High-priority alert based on source significance and keyword relevance.
Context
What happened: A developer released agent-usage-guard, a Claude Code mod that reads live plan usage and holds subagent fan-out, heavy-context churn, and retry loops before spend occurs.
Key numbers: Defaults include approval at 80% of a plan window and refusal at 95%, 4 concurrent agents, 12 agent starts per 10 minutes, 10M subagent tokens per 10 minutes, and refusal after 3 identical tool
Why it matters: The guard targets the multiplied failure mode of large context, rapid tool loops, parallel agents, and blind retries, and it fails open so a broken hook does not break the session.
What changed
A developer released agent-usage-guard, a Claude Code mod that reads live plan usage and holds subagent fan-out, heavy-context churn, and retry loops before spend occurs.
Why it matters
The guard targets the multiplied failure mode of large context, rapid tool loops, parallel agents, and blind retries, and it fails open so a broken hook does not break the session.