RepoJournal
Safety Evals Safety Evals
39 wires and counting

$ follow Safety Evals

Keep up with Safety Evals in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-11-06
stories 17

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

Exec_remote orphans and lost hooks fixed, shard logs merge

By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology

inspect_ai shipped three operator-visible fixes that change how cancelled tasks, remote commands and multi-worker runs behave, plus the merge core that turns a sharded run's per-worker logs back into one eval log.

fix: on_task_end hook not called when a task is cancelled or fails (#6273) UKGovernmentBEIS/inspect_ai

by Ransom Richardson

Tasks that end by cancellation or failure now fire on_task_end, so a hook you rely on for teardown, scoring, or cleanup no longer silently skips those runs. The hook was documented to bracket task_run but was only called on the normal completion path, even though every aborted task still wrote a final log.

fix: exec_remote no longer leaves a command running when its start is cancelled or times out (#6275) UKGovernmentBEIS/inspect_ai

by Ransom Richardson

exec_remote no longer leaves a command running in the sandbox when its start request is cancelled or times out. Previously the job's pid was known only from the start reply, so a lost or timed-out reply left a running command with no handle to stop it.

feat: merge a sharded run's per-worker logs into one eval log (#5902) UKGovernmentBEIS/inspect_ai

by Ransom Richardson

A task split across workers writes one shard log per process under <dir>/<name>.shards/<k>/, and the merge core and public API to fold those into a single eval log with whole-task results has landed. This is the first shipped step of the eval-sharding design, so plan on shard-to-log merging being available rather than needing to stitch logs yourself.

fix: exec_remote returns fast commands without a fixed 5 s poll wait (#6313) UKGovernmentBEIS/inspect_ai

by Eric Patey

Fast remote commands return about one request after they exit instead of sitting through a fixed 5 s sleep; the release notes' \"sleep 1\" example drops from about 6.3 s. The sandbox tools server now accepts an optional wait on exec_remote_poll (0 to 60 s, default 0).

fix: read ctl --file stdin with click.open_file (#6336) UKGovernmentBEIS/inspect_ai

by Eric Patey

Smaller and host-side only: the --file - path now reads UTF-8 stdin through click.open_file after click 8.5.0 deprecated click.get_text_stream and broke the asyncio canary's type checks.

Quick answers

What shipped in Safety Evals on November 6, 2026?
inspect_ai shipped three operator-visible fixes that change how cancelled tasks, remote commands and multi-worker runs behave, plus the merge core that turns a sharded run's per-worker logs back into one eval log. In total, 8 commits and 9 pull requests landed.
Who contributed to Safety Evals on November 6, 2026?
2 developers shipped this update, including Ransom Richardson and Eric Patey.
What were the notable Safety Evals updates?
fix: on_task_end hook not called when a task is cancelled or fails (#6273), fix: exec_remote no longer leaves a command running when its start is cancelled or times out (#6275), and feat: merge a sharded run's per-worker logs into one eval log (#5902).