Skip to content

v4.2.0 — instruction precedence, verification calibration & deploy CLI safety

Choose a tag to compare

View all tags
@oleg494 oleg494 released this 06 Sep 06:50
· 22 commits to master since this release
v4.2.0
09c9803

What's Changed in v4.2.0

Instruction Precedence & Authorization Gate

  • Clarified instruction hierarchy in AGENTS.md and OPS.md (Host system/developer instructions > User instructions > Kit skills). When skill guidance conflicts with user directions, agents follow the user and state the departed skill.
  • Replaced the brainstorming unconditional HARD-GATE with an AUTHORIZATION-GATE: agents proceed directly on authorized reversible local work without halting for redundant re-approval. Pauses remain reserved for irreversible, external, destructive, money/auth/privacy actions, or outcome-changing ambiguity.

Verification Scope Calibration

  • Calibrated verification discipline across superpowers, SKILL_RUNTIME, test-driven-development, dispatching-parallel-agents, verification-before-completion, and testing-discipline.
  • Scoped checks appropriate to the touched components by default; full-suite re-runs on untouched components are no longer required for bounded changes, and unrun suites are never claimed.

Deploy CLI Safety Boundary

  • Added argparse CLI handling to scripts/tools/deploy.py (--canonical, --dry-run, --help).
  • Commands like python scripts/tools/deploy.py --help now safely display usage without triggering unintentional host deployments or router modifications.
  • Added comprehensive regression coverage in tests/test_deploy_cli.py.

Tooling & Runner Resilience

  • Stdlib fallback: Restored full regex-based YAML frontmatter fallback parsing in memory/db-tools/lint_wiki.py and scripts/doctor.py so CI runners and fresh installs without PyYAML pass all hygiene and specification tests cleanly.
  • Cross-drive testing: Handled Windows cross-drive working directory changes (cd /d) in tests/test_findings_lifecycle.py.
  • Backup reliability: scripts/tools/backup_memory.py now exits with status 1 when any database snapshot is skipped.
  • Integrity manifest: Updated byte-level hashing to preserve invalid UTF-8 byte distinctions while maintaining newline equivalence.

Evals & Traps Expansion

  • Expanded trap-suite from 24 to 26 scenarios with authorized-work-proceeds and calibrated-testing.

Licensing & Attribution

  • Preserved license: Proprietary on skills/windows-encoding-fixes with explicit, honest disclosure in README.md (no false claims of a 100% MIT bundle).

Full Commit Diff: v4.0.3...v4.2.0

Assets 2
Loading