OpenAI: your Sol-era AGENTS.md and skills now slow Astra down
OpenAI's official fix: cut the prescriptive rules, make skills route instead of lecture, and define 'done' up front. Plus a no-bloat LiteLLM alternative.
Copy markdown
Why your old config now backfires
Astra handles nuance better, so the elaborate recipes, blanket 'read architecture.md before every edit' rules, and always-run test mandates you wrote for Sol and Luna now trigger premature stops and wasted passes. OpenAI's guidance: delete them and let Astra decide what to read and when to verify.
Make skills routers, not manuals
Swap long skill descriptions for narrow triggers -- 'Use when adding or changing a migration' -- that point to docs loaded only when needed. Fat descriptions get truncated by context compaction, so an over-stuffed skill can go invisible mid-run.
Define 'done' before you hit go
For long autonomous tasks, spell out the loop up front -- implement, run, inspect, fix, verify -- instead of vague 'keep going' notes. And loosen permission language: Astra is aligned enough that Sol-era guardrail phrasing makes it halt too early.
Audit for conflicts first
Before editing anything, ask Astra to flag where your AGENTS.md and skills clash on autonomy, approval, and completion. Overlapping or contradictory rules are the top reason an agent stops when you wanted it to finish the job.
Elsewhere: a no-bloat gateway, cheaper DeepSeek cache
'litelm' -- a stripped-down LiteLLM alternative -- hit the Hacker News front page today for teams who want a proxy without the sprawl. And DeepSeek V4.1-Flash reportedly cleared 1 trillion tokens in its first 24h on OpenRouter, with cache reads near $0.006 per million -- cheap fuel for context-heavy agent loops.