All notable changes to Harness² CLI are documented here. The format follows Keep a Changelog, and the project uses Semantic Versioning.
The first release of Harness² CLI, a general-purpose version of
google-research/harness2 (forked at
1cfe254, with upstream's fixes merged through 6868817). The method and its prompts are
unchanged; everything around them is new.
- One
harness2command with subcommands:init,doctor,task,run,status,stop,diff,eval,apply,judge,report,runs,config,chat,guide,login,setup,demo,bridge,completion. - Your own tasks:
.harness2/tasks/<id>/task.mdwith optionaltask.toml, run in a clean git worktree of your repository or in a scratch directory, graded optionally by acheckcommand or arubric.md. - Claude Code and command executors. A project's own harness (CLAUDE.md,
.claude/skills,.claude/agents,.claude/scripts, hooks; or AGENTS.md,.agents/,.codex/hooks.json) is imported as the starting harness. - Subscriptions:
claude-login/<model>runs on Claude Code's own login andchatgpt-login/<model>on a Codex login kept for harness2 (harness2 login codex), with personal settings, hooks, plugins and instructions kept out of the runs. - Any model provider: Anthropic, OpenAI, OpenAI- or Anthropic-compatible servers, Vertex AI
and Bedrock, through provider profiles;
harness2 bridgeserves any of them to Codex. - Layered settings: defaults,
~/.config/harness2/config.toml,.harness2/config.toml, profiles,HARNESS2_<SECTION>__<KEY>variables and flags, withconfig show --origin. - One final harness per run, built from every task's adopted harness and lesson;
harness2 evalcompares it with the starting harness, andharness2 applywrites it into the project, keeping edits made since the run started. - Run control:
--dry-runcall estimates,--budget-usdand--max-callslimits, clean stops at a provider's usage limit, Ctrl-C andharness2 stopwith resume,--detachbackground runs, relocatable run directories,events.jsonl. - Use by AI agents:
--jsonfor every command including errors, flags for every prompt, andharness2 guide(also installable as a Claude Code skill). - Interactive use: the
initwizard,status --watch,apply -i,--selector human,harness2 chat. - CI on Linux and macOS with Python 3.12 and 3.13.
- The distribution is
harness2-cli; the command and the import package stayharness2. - The agent CLIs are the ones installed on the system instead of pinned downloads.
- Codex harness scripts live in
.agents/scripts/, so they never collide with a project'sscripts/.
- The seven
harness2-*scripts' old flags (--bench,--substrate); the script names remain as aliases of the new subcommands. - The Vertex-only configuration in
.envfrom the core package. JobBench, LAB and WorkBuddy-Bench moved, with that configuration, to the harness2-benchmarks plugin.
Merged from upstream:
- A finished task or rollout is reused only while its prompt and files are unchanged.
- A failed improver session is reported as an error and retried on the next run.
- A partly finished task resumes the modes that did finish.
- A re-run that fails clears that mode's old selection, and a failed re-judge keeps a score only for the same selection.
- A grade without a pass count stays ungraded instead of counting as zero.
- A rejected harness edit keeps its files for the retry.
- An OpenCode session error is no longer cached as an answer.
- Benchmarks plugin: JobBench isolates each rollout's installs, temporary files and workspace, notes a non-zero exit in the trajectory, and leaves a judge worker failure ungraded; LAB rejects conflicting judge verdicts and keeps relative artifact paths for grading.
- Agents no longer inherit provider API keys, AWS credentials or host tokens such as
GITHUB_TOKENfrom the environment; each keeps only its own provider's key (merged from upstream). See Secrets.