Rule precedence

Rule frontmatter separates instruction ordering from delivery and severity:

Rule precedence

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

Rule frontmatter separates instruction ordering from delivery and severity:

FieldValuesDefault
`precedence``critical`, `high`, `medium`, `low` (the `RULE_PIN_TIERS` vocabulary)`medium`
`overridable`boolean`true`
`enforcement``critical`, `high`, `medium`, `low`optional

Every shipped kernel/core rule declares both ordering fields. `no-attribution`, `anti-laziness` and `human-authorization` ship critical precedence and `overridable: false`. The former `block` enforcement spelling on civic-action and network-analysis rules is normalized to `critical`; their directives remain in their source bodies.

A skill follows its bound rules. Between rules, higher precedence wins; at equal precedence, narrower scope wins. When resolving a conflict, name the rule followed. Platform constraints stay absolute and the general context layer order is unchanged. Precedence is instruction-level guidance; it does not execute rules or resolve arbitrary semantic contradictions at runtime.

Optional project ordering lives in `.aiwg/aiwg.config`, under the existing `rules` object:

{
  "rules": {
    "precedence": {
      "citation-policy": { "tier": "high" },
      "project-style": { "tier": "low", "project_context": true }
    }
  }
}

Each non-empty override object accepts `tier`, `project_context`, or both. `project_context: true` lets WORKSPACE.md project context outrank that named rule; `false` retains the normal rule ceiling. IDs must resolve to exactly one rule, using its slug, source name, or stable discovery ID. Unknown or ambiguous rules are errors. Duplicate aliases for the same rule in one scope are errors. An override of a non-overridable rule always fails with its name, even when assigning its existing tier or setting `project_context: false`.

Use AIWG to complete this documented outcome: Each non-empty override object accepts tier, project_context, or both. project_context: true lets WORKSPACE.md project context outrank that named rule; false retains the normal rule ceiling. IDs must resolve to exactly one rule, using its s
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

The same existing user config surface supports `rules.precedence`. User entries apply first; project fields override user fields for the same rule. An omitted field retains the lower layer's value. `precedence_source` identifies the last scope contributing an override (`rule`, `user`, or `project`). Ordering changes never alter activation or pinning. Regenerate context after changes to refresh the configured ordering listed in WORKSPACE.md. Both generated WORKSPACE.md and AIWG.md describe the rule/skill resolution policy.

Rule lint validates all three metadata fields. Its conflict warning is deliberately conservative: rules must declare matching `subject` or `subjects` metadata and contain literal opposite directives, such as `Always add attribution.` and `Never add attribution.`, at equal effective precedence. Other semantic conflicts still require review; no warning is not proof of consistency.