Han Feedback: han-feedback (2026-06-17)

[oppegard]

Han Feedback — 2026-06-17

Skills used: han-feedback:han-feedback
Agents used: none
Context: Reporting a Codex compatibility problem after seeing errors that Codex could not use the sonnet model while working with Han skills.
Outcome: A focused maintainer issue body was prepared for testdouble/han.
Affected skills observed: han-planning:plan-a-feature, han-planning:plan-a-phased-build, han-planning:plan-work-items, han-planning:plan-implementation, and han-plugin-builder:agent-builder or its guidance references.

---

What worked well

• The model-selection assumptions are easy to find in the repo. Direct SKILL.md references explicitly say to pass model: "sonnet", which makes the likely failure mode concrete instead of speculative.
• The affected instructions are concentrated in a small set of planning skills and model-selection guidance, so the fix should be reviewable as a targeted compatibility pass.
• The guidance already treats model tier as an abstraction (opus, sonnet, haiku, inherit), which gives maintainers a clear path to map tiers per host environment instead of hard-coding Claude-specific model names everywhere.

---

What didn't work

• Some Han skills instruct Codex to launch sub-agents with model: "sonnet". In Codex, that model override can fail because sonnet is a Claude-specific model tier/name rather than a Codex-supported model identifier.
• The failure appears before the skill can do useful work, so the user sees a tooling/model compatibility error instead of a graceful fallback or host-specific mapping.
• The instructions do not appear to distinguish Claude Code from Codex when specifying agent model overrides. A portable instruction such as "use the host's equivalent mid-tier model" or "omit the model override outside Claude Code" would avoid this class of failure.
• The current behavior makes otherwise useful Han skills feel brittle in Codex even when the underlying workflow is not Claude-specific.

---

Overall

This looks like a cross-host compatibility bug in Han's model-selection instructions rather than a problem with the planning workflows themselves. Han skills can keep the opus/sonnet/haiku tier language as design intent, but Codex needs either a supported model mapping or instructions to avoid passing Claude-specific model overrides. The issue is especially visible in planning skills that require every sub-agent dispatch to pass model: "sonnet".

---

Rating

Dimension	Score
Output accuracy	3/5
Evidence discipline	4/5
Finding signal-to-noise	5/5
Output length vs. decision count	4/5
Turn efficiency	3/5

[mxriverlynn]

@oppegard 🤦🏻‍♀️ i completely forgot i had models set in those. thanks for the feedback on this! will do some digging as soon as i have time, and either find a way to generalize it, or just remove the model specifier entirely

[oppegard]

No worries! I love the idea of delegating to a cheaper model when appropriate. And, it looked to me like Codex tried to follow the instruction to use sonnet, so I think the capability is there.