BUG: Custom agent definitions fail with 'Not logged in' on OAuth (v2.1.81)

Bug Description

All custom agent definitions in ~/.claude/agents/definitions/*.md fail immediately with "Not logged in · Please run /login" when spawned as subagents. Built-in agents (general-purpose, Explore, Plan) work fine.

Environment

• Claude Code version: 2.1.81
• OS: macOS (Darwin 25.3.0)
• Auth method: OAuth (claude.ai subscription, no API key)
• Shell: zsh

Steps to Reproduce

1. Create a custom agent definition at ~/.claude/agents/definitions/MyAgent.md:

---
name: MyAgent
description: Test agent
model: opus
---
Respond to prompts.

1. Start a Claude Code session (OAuth authenticated)

1. Spawn the custom agent:

Use the Agent tool with subagent_type: "MyAgent" and prompt: "Say hello"

1. Result: Immediate failure — "Not logged in · Please run /login" with 0 tokens, 0 tool uses, ~50ms

1. Spawn a built-in agent (general-purpose) with same prompt → works fine

What Was Tested

| Agent Type | Definition | Result |
|-----------|-----------|--------|
| general-purpose | Built-in | ✅ Works (14k tokens) |
| general-purpose + model: opus | Built-in + override | ✅ Works |
| general-purpose + isolation: worktree | Built-in + override | ✅ Works |
| Architect | Custom .md (opus, worktree) | ❌ Fails (0 tokens) |
| Engineer | Custom .md (opus, worktree) | ❌ Fails (0 tokens) |
| Designer | Custom .md (opus) | ❌ Fails (0 tokens) |
| Algorithm | Custom .md | ❌ Fails (0 tokens) |
| Intern | Custom .md | ❌ Fails (0 tokens) |

Note: Engineer and Designer worked earlier in the same session but failed later, suggesting OAuth token propagation to custom agent subprocesses degrades over time. Built-in general-purpose continues to work throughout.

Expected Behavior

Custom agent definitions should authenticate using the same OAuth session as the parent process, just like built-in agents do.

Additional Context

• This worked reliably before (user reports it worked as recently as March 23, 2026)
• v2.1.81 (March 20) included an OAuth token race condition fix — this may have introduced a regression in how tokens are propagated to custom agent subprocesses
• /login in the parent session does not fix the issue for custom agents
• Related issues: #16594 (hardcoded model names in built-ins), #20931 (custom agent loading)
• The ANTHROPIC_API_KEY env var is NOT set (pure OAuth auth)
• CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is enabled in settings