1.8.8b12 — the agent cannot read the answers

An opt-in preview. Stable remains 1.8.7; the Personal Wiki becomes long-term supported in 1.9.0.

A last walk through 1.8.8b11 as published confirmed the second round of fixes live: stuck browser reads end at their deadline and name the Keychain, a device that joins right after the host starts sees the next turn, a host restart gives TurnLostError, and co wiki start asks again after its runner changes. It also found one thing that made a new feature's results untrustworthy.

Benchmarks you can trust

• During co eval run the agent under test searched the project, found .co/benchmarks/<name>.yaml and read the expected outcomes — in four of five cases on the co create template. A tool call that names the benchmark or earlier eval runs is now refused for the length of the run, and an attempt whose tool results contain the expectations anyway is marked INVALID: not judged, never a pass, and the run exits 1 (#1712).
• co benchmark list > example.yaml now captures only the YAML.

Smaller promises kept

• co browser tab ls with no browser running says so and starts nothing; a command against a frozen daemon gives up in about 40 seconds; a forced close removes the socket and lock files it leaves.
• The paid browser engine keeps its profile inside CO_BROWSER_PROFILE_DIR (in onion/) when that is set, instead of the real home directory.
• co wiki show me shows your page; each notebook's schedule has its own launchd label, so one notebook's stop cannot remove another's job.
• co create . suggests a real name. docs/network/host.md says to sign in before the Quick Start and that the default trust level admits contacts only.

Still open

• The web client (@connectonion/react 0.4.4-rc.6) must send request_id before a session open on two browsers can answer approvals.
• co auth status still signs you in, as on 1.8.7; its fix (#1693) is in review.

Install

python -m pip install --upgrade 'connectonion==1.8.8b12'
co --version