Install Harness 1.5.1¶
Harness 1.5.1 is published on PyPI. Use it to plan work, inspect evidence and return to saved progress. The browser companion helps you save and review requests; the CLI handles execution under its separate scope and approval rules.
Use Python 3.10 or later and a fresh environment. Keep Harness separate from Attune AI, which pins a different MCP SDK version:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install "attune-harness==1.5.1"
python -c "from importlib.metadata import version; print(version('attune-harness'))"
python -m attune_harness
attune-harness --help
Expected result: Version 1.5.1, a verified deterministic arithmetic receipt,
and the command catalog. The example makes no model call and measures no model
quality. For a managed CLI, you can instead use
uv tool install 'attune-harness==1.5.1', then inspect uv tool list --show-paths.
On Windows, use py -3 -m venv .venv, then run Python commands with
.\.venv\Scripts\python.exe. For example:
.\.venv\Scripts\python.exe -m pip install "attune-harness==1.5.1"
.\.venv\Scripts\python.exe -m attune_harness
This is reference syntax, not a new Windows shell qualification. Keep task records outside checkouts used for file effects and use canonical absolute paths.
Choose a useful next step¶
- Get started: check your installation and obtain the release examples.
- Browse all goals: choose from 21 goals and six testing FAQs, with prompts and review points.
- Start and continue work: save answers and review the request in a browser form.
- Plan and accept work: inspect scope and acceptance criteria before execution.
- Research a question: retrieve local sources and check a supported link claim.
- Save and resume: return to saved answers and inspect work before continuing.
Preparing a request, saving answers, accepting intent and authorizing execution are separate decisions. The goals worksheet records proposed wording; it does not accept a spec, dispatch a model or implement the example app. Installing the runtime does not install a host plugin or skill. Follow the selected host's setup instructions separately.
Attune AI is deprecated. Use Harness for new work. Before changing an existing setup, read the current migration guide. Preserve working environments while validating the journeys you need. Deprecation does not establish saved-work migration or complete host or Windows parity.
Work on Windows within the documented limits¶
Windows users can install Harness and use its documented CLI workflows and browser forms. Release qualification passed the hosted Windows Server source and installed-wheel checks. Forms save and review intent; they do not build. Native planning and building remain experimental.
Full Windows parity remains deferred. Hosted Server checks do not qualify Windows 11 desktop, all shell journeys, complete storage/recovery acceptance or live native-memory reliability. The saved-store contract still uses a two-second acquisition limit, eight writers and one save per caller; durability barriers and qualification assertions remain intact. A passing software check does not close those broader acceptance gaps. Read the release notes for the supported components and remaining limits.
Inspect completed assessment evidence¶
For an existing completed document assessment, replace TASK with its saved
task directory and inspect:
attune-harness status TASK --format markdown
attune-harness status TASK --format html > assessment-report.html
Open the generated HTML report. Both formats present the accepted goal and criteria, checker outcomes, retrieved-source identities and separately attributed participant judgments. A successful status command does not mean the assessment passed. Inspection does not rerun checks, refresh evidence or establish semantic acceptance. Other profiles and incomplete assessments keep their existing presentation. This addition is absent from published 1.5.0.
Read the inspection procedure.
Check the release and its evidence¶
Published 1.5.1 uses source commit
ebb5b03b8a4398325df7ddabd28c08186451452e. The release record
names the public wheel and sdist hashes, qualification and publishing runs, and
the scope of the post-publication installation checks. The released changelog
records the maintenance changes. Software and installed-artifact checks do not
establish model quality, independent user acceptance or full platform parity.
Keep older evidence in context¶
The 1.5.0 guide and verification record retain their original environment, walkthrough outcomes and limits. Goal screenshots remain explicitly 1.5.0 observations. Illustrative outputs remain labeled, and saved example wording is unchanged.
The historical preparation record and preparation guide retain the candidate artifact and its open gates at that time. They are historical records, not the current installation procedure. Historical 1.3.0 and candidate tutorials remain available in the repository under their original identities.