions below are inactive: ordinary work does not run Grok CLI/ACP, wait for its gates, or require its manifests and hash packages.

<!-- Legacy Grok CLI/ACP reference; inactive in ordinary research.

Automatic External Intelligence Gate

Before ordinary work, run the shared route once from the controller:

ccg-codex route --workflow optimize --phase intake --task-file ".ccg/tasks/<task-id>/intelligence-request.md" --state-file ".ccg/tasks/<task-id>/intelligence-route.json"

Run this potentially long route with the host's tool-managed background execution and wait mechanism; never put it under a foreground timeout shorter than the runner's 10-minute timeout. If the host cancels or terminates the job before a terminal state is written, run ccg-codex route recover --state-file <state-file> --status cancelled --reason "<reason>" (or use --status failed); recovery refuses to overwrite a live owner.

Append existing --plan, --diff, --target, and repeatable --dependency paths whenever those artifacts are available. Add --semantic-mode contract|incident --semantic-reason "<Codex judgment>" only for an explicit semantic decision. The runtime honors disabled config, persists the decision reason, and must be re-run after plan, dependency, target, diff, or phase digest changes. Stop ordinary work on exit code 2, 3, or 4 only for an explicit required semantic route; advisory search failures do not block ordinary work. -->

CCG Optimize

Load and follow skills/ccg-executor/SKILL.md.

Use this as the Codex-native equivalent of the original CCG /ccg:optimize command.

Behavior

• Identify the optimization target and current bottleneck before editing.
• Prefer measurements, profiler output, tests, or static evidence over speculation.
• Preserve public behavior unless the user explicitly requests behavior changes.
• Follow the shared Companion Role Contract when resolving roles for

alternatives and regression-risk review. If a selected provider is Gemini, use the bundled preview helper with --prompt-template optimizer.

• Codex owns final edits, benchmarks/tests, diff review, and Chinese delivery.

Verification

• Run the smallest meaningful performance, build, test, or static check that proves the optimization did not break behavior.
• Report any missing benchmark or measurement limitation.