Sub-Agent Issues — Extended Gotcha List

The main SKILL.md calls out the single most dangerous gotcha (the context: fork semantic inversion). The issues below are additional papercuts and documentation bugs that a sub-agent orchestrator should be aware of.

1. context: fork is semantically inverted

fork creates an ISOLATED blank context, not a copy of the parent. If you rely on the English meaning you will ship a broken orchestrator.

• GitHub issue: anthropics/claude-code#20492

2. AskUserQuestion is silently unavailable in sub-agents

The tool is listed as available but calls fail or return empty. Sub-agents that rely on human-in-the-loop clarification will freeze or loop.

• GitHub issue: anthropics/claude-code#34592

3. Sub-agents silently default to a different model than the parent

If you do not pin model in the AgentDefinition, the sub-agent may run on a different (often weaker) model. Symptoms look like a capability regression that turns out to be a routing regression.

• GitHub issue: anthropics/claude-code#5456

4. Filesystem skill discovery bypass

Sub-agents can currently pick up skills directly from .claude/skills/ on disk, but the documentation says they can't. This works today, may break tomorrow, and should not be used as a production channel for passing skills.

• GitHub issue: anthropics/claude-code#32910

5. Parent-context inheritance (feature request, still OPEN)

17 upvotes as of writing. Until this lands, every sub-agent starts fully blank except for CLAUDE.md, tools, and MCP servers — which is the premise of this whole skill.

• GitHub issue: anthropics/claude-code#12790

6. Silent billing routing to a pricier model

Some sub-agent invocations have been observed routing through higher-priced models without warning, producing unexpectedly large bills.

• GitHub issue: anthropics/claude-code#39903

---

These references back the core SKILL.md. If you close any of the above tracking issues with resolutions that change the inheritance model, also update the skill body so future orchestrators don't act on stale assumptions.