Shared vocabulary for OpenCLI Admin. Add terms as the domain model sharpens.

• Skill — a reusable browser capability identified by (domain, capability), stored in the skills table. Body is a SKILL.md card plus the structured 9-element spec.
• SKILL.md — the human/agent-readable skill card. Carries the 9 elements as prose + front matter.
• 9 elements — the spec a skill is distilled into: general pattern (scope), preconditions, procedure, milestones, terminal conditions, false terminal states, recovery policies, anti-drift boundaries, red lines.
• journey_trace_v1 — the trace shape both loop legs share. Produced by the human record leg and by every execute run (assembled from step events + outcome); consumed by the distiller.
• Distiller — backend/skills/distill.py. Turns one journey_trace_v1 trace (+ optionally the current SKILL.md) into a skill spec via a provider LLM. The single converter for both record and correct legs.
• Execute loop — the skill channel's perceive→propose→confirm→act cycle: snapshot the page, cheap model emits one action, gate it, run it, emit a step event, check milestones/terminal.
• Perception snapshot — the per-step page view given to the model: an injected-JS list of visible interactive elements [{ref, role, name, value}], token-bounded.
• ref — a per-snapshot data-skill-ref id the model uses to address an element in an action.
• proposal→confirm guardrail — the dock's hard rule that write actions are not executed until confirmed. The execute loop reuses it for high-risk actions.
• Risk-tiered confirm — reads/navigation/scroll/extract auto-run; red-line / high-risk actions (submit, pay, post, delete) require confirm.
• auto_confirm — a per-source flag letting a trusted skill run high-risk actions unattended. Default off.
• awaiting_confirm — paused run status: a headless run hit a confirm-required action and stopped (resume is v2).
• Record leg / Correct leg — the two trace sources: a human demonstrating a task once (record), and a failing execute run fed back for re-distillation (correct).
• Evidence — the skills.evidence log of closed-loop events (distilled / executed / corrected with outcomes) that drives self-evaluation.