Codex

Codex — the GPT-5.6 Execution Lane

Run a benchmark-selected GPT-5.6 task through the Codex CLI (codex exec) and return the result. This is the OpenAI execution lane — the general-purpose lane for work the model-selection policy sends to GPT-5.6, and the canonical owner of general codex exec mechanics — when the CLI changes, update here first. GPT selections are reachable only through this CLI; the Agent tool's model parameter covers Claude models only.

Under Claude Code, this skill runs only on explicit invocation or cross-provider escalation, never as the automatic default. The harness-native model lane under Claude Code is the Anthropic lane (Opus 5). This skill is a deliberate cross-provider tool — codex review as a second-opinion, codex exec for a GPT-specific constraint — not a routing default.

Two flows keep their own specialized codex integration — route to them instead of re-implementing here:

Phase 1: DECIDE — does this task belong on GPT-5.6?

Policy mirror — canonical copy: /do SKILL.md, Model Selection (edit there first, then here). Rankings, higher = better; cost = avg USD per task, written as a plain number (slash-command templating corrupts dollar-digit sequences in injected skill bodies), what the owner actually pays.

Run deterministic work as scripts, not through Codex. The /do model policy selects the lane and passes model plus effort. Legacy GPT-5.5, all Luna choices, and the other non-default GPT-5.6 settings are manual-only; do not substitute them automatically. Luna max, for example, saves 0.44 USD versus Sol high but consumes 45k more output tokens and 65 more steps for two fewer Pass@1 points. Consult the canonical table in /do SKILL.md.

These are defaults, not limits. Standing permission to escalate when output misses the bar applies within the policy; max still needs an explicit override. For anything that ships, intelligence > taste > cost; cost is a tie-breaker only.

Gate: task has a GPT-5.6 policy selection. Otherwise route to scripts or the policy's Claude pick and stop here.

Phase 2: WRAP — how GPT-5.6 runs from this harness

Wrapper symmetry: the wrapper is needed for whichever model family is NOT the current harness.

• Under Claude Code (current default): GPT-5.6 runs through a wrapper — either the dispatched agent runs codex exec via Bash with a self-contained prompt, or a thin Claude wrapper agent (model: "sonnet", low effort) writes the self-contained codex prompt, runs it, and returns the result.
• Under the Codex harness: Claude models require the wrapper instead.
• Claude models under Claude Code need no wrapper — just the Agent/Workflow model parameter.

Pick the direct-Bash form when the calling agent already holds the task context; pick the thin wrapper agent for fan-out (one wrapper per data source) so the orchestrator stays lean.

Availability check first: command -v codex — when absent, fall back to the policy's Claude pick (model: "sonnet" for mechanical work) and tell the user in one line which lane ran.

Phase 3: PROMPT — write a self-contained prompt

Codex runs in its own process with no conversation history. The prompt must carry everything:

1. Context — one short paragraph: what the repo/data is, what state matters.
2. Task — the concrete operation, with file paths relative to the working directory. Let codex read files itself; embedding large content wastes tokens and loses formatting.
3. Output format — the exact structure to return (table, JSON, diff), so the wrapper can consume it without a second pass.

Prompt hygiene (hard rule): codex prompts leave the machine. Send only public content — secrets, credentials, and private component names (anything from the installed index in ~/.claude/vexjoy/index/, overlay skills, or other local-only inventories) stay out. Run the deterministic scan on the prompt text before executing:

printf '%s' "$PROMPT" | rg -n "Bearer|Authorization|token|secret|api[_-]?key|password|PRIVATE KEY" && echo "HYGIENE VIOLATION"

On a hit or a private component name: scrub the flagged content when the task survives without it; otherwise reroute the task to a Claude model. A bare refusal is not an outcome.