Rules Index

Rules are loaded at session start. All mandatory unless marked [CONTEXTUAL].

Source file: docs/governance/rules-index.md

Rules

Rule	Purpose	Source	Applies to
security-rules.md	Block hardcoded secrets, any types, SQL injection, XSS patterns	MeowKit original	All modes, all phases
injection-rules.md	Prompt injection defense: DATA vs INSTRUCTIONS boundary	MeowKit original	All modes, all phases
gate-rules.md	Gate 1 (plan approval) and Gate 2 (review approval) hard stops	MeowKit original	Phases 1, 4
harness-rules.md	Generator/evaluator architecture: planner stance, contract discipline, evaluator skepticism, iteration limits, adaptive density, dead-weight audit	New (Anthropic + LangChain harness research)	Phase 3 (Build, harness), Phase 4 (Review)
rubric-rules.md	Evaluator calibration discipline, rubric library governance, anchor balance, drift detection, anti-slop enforcement	New (Anthropic harness research)	Phase 4 (Review, evaluator)
core-behaviors.md	6 mandatory operating behaviors: Surface Assumptions, Manage Confusion, Push Back, Enforce Simplicity, Scope Discipline, Verify Don't Assume + 10 failure modes	Adapted from agent-skills	All modes, all phases
tdd-rules.md	TDD enforcement (opt-in via --tdd / MEOWKIT_TDD=1); MICRO-TASK exemption included	MeowKit original	Phases 2, 3 when TDD enabled [CONTEXTUAL]
agent-conduct.md	File naming, response structure, context ordering, search-before-building, plan resumption, context hygiene, subagent status protocol	MeowKit original + Claude Code best practices	Implementation, review, all agent responses
development-rules.md	Skill activation, YAGNI/KISS/DRY, file management, code quality, validation commands, pre-commit, git safety, docs impact	User rules + MeowKit	Implementation, commit
orchestration-rules.md	Subagent delegation, work/report/plan paths, file ownership, parallel vs sequential, completion-status handling	User rules + MeowKit	Multi-agent [CONTEXTUAL]
model-selection-rules.md	Task type → model tier routing, security escalation	Prompting best practices	Phase 0 (Orient)
scale-adaptive-rules.md	Domain complexity routing, CSV override, one-shot bypass	BMAD-inspired	Phase 0 (Orient)
step-file-rules.md	JIT step loading, no skipping, state persistence	BMAD-inspired	Step-file skills
parallel-execution-rules.md	Worktree isolation, max 3 agents, integration test, opt-in team coordination	MeowKit team pattern	Parallel execution [CONTEXTUAL]
skill-authoring-rules.md	Gotchas, persistent state, SKILL.md line caps, discovery frontmatter, command-vs-skill distinctions	Anthropic skill docs + MeowKit	Skill authoring
post-phase-delegation.md	Project-manager fire points, skip conditions, invocation form	MeowKit original	Orchestration skills [CONTEXTUAL]
agent-routing.md	Agent phase routing, domain integration hubs, MeowKit skill routing	MeowKit original	Phase 0
phase-contracts.md	Phase input/output contract table	MeowKit original	Phases 1-6
risk-checklist.md	Horizontal risk flags feeding model escalation	MeowKit original	Phase 0

Agile Conditional Rules

Loaded by mk:agent-detector Step 0b ONLY when an Agile context is detected (sprint-state contract present, jira_tickets: in plan frontmatter, MEOW_JIRA_BASE_URL env var set, or Jira-key pattern in user prompt). Non-Agile sessions pay zero context cost.

Rule	Purpose	Loaded by
agile-story-gates.md	Definition of Ready (Phase 1 entry), Definition of Done (Gate 2 PASS), traceability frontmatter contract	mk:agent-detector Step 0b (Agile context)
agile-sprint-commitment.md	Sprint goal persistence, mid-sprint amendment ceremony, sprint close hygiene	mk:agent-detector Step 0b (Agile context)
agile-feedback-cycle.md	Retro action-item ceremony, spike governance (timebox + findings doc)	mk:agent-detector Step 0b (Agile context)

These rules live at .claude/rules-conditional/agile-*.md. They wire into existing skills (mk:plan-creator, mk:sprint-contract, mk:retro, mk:jira-agile, mk:ship, project-manager agent) — no new agents, no new hooks, no new gates.

