One feature. Nine fixes. That's the ratio. Read the ticker above and the story tells itself.
The Silent Failures
Three bugs that never threw an error. A stuck ACP turn starved an entire Discord channel. Heartbeat sessions ballooned past 200K messages because the compaction guard couldn't distinguish boilerplate from conversation. Memory flushes wrote the same content 4–7 times.
Operators blamed themselves. “Channel must be quiet today.” “Agent's just being thorough.” The platform was lying. Now it isn't.
Boundaries Closed
Telegram and iMessage runtimes now route through the plugin SDK. The doctor command got split into provider modules. And jarimustonen's pluggable memory API — the day's lone new feature — lets third-party plugins inject their own system prompt sections instead of inheriting hardcoded ones.
Four PRs of pure structural work. No behavior changes for existing users. Just cleaner seams for the next thing.
Edge Users, Finally
A one-line fix gives 1,500 Matrix users voice bubbles instead of generic file downloads. Linux nvm users stop getting silent TLS failures on every HTTPS request. Discord's resume picker stops expiring the instant it appears.
None of these groups drive hype cycles. All of them depend on the software actually working.
The Ratio
Most open-source AI projects ship the opposite mix. Features first, fixes later, infrastructure never. The typical ratio in trending AI repos runs closer to 7 features per 3 fixes. OpenClaw just posted 9:1 going the other way.
That discipline is rare. It is also fragile. Feature requests pile up. Contributors want to build new things, not audit compaction guards. Whether OpenClaw holds this posture through the next wave of demand is the question worth watching.
For DeployClaw users, every fix is already live. All ten PRs are linked in the merge log above.
March 21 was the day OpenClaw chose plumbing over paint, and the project is stronger for it.