You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Feature]: 클로드 코드 서브에이전트 호출이 소넷(플레이스홀더용)으로 보이는게 너무 헤깔려양 #252
Thanks for the report — the confusion is understandable. When a subagent turn shows "Sonnet", it can be a placeholder label rather than proof of which model actually served the request, which makes placeholder-labeled calls hard to distinguish from genuine Sonnet calls. That ambiguity is exactly the problem you're describing.
Your suggestion — using an explicitly low-tier placeholder such as Haiku so it is less likely to be mistaken for a genuine Sonnet call — is a fair UX enhancement. Keeping this open as an enhancement request for the display/labeling backlog. If you can capture a case where the routed model demonstrably differs from what was executed (not just labeled), please attach it; that would upgrade this from a labeling issue to a routing bug.
Implemented on dev at 71ebf77b (feature f103d6cd, docs sync f39401d6 + 7b1e497c).
What changed:
Claude Code subagent guidance placeholder is now haiku instead of sonnet in src/claude/agents-inject.ts (NO_MODEL_ARG).
Docs-site guidance was synced across en/ja/ko/ru/zh-cn so docs no longer reintroduce the confusing Sonnet placeholder.
Existing inject tests assert the new placeholder guidance.
Routing note:
The placeholder string does not drive proxy routing. Real model selection remains pinned by opencodex (ocx-route); the model argument remains inert for these injected agents. The change is deliberately a display/UX disambiguation so placeholder-labeled turns are no longer easy to confuse with genuine Sonnet calls.
Activity
lidge-jun commentedon Jul 22, 2026
Thanks for the report — the confusion is understandable. When a subagent turn shows "Sonnet", it can be a placeholder label rather than proof of which model actually served the request, which makes placeholder-labeled calls hard to distinguish from genuine Sonnet calls. That ambiguity is exactly the problem you're describing.
Your suggestion — using an explicitly low-tier placeholder such as Haiku so it is less likely to be mistaken for a genuine Sonnet call — is a fair UX enhancement. Keeping this open as an enhancement request for the display/labeling backlog. If you can capture a case where the routed model demonstrably differs from what was executed (not just labeled), please attach it; that would upgrade this from a labeling issue to a routing bug.
devlog: 260723_issue_fixes WP0 roadmap — diff-level docs for #315/#311/…
fix(claude): recommend haiku as the inert subagent model placeholder …
docs(claude-code): sync sonnet->haiku placeholder guidance across loc…
devlog: 260723_issue_fixes loop close-out (DONE — #315/#311/#252)
lidge-jun commentedon Jul 23, 2026
Implemented on
devat71ebf77b(featuref103d6cd, docs syncf39401d6+7b1e497c).What changed:
haikuinstead ofsonnetinsrc/claude/agents-inject.ts(NO_MODEL_ARG).Routing note:
ocx-route); themodelargument remains inert for these injected agents. The change is deliberately a display/UX disambiguation so placeholder-labeled turns are no longer easy to confuse with genuine Sonnet calls.Closing as completed on
dev.devlog: 260723_issue_fixes WP0 roadmap — diff-level docs for lidge-ju…
fix(claude): recommend haiku as the inert subagent model placeholder …
docs(claude-code): sync sonnet->haiku placeholder guidance across loc…
devlog: 260723_issue_fixes loop close-out (DONE — lidge-jun#315/lidge…