example

Example run — bug-rca-fix

This directory contains the real captured output of running the bug-rca-fix template end-to-end against gpt-4o via OpenAI.

How this was generated

relay init --template bug-rca-fix
# write context.md describing the bug
relay run --loop --backend openai
relay distill

What's here

• artifacts/context.md — the bug report fed to the loop (a YAML config loader silently coerces the string "false" to a truthy Python value)
• artifacts/repro.md — minimal reproduction (rca_reproducer)
• artifacts/hypothesis.md — root cause analysis (rca_hypothesizer)
• artifacts/plan.md — fix plan (planner)
• artifacts/plan_review.md — reviewer's verdict (APPROVE on first pass on this run)
• artifacts/build_log.md — what the implementer changed
• artifacts/audit.md — auditor's verification (APPROVE)
• LESSONS.md, lessons.json — what relay distill produced from this run plus a second deliberately-degraded run on a different bug. The second run hit the iteration cap (reviewer kept rejecting on real grounds: missing failing test, vague rollback) — those rejections are exactly what the planner inherits next time.

Notes on the captured artifacts

The artifacts are unedited raw model output. You may notice:

• repro.md is wrapped in ```markdown fences — that's the model's habit, not a bug in agent-relay. We leave it untouched so you see what gpt-4o actually produces.
• The reviewer approved on first pass for this particular bug. That's realistic when the context names files, expected behaviour, and constraints clearly. When the context is sparse, the reviewer rejects — see the LESSONS.md entries from the second run for examples of that.