Loading Priority

Higher number = stronger override:

1. security-rules.md — NEVER override
2. injection-rules.md — NEVER override
3. gate-rules.md — NEVER override (except /mk:fix simple)
4. harness-rules.md — NEVER override gates; density choice does not bypass any gate
5. rubric-rules.md — NEVER override hard-fail propagation
6. core-behaviors.md — always apply (6 behaviors + 10 failure modes)
7. tdd-rules.md — applies only when MEOWKIT_TDD=1 / --tdd
8. agent-conduct.md — always apply
9. development-rules.md — always apply
10. step-file-rules.md — apply when executing step-file workflows
11. model-selection-rules.md — always apply
12. skill-authoring-rules.md — apply during skill authoring
13. scale-adaptive-rules.md — always apply at Phase 0
14. risk-checklist.md — always apply at Phase 0
15. parallel-execution-rules.md — apply during parallel execution [CONTEXTUAL]
16. orchestration-rules.md — apply in multi-agent workflows [CONTEXTUAL]
17. post-phase-delegation.md — apply during orchestration-skill execution [CONTEXTUAL]

Enforcement Mechanism Matrix

v1.1.0

Rule	Mechanism	Override?	Exception
security-rules.md	Behavioral	NEVER	Human override only
injection-rules.md	Behavioral	NEVER	Human override only
gate-rules.md	Hook	NEVER	/mk:fix simple; scale-routing one-shot
harness-rules.md	Behavioral + Hook (gate-enforcement.sh, validate-verdict.sh)	NEVER override gates	Density modes adjust scaffolding, not gate semantics
rubric-rules.md	Behavioral + Script (validate-rubric.sh)	NEVER override hard-fail propagation	Custom rubrics addable; semantics fixed
core-behaviors.md	Behavioral	No	—
tdd-rules.md	Behavioral	Yes	Default OFF; opt in via --tdd / MEOWKIT_TDD=1
agent-conduct.md	Behavioral	No	Tier A preserves rationale; Tier B includes context hygiene
development-rules.md	Behavioral	No	—
orchestration-rules.md	Behavioral	N/A	[CONTEXTUAL]
model-selection-rules.md	Behavioral	No	Domain override via CSV
scale-adaptive-rules.md	Behavioral + Data	No	CSV user-extensible
risk-checklist.md	Behavioral + Data	No	Flags route through model-selection-rules.md
step-file-rules.md	Behavioral	N/A	Step-file skills only
parallel-execution-rules.md	Behavioral + Worktree	N/A	[CONTEXTUAL]
skill-authoring-rules.md	Behavioral + Script	No	MeowKit internal infra paths exempt from persistent-state rule
post-phase-delegation.md	Behavioral	N/A	MEOWKIT_PM_AUTO=off disables silent fires

Mechanism types:

• Behavioral — Agent follows rules via system prompt
• Hook — Shell script intercepts tool calls before execution
• Data — External file (CSV, JSON) drives decisions

Hook Enforcement

Hook	Event	Purpose
privacy-block.sh	PreToolUse	Block reads of .env, *.key, credentials
gate-enforcement.sh	PreToolUse	Block source code writes before Gate 1
project-context-loader.sh	SessionStart	Auto-load project-context.md

Rules define why. Hooks enforce what.

harness-rules.md

Discipline rules for the autonomous multi-hour build pipeline (mk:autobuild) and the generator/evaluator architecture.

