claude-code - 💡(How to fix) Fix Sessions auto-archiving every 30-90 min when launched from specific cwd bucket (since 2026-05-09)

Summary

Claude Code sessions launched from /Users/jason/.openclaw/ are being auto-archived in FleetView every 30-90 minutes of activity, dropping the mobile bridge. Identically-configured sessions launched from sibling working directories under /Users/jason/.openclaw/workspace/ survive normally for hours. The behavior started on or about 2026-05-09 and is now happening multiple times per day on Fuse-development work specifically.

This is a re-report extending the 2026-05-12 internal write-up — issue has persisted through 2026-05-15 with no improvement.

Environment

• claude-code 2.1.121 (Homebrew, installed 2026-04-27, unchanged since)
• macOS Darwin 25.3.0 / Apple Silicon
• Model: claude-opus-4-7[1m] (1M context, xhigh effort; also reproduced at medium)
• Mobile bridge attached (Claude Code mobile app)
• MCP servers: obsidian only

Repro

1. Launch Claude Code from /Users/jason/.openclaw (NOT the /workspace/ subdir)
2. Begin Fuse-development work (heavy tool use, large MCP/skill load)
3. Within 30-90 min the session disappears from FleetView Active and the mobile bridge loses access
4. CLI session itself often remains usable until the next turn fails

Differential — identical config, different behavior

Sessions that survive normally (run for hours):

• APEX session at cwd /Users/jason/.openclaw/workspace/...
• FOV session at cwd /Users/jason/.openclaw/workspace/...
• Alpha session at cwd /Users/jason/.openclaw/workspace/...
• This very report is being filed from a workspace-bucket session that has survived 2+ hours of heavy Fuse work today (2026-05-15).

Sessions that get archived:

• Fuse sessions launched from cwd /Users/jason/.openclaw/ (parent dir, different bucket string)

Evidence — transcript churn (parent bucket only)

~/.claude/projects/-Users-jason--openclaw/
  2026-05-12 01:05  3.8 MB session
  2026-05-12 02:16  8.0 MB session  ← archived mid-work
  2026-05-12 03:01  1.6 MB session  ← archived mid-work
  2026-05-12 03:06  1.7 MB session  ← archived mid-work
  2026-05-12 03:10  1.2 MB session

2026-05-08 had 7 sessions across 3 hours (same pattern). 2026-05-15 (today): previous Fuse session "stopped responding" mid-flight, presumed archive event.

Hypothesis

Server-side policy keyed off project-dir bucket name or transcript density per bucket. The -Users-jason--openclaw bucket has ~83 stored transcripts; the workspace sibling buckets have 1-52 each. The dense bucket appears to trip something the sparse ones don't.

Local config signals (~/.claude.json cachedGrowthBookFeatures)

• tengu_bridge_repl_v2_config.teardown_archive_timeout_ms = 1500
• tengu_bridge_repl_v2_config.session_keepalive_interval_ms = 0
• tengu_bridge_repl_v2_config.session_keepalive_interval_v2_ms = 0
• tengu_bridge_repl_v2_config.non_exclusive_heartbeat_interval_ms = 180000
• tengu_bridge_repl_v2_config.http_timeout_ms = 10000
• tengu_lapis_finch = True
• tengu_shale_finch = True

When it started

On or about 2026-05-09. No claude-code binary update has happened since 2026-04-27; no ~/.claude/settings.json or workspace settings.local.json touched in 14+ days. The behavior change correlates only with GrowthBook feature-flag refreshes, which write to ~/.claude.json on every launch.

Ask

Please look at server-side session-archive policy / feature-flag changes for account [email protected] (or [email protected]) since 2026-05-09:

1. Which flag or policy is triggering archive on this account/cwd?
2. Why only the -Users-jason--openclaw bucket and not sibling workspace/* buckets?
3. Can keepalive be enabled for heavy multi-hour Fuse-development sessions, or teardown grace raised from 1500 ms?

Impact

Workflow blocker is operationally severe — each archive event drops the mobile bridge, breaks remote control from my phone, and forces a restart that loses session context. Losing ~2 hours of compounding work per day to this since 2026-05-09. Alpha launch shipped 2026-05-15 with this still degrading productivity.