$ the-wire · showcase
Inspect 0.3.282 scopes tool state per agent; sentinel gains step history and human decisions
By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology
Inspect 0.3.282 makes concurrent agents in a single sample keep separate tool state, and the sentinel host side now surfaces history, human decisions and escalations as step and host parameters, which changes the ask_human signature for anyone running inspect_sentinel.
v0.3.282 UKGovernmentBEIS/inspect_ai
Tool state is now scoped so agents running concurrently inside one sample keep separate tool state; previously they shared it. This is a breaking change: published as v0.3.282 alongside conditional S3 writes and a local lock helper for sharding. If you run parallel agents within a sample, pin or upgrade deliberately and re-check results from runs that relied on shared state.
fix(sentinel): human() shows the escalations inspect_sentinel now passes to ask_human UKGovernmentBEIS/inspect_ai
With meridianlabs-ai/inspect_sentinel#97 installed, escalations no longer live on step payloads, so the host's old ask_human(step, choices) read of step.escalations breaks. human() now passes them as Host.ask_human(step, escalations, choices), and a protocol receives earlier sequential() links' escalations as a third parameter.
feat(sentinel): step.history holds every message the agent sent; the agent loop's list is step.agent_messages UKGovernmentBEIS/inspect_ai
Every sentinel step now carries history: every message its agent instance sent to the model, in the order sent. The field previously called history is now agent_messages, and sent is now history, a breaking rename for sentinel 0.2.0, though inspect_ai's own API does not change.
feat(sentinel): steps carry a person's earlier human() decisions in the conversation UKGovernmentBEIS/inspect_ai
Steps now fill step.human_decisions, so a monitor rendering the conversation into its prompt can see which past tool calls a person already reviewed through human() instead of flagging them again and re-asking the operator. step.escalations still covers only the current step.
docs: clarify responsibility for Docker security updates (#5859) UKGovernmentBEIS/inspect_ai
The docs now state that Inspect's version checks confirm Docker feature support but do not verify security updates, and make explicit who owns patching Docker, so unpatched host vulnerabilities stay on your side of the line.
Action items
- → Upgrade inspect_ai to 0.3.282 and verify tool state isolation if you run concurrent agents in one sample UKGovernmentBEIS/inspect_ai [plan]
- → Move inspect_sentinel host integrations to ask_human(step, escalations, choices) before installing sentinel with #97 UKGovernmentBEIS/inspect_ai [plan]
- → Rename step history to agent_messages and read the as-sent list from history before sentinel 0.2.0 UKGovernmentBEIS/inspect_ai [plan]
