Auto-archive of inactive sessions breaks multi-agent / multi-project workflows

Summary

Claude Code's session auto-archive behavior makes the tool effectively unusable for users who work on multi-agent or multi-project development, where running multiple parallel Claude Code sessions is essential.

Reproduction

1. Open two or more Claude Code sessions in the desktop app (e.g., one per project, or multiple agents within the same multi-agent system).
2. Switch focus between sessions.
3. Each non-focused session is auto-archived within ~1.5 seconds (teardown_archive_timeout_ms = 1500 in tengu_bridge_repl_v2_config).
4. When returning to a previously-focused session, the dialog "このセッションはアーカイブされています。会話を続けるには、アーカイブを解除してください。" / "This session has been archived. Click to unarchive." appears, requiring a manual click to resume.
5. This happens dozens of times per day, completely breaking the workflow.

Why this is critical for multi-agent development

Multi-agent systems require running and debugging several coordinated agents simultaneously (e.g., a planning agent, a design agent, a value-evaluation agent, each in its own context window). Claude Code's current design implicitly assumes a single active "task" at a time — which is incompatible with the increasingly common workflow of multi-agent development.

Expected behavior

Either of the following would resolve the issue:

• Option A: Do NOT auto-archive a session unless the user explicitly closes it.
• Option B: Make the timeout configurable (e.g., in ~/.claude/settings.json) with a way to mark sessions as "always-keep-alive".

Additionally, the UI should clearly distinguish between (a) sessions the user has finished and (b) sessions kept in the background waiting to be resumed.

Comparison with peer tools

OpenAI's Codex CLI does not exhibit this behavior. Multiple Codex sessions can run concurrently without auto-archive interruption. This demonstrates that the multi-session model can be implemented without forcing users through manual unarchive cycles. The current Claude Code design is therefore a deliberate design choice rather than a technical constraint, and one that puts Claude Code at a competitive disadvantage for the exact workflow (AI-agent development) that Anthropic actively encourages.

System info

• macOS Darwin 25.5.0
• Claude Code version 2.1.142 (Desktop app, claude-desktop entrypoint)
• Detected feature flags from ~/.claude.json:
• tengu_ccr_bridge_multi_session: true
• tengu_bridge_repl_v2_config.teardown_archive_timeout_ms: 1500
• multisession_poll_interval_ms_at_capacity: 60000
• session_keepalive_interval_ms: 0

Impact

Multi-agent development is a fast-growing use case that Anthropic itself actively promotes (Claude Agent SDK, MCP, A2A protocols, sub-agents, etc.). The current archive design actively penalizes this workflow, requiring users to either:

• Constantly click "unarchive" dozens of times per session
• Or limit themselves to a single Claude Code instance at a time, defeating the purpose of multi-session support

For users building multi-agent systems or working across multiple projects simultaneously, this turns Claude Code from a productivity multiplier into a usability bottleneck.