ai-safe2-framework

Continuous Local Evidence

safe2 evidence watch turns the one-shot agent-input change monitor into an explicit local polling runner. It inventories tracked skills and harness configuration, rescans changed skill packages, preserves an immutable report for every run, and atomically advances a validated state file.

# Establish the first explicit baseline and report
safe2 evidence watch . \
  --state .safe2/watch-state.json \
  --evidence-dir .safe2/evidence/changes

# Continue until interrupted
safe2 evidence watch . \
  --state .safe2/watch-state.json \
  --evidence-dir .safe2/evidence/changes \
  --continuous --interval 60

# CI or scheduled bounded run
safe2 evidence watch . \
  --state .safe2/watch-state.json \
  --evidence-dir .safe2/evidence/changes \
  --continuous --max-runs 3 --interval 10 --strict

The first run normally returns hold because all tracked configuration is new. An unchanged validated baseline returns approve; changed agent configuration returns hold; a static skill gate can return hold or reject. --strict preserves the report and then exits nonzero on either review state.

This is detection and evidence preservation, not interception or enforcement. Polling can observe a file only after it reaches the declared filesystem root. It cannot see text pasted only into an agent context, prevent installation, prove authorship or intent, or guarantee that every harness uses the scanned copy. Use a harness pre-install/pre-load hook for prevention and treat this runner as an independent, local evidence layer. It exports no content or telemetry by default.