Prompt file imported from redhat-community-ai-tools/harness-eval (
.cursor/commands/harness-autonomy.md). Copyright stays with the author.
Harness Autonomy
Run the decidable checks an auto-merge policy can trust: block and policy rules only, no heuristics, no LLM. Exit 0 PASS, 1 FAIL, 2 REVIEW_REQUIRED.
Instructions
- Run the check on the current project:
uvx --from harness-eval harness-eval harness-autonomy . --format json
If a trusted policy file exists (commonly .harness-eval/autonomy-policy.yaml, protected by CODEOWNERS), add --policy <file>. Never edit that file to make a run pass.
If uvx is not available, fall back to pip install harness-eval and use harness-eval directly.
-
Report, in order: the verdict and exit code; blocking findings (defects to fix); policy findings (facts a person or the policy file must accept); waived findings and their recorded reasons; coverage (components, rules run, rules not applicable and why, skipped layers, baseline suppressions); and the evidence block (fingerprint, revision, rules digest, config digest).
-
Do not soften a FAIL or upgrade a REVIEW_REQUIRED. The value of this check is that it is reproducible.
