Tenet Governance

Shared reference for the governor plugin. The three user-invocable skills (governor-bootstrap, governor-manage, governor-verify) load this for the tenet format spec and shared vocabulary.

What tenets are

Tenets are architectural constraints that require human judgment to verify. They are not linting rules.

Tenet format in AGENTS.md

## Tenets

CRITICAL: These tenets are MANDATORY and MUST be followed in all work on this codebase.

### T1. <Name>

<Description: 2-4 sentences explaining the constraint and rationale>

**Severity:** <critical | high | medium | low>

### T2. <Name>

<Description>

**Severity:** high

## Tenet Exceptions

Approved exceptions. Each must have justification.

| File | Tenet | Reason | Approved |
|------|-------|--------|----------|
| `src/legacy/adapter.go` | T1 | Legacy integration | 2024-01-15 |

No evidence section. Evidence is gathered at validation time and shown to the user for review, but never persisted. Persisted evidence rots quickly and provides little benefit — the codebase is the source of truth.

Severity levels

Default if severity is omitted: high.

Exception syntax

Table (AGENTS.md): See format above.

Inline (in code):

// governor:ignore T1 - Legacy adapter, tracking in #123

Language-specific comment syntax:

• Go/Java/C#/JS/TS: // governor:ignore T1 - reason
• Python: # governor:ignore T1 - reason
• HTML: <!-- governor:ignore T1 - reason -->
• CSS: / governor:ignore T1 - reason /

Two distinct operations

Validation (bootstrap/manage): "Is this proposed tenet grounded in reality?"

• Search the codebase for supporting/contradicting patterns
• Verdicts: SUPPORTED, WEAKEVIDENCE, NOTSUPPORTED, CONTRADICTED
• Run at add-time, and re-run on demand from the manage skill
• Findings are shown to the user, never written to AGENTS.md

Verification (verify): "Does this code follow the existing tenets?"

• Analyze code against tenets in AGENTS.md
• Output: COMPLIANT or VIOLATED with confidence scores

Confidence scoring (verification)

Good tenet characteristics

1. Actionable — "X must/must not Y" format
2. Grounded — A pattern actually present (or absent) in the codebase, not aspirational
3. Architectural — About structure, not style
4. Persistent — Unlikely to change frequently
5. Verifiable — Compliance determinable with reasonable effort

Parsing rules

When parsing AGENTS.md:

1. Match ## Tenets heading
2. Extract tenets matching ### T<N>. <Name> headers
3. Per tenet:

- Next paragraph = description - Severity: line = severity (default: high)

1. Match ## Tenet Exceptions heading
2. Parse exception table rows: file, tenet, reason, approved

Pre-flight check (all operations):

test -f AGENTS.md && grep -q "^## Tenets" AGENTS.md && echo "ok" || echo "missing"

If missing, direct the user to run the bootstrap skill.

Reference documentation

• [Tenet Format](reference/tenets.md) — Full format spec, parsing edge cases, malformed-file handling
• [Verification Patterns](reference/verification-patterns.md) — Language-specific detection patterns
• [Examples](reference/examples.md) — Good/bad tenets, example AGENTS.md, validation output samples