openclaw - 💡(How to fix) Fix Add native sessions cleanup support for orphan/unindexed transcript archive/prune [1 comments, 2 participants]

Problem

openclaw sessions cleanup currently supports --fix-missing, which removes session-store entries whose transcript files are missing. It does not appear to expose the inverse cleanup path: transcript files under ~/.openclaw/agents/<agentId>/sessions/*.jsonl that are no longer indexed by sessions.json.

This leaves historical/orphan transcript files discoverable by grep and lifecycle inventories even though they are not live session rows. In one production cleanup, 23 orphan/unindexed transcripts had historical anthropic/claude-sonnet-4-6 first-model entries. Native openclaw sessions --all-agents --json showed 0 indexed target rows and 0 Anthropic model/provider/override rows, but filesystem scans still flagged them as stale pins until manually quarantined.

Desired native behavior

Add a native, audited cleanup mode such as one of:

• openclaw sessions cleanup --archive-orphans --all-agents --dry-run
• openclaw sessions cleanup --prune-unindexed --all-agents --dry-run
• openclaw sessions cleanup --orphan-transcripts --archive <dir> --all-agents

Required behavior:

1. Enumerate configured/discovered agent session stores.
2. Build the indexed transcript/sessionId set from sessions.json.
3. Identify core transcript files not indexed by the store, excluding trajectory/checkpoint/deleted marker files.
4. Dry-run prints exact file paths, agentId, sessionId, mtime, size, sha256.
5. Enforce mode either archives to a supplied quarantine directory with manifest or prunes only with explicit confirmation flag.
6. Never edits transcript JSONL content.
7. Runs through the same Gateway/session-store safety model where applicable, or clearly marks filesystem-only archive semantics.

Why this matters

Without native support, operators have to write ad hoc filesystem scripts for a session-lifecycle maintenance case that is adjacent to existing native session cleanup. That creates avoidable substrate drift and makes “live model pin” vs “orphan historical transcript” harder to distinguish.

Evidence from production cleanup

• OpenClaw version: 2026.5.3-1 (2eae30e).
• Existing help surface checked:
  • openclaw sessions --help
  • openclaw sessions cleanup --help
• Related available option: --fix-missing removes store entries whose transcript files are missing.
• Missing option: no orphan/unindexed transcript archive/prune command.
• Production finding:
  • 23 orphan/unindexed core transcripts with first model anthropic/claude-sonnet-4-6.
  • 0/23 appeared in openclaw sessions --all-agents --json.
  • 0 model/provider/override Anthropic rows in current session index.
  • 0 active file handles.
  • Manual quarantine required to clear filesystem-only stale detector signal.