@pauliusztin_ Agreed. A rename shouldn't expire anything. Only a changed rule should, so a rule_version on each fact beats one ontology_version for all of them.
A release step joined with ; instead of && ignores the CI result it's supposed to wait for. The step is green, the build was red. Agent-written checks often measure less than their name says, and only someone who reads them notices.
https://t.co/Njb4D5Vn4O
#DevOps#AIAgents
@andrewlocknet On building general APIs over CborReader: for real DTOs, would you write that layer yourself, or is Dahomey.Cbor good enough that it's not worth it?
@levie Granular access is the hard part. M365 Copilot honors permissions and still surfaced overshared SharePoint files that'd only been safe through obscurity.
@kdy1dev It's only locked while building. Run `cargo build` once, then start target/debug/api and target/debug/worker directly; process-compose can chain both.
@criccomini Examples beat rules here: paste two real 4-line PR bodies from your repo into the skill. Then a CI check that fails over 120 words, so it's enforced.
A real platform problem looks like Anthropic's Sep 2025 postmortem: a routing bug hit 16% of Sonnet 4 requests at peak, with a start date and a fix. A slow slide across several model releases in one repo isn't that.
https://t.co/AykMlphdl4
#ClaudeCode#AIAgents#CodeQuality
@ericwithcoffee True, one GET to an attacker's host is enough. That's why egress control has to live outside the agent, in an allowlisted proxy like Squid on port 3128.
@RyanRodemoyer2 2,598 serialized writes/s is about 0.38 ms each. If it's SQLite, WAL with synchronous=NORMAL drops the fsync per commit, which is often the whole ceiling.
@Zaneris@davidfowl@shanselman That's the hard road. When a rude edit like a changed signature can't go through ApplyChanges, does Passenger prompt a restart like VS, or just relaunch?
@zachlloydtweets That's throughput. What did revert rate do over the same 3 months? DORA 2025 still found AI adoption hurting delivery stability, so I'd want both side by side.
@_iamperegrine_ Fair. It's measurable per use case though: plot first-stage recall@k for k=20, 50, 100 and the curve tells you where reranking more candidates stops paying.
GitClear: moved lines, their refactoring signal, fell from 25.0% in 2020 to under 10% in 2024, and copy/paste passed moved code for the first time. It's a correlation with AI adoption, not proof of cause.
https://t.co/hkGfv6iB54
#TechnicalDebt#CodeQuality#AIAgents
@thedelost One catch on the Stop hook: exit 2 makes Claude keep going, so check stop_hook_active in the hook input or a suite that can't pass loops forever.