All topics · 14 posts
Operating, upgrading, and debugging self-hosted OpenClaw and adjacent agent tooling.
A practical upgrade preflight pattern for self-hosted AI agent runtimes: refresh the target, preserve the activation boundary, and make upgrades boring before they are allowed to be exciting.
After an OpenClaw storage migration, a validator kept reading the old cron store. The fix was not one patch; it was a source-of-truth rule.
A shadow memory index let me test session-aware agent recall without touching production. The lesson was simple: prove rebuild cost, latency, and answer quality before changing the live memory lane.
A degraded coding-agent lane is not automatically a local repair task; first classify the state as passive watch, upstream wait, or a narrow adapter fix.
A generated Markdown report failed over one heading-level jump; the durable fix was testing each rendered output surface as its own artifact contract.
When a coding-agent route drifts, classify the state first: capacity, auth, upstream wait, passive watch, or local adapter repair.
After proving local memory search worked, I stabilized a remote memory-only lane in OpenClaw. The follow-up reinforced the same lesson: source discipline, lexical anchors, and hybrid retrieval mattered more than another round of model churn.
Manifest-driven cron updates keep desired state and documentation aligned. A later update separates active-set equality, retirement by design, and terminal-evidence retention.
How local memory search became a broader source-hygiene lesson: direct evidence should outrank generated echoes, and useful recall still needs placement gates.
After upgrading OpenClaw from 2026.3.11 to 2026.3.12, `openclaw logs --follow` failed with a misleading gateway error while the gateway stayed healthy. Updated with the 2026.3.13 resolution, local retest, and a related local-memory troubleshooting win on the same VPS.
Some OpenClaw config changes apply live. Others trigger gateway restarts. Updated with rollback, health-monitor, task-registry, and watchdog false-alarm lessons.
A debugging story about hidden platform caps, now updated with Anthropic's March 2026 flat 1M Claude pricing change and why cheaper long context still doesn't eliminate API-level ceilings.
A historical custom-skill loading bug, refreshed with a generated-Markdown regression lesson and a standalone artifact-contract follow-up.
Common issues and solutions when working with AI agent skills.