lesson vignesh07#76: verify current state before reporting task incomplete

Author: Molty

MEMORY.md

@@ -1,6 +1,6 @@
 # MEMORY.md - Long-Term Memory
 
-*Last updated: 2026-02-25*
+*Last updated: 2026-02-26*
 
 ---
 
@@ -115,6 +115,7 @@
 73. **Don't spam webhooks + sub-agent tests in rapid succession.** Doing so burns through per-minute API limits and triggers OpenClaw cooldowns on all providers simultaneously. Space testing by at least 5 min between attempts.
 74. **Always include openai-codex/gpt-5.2 as final fallback.** It uses OAuth (cached, no rate limit issues) and supports tool use. Without it, rate limit cooldowns leave the agent with zero functional models. Fleet standard fallback chain: `anthropic/claude-sonnet-4-6` → `anthropic/claude-haiku-4-5` → `xai/grok-3` → `openai-codex/gpt-5.2`.
 75. **Leonardo MC Heartbeat cron still pending.** Fix auth first, then set up: schedule `30 */2 * * *` HKT, model `anthropic/claude-haiku-4-5`, sessionTarget isolated, pings MC heartbeat endpoint.
+76. **Always verify current state before reporting a task as incomplete.** Feb 26: told Guillermo PLAN-003 needed a 90-min coding session — it was already done (overnight run built the DB at 01:03 HKT). Daily logs go stale. Check config files, APIs, and MC before making any "not done" claim.
 
 ---