## Summary 1. **Credit spike (main report):** After switching to Claude Fable 5, a short factual question ("when does the Fable 5 promo credit expire?") triggered automatic invocation of the `claude-api` skill, which appears to have loaded its entire large reference document (SDK usage, Managed Agents, tool-use patterns — tens of thousands of tokens) into context, unrelated to the question asked. The skill match appears to have been triggered by the keyword "Fable" alone. Because this happened o

Skill auto-invocation loads massive unrelated reference doc on premium model, causing rapid unexplained cost spike + silent Fable to Sonnet fallback quality cliff

Status Open

Reported on v2.1.212

Maintainer reply None cached

Activity 0 comments · opened Jul 21, 2026

Credit spike (main report): After switching to Claude Fable 5, a short factual question ("when does the Fable 5 promo credit expire?") triggered automatic invocation of the claude-api skill, which appears to have loaded its entire large reference document (SDK usage, Managed Agents, tool-use patterns — tens of thousands of tokens) into context, unrelated to the question asked. The skill match appears to have been triggered by the keyword "Fable" alone. Because this happened on the most expensive tier (Fable 5, 3x+ Sonnet pricing), roughly $20 of credit was consumed within minutes with no real work done. There was no way to see what was charged for or why, inside the session UI.

Silent model fallback: On a paid plan, the Fable→Sonnet fallback (usage cap reached) happened silently with no warning, so the user only noticed after quality had already degraded.

Quality drop after fallback: After falling back to Sonnet 5, the same session saw repeated PowerPoint COM automation argument errors (5+ retries), a failed script that deleted an already-generated output PDF, an attempted slide text edit without user approval, and repeated retries against the same file-lock error.

Skill auto-matching should not fire on incidental keyword overlap (here, "Fable" matching in a pricing question) when the request has nothing to do with the skill's actual domain.
Before loading a large reference doc into context — especially on a premium-tier model — surface a cost-impact warning first.
On model fallback (e.g. usage cap reached), give an explicit notice + a choice (wait vs. switch), rather than switching silently.
Show real-time, itemized credit/token usage so users can see what was charged for.
Investigate whether Sonnet 5 quality degrades in long, multi-turn, tool-heavy sessions.

Claude Code 2.1.212, Windows 11, models: Claude Fable 5 → Claude Sonnet 5