| name | codex-spark-delegation |
| description | Delegate coding tasks to Codex (gpt-5.3-codex-spark) via `codex exec` while acting as PM: write work-order prompts, dispatch, verify results from the working tree, and iterate with `codex exec resume`. |
Codex Spark Delegation
Act as the PM: plan and review in this session, and delegate the actual
code editing to Codex's fast implementation model (gpt-5.3-codex-spark)
via codex exec. Unlike cursor-composer-delegation (UI automation,
fire-and-poll), codex exec is a synchronous CLI: it blocks until the
task finishes, returns the agent's final message on stdout, and edits the
working tree directly. No IDE, no Accessibility API, no polling.
Use codex-exec-review for reviews (read-only). This skill is for
implementation (writes).
Dispatch Template
command codex exec \
-m gpt-5.3-codex-spark \
--sandbox workspace-write \
-C /path/to/repo \
-o /tmp/codex-spark-last-message.txt \
"<work-order prompt>"
-m gpt-5.3-codex-spark — the fast implementation model. Do NOT
omit it; the config default is a different (slower, pricier) model.--sandbox workspace-write — real containment, verified empirically:
under --sandbox read-only a file-write attempt failed with
Operation not permitted, and the run header showed
sandbox: read-only.- CAVEAT: this user's shell has
alias codex='codex --yolo'
(common.sh). That alias silently overrides --sandbox/approval
flags — a run through the alias showed sandbox: danger-full-access
even with --sandbox read-only passed explicitly. Always invoke as
(bypasses the alias) when running this
skill's commands from an interactive shell, and check the run
header's line to confirm the mode actually took effect.
in is unrelated to sandbox
enforcement — it only gates loading project-local config.
