Skip to content

v0.29.0: Kilo Code, MiniMax Code, Grok Build and Agent Zero

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Oct 00:33
v0.29.0
2826cf6
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Four new harnesses

HarnessRouter 0.29.0 adds four agent harnesses, bringing the open-source edition to twenty. Each installs on first start, runs behind the same Responses API, and takes the same plugins, skills, MCP servers, tool policy and browser as every other base.

Harness Base id Upstream License
Kilo Code kilo Kilo CLI 7.8.1 MIT
MiniMax Code minimax MiniMax Code 0.5.4 MIT
Grok Build grok Grok Build 1.0.41, by SpaceXAI Apache-2.0 source, SpaceXAI's own binary
Agent Zero agentzero Agent Zero v2.13 MIT 
docker run -d -p 3000:3000 -v hr-data:/data harnessrouter/harnessrouter:0.29.0

All four are in the default HR_BACKENDS. Agent Zero is the large one (about 640 MB on the data volume); leave it out of HR_BACKENDS if you do not want it.

Measured

Every scenario was driven through the console on a self-hosted instance. Full model catalog, five scenarios per harness and model (first turn, follow-up, model switch, file artifact, recall after a sandbox recycle):

Harness Models Scenario runs passed One conversation across every model family Plugins, custom harness, browser
Kilo Code 59 291 of 295 13 of 13 pass
MiniMax Code 57 283 of 285 14 of 14 pass
Grok Build 56 279 of 280 14 of 14 pass
Agent Zero 56 267 of 280 14 of 14 pass

The rows are in docs/support-matrix.md and the account of every failure in docs/support-matrix-notes.md ("Four bases in one release"). A model that failed twice the same way on a harness is not offered on it.

Fixes that reach every harness

  • A conversation handed from one model to another no longer breaks on what the earlier model left behind: an image in a tool result, a tool call with an empty id, or two tool calls with one id.
  • A tool whose schema has oneOf, anyOf or allOf at its root now works on OpenAI models as well as Claude.
  • A provider that stops answering mid-stream is reported as a failure with its reason, not as an empty answer.
  • Per-turn configuration that can carry MCP credentials stays out of session checkpoints, and OpenCode's direct Anthropic turns go through the relay, so the agent never holds the connection's key.

Thanks

The four harnesses and the suite work behind them are by @ZixiaoL (#362, #363, #364, #365, #366), landed together in #373.

Full Changelog: v0.28.7...v0.29.0

Contributors

ZixiaoL
Assets 2
Loading