openclaw - 💡(How to fix) Fix [agents/loop-detection]: downgrade Loop warning log level to info for isolated sessions with lifetime <= cron interval

Version: v2026.4.19-beta.1

Symptom

[agents/loop-detection] Loop warning: exec called Ntimes fills gateway.err.log with ~50 warnings per 48h, roughly 4 per 15min cron tick. The warnings come from isolated-session cron jobs (each run is a fresh ephemeral session), not from stale or polluted interactive channel sessions. The loop detector flags repeated exec calls within a run as potential runaway, but for isolated sessions with lifetime shorter than the cron interval, re-tries within the same run are the intended behavior (retry-on-fail, exhaust-before-fallback pattern).

Current behavior

Loop detector emits warn level for all loop warnings regardless of session context. This produces substantial noise in gateway.err.log for legitimate isolated-session use.

Proposed fix

In agents/loop-detection.ts (or wherever the warning is emitted), check the session context:

const shouldWarn = (
  sessionContext.target !== "isolated"
  || sessionContext.lifetimeMs > cronIntervalMs
);
const level = shouldWarn ? "warn" : "info";
logger[level](`Loop warning: ${tool} called ${n} times`);

Where cronIntervalMs comes from the registered cron job metadata (not available? use a hardcoded 5min default as a cheap proxy).

Rationale: isolated sessions cannot accumulate cross-run state, so a "loop" within a single isolated run is bounded by that run's lifetime. If the run completes within its cron interval, the loop is self-bounded and not a leak.

Reproduction

# 1. Register an isolated-target cron job that issues 5+ exec calls
#    (any retry-heavy job, e.g., Code Red Watchdog is a known example)

# 2. Run for 1h

# 3. Count warnings
grep -c '\[agents/loop-detection\] Loop warning' ~/.openclaw/logs/gateway.err.log
# Expect 4+ warnings per hour; all are false positives for isolated sessions

Acceptance

Post-fix sample shows warn-level loop-detection entries drop by >80% for isolated-session cron jobs. Info-level entries still present for audit trail, but not in the err log.

Alternative (if upstream declines)

Accept as out-of-scope; downstream gateway operators can tail -f gateway.err.log | grep -v '\[agents/loop-detection\]' for local log filtering.