ask-codex

# Ask Codex Use local `codex exec`. Cloud tasks use `codex-cloud`; specialized review tasks use the corresponding `ask-codex-for-*` skill. If `codex` is not on PATH, use `ask-codex-mcp` instead. ## Task contract Each invocation starts a separate session. Include the goal, necessary context and prior decisions, allowed scope, completion condition, and output format. State whether edits and network access are allowed. For questions and reviews, keep repository files read-only. Allow relevant read-only inspection rather than forbidding the commands needed to read the target. Preserve existing work. Implementation authorization does not grant destructive Git/filesystem operations, commits, pushes, or unrelated dependency changes. Carry forward authorization already established by the user and continue authorized work through verification. Codex already loads installed skills, so refer to a skill by name in the prompt and do not pass its directory with `--add-dir` or reference skill file paths. Every such prompt must tell Codex to stop and report the missing skill by name if it is not available; do not invent its standard or claim a complete review. ## Invocation and access - Set `-C` for repository work and `-s` explicitly: `read-only` for analysis, `workspace-write` for authorized implementation. - Never use `danger-full-access` or `--dangerously-bypass-approvals-and-sandbox`. Ask only if the task requires access beyond the authorized boundary. - Use narrow `--add-dir` paths when necessary. Use `--skip-git-repo-check` only outside a Git repository. - Prefer `--ephemeral`; capture the final response in a unique `mktemp` output file. Check exit status and non-empty output before reporting success; clean up the file afterward. - Keep credential stores, private keys, secret files, and unrelated home-directory data outside scope unless the task already authorizes their use. Never put secrets in prompts or captured output. Ask only for sensitive access not already authorized. - Default review/Q&A to no network. Permit network for requested research or necessary documentation/package lookup and state that policy in the prompt. `workspace-write` limits writable locations, not the damage possible inside them. Preserve the task's action constraints in the delegated prompt. ## Model Use `gpt-5.6-sol` with medium effort for ordinary work and high for reviews. Honor a user-selected model or effort. Set effort explicitly. ```bash OUT="$(mktemp -t codex-last-message.XXXXXX)" if codex exec -m gpt-5.6-sol -c model_reasoning_effort="medium" \ -C /path/to/repo -s read-only --ephemeral -o "$OUT" \ "<self-contained task, scope, action/network boundaries, and completion condition>" then if [ -s "$OUT" ]; then cat "$OUT"; else echo "Codex returned no final output" >&2; fi else echo "Codex execution failed" >&2 fi rm -f "$OUT" ``` ## References - Execution failure or timeout: [fallback.md](references/fallback.md). - Continuing an existing session: [resuming.md](references/resuming.md). - Task-specific invocation examples: [examples.md](references/examples.md).