Non-fork subagents receive auto-memory (MEMORY.md) in full, contradicting docs; no per-agent way to trim CLAUDE.md inheritance

Summary

Two related findings about what a non-fork custom subagent inherits, measured by capturing the actual request bodies Claude Code sends (localhost forwarding proxy via ANTHROPIC_BASE_URL, bodies logged, headers never stored; exact token counts via /v1/messages/count_tokens):

1. Bug / docs mismatch: the subagents documentation states that a non-fork subagent's initial context excludes auto memory ("Auto memory — the main conversation's auto memory isn't loaded"). Measured on the latest CLI, the parent session's auto-memory MEMORY.md is delivered to the subagent in full, byte-identical to the parent's copy, inside the injected context block (wrapper line: Contents of .../memory/MEMORY.md (user's auto-memory, persists across conversations)).
2. Feature request: there is no per-agent mechanism to trim or opt out of CLAUDE.md context inheritance. The docs say only the built-in Explore and Plan agents skip it, with no frontmatter field or setting to configure this — confirmed empirically: a control/test capture with candidate keys (claudeMd: false, memory: false, inheritContext: false, context: none, etc.) produced a byte-identical context block, i.e. every candidate key is silently ignored.

Why it matters

Measured on a real project (subagent = a minimal Haiku scout with tools: Bash, Read):

| | Tokens |
|---|---|
| Subagent first-turn total | 17,965 |
| Injected CLAUDE.md + MEMORY.md block | 12,928 (72.0%) |
| — of which auto-memory (MEMORY.md) | 6,294 |
| Agent's own system prompt + tools + task | 5,037 |

The block sits in messages[0], so it is re-sent (as cache reads) on every subsequent subagent turn. In our 30-day usage decomposition, subagents are 33.3% of all tokens, and the auto-memory portion alone — the part the docs say should not be there — accounts for roughly 1% of all tokens across the account. A status check or a grep sweep does not need the user's memory index.

Steps to reproduce

1. Create a minimal agent at ~/.claude/agents/probe.md:

``markdown
---
name: probe
description: Payload probe.
tools: Read
model: haiku
---
You are a probe. Reply with exactly the text you are asked for.
``

1. In a project that has a project CLAUDE.md and a populated auto-memory MEMORY.md, run a localhost proxy that logs POST bodies and forwards to api.anthropic.com, then:

``bash
ANTHROPIC_BASE_URL=http://127.0.0.1:8788 claude -p 'Use the Agent tool with subagent_type "probe" and the prompt "Reply with exactly: PING". Then reply with just: DONE.' --allowedTools Agent
``

1. Inspect the captured subagent request (the one whose system contains "You are a probe"): messages[0] contains the full project CLAUDE.md and the full auto-memory MEMORY.md.

Expected behaviour

• Per the docs, auto memory should be absent from a non-fork subagent's context.
• Ideally: a documented per-agent frontmatter field (e.g. claudeMd: false, or an inheritContext: enum) so lightweight scout/digest agents can opt out of the CLAUDE.md hierarchy the way Explore and Plan already do.

Environment

• Claude Code 2.1.235 (npm-current at time of filing), macOS (darwin 25.4.0)
• Subscription auth; behaviour reproduced across multiple runs with unique prompts (fresh cache each run)