verify-a3

name	verify-a3
description	Verify adoption criterion A3 (Feedback Loop Closure) in the current project. Reports fulfillment level 0–3.
allowed-tools	Bash Read

Verify A3 — Feedback Loop Closure

Criterion Definition

Level	Description
0	Agents produce output and hand it directly to the human; no self-verification
1	Agents run basic checks (compilation, syntax) but do not iterate on failures before presenting results
2	Agents run the full test suite and linter, iterate on failures, and present only passing results to the human
3	Agents have access to CI pipeline results, production monitoring, and MCP-connected tool outputs; verification is fully automated and includes post-deployment checks

Evidence to Gather

• Read CLAUDE.md or AGENTS.md and look for sections that instruct agents to run tests, linters, or other quality checks before presenting results.
• Check the .claude/skills/ directory — read skill files and assess whether they include verification steps (build, test, lint) before reporting results.
• Check the MCP server configuration (.claude/settings.json, .mcp.json) for any CI pipeline or production monitoring servers that give agents programmatic access.
• Check .claude/settings.json for hooks that automatically run quality checks.
• Look for a test suite (test directories or test files) that an agent could run.

Instructions

Gather the evidence described above and determine the fulfillment level for A3.

This criterion measures the degree to which agents close their own feedback loops before handing work to a human. Look for evidence that the workflow (skills, CLAUDE.md, AGENTS.md) instructs or enables agents to verify their own output.

Scoring guide:

• Level 0: No verification guidance in agent context files or skills. Agents produce output and pass it directly to humans with no self-check.
• Level 1: Agent context files mention compilation or syntax checks, or skills run basic build commands, but there is no instruction to run the full test suite or iterate on failures.
• Level 2: Agent context files or skills explicitly instruct agents to run the full test suite and linter before presenting results, and to iterate on failures until they pass. A CLAUDE.md that says "always run npm test before completing" qualifies.
• Level 3: Level 2 is met AND agents have programmatic access to CI pipeline results (via GitHub/GitLab MCP server) AND production monitoring access (via observability MCP server) enabling post-deployment verification without human involvement.

Report in exactly this format:

A3 — Feedback Loop Closure

• Level: [0 / 1 / 2 / 3]
• Rationale: [one or two sentences citing the specific evidence]