Take #784's skill commit only: gpt-6-prompting, corrected for this fork

[Edo771977]

Upstream #784 shows as 93 files and +11799/−2511, but that is not the change: its branch sits 42 commits ahead of upstream main (release 1.6.1, GPT-5.6/GPT-6 model support, broker and reconciliation work, a vitest migration). The PR's own commit is the last one — the skill swap, 9 files — and that is all this takes, cherry-picked with its author intact.

355/355, tsc clean.

The import

skills/gpt-5-4-prompting/ is replaced by skills/gpt-6-prompting/: model selection for Luna / Sol / Astra, brief blocks (autonomy, repo policy, verification states, progress updates, output contracts) and per-model recipes. The subagent and codex-cli-runtime point at the new name, and the commit updates the tests that assert those names itself.

Its own CHANGELOG entries were not taken: they describe model aliases and runtime fixes from the other 41 commits, which this does not import.

Three claims that were not true here

The skill is instructions the rescue subagent follows literally, so a wrong statement in it produces wrong commands.

in the imported text	here
recipes launch with --model luna, --model sol, --model astra	MODEL_ALIASES holds exactly one entry, spark, and normalizeRequestedModel() passes anything else through untouched — each of those would have reached Codex as the literal word and failed. They now use the full slugs from the skill's own table, which need no alias table
"The Codex plugin caps every model at high … Do not pass xhigh"	VALID_REASONING_EFFORTS includes xhigh. The advice to stay at or below high is good and stays, as advice with the cost tradeoff named, not as a policy this fork enforces
Astra above medium "the plugin rejects it"	nothing here is per-model; reworded to say why it is wasteful instead

A test now reads MODEL_ALIASES and VALID_REASONING_EFFORTS out of codex-companion.mjs and checks every --model/--effort the skill names against them, so a later sync with the upstream skill cannot quietly reintroduce a model this fork would forward verbatim or an effort it would reject.

sabotage	the guard says
a launch line back to --model sol	"launches with --model sol, which is neither an alias of this fork nor a full slug"
an --effort max	"names an effort this fork rejects: max"

Worth knowing

The skill's canonical source is a personal repo (andrewmaspero/codex-skills) with a "keep in sync" note, so this content lives elsewhere and will drift. The guard above is what makes the next sync safe to do quickly.

Its usefulness also depends on gpt-6-* being selectable by the installed Codex CLI. Nothing breaks if it is not — the guidance simply goes unused, and spark keeps working — but that is the thing to check before leaning on the recipes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UXfvnjSC72HsM6EEPVt2Tg

---

Generated by Claude Code