grok-plugin-cc - Ben Bishop

A Claude Code plugin that delegates code review, adversarial review, and coding rescue tasks to Grok through xAI's official Grok Build CLI.

grok-plugin-cc command list covering setup, review, adversarial review, rescue, and background job commands, plus sample grok CLI JSON output

Problem

How do you bring Claude Code review and rescue workflows to a hosted frontier model without reimplementing an agent runtime, while keeping billing, safety gates, and setup explicit?

Approach

The plugin is a thin broker between Claude Code commands and xAI’s official grok CLI in headless mode (grok -p ... --output-format json). Setup stores only the selected model id; authentication stays with grok login. The earlier attempt to drive Grok through Codex custom provider overrides was abandoned: Codex 0.142.0 requires wire_api=responses, and xAI rejected Codex’s agent tool declaration at /v1/responses. The current harness uses Grok Build directly.

/grok:setup lists CLI models, configures a default, and smoke-tests the headless path. /grok:review and /grok:adversarial-review run with a read-oriented sandbox, explicit write/edit deny rules, structured JSON findings, and schema validation. /grok:rescue is confirmation-gated and billed, then followed by post-hoc diff-safety checks.

Background job commands (/grok:status, /grok:result, and /grok:cancel) reuse the same job store pattern as the local-model sibling plugin.

Outcome

A hosted-Grok counterpart to local-model-plugin-cc that keeps Claude Code as the command surface, Grok Build as the execution engine, and safety/billing caveats visible in the workflow.

Architecture

Claude Code commands call broker scripts in the plugin package, which spawn the grok CLI with mode-specific sandbox, schema, deny-rule, and max-turns flags. Automated tests use Node’s built-in test runner and a fake grok binary so CI never needs a live xAI account.