Glossary

A one-stop decoder for the vocabulary in reports and docs. You only need two terms to start: lifecycle dimensions and report. Everything below is progressive detail you load when a task needs it.

The mental model in three lines

• A harness wraps a target so an agent can run a bounded, recoverable loop: understand → change → validate → repair → re-validate → state residual risk.
• Better Harness starts from the Agent Work Loop, qualifies it with available session and project evidence, guards changes while they happen, and feeds what it learns back into rules.
• Session evidence changes confidence and coverage, not the model.

Core concepts

Term	What it means
Harness	The engineering environment around an agent that makes a change loop bounded and recoverable.
The loop	understand context -> bounded change -> choose validation -> interpret failure -> repair -> re-run validation -> state residual risk
Agent Work Loop	The default /better-harness model: what the Harness supports, what agents actually did when observable, where a task loop lost control, and what should improve next.
Task Episode	One user goal with one acceptance boundary; the review unit for behavior claims.
Software Fluency	The static project lens used for an explicit repository-only score and the independent project evidence pass.
Progressive disclosure	Reveal context to an agent by task, not all at once.

Lenses and models

Term	What it means
Five lifecycle dimensions	Task Understanding, Controlled Execution, Change Validation, Reliable Delivery, Learning Capture.
Five software capabilities	Context Map, Environment Readiness, Fast Feedback, Quality Gates, Change Safety.
AI Readiness Ladder	The L1–L5 maturity scale (Awareness → Assisted → Structured → Spec-Governed → Closed-Loop).
Style	The visual framing of a report (analyst, audit scorecard, consulting deck, dashboard, …).
Output mode	The rendered form of a report: Qoder Canvas, HTML visual, or Markdown.

Evidence and scoring

Term	What it means
Evidence boundary	The rule that separates static file evidence from executed command, CI, runtime, or UI evidence; unverified areas cap confidence.
Evidence state	Present, Wired, Exercised, Outcome-supported, Missing, Unobserved, or Not applicable — see Agent Work Loop.
Confidence	Low/Medium/High rating bound to how much was actually executed vs. only read.
Change confidence	Whether an AI-generated change is ready to land, judged by blast radius, sensitive paths, size, and validation.

The action loop (report → change)

Term	What it means
Handoff	A row-scoped next step inside a report (draft a fix, schedule a follow-up), not a dead-end score.
Repair plan	A bounded fix plan for one finding, drafted via /better-harness repair-plan without writing report artifacts.
Loop Engineering	The domain that decides whether repeated work exists and which durable owner (skill, hook, script, automation, rule) should hold it.
Loop Discovery	The routing gate that proves a loop from evidence and picks the smallest durable owner.
Schedule-ready	A finding stable enough to become a recurring /schedule /better-harness follow-up, with cadence, validation, and a stop condition.

Extension and hosting

Term	What it means
Skill	A repeatable agent workflow defined by SKILL.md frontmatter plus a concise workflow.
Host adapter	Per-host discovery and evidence-shape glue; keeps the engine host-neutral.
Host shell	Thin host metadata (.claude-plugin/, .qoder-plugin/, .cursor-plugin/, .codex-plugin/) that exposes canonical behavior without owning product logic.
Canonical owner	The single directory that owns a behavior's product judgment; host shells and mirrors point back to it.

The full glossary with owner links lives in docs/glossary.md.