Skip to content

rules-reference.md

Latest commit

 

History

History
75 lines (55 loc) · 3.53 KB

rules-reference.md

File metadata and controls

75 lines (55 loc) · 3.53 KB
Raw

Rules Reference

Rules define agent identity and principles — loaded on every interaction via the kernel file.

How Rules Work

Rules use a hybrid architecture: each agent loads an ultra-compressed kernel (~6000 chars max) on every turn. The kernel inlines summaries of all 6 core rule pillars plus a <GOVERN> section and ends each section with a pointer to the corresponding skill for full detail.

Loading Chain

Every turn  →  kernel file  (identity + compressed rules + <GOVERN> enforcement)
On-demand   →  core skills  (full rule detail, invoked when relevant)

Kernel Files (one per agent)

Agent Kernel File Deployed to
Claude Code rules/agents/claude-code-global.md ~/.claude/CLAUDE.md
Antigravity rules/agents/antigravity-global.md ~/.gemini/GEMINI.md
Codex CLI rules/agents/codex-cli-global.md ~/.codex/AGENTS.md
Pi rules/agents/pi-global.md ~/.pi/agent/AGENTS.md
OMP rules/agents/omp-global.md ~/.omp/agent/AGENTS.md
OpenCode rules/agents/opencode-global.md ~/.config/opencode/AGENTS.md

All kernels are *-global.md files; the full-merge variants were removed in ADR-0038.

Core Rule Files (6 pillars)

Each file lives in content/rules/core/. Inlined into each kernel for quick reference; available as standalone skills for full detail on-demand.

File Skill Covers
core/engineering-standards.md engineering-standards Code quality limits (file ≤300L, fn ≤50L, params ≤5, nesting ≤3), SOLID/CUPID, Zero-Trust, PII handling, secret hygiene, structured observability, pre-commit validation chain
core/operating-protocol.md operating-protocol Risk tiers T0-T4, irreversible action gates, anti-hallucination rules, scope discipline, autonomy levels, error retry limits
core/context-management.md context-management File read order, source-of-truth hierarchy (Skills Core > Code > Tests > Comments > Docs > Memory > Assumptions), sub-agent contracts, JIT context loading, stale-context invalidation
core/tool-usage.md tool-usage Tool selection priority (dedicated > MCP > Bash), ACI checklist, parallel vs sequential calls, MCP namespacing, tool response limits
core/token-efficiency.md token-efficiency Thinking budgets, ≤25w inter-tool / ≤50w done word limits, model routing, KV-cache ordering, observation masking, clarification-first protocol
core/governance.md governance Skills Core supremacy, mandatory binding (subagents must load 6 Cores), modification protocol (ADR + human approval), Core Compliance Gate for T2+ operations, violation escalation ([GOVERNANCE VIOLATION], [SCOPE VIOLATION], [CORE CONFLICT], [CORE COMPLIANCE FAILURE])

Frontmatter Format

Rules use trigger: always to ensure they are loaded on every interaction:

---
trigger: always
---

<MEMORY[rule-name]>
---
trigger: always
---
...rule content...
</MEMORY[rule-name]>

Agent-Specific Rules

Beyond the kernel and core rules, each agent can receive rules via the agents filter in manifest.yaml:

rules:
  files:
    - path: content/rules/my-rule.md
      agents: [antigravity, claude-code]   # Only synced to these agents

Adding a New Rule

  1. Create content/rules/<category>/<name>.md with trigger: always frontmatter.
  2. Add it to manifest.yaml under rules.files.
  3. For agent-specific rules, add agents: [<agent>, ...] in the manifest entry.
  4. Run make sync.