Concepts

The vocabulary Foyer uses — intent, signals, repair prompts, and how a verification runs.

Foyer's PR comment and CLI output use a small vocabulary repeatedly. This page defines each term once, in order of when you'll meet it.

Verification

A single run of the merge-readiness gate against one PR head SHA. Triggered by a GitHub webhook (pull_request: opened | synchronize | reopened) or manually via the CLI. A verification produces one PR comment and, on failure, one repair prompt per failed signal.

Signal

One verification check. Each signal returns passed | warning | failed | unavailable | unsupported_on_provider plus a summary, evidence rows, and (when failed) a repair prompt. The eleven shipped signals are catalogued at Signals.

unavailable means the signal couldn't run (e.g. no test runner detected, abort during install). unsupported_on_provider means the signal is GitHub-only and you're on GitLab. Both are real, deliberate values — Foyer never silently skips a signal.

Intent

The PR's stated requirements, captured from one of:

A Linear issue linked to the branch (foyer review --linear ENG-123, or auto-detected by branch name).
A GitHub issue referenced in the PR body.
The PR body itself.
A pasted spec.
A plan artifact attached via foyer plan attach or foyer plan save.

Intent is split into atomic, verifiable requirements by an LLM at intent-build time. Each requirement is scored independently.

`scope_alignment`

The signal that scores the diff against the PR's intent. For each requirement, the judge decides: was this implemented (met), partially implemented, or missed. The verdict is one of exact, superset, subset, divergent. A divergent verdict is the most common reason a PR comment goes red.

This is Foyer's headline differentiator vs. CodeRabbit / Greptile / Bugbot. They review code in isolation; Foyer reviews whether the code did what you asked. Scope validation is the gate — the other ten signals back it up.

Evidence

Per-signal proof attached to the result. Five kinds: log, screenshot, command, diff, repro. The PR comment renders evidence rows under each signal. The repair prompt's context.evidence mirrors them so the fixing agent doesn't need to re-investigate.

Repair prompt

Every failed signal carries one — a structured handoff the agent can apply directly. Shaped:

interface RepairPrompt {
  prompt: string;
  context: Record<string, unknown>;
}

The prompt is a plain-markdown instruction that any coding agent (Claude Code, Codex, Cursor, Windsurf, …) can consume. The context carries failed-signal evidence, head SHA, file paths, and reproduction commands. Drop it into your agent, the agent fixes the issue, you push, next verification confirms green.

Audience

Foyer supports two output voices:

technical (default) — runner summaries assume the reader writes code.
non_technical — same signals, plain-English summaries. Used for "show me what's wrong" views for non-engineering stakeholders.

The repair prompt always reads the technical summary; only the PR comment swaps copy.