← ClaudeAtlas

cdx-sollisted

Use when the user invokes /cdx-sol or wants to send a task to GPT-5.6-SOL (OpenAI's Codex model) from Claude Code stably and token-efficiently. Wraps codex-companion via cdx-sol.mjs — background+poll for 120s-timeout safety, read-only by default, effort-tier + terse-output token control. Runs on the ChatGPT subscription (no per-token API billing).
dT-Tb-labs/babel-orchestration · ★ 0 · AI & Automation · score 72
Install: claude install-skill dT-Tb-labs/babel-orchestration
# cdx-sol — stable, token-efficient GPT-5.6-SOL access Send a task to GPT-5.6-SOL through the local Codex subscription. Use for second-opinion review, independent diagnosis, or rescue work where a non-Claude model helps. ## Invoke ```bash node "$HOME/.claude/skills/cdx-sol/cdx-sol.mjs" --tier <quick|normal|deep> --cwd "<repo-abs-path>" "<prompt>" ``` Use Bash tool `timeout: 600000` (10 min). The wrapper launches a background Codex job, polls internally, and prints only SOL's final answer — one round-trip, no progress spam. Read-only sandbox by default. ## Token discipline (why this skill exists) Three levers, all applied by default: 1. **Prompt (IN):** write the `<prompt>` caveman-compressed — drop filler, keep technical substance. Always include: the `--cwd` repo path, the exact target files, and the success criteria. Under-context makes SOL ask back = a wasted round-trip, so compress filler, never substance. 2. **Reasoning (the big lever = subscription burn):** pick the tier. - `quick` (effort low) — lookups, small focused checks. - `normal` (effort medium, default) — most reviews / diagnosis. - `deep` (effort high) — hard multi-file reasoning only. Costs the most subscription quota and time (a high-effort run can take 5–6 min). 3. **Output (OUT):** the wrapper appends a terse directive to every prompt, and offloads any output over ~24k chars to `<cwd>/.sol/` (returning a head + file path) so a big dump never floods context. ## Safety - **Read-only by defa