Skip to content

v0.211.1

Choose a tag to compare

@github-actions github-actions released this 20 Oct 01:11
· 100 commits to main since this release
ed0884e

protoAgent v0.211.1

A patch release of five fixes. /goal can now be verified by what the agent's tools did. Image generation, transcription and embeddings reach the gateway you set up in the wizard. vega-lite charts stop reporting a false "Script error". An answered form's reply no longer appears twice after a reload. The chat keeps following the stream when tall cards grow.

Upgrade notes

  • No action needed for the gateway fix (#4400). If an earlier version wrote the App default http://gateway:4000/v1 into host-config.yaml, the next config reload replaces it with the endpoint your chat model actually uses.

Fixed

  • The /goal verifier sees the turn's tool calls (#4399). It was always handed Recent tool calls: (none), so a goal proven by tool results (write a note, then read it back) never passed and looped until it was marked unachievable. It now gets each tool call of the current turn paired with its result. A verifier reply made of content blocks (text plus thinking) is no longer misread as an evaluator error.
  • Image generation, transcription and embeddings use the setup wizard's gateway connection (#4400). They read only the legacy model.api_base, which still pointed at the App default, so they failed with ConnectError while chat worked. They now use the same connection as the chat model, and the egress allowlist covers it.
  • vega-lite charts render on the first try (#4401). The chart frame's CSP refused the chart stylesheet, and that refusal was reported as "Script error" with a failed render verdict. The chart had drawn fine, but check_artifact reported a failure, so the agent retried and then fell back to Chart.js. The stylesheet is now allowed, and a stylesheet or image that fails to load no longer counts as a render failure.
  • An answered agent form no longer doubles the reply after a reload (#4403). Only the turn after your answer carries the agent's reply. The answer's labelled card text is kept in the server transcript, so a chat opened in another browser shows Plan: Plan C instead of the raw resume JSON.
  • The chat keeps following the stream when cards grow (#4404). Tall cards (a diagram, a chart, an inline artifact) used to leave the view parked on the first card, with later turns landing off-screen. Following now stops only when you scroll up yourself. The jump button lands at the real bottom, and every send brings you back to the latest turn.

What's Changed

  • fix(artifact): vega-lite charts no longer fail check_artifact with "Script error" by @mabry1985 in #4401
  • fix(chat): an answered form's reply no longer doubles on reload; the card's text reaches the durable transcript by @mabry1985 in #4403
  • fix(goals): the llm goal verifier sees the turn's tool calls by @mabry1985 in #4399
  • fix(llm): image gen / transcription / embeddings reach the wizard's gateway connection (B1) by @mabry1985 in #4400
  • fix(chat): the transcript keeps following the stream through late card growth; a send always re-pins (B6) by @mabry1985 in #4404

Full Changelog: v0.211.0...v0.211.1