Glossary

Quick definitions for every term you'll encounter in the docs, the dashboard, and the MCP tool descriptions. For a longer mental model, start with Core concepts.

Acceptance checklist: Markdown file (typically docs/E2E_TEST_CHECKLIST.md) listing the testable claims that must be demonstrably true before a release ships. Read by codeloop_gate_check to evaluate acceptance_criteria_met.
Activation chain: The path from agent prompt → user rule → MCP server → verify tool. codeloop doctor validates every link.
Artifact: Everything written to artifacts/runs/<run_id>/— the full evidence record for a run. See Core concepts.
Baseline: The canonical screenshot for a screen × viewport, stored at .codeloop/baselines/.... Future visual reviews diff against this. See Visual review.
Confidence score: Single 0–100 number summarising the run's evidence. 94+ is the default “ready for review” bar.
Design reference: A Figma frame or local PNG that the coded UI must match. Compared by codeloop_design_compare.
Diagnose: The classification step that turns raw verify failures into prioritised issues with concrete repair tasks for the agent.
Doctor: npx codeloop doctor— one command to validate every layer of the activation chain plus the runtime tooling for your OS.
Gate: One of the binary checks (build_passes, required_tests_pass, zero_critical_issues, visual_regression_threshold, acceptance_criteria_met) evaluated by codeloop_gate_check.
Gate check: The terminal step in the loop. Aggregates every gate, returns a confidence score plus a recommendation: ready_for_review, continue_fixing, or blocked.
Integration check: Re-runs verify across multiple sections to confirm Section A still works after Section B ships. Reports per-section confidence.
Interaction: A single UI action (click, type, swipe, ...) issued through codeloop_interact while a recording is open.
Key frame: One of the ~15 motion-validated frames extracted from a recording by codeloop_interaction_replay. Sampled at the points of greatest visual change.
Local mode: CODELOOP_MODE=local— tells the CLI / MCP server to use a self-hosted backend with no public auth or metering. See Self-host runbook.
Master spec: docs/specs/_master.md— the top-level “what we are building” document for a multi-section project.
MCP: Model Context Protocol — the open standard CodeLoop uses to expose its tools to agents like Cursor and Claude Code.
MCP tool: A single function callable by an agent over MCP. CodeLoop ships 29.
Plugin: JSON entry in .codeloop/plugins.json that wires any CLI test runner into the verify suite. See Plugin SDK.
Recording: Video capture of the running app while the agent drives interactions. Stored at artifacts/runs/<run_id>/videos/....
Replan: codeloop_replan— recomputes the section dependency graph after a spec edit so the agent picks up from the right point.
Replay: codeloop_interaction_replay— post-processes a recording: extracts key frames and aligns them to app log lines.
Run: One invocation of codeloop_verify. Identified by a stable run_id that flows through the rest of the loop.
Run artifact: See Core concepts. The directory + JSON contract that decouples the agent from downstream consumers.
Section: A bounded slice of a multi-section project — one screen, one feature, one milestone — with its own spec and acceptance criteria.
Section state: One of pending, in_progress, completed, blocked. Read / written via codeloop_section_status.
Self-host: Run the entire CodeLoop stack on your own infrastructure. See Self-host runbook.
Spec: The markdown document describing what the project (or a section) should do. Read by gate check to evaluate acceptance_criteria_met.
Threshold: Configurable bar for a gate — e.g. visual_review.threshold = 0.02 (2 % pixel diff allowed) or gate_check.min_confidence = 0.94.
User Rule: Cursor's global rule that tells the agent to always call CodeLoop. Pasted once under Settings » Rules » User Rules. See Quick Start.
Verified by CodeLoop badge: SVG badge published by the GitHub Action with the latest gate score and a link to the showcase page for the run.
Verify: The first step in the loop. Runs build + lint + tests + screenshots and produces a structured pass/fail report.
Visual review: The screenshot diff step. See Visual review.
Workflow enforcement: codeloop_check_workflow— guardrail that fails loudly when the agent tries to declare done without the prerequisite tools called.