felmonon/Agent Reliability Harness

Validate agent tool-use traces against a reliability policy and (optionally) compare against a saved baseline to gate the workflow on real regressions. Runs fully locally: no telemetry, no network calls, no secrets required.

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Stale
License
None
namedescriptionrequireddefault
policyPath to the policy JSON fileyes
tracesSpace-separated trace file paths or globsyes
formatTrace format: auto | arh | openai-chat | anthropic-messagesnoauto
baselineOptional path to a baseline JSON report (from a previous 'arh validate --json-out'). When set, the action runs 'arh compare' and gates on regressions instead of absolute failures. no""
fail-onGate mode for compare: regressions | failures | nevernoregressions
max-score-dropOptional per-trace score-drop limit for the regressions gateno""
fail-underMinimum passing score for validate (0-100)no70.0
output-dirDirectory for report artifactsnoarh-reports
namedescription
json-reportPath to the JSON validation report (next baseline)
sarif-reportPath to the SARIF report (upload with codeql-action/upload-sarif)
junit-reportPath to the JUnit XML report
compare-reportPath to the comparison JSON (only when baseline was given)