2026-03-09 — 2026-03-13
9 changes
OpenClaw became a more predictable system this week through 9 stability fixes that eliminated silent failures and configuration drift. The most critical fix protected bootstrap files like MEMORY.md and SOUL.md from accidental deletion during AI memory flushes, preventing workspace corruption that previously required manual git recovery. Telegram exec approvals were completely non-functional with "unknown approval id" errors blocking all elevated operations until we fixed approval ID handling and thread routing. Web search citations were disappearing when using Perplexity through OpenRouter, degrading research quality for users who depend on source tracking. We also stopped duplicate messages in Telegram during provider timeout scenarios and fixed the Sessions table layout breaking on mobile screens. The pattern across all fixes was clear: making broken configurations fail fast and explicitly rather than degrading silently. Two infrastructure improvements support this direction - cron jobs now track structured failure reasons (auth, rate limit, billing) and bug reports collect provider routing details to diagnose vanishing issues. The system is now more reliable but also more transparent about what's broken when things go wrong. We'll use the improved error classification next week to build smarter retry logic for provider routing failures.
Social posts
Fixed 9 silent failure modes in OpenClaw this week. The worst: AI memory flushes could accidentally delete MEMORY.md and SOUL.md files, requiring manual git recovery. Now bootstrap files are protected and memory operations are restricted to daily append-only files.
Why do Telegram exec approvals fail with 'unknown approval id' errors? Because approval ID normalization was broken and responses landed in wrong chat threads. Sometimes the simplest features hide the messiest edge cases.
Should systems fail silently or loudly? We're moving OpenClaw toward failing fast and explicitly. Better to catch broken configurations at startup than discover them during critical operations when users need reliability most.
Spent this week fixing 9 silent failure modes in OpenClaw, from Telegram exec approvals breaking with 'unknown approval id' errors to web search citations disappearing with certain providers. The pattern was clear: our system was degrading gracefully when it should have been failing loudly. We're now making broken configurations explicit rather than letting them cause mysterious runtime issues.
The most dangerous bug we fixed this week: AI memory flushes could accidentally overwrite workspace bootstrap files like MEMORY.md, requiring manual git recovery. It's a reminder that even append-only operations need careful boundaries. Now memory flushes are restricted to daily files and bootstrap files are treated as read-only.