example

Example run — rfc-then-implement

This directory contains the real captured output of running the rfc-then-implement template end-to-end against gpt-4o via OpenAI.

How this was generated

relay init --template rfc-then-implement
# write context.md describing the feature request
relay run --loop --backend openai

What's here

• artifacts/context.md — feature request (cursor-based pagination on a public list endpoint)
• artifacts/rfc.md — the architect's draft RFC
• artifacts/rfc_review.md — reviewer's verdict
• artifacts/build_log.md — implementer's record
• artifacts/audit.md — auditor's verification

On the empty LESSONS.md

Lessons compile from rejection sections (## Required Changes, ## Catches, ## MUST_FIX) across past runs. This particular run went APPROVE → APPROVE → APPROVE — there were no rejections to distill, so relay distill correctly emitted an empty result.

To see populated lessons output, look at templates/bug_rca_fix/example/LESSONS.md — that template's example includes a second run that hit the iteration cap, so the reviewer's ## Required Changes bullets compiled into real warn-level lessons.

The compounding mechanism is identical for rfc-then-implement: when reviewer or auditor reject, those bullets become lessons the next architect / implementer reads in their prompt.