Using Skills Inside Harnesses
Reference documentation for Using Skills Inside Harnesses in Adaptive Skills.
Goal
Show how an ordinary skill runs inside a declared Agent Harness Contract (AHC) without changing the skill itself. The harness sets the envelope; the skill executes within it.
The AHC is declared in AletheIA (docs/contracts/agent-harness-contract.md); a fill-in template
lives at ../templates/agent-harness-contract.yaml.
How the pieces meet
allowed_skills/blocked_skills— the harness names which skills may run. The skill does not decide this; the contract does.gates— when a skill’s core move would write, delete, call out, or change structure, the matching gate (before_write,before_delete,before_external_call,before_structural_change) applies before the move commits.sensors— the sensor-before-judgment rule: a skill should run its computational sensors (tests, linters, type checks) before consulting an inferential one (a review skill).observability— the skill’s decisions and evidence feed the harness trace (see../templates/harness-trace-summary.md).rollback/human_review— for risky or irreversible work, the skill’s output stays a draft until the contract’s human review clears it.
What does not change
- The skill’s 11 sections, triggers, and verification stay exactly as authored.
- The skill never embeds tool permissions, autonomy, or rollback — those live in the contract.
- A skill that was safe under a light harness stays correct under a strict one; only the envelope tightens.