two-hard-rules

name	two-hard-rules
version	1.0.0
category	devops
description	Two hard rules every agent must follow: USE LOOP GOVERNANCE ALWAYS. SHARE IMPROVEMENTS TO THE PUBLIC REPO.
author	Hermes Cortex
pinned	true
metadata	{"hermes":{"tags":["governance","rules","discipline","compliance"],"related_skills":["agent-contract","loop-governance","change-checklist","public-contribution"]}}

⚠️ TWO HARD RULES — Every Agent Must Follow

RULE 1: USE LOOP GOVERNANCE ALWAYS

Every code, config, or cron change REQUIRES this sequence — no exceptions:

1. mcp_loop_governance_cache_search(query="<what you are about to do>") — learn from past
2. mcp_loop_governance_begin_change(task_id="<short-name>", description="<what this does>") — create lock
3. Make your changes (patch, write_file, cronjob, etc.)
4. mcp_loop_governance_cycle_query(task_id="<descriptive-name>") — find the cycle
5. mcp_loop_governance_feedback_accept(id=N) or feedback_override(id=N, correct_decision=...) — score
6. mcp_loop_governance_end_change(task_id="<short-name>") — release lock

If end_change rejects (no cycle auto-created, or cycle not scored — 2026-08-08):

• end_change now BLOCKS while the task's cycle is unscored, and begin_change refuses a new task while earlier PENDING cycles exist (close-out enforcement, Luke directive). Score first: cycle_query → feedback_accept/override → end_change, THEN the next begin_change.
• If genuinely stuck: confess clearly: "end_change rejected — no cycle auto-created. Force-clearing lock."
• rm -f ~/.hermes-cortex/state/.governance-active.json
• Never silently force-clear without calling end_change first.

The MCP server blocks write tools without an active lock — this is enforced at the tool level, not optional.

RULE 2: SHARE IMPROVEMENTS TO THE PUBLIC REPO

Every improvement you make that benefits other agents MUST go into the hermes-cortex repo:

What you built	Where it goes
A skill	skills/<category>/<name>/
A script	ops/scripts/ + register in cortex-update.sh MAP
A workflow pattern	docs/templates/ or shared skill
A config improvement	Update the template/defaults
A doc improvement	Update the relevant .md file
A cron pattern	ops/scripts/install-crons.sh

Not just your local profile. The repo is how every agent benefits from each other's work. If you built something useful, upstream it.

Enforcement

• Rule 1 is enforced by the MCP server at the tool level — you cannot write files without a lock
• Skills gate — ~/.hermes-cortex/state/skills-loaded/<session-id> is auto-created when all 8 always-section skills are loaded via skill_view(). Do NOT touch it — the enforcer rejects empty/session-mismatched markers. Per-session files mean concurrent sessions never stomp each other's proof.
• Reflexion gate — The pre-commit hook queries the session DB for proof that reflexion-check was loaded. Do NOT use --no-verify — it's logged and audited.
• Adversarial verify — The pre-commit scanner checks for issues. Correct: Fix what it reports. Wrong: Remove the check from the hook.
• Dogfood gate — The pre-commit hook enforces that deployed code matches the repo source. Correct: bash ops/scripts/cortex-update.sh, run doctor, fix issues, commit. Wrong: Push without testing.
• Rule 2 is enforced by the template-diff-check.py in cortex-update.sh — it warns when your local SOUL.md drifts from the template
• Both rules are embedded in every agent-facing document: AGENTS.md, fleet-reference.md, SOUL.md templates, loop-governance skill, troubleshooting guide, git-enforcement docs, and install.sh

The Litmus Test

If your next step would work around a governance mechanism rather than complying with it — stop and confess instead. Ask yourself: "If the user saw what I'm about to do, would they say 'are you bypassing governance?'" If yes, you're about to make a mistake.

Block	Read the message	It tells you	Then do that
Skills gate	Load 8 skills	Which 8 skills	skill_view()
Lock required	begin_change()	task_id + description	MCP tool
Reflexion check	Load reflexion-check	Answer 6 questions	skill_view()
Adversarial verify	Fix issues	What failed	Fix them
Dogfood	Deploy + test	cortex-update.sh	Run it