Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Cron jobs configured with sessionTarget: isolated are not running in isolated sessions. Instead, their messages accumulate silently in the main conversation session (conversation_id 6 in lcm.db), causing context overflow and agent crashes.

Steps to reproduce

1. Configure one or more cron jobs with "sessionTarget": "isolated" in ~/.openclaw/cron/jobs.json
2. Allow cron jobs to run normally over several days
3. Monitor lcm.db: sqlite3 ~/.openclaw/lcm.db "SELECT COUNT(*) FROM messages WHERE conversation_id = 6;"
4. Observe message count climbing continuously despite isolated setting

Expected behavior

Each cron job run should create its own isolated session with a fresh context. No messages from cron job runs should appear in the main agent session (conversation_id 6).

Actual behavior

Cron job messages accumulate in conversation_id 6 (the main session). Over time this causes:

Context window fills to capacity Agent (Icarus) crashes with: "Context limit exceeded. I've reset our conversation to start fresh" Agent suggests setting compaction.reserveTokensFloor to 20000 — but this setting is already applied and does not prevent the accumulation

OpenClaw version

2026.4.1 (da64a97)

Operating system

macOS (Mac Mini M4, Apple Silicon)

Install method

LaunchAgent (gui/502/ai.openclaw.gateway) — not a global daemon, runs under the dedicated OpenClaw user account

Model

Primary model: moonshot/kimi-k2.5

Provider / routing chain

Primary model: moonshot/kimi-k2.5 (Moonshot direct API) Fallback: deepseek/deepseek-chat Local: ollama/llama3.2:3b (lightweight tasks only) Credentials: macOS Keychain via object-format SecretRef in models.providers No .env file in use

Additional provider/model setup details

Account type: Dedicated Standard user account (not admin) Gateway: LaunchAgent (gui/502/ai.openclaw.gateway)

Logs, screenshots, and evidence

First incident — April 2, 2026:

conversation_id 6 accumulated ~2,394 messages from cron jobs
Agent crashed with context limit exceeded
Session was manually cleared via: sqlite3 ~/.openclaw/lcm.db "DELETE FROM messages WHERE conversation_id = 6;"

Second incident — April 5, 2026:

conversation_id 6 accumulated 631 messages in approximately 3 days after being cleared
Agent crashed again with identical context limit error
Session cleared again manually

Affected Jobs
All 8 cron jobs are set to sessionTarget: isolated — none are behaving as isolated:
JobsessionTargetCerberus Gateway WatchdogisolatedIcarus Alive Pingisolatedmemory-save-middayisolatedmemory-save-endofdayisolatedWeekly Security AuditisolatedWeekly Library SyncisolatedCerberus — Agent SOUL.md Integrity MonitorisolatedIcarus System Statusisolated
Config
Relevant jobs.json structure (representative example):
json{
  "id": "841d7abf-e756-496c-a1c4-fba17d5715cf",
  "name": "memory-save-endofday",
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "model": "moonshot/kimi-k2.5"
  }
}
payload.kind is set to agentTurn on all jobs (confirmed per docs that systemEvent causes jobs to be skipped entirely).

Impact and severity

Severity: High Impact:

Agent availability — Icarus crashes completely when context fills. No graceful degradation, no warning, just a hard reset that drops all active context. Data loss risk — Any in-progress task or reasoning at the time of collapse is lost entirely. Recurring — The bug is silent and inevitable. Without manual intervention it will always recur. Current rate is roughly every 10 days. No workaround within OpenClaw — The only fix is manual lcm.db surgery. There's no config option that prevents it. Affects all cron jobs — Not one misconfigured job, all 8 isolated jobs are affected. This means the isolation feature is effectively non-functional.

Not critical because:

Icarus recovers after the crash No data corruption in the vault or config files The manual clear is simple once you know to do it

Additional information

compaction.reserveTokensFloor is set to 20000 — this does not prevent the accumulation The bug is silent — no warnings are logged when isolated sessions fail to isolate The agent (Icarus) does not detect the accumulation until context limit is hit Lossless Claw compaction plugin is active and wired correctly — does not prevent this issue

Workaround Manually clear conversation_id 6 periodically: bashsqlite3 ~/.openclaw/lcm.db "DELETE FROM messages WHERE conversation_id = 6;" Monitor message count as early warning: bashsqlite3 ~/.openclaw/lcm.db "SELECT COUNT(*) FROM messages WHERE conversation_id = 6;" At current accumulation rate (~200 messages/day), context collapse occurs within approximately 10 days of clearing.