Skip to content

v0.40.1

Choose a tag to compare

View all tags
@github-actions github-actions released this 09 Nov 12:05
v0.40.1
6e64853
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Bug fixes: tool switches, instructions on a continued task, one memory per agent, and new setups on Harness Studio.

Upgrading

Nothing to do. Existing setups keep their tool switches; they now take effect as the console describes.

Fixed (#659)

  • Tool switches (Setup Settings › Tools). A switched-off tool is now withheld from the model through each agent's own setting:

    • Qwen Code: permissions.deny. Its background memory, dream and skill agents go off while a file or shell tool is off.
    • Gemini CLI: tools.exclude.
    • Cline: disabledTools. apply_patch goes with Editor, and Cline's team tools go off with any tool, since a teammate starts with every tool.
    • Codex: its shell tool, through features.shell_tool.
    • Hermes: whole toolsets.
    • dsh: single-tool rows.

    Where an agent can only be asked not to use a tool, the switch now says so: "This agent can't switch it off. Off asks it not to use it, and it may still use it."

  • Claude Code writes its tool settings fresh every turn, so a tool switched back on is available again on the next task.

  • An edit to a setup's instructions reaches a task you continue. Hermes, fast-agent, Kimi Code, Grok Build, MiMo Code, dcode, OpenHands and Qoder used to keep the instructions a conversation started with. Hermes and fast-agent now rebuild from the current instructions. The others are given the current instructions with the next message, marked as replacing the earlier copy. This happens only when the instructions actually changed.

  • One memory per agent. With this instance's memory attached, Claude Code's auto memory and Qwen Code's managed memory no longer keep a second, private memory beside it. Codex's memories and Gemini CLI's experimental auto memory are off too.

  • Harness Studio shows a setup built while you were away. If you left the build page before the build finished, the new setup now appears on Harness Studio a few seconds after it's saved, without a reload.

  • Codex: an option Codex no longer reads is no longer written into its configuration.

Checked

Real CLI turns inside the release candidate's image against a recording stand-in model (scripts/tools-off-check.py, scripts/continued-instructions-check.py), compared with 0.40.0:

  • With every tool switched off:
    • Qwen Code, Gemini CLI and Cline offer none of them.
    • Codex, Hermes and dsh offer only what they can't switch off, and the agent is asked not to use it.
  • An edit to the instructions reaches the next turn of a continued task on every agent, and the conversation still remembers its first turn.
  • With this instance's memory attached, Claude Code and Qwen Code keep no second memory.

Real turns on a running instance:

  • the memories end-to-end check, 24 of 24;
  • the five usual scenarios (first turn, follow-up, model switch, artifact, recycle) on Codex, OpenCode, Gemini CLI, Claude Code, Pi, Qwen Code and Roder;
  • a setup built while you're on another page shows on Harness Studio a few seconds after it's saved.

Release gate: on a fresh volume on amd64 and arm64, every default harness installs, a second start installs nothing, and the credential canary is clean. The release's tree is identical to the candidate's.

Full Changelog: v0.40.0...v0.40.1

Assets 2
Loading