Mandatory verification checklist that every agent MUST load before closing a governance cycle (end_change). Referenced by AGENTS.md as a pre-close requirement.
Mandatory verification checklist that every agent MUST load before closing a governance cycle (end_change). Referenced by AGENTS.md as a pre-close requirement.
Load this skill (skill_view(name="change-checklist")) before calling end_change(). Not optional.
Run Phase 5 (Final Verification) before Closing the Cycle. Do not proceed to Closing until every Phase 5 item passes, including the adversarial scan.
Adversarial scan is not optional for code changes. Run change-checklist Phase 1.5 before closing: python3 ops/scripts/quality/adversarial-verify.py --file <changed-file> --level A2 --gate for every changed script (A4 for security/guard/hook/enforcer paths). Must exit 0. If it found findings, fix them. Do not close with findings. "0 findings" from static scan is NOT a pass — also execute the changed path against boundary inputs and attack its implicit assumptions.
Pre-Work (Before begin_change)
mcp_loop_governance_cache_search(query="<what you are about to do>") — learn from past cycles
Read this repo's AGENTS.md and SOUL.md for agent-specific rules
Run mcp_loop_governance_config_show() to know current thresholds
Phase 1 — Tests
Run full test suite (pytest tests/ -x --tb=short)
All tests pass (report exact count)
If tests fail: fix before closing cycle. Never close with failing tests unless user explicitly waives.
Phase 1.5 — Adversarial Verification (MANDATORY — no bypass)
Run the static gate on every changed script file: python3 ops/scripts/quality/adversarial-verify.py --file <changed-file> --level A2 --gate
A4 is MANDATORY for security/guard/hook/enforcer files (anything under plugins/, hooks/, mcp-servers/, ops/scripts/manage/, ops/scripts/cortex_doctor/, ops/scripts/quality/, tests/, plus pre-commit-score and cortex-update.sh)
Critical/high findings → block: fix before proceeding. No --no-verify.
Verifier step — "0 findings" is NOT a pass: execute the changed path with boundary inputs (-1, 0, None, empty, whitespace, inf/nan, non-ASCII) and attack the code's implicit assumptions (deployed file == loaded module, input already validated, etc.)
Record what was checked: finding IDs, boundary inputs run, assumptions violated, exit codes
Phase 2 — Multi-OS Compatibility
If paths changed: verify no Linux-only paths (/home/, /etc/) used on macOS, and no macOS-only paths used on Linux
If service files changed: verify LaunchAgent plists for macOS AND systemd service units for Linux
If scripts changed: verify shebangs point to python3.12 (macOS ships Python 3.9 which breaks PEP 604; all Hermes projects require 3.12+)
~/.local/bin/python3.12 is the canonical macOS Hermes Python
Phase 3 — Multi-Role Verification
Agent perspective: does the change work via tools (MCP, terminal)?
User perspective: does the change behave correctly when the user runs it?
If cron: verify the job creates/updates correctly
If config: verify the config is valid (YAML/JSON syntax, nginx -t, etc.)
Phase 4 — Documentation
Did the change affect any public interface? Update docs/ or AGENTS.md accordingly
Does the change need a CONTRIBUTING.md update?
Is there an ops/ script or config that should be documented in docs/operations-reference.md?
Phase 5 — Final Verification
Run cortex-doctor.py or equivalent health check
Adversarial scan passed — MANDATORY for every changed script (change-checklist Phase 1.5): python3 ops/scripts/quality/adversarial-verify.py --file <changed-file> --level A2 --gate per file (A4 for security/guard/hook/enforcer paths). Must exit 0. The pre-commit hook runs this gate too — no --no-verify, no bypass.
Verify all services restart cleanly (launchctl list / systemctl status)
Verify symlinks are valid
Verify no stale paths remain (grep -r '/home/luke\\|${CORTEX_HOME}\\|${CORTEX_DEPLOY}' on changed files)
git add <files> && git commit -m "<descriptive message>" — through the pre-commit hook (it runs the adversarial gate + scoring; --no-verify is a logged, audited bypass — never use it to ship a hook/gate-rejected change)
git pull --rebase origin main && git push origin main
If Scoring Is Blocked
Run bash ~/.hermes-cortex/tools/loop-governance/verify.sh if it exists, or python3 -c "from loop_db import LoopDB; db = LoopDB(); print(db.get_summary_stats())"
If fix takes > 2 min, record manually later
Never skip entirely — governance-auditor cron flags unscored changes
SKIP_SCORE=1 has been removed — the bypass no longer exists. git commit --no-verify is a logged, audited bypass — fix the blocker and commit through the hook instead.
Install
To install this skill (after cloning hermes-cortex):
