Skip to content

v0.8.2

Choose a tag to compare

View all tags
@m2dumpling m2dumpling released this 28 Oct 05:13
v0.8.2
ed5ab80
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Pawbot v0.8.2

Published on GitHub
and PyPI.

This release fixes misleading subagent completion messages and adds a small,
reproducible workspace evaluation suite for developers.

Known issue

If the first native TUI download fails because of a TLS certificate or network
error, the launcher can incorrectly report that the archive is not published.
Check the connection and certificate trust before retrying, or temporarily use
pawbot agent --classic. The archive and checksum are available in this release.

Fixed

  • A subagent that reaches an execution limit, returns no answer, or fails a
    requested task verification is reported as incomplete instead of successful.
    Its stop reason and partial result remain available for both synchronous and
    background delegation. The change does not retry work automatically or undo
    side effects already performed by a tool.

Developer evaluation

  • Six controlled workspace tasks exercise file reading, code repair, config
    migration, log analysis, and read-only behavior. Each trial uses a fresh
    temporary workspace, and independent checks inspect the resulting files and
    tool trajectory. Four cases are for development and two are held out from
    tuning; the holdout tasks are public and are not a benchmark.
  • A recorded small-sample run using a real model passed 18 of 18 trials. This
    measures those fixed tasks only, not production accuracy. The workspace
    suite does not exercise subagent delegation; dedicated regression tests cover
    the subagent fix. See the experiment evidence
    for conditions, token usage and limitations.
  • The Promptfoo guide is in the source
    repository and the evaluation directory is not bundled with the wheel.
    Temporary workspaces are not operating-system security sandboxes; execute
    generated code only in an appropriately controlled environment.

Release validation

The release workflow verifies the five native TUI archives and their checksums.
Python distributions are built and tested in a separate prepare run; the
publishing run verifies and uploads those exact distributions. See the
publishing guide for the release and installation checks.

Assets 12
Loading