Cross-Harness Review

Cross-Harness Review is a Grok plugin and Node CLI for obtaining a bounded, read-only second opinion from external coding harnesses. Grok remains the decision maker: every reviewer result is untrusted evidence that must be verified locally before any change is made.

v0.2.0 at a glance

• Supported reviewer IDs: claude, codex, and opencode.
• One Node core owns discovery, configuration, routing, bounded execution, and normalized v2 results. PowerShell and POSIX scripts are forwarding shims.
• Prompts are sent on stdin, output is bounded, and review scope is enforced by a host-computed Git snapshot.
• Direct host test execution is opt-in and fails closed unless the user policy, project allowlist, and adapter capabilities all permit it.
• Cursor is deliberately deferred. Its preparatory source remains in the repository but is not discoverable, configurable, assignable, or accepted by the public result schema.

Requirements

• Grok CLI with plugin support.
• Node.js 20 or later.
• Git on PATH.
• At least one authenticated reviewer CLI: Claude Code (claude), OpenAI Codex (codex), or OpenCode (opencode).

Install

npm package

npm install -g @sidler289-code/cross-harness-review
grok plugin install "$(npm root -g)/@sidler289-code/cross-harness-review" --trust
grok plugin enable cross-harness-review

GitHub checkout

grok plugin install https://github.com/sidler289-code/GrokBuild-outsideHarness.git --trust
grok plugin enable cross-harness-review

Verify the installation:

cross-harness-review --version
grok inspect

The CLI should report 0.2.0. Grok should list the two skills cross-harness-review and cross-harness-auto, with no hooks, agents, or active MCP server.

v0.2.0 workflow

Discover installed reviewer CLIs, persist an explicit role mapping, and verify it before an audit:

cross-harness-review detect --json
cross-harness-review setup --plan claude --code codex --tests opencode --json
cross-harness-review roles --json

The example assigns three roles explicitly. You may use one, two, or three of the supported IDs, subject to their capability gates. setup does not enable direct test execution by default. Do not pass --enable-tests until the selected adapter has verified structured events, approved-command restriction, and direct execution.

Run reviews

cross-harness-review audit plan --plan-file docs/plan.md --repo . --json
cross-harness-review audit code --plan-file docs/plan.md --repo . --scope uncommitted --json
cross-harness-review audit security --repo . --scope base:main --json
cross-harness-review audit tests --repo . --json

From Grok, use the explicit slash workflow, for example:

/cross-harness-review code --uncommitted

Natural-language requests may invoke the auto skill when appropriate. Reviewer output never has authority to edit the working tree.

Configuration

| Variable | Purpose | |---|---| | CROSS_HARNESS_CONFIG | Absolute user-config path override | | CROSS_HARNESS_CLAUDE_BIN | Explicit Claude executable | | CROSS_HARNESS_CODEX_BIN | Explicit Codex executable | | CROSS_HARNESS_OPENCODE_BIN | Explicit OpenCode executable |

A broken explicit override fails closed; it does not fall back to another binary. On Windows, shell wrappers (.cmd, .bat, .ps1) remain rejected by the argv-only subprocess runner.

Security boundary

• Claude uses a restricted tool set and empty MCP configuration.
• Codex uses a read-only, ephemeral sandbox and ignores user config and rules.
• Prompts travel through stdin, not argv; output and diagnostics are capped.
• No permission or sandbox bypass flags are used.
• Findings outside the selected Git scope are downgraded to verification: out_of_scope.

See SECURITY.md, the v0.2.0 verification record, and the host smoke record for the evidence boundary and release details.

Local validation

grok plugin validate .
powershell.exe -NoProfile -ExecutionPolicy Bypass -File tests/probe.Tests.ps1
npm.cmd run test:node
npm.cmd pack --dry-run

License

MIT ? see LICENSE.