claude-code-guide subagent cannot be invoked: fixed ~214k token overhead exceeds the 200k context limit

Summary

The built-in claude-code-guide subagent cannot be dispatched. Its own system prompt plus tool definitions total roughly 214,300 tokens, which exceeds the 200,000 context limit before any user input is added. Every invocation fails immediately, regardless of prompt size.

Reproduction

Dispatch claude-code-guide via the Agent tool with any prompt.

Observed twice in one session with deliberately different prompt sizes:

| Prompt | Reported request total | Implied fixed overhead |
|---|---|---|
| ~668 tokens | ~214,964 | ~214,296 |
| ~13 tokens | ~214,381 | ~214,368 |

Verbatim error from the second, minimal attempt (prompt was the single sentence "What is a Claude Code hook? Answer in one sentence."):

Agent terminated early due to an API error: Prompt is too long · the request is
~214381 tokens (limit 200000) but this conversation is only ~13 tokens — the rest
is system prompt, tool definitions, and attachment content. A single-exchange
conversation cannot be compacted; reduce attached files/tools or start with less
context.

The delta between the two totals (~583 tokens) tracks the prompt-size difference almost exactly, which indicates the overhead is constant and independent of input.

Expected

The agent dispatches and answers, or its definition is trimmed to fit within the context limit.

Actual

The agent terminates with an API error before performing any work. The error's own suggestion — "reduce attached files/tools or start with less context" — is not actionable from the caller side, since the overhead is in the agent's definition rather than the caller's prompt.

Impact

claude-code-guide is the agent specialized for answering questions about Claude Code features — hooks, settings, slash commands, MCP, the Agent SDK. That is precisely the case where it would be reached for, and it is unavailable.

Workaround: a general-purpose agent with web search against the public Claude Code documentation produced a usable answer, at the cost of rediscovering the docs rather than drawing on the specialized agent's grounding.

Environment

• Claude Code, desktop app
• macOS (Darwin 27.0.0)
• Model: Claude Opus 5
• Observed 2026-08-07