pass^k, not pass@k

Most benchmarks ask if your agent succeeded once. thaghr's headline metric asks if it succeeds every time, the question that actually matters once real users depend on it.

Fault injection at the transport layer

No framework adapter to write. thaghr intercepts HTTP calls underneath any SDK, deterministic under a seed, and faults never touch your real API bill.

Gate CI on reliability

--fail-under turns a campaign into a build gate. Ship a regression in your agent's reliability, and the pull request fails before it merges.