The intent compiler turns raw intent into routed, verifiable internal work.
It does not execute by itself.
raw intent -> classified context -> action plan -> routed skill -> verification -> promotion
Each intent should preserve:
- source
- role
- trust level
- desired outcome
- authority boundary
- promotion target
Intents may come from:
- Alex
- system-generated observations
- Codex runs
- automation reports
- verifier failures
- skill excavation
- cookbook reviews
- friction reviews
- local-only notes
Classify intent as:
- law
- instruction
- evidence
- taste
- hypothesis
- residue
Unknown or untrusted intent cannot become law.
The intent compiler may output:
- action candidate
- skill candidate
- validator backlog item
- proof draft
- automation report
- stop report
- check-in report input
Only classified and admitted intent can drive execution.
Local-only intent may guide alignment but must not be emitted into tracked docs or runtime packs.
Public, revenue, deployment, billing, DNS, secrets, production, and outreach work remains approval-gated.
The later compiler primitive should be one of:
IntentRecord
ActionPlan
It should include:
- intent
- source
- role
- trust level
- authority boundary
- route
- verification
- promotion target