v0.14.0

0.14.0 (2026-10-24)

⚠ BREAKING CHANGES

• move questions to the eval's root and ask them of each attempt (#543)
• gateway: vercel() for Claude Code and Codex, and gateway-routed judges (#503)
• egress: default to 'proxy', which sends everything through the egress gateway (#501)

Features

• checks: retry c.exec on a non-zero exit with retries and interval (#542) (311cd92)
• cli: estimate cost as a range in reeval run --dry-run (#519) (4e18100)
• egress: default to 'proxy', which sends everything through the egress gateway (#501) (0cc02cb)
• gateway: add the gateway list with direct(), plan-time resolution and pins (#502) (22e4d5a)
• gateway: run Goose, OpenCode, fx and Grok Build on Vercel AI Gateway (#504) (4d89bed)
• gateway: vercel() for Claude Code and Codex, and gateway-routed judges (#503) (9128d66)
• language-runner: add a record call for checks made in the eval's own language (#511) (c0dfacf)
• move questions to the eval's root and ask them of each attempt (#543) (c5f0ab7)

Bug Fixes

• engine: keep a worker up until the main process has its last message (#524) (cfa2fd5)
• install: detect musl from the libc in use (#522) (d9990e1)
• install: keep the user PATH REG_EXPAND_SZ on Windows (#523) (a12473b)
• judge: explain decision-model refusals, follow the attempt's judge in captures and scale Gemini clicks (#544) (d2f9bc9)
• language-runner: keep a remote body's checks in its attempt record (#507) (9f094e0)
• python: name the wire types' schema sources with POSIX paths on every host (#528) (eee7c3d)
• ui: count a study as not gated, never pass (#541) (c2442eb)
• usage: count reasoning tokens once, as part of output (#525) (4029d3d)