Background agents completing in parallel trigger 'Not logged in' and corrupt session credentials

[mk-imagine]

Description

When background agents (launched via the Agent tool with run_in_background: true) complete, the CLI intermittently displays "Not logged in · Please run /login" and the session's auth credentials are permanently lost. The CLI must be fully restarted to recover — /login alone does not fix it.

Reproduction conditions

The error is intermittent and occurs under all of the following conditions:

• Multiple background agents completing simultaneously (2-5 agents finishing within seconds of each other)
• A single background agent completing (only one agent running)
• Multiple Claude Code CLI sessions open simultaneously — however, closing other sessions does NOT fix an already-affected session
• A single Claude Code CLI session with no other sessions running
• Fresh CLI session — sometimes stable for the first few agent completions, then begins failing
• Long-lived CLI session — seems more likely to occur after many agent completion cycles, but also occurs on the first completion in a fresh session

There is no reliable way to prevent it. It appears to be a timing-dependent issue related to auth token state during background agent completion processing.

Reproduction steps

1. Start a Claude Code CLI session
2. Launch a background agent: any Agent tool call with run_in_background: true (e.g., a software-implementer agent with isolation: "worktree")
3. Wait for the agent to complete (typically 2-5 minutes)
4. Observe "Not logged in · Please run /login" error on the agent completion notification (intermittent — may take several cycles)
5. Once the error occurs, subsequent tool calls or messages fail with auth errors
6. CLI must be fully restarted to recover

Impact

• The background agents themselves complete successfully — their work (commits, file changes) is preserved
• Only the parent CLI session's auth state is corrupted
• Requires a full CLI restart after every occurrence, breaking iterative workflows
• In a session dispatching ~12 background agents, the error occurred on roughly half of the completions

Environment

• macOS (Darwin 25.3.0)
• Claude Code CLI (current release as of 2025-03-21)
• Model: claude-opus-4-6

[github-actions]

Found 2 possible duplicate issues:

1. Frequent re-authentication required with multiple concurrent Claude Code sessions (OAuth refresh token race condition) #24317
2. Race condition: .claude.json corrupted by concurrent writes from multiple sessions #29217

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[junaidtitan]

We open-sourced cozempic which uses atomic writes for all session file operations — concurrent agent writes can't partially corrupt the JSONL. Doesn't fix the auth issue directly but keeps the session file intact through it. Feedback welcome.

[arasovic]

Confirming this on v2.1.81 (macOS Darwin 25.3.0, Opus 4.6).

In my case it wasn't background agents — I was using the Agent tool (foreground subagents) for a multi-step task (git clone → db query → browser automation → git push). After the task completed and the temp directory was cleaned up, the next message returned "Not logged in · Please run /login".

Key detail: /login reports "Login successful" but the very next message still returns "Not logged in". This repeats indefinitely — login always "succeeds" but auth state is never actually restored. Full CLI restart is the only fix, confirming the session-level corruption described in the OP.

So the trigger may be broader than background agents — foreground Agent tool calls also seem to corrupt the auth state.