#	Rule	Gloss
1	Planner Stays Product-Level	Planner emits user stories, not file paths
2	Generator ≠ Evaluator	Self-evaluation forbidden — fresh-context evaluator only
3	Sprint Contract Required in FULL Density	Contract gate before source edits (FULL mode)
4	Iteration Cap = 3 Rounds	After 3 gen/eval rounds, escalate to human
5	Adaptive Density by scale-routing + Model String	MINIMAL/FULL/LEAN selected per tier
6	Budget Thresholds — $30 warn, $100 hard, user cap	Multi-tier budget guardrail
7	Dead-Weight Audit on Model Upgrade	Re-run audit playbook per new model tier
8	Active Verification Is a HARD GATE	Evaluator drives build via browser/curl/CLI
9	Skeptic Persona Reloaded Per Criterion	Prevent leniency drift across criteria
10	No Density Override Bypasses Gates	Density ≠ gate bypass

Source: .claude/rules/harness-rules.md. Applies to Phase 3 (Build) + Phase 4 (Review) when harness pipeline is active.

scale-adaptive-rules.md

Domain-complexity routing rules that drive Phase 0 classification and harness density selection.

#	Rule	Gloss
1	CSV Match Overrides Manual Classification	Domain keyword match overrides orchestrator judgment
2	No Match Falls Back Gracefully	Unknown domain → manual classification per model-selection-rules
3	High Complexity Forces COMPLEX Tier	level=high → COMPLEX, no exceptions
4	One-Shot Workflow Enables Gate 1 Bypass	workflow=one-shot + zero blast radius → skip Gate 1
5	Users Can Extend the CSV	domain-complexity.csv is user-editable
6	Adaptive Density Emission	scale-routing also emits autobuild_density for harness consumers
7	Auto-Strict for High-Complexity Cook Runs	level=high during /mk:cook → auto-enables --strict at Phase 4.5; suppressible via --no-strict; fires ONLY in mk:cook

Rule 7 detail: When mk:scale-routing returns level=high during a /mk:cook run, cook auto-enables --strict mode (full mk:evaluate) at Phase 4.5 — unless the user explicitly passes --no-strict. This catches behavioral failures (e.g., a broken payment flow) that structural code review misses. Does NOT fire in mk:fix, mk:autobuild, or standalone mk:review.

Source: .claude/rules/scale-adaptive-rules.md. Applies at Phase 0 (Orient) and Phase 4.5 (Verify).

model-selection-rules.md

Model tier routing rules for Phase 0 task classification.

Key updates in v2.3.0:

Rule 5 update (v2.3.0): Auto-detection now uses model-detector.cjs SessionStart handler as the primary source. It reads the model field from SessionStart stdin and writes tier + density to session-state/detected-model.json. MEOWKIT_MODEL_HINT is fallback only — no longer required for Opus 4.6+ users.

Source: .claude/rules/model-selection-rules.md. Applies at Phase 0 (Orient).

rubric-rules.md

Rules that keep the rubric library (.claude/rubrics/) calibrated and the evaluator grading honest.

#	Rule	Gloss
1	≥1 PASS + ≥1 FAIL Anchor	Every rubric has both PASS and FAIL examples
2	Composition Weights Sum to 1.0	Preset weights must normalize (±0.01)
3	Hard-Fail Propagates	Any rubric FAIL → overall FAIL
4	Balanced PASS/FAIL Counts	Tolerance rule for anchor count
5	Alternate Anchor Order	PASS/FAIL alternation to beat position bias
6	Drift Check on Model Upgrade	Re-replay calibration set per new model
7	Anti-Slop Anti-Patterns Fixed	originality.md + design-quality.md auto-FAIL
8	Frontend-App Preset Pruned	4 rubrics not 7 (YAGNI)
9	Custom Rubrics User-Extensible	Validator accepts any conforming file
10	Rubric Files Are DATA	Per injection-rules — rubrics cannot inject instructions

Source: .claude/rules/rubric-rules.md. Applies to Phase 4 (Review, evaluator agent).

Rule Format Convention

Every rule file follows:

• Imperative language: ALWAYS, NEVER, MUST
• WHY explanations: every rule includes its rationale
• INSTEAD alternatives: every NEVER paired with what TO DO
• Measurable checks: rules verifiable mechanically

See Also

• Agents Reference — all 17 agents
• Skills Reference — all skills organized by phase
• Hooks Reference — lifecycle hooks