$ the-wire · showcase
Viewer plays terminal sessions, Kimi K3.1 Pro lands in 0.3.284
By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology
inspect_ai 0.3.284 ships support for Moonshot's Kimi K3.1 Pro and K3.1 Pro Highspeed plus an in-viewer asciicast player, while a fix for selective .eval reads addresses logs that currently fail to open at all.
chore(main): release 0.3.284 UKGovernmentBEIS/inspect_ai
Release 0.3.284 is tagged on PyPI and npm with the Moonshot model support included, so upgrading is the shortest path to both changes below. The publish jobs wait on environment approval, so the tag may lead the artifacts by a short window.
feat(moonshot): support Kimi K3.1 Pro and K3.1 Pro Highspeed UKGovernmentBEIS/inspect_ai
`moonshot/kimi-k3.1-pro` and `moonshot/kimi-k3.1-pro-highspeed` now resolve to their own entries; previously both lookups fuzzy-matched the Kimi K3 record and reported the model as "Kimi K3" in your results. If you route evals to Moonshot, model names and per-model reporting for these two endpoints are now correct.
feat: viewer plays asciicasts stored in the sample store under asciin… UKGovernmentBEIS/inspect_ai
Tools that record terminal sessions (for example `inspect_ranges`' `terminal()`) can write asciicast v2 documents into the sample store under `asciinema:`-prefixed keys, and the viewer's Metadata tab now plays them in a Terminal Sessions card, one player per session sized from the cast header. Those entries are kept out of the Store record tree so they no longer clutter it.
fix: reading .eval logs with exclude_fields failing on very large float values UKGovernmentBEIS/inspect_ai
In an `.eval` log holding a valid JSON number too large for a double, such as `1e309`, `exclude_fields` reads raised `ijson.common.IncompleteJSONError: parse error: numeric (floating point) overflow` while a full read returned `inf`. The streaming readers now fall back to `json` for this yajl2_c error, matching the existing fallbacks for NaN/Infinity and integers above 2**63 - 1.
test: move tests and docs off retired OpenAI models UKGovernmentBEIS/inspect_ai
OpenAI retired `gpt-4`, `gpt-3.5-turbo`, `o3-mini`, `o4-mini` and `o1-pro` on 2026-10-23 and the API now returns 404 `model_not_found` for them; live tests and docs examples move off those IDs. If any of your inspect_ai tasks or copied doc examples still name them, they will fail the same way.
Action items
- → Upgrade inspect_ai to 0.3.284 to resolve moonshot/kimi-k3.1-pro and kimi-k3.1-pro-highspeed as their own models UKGovernmentBEIS/inspect_ai [plan]
- → Replace gpt-4, gpt-3.5-turbo, o3-mini, o4-mini and o1-pro in inspect_ai task and example configs before the next run UKGovernmentBEIS/inspect_ai [plan]
