Codex delegates tasks to Claude Code through a bundled Node runner, then checks the results. The skills are instructions for the calling agent: choose controls, run Claude in the right project, inspect artifacts, verify the work, and resume when needed.
The inverse of cdx, which lets Claude Code drive Codex.
| Skill | Use |
|---|---|
driving-claude |
Invocation, controls, results, failure routing |
claude-task |
One delegated task |
claude-review |
Findings against a specific diff |
claude-fleet |
Independent workers and synthesis |
claude-session |
Resume, fork, and restore launch configuration |
claude-setup |
Diagnose installation, auth, and runner failures |
prompting-claude |
Compose prompts and select model-specific guidance |
claude-structured-output |
Consume schema results and recover failed reports |
Use $claude-review in Codex or describe the task, such as “get Claude's opinion on this diff.” The model reference covers Opus 5.5, Fable 5.1, Sonnet 5, and direct API migration. Its recommendations are dated and distinct from measured results.
- Authenticated Claude Code on PATH. This refresh inspected version 2.1.280.
- Node.js for the runner and tests.
- Codex with plugin support.
Install from the Nuko Nova marketplace:
codex plugin marketplace add nuko-nova-dynamics/marketplace
codex plugin add cld@nuko-nova-toolsFor an existing installation, refresh the catalog first with codex plugin marketplace upgrade nuko-nova-tools, then run the add command. Open a new Codex task after updating so its skills reload.
node /path/to/cld/scripts/claude-run.mjs --sandbox ro \
--cd /path/to/target-repo \
--schema /path/to/cld/schemas/review-findings.schema.json \
-- "Review the current diff for actionable defects. Do not edit. Cite source evidence."The runner prints status, session ID, model usage, token/cost estimates, permission denials, and a bounded final-message excerpt. Each run saves raw stdout/stderr, its terminal result, final message, and a run.json lifecycle record. The record retains safe launch context; it does not copy prompts, raw flags, inline settings, or environment values. Raw provider output may still contain sensitive content. Exit zero requires a successful terminal result; --schema also requires Claude Code's structured_output. The runner relies on Claude Code for schema validation and does not treat JSON recovered from prose as validated.
| Preset | Behavior |
|---|---|
ro |
Default permission mode, direct file-edit tools denied, selected read/shell operations pre-approved |
write |
File edits accepted and shell pre-approved |
full |
Claude permission bypass, requiring explicit user authorization |
The historical --sandbox name describes permission presets, not filesystem isolation. Inherited configuration, hooks, shell operations, MCP servers, extra allowances, and raw flags affect what can run. See the flag map before extending permissions.
Use absolute runner/schema paths and an explicit target --cd. On resume, restore required launch flags. Output is written to artifact files as it arrives; the console prints its final summary on exit. Parsing is capped at 16 MiB of stdout: larger output fails explicitly while the raw file remains available. SIGINT and SIGTERM are forwarded to Claude, with forced shutdown after five seconds. Interrupted runs exit nonzero and save available artifacts. POSIX cancellation targets the owned process group; Windows cancellation targets the direct child. --scratch now selects a parent folder: every call creates a unique cld-run-* child. Use the printed artifact paths instead of assuming files sit directly under the supplied parent. This prevents concurrent workers from replacing each other's results.
node --test tests/runner.test.mjs
node tests/smoke.mjsOffline tests use real fake-CLI processes and cover terminal results, schema presence, permissions, cancellation, concurrent artifact isolation, incremental files, parse limits, and secret exclusion from run metadata. GitHub Actions runs them on Linux and macOS. Live smoke tests consume account usage, run plain/schema/verbose probes, and preserve their artifacts. Each probe sets a $0.50 Claude Code budget and a five-turn limit; those controls are not billing guarantees. To probe a specific model:
CLD_SMOKE_MODEL=claude-opus-5-5 node tests/smoke.mjsThe run ownership decision describes the execution module and its limits.
The September research record lists sources, decisions, and validation limits.
Apache-2.0