Summary
Since the GPT-5.6 rollout, codex login + a cache refresh makes gpt-5.6-sol the default Codex model (family: Sol/Terra/Luna). But the codex-rescue subagent — the only place in the plugin that reshapes a user request into a Codex-bound prompt — still points at a gpt-5-4-prompting skill. GPT-5.6's official guidance is materially different from the 5.4 era (leaner, outcome-first contracts; 5.6 is more concise by default), so prompts get shaped in the wrong direction for the model that actually runs them.
Ref: https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6
Where (plugin v1.0.5)
1. Stale prompting skill (primary)
agents/codex-rescue.md:
- frontmatter:
skills: [codex-cli-runtime, gpt-5-4-prompting]
- body: "You may use the
gpt-5-4-prompting skill only to tighten the user's request into a better Codex prompt before forwarding it."
skills/ ships only gpt-5-4-prompting — there is no gpt-5-6/gpt-5-5 variant.
2. Effort hint omits current tiers (secondary, cosmetic)
commands/rescue.md (arg hint):
--effort <none|minimal|low|medium|high|xhigh>
The GPT-5.6 API supports max, and the Codex CLI exposes ultra for Sol/Terra — neither is discoverable from the hint.
Impact
Every /codex:rescue that uses the agent's prompt-tightening step optimizes the forwarded prompt against 5.4-era guidance while gpt-5.6-sol executes it. Silent quality regression, not an error.
Suggested fix
- Add a
gpt-5-6-prompting skill (or rename/update the existing one) reflecting the official 5.6 guidance, and update the skills: list + body reference in agents/codex-rescue.md.
- If a single skill should cover multiple generations, make its scope explicit (e.g.
codex-prompting) so it doesn't pin to one model version.
- Update the
--effort hint in commands/rescue.md to include max (and note ultra for Sol/Terra where applicable).
Environment
- Plugin
codex v1.0.5 (openai/codex-plugin-cc)
- Codex CLI 0.144.1
- Default model
gpt-5.6-sol, effort high
Summary
Since the GPT-5.6 rollout,
codex login+ a cache refresh makesgpt-5.6-solthe default Codex model (family: Sol/Terra/Luna). But thecodex-rescuesubagent — the only place in the plugin that reshapes a user request into a Codex-bound prompt — still points at agpt-5-4-promptingskill. GPT-5.6's official guidance is materially different from the 5.4 era (leaner, outcome-first contracts; 5.6 is more concise by default), so prompts get shaped in the wrong direction for the model that actually runs them.Ref: https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6
Where (plugin v1.0.5)
1. Stale prompting skill (primary)
agents/codex-rescue.md:skills: [codex-cli-runtime, gpt-5-4-prompting]gpt-5-4-promptingskill only to tighten the user's request into a better Codex prompt before forwarding it."skills/ships onlygpt-5-4-prompting— there is nogpt-5-6/gpt-5-5variant.2. Effort hint omits current tiers (secondary, cosmetic)
commands/rescue.md(arg hint):--effort <none|minimal|low|medium|high|xhigh>The GPT-5.6 API supports
max, and the Codex CLI exposesultrafor Sol/Terra — neither is discoverable from the hint.Impact
Every
/codex:rescuethat uses the agent's prompt-tightening step optimizes the forwarded prompt against 5.4-era guidance whilegpt-5.6-solexecutes it. Silent quality regression, not an error.Suggested fix
gpt-5-6-promptingskill (or rename/update the existing one) reflecting the official 5.6 guidance, and update theskills:list + body reference inagents/codex-rescue.md.codex-prompting) so it doesn't pin to one model version.--efforthint incommands/rescue.mdto includemax(and noteultrafor Sol/Terra where applicable).Environment
codexv1.0.5 (openai/codex-plugin-cc)gpt-5.6-sol, efforthigh