Live rehearsal history
This project has been run for real multiple times during development — not just described — through the actual GitHub webhook delivery path (not a direct function call or synthetic payload), against both a local tunnel and the deployed Render service. This page is the log of those runs: which PR, which path was exercised, and the measured result. It's the evidence that the numbers quoted elsewhere in this guide (the 15-second review target, a redeploy's real duration) are measured, not assumed.
| PR | Path exercised | Result |
|---|---|---|
| #2 | Direct orchestrator.run_review() call (step 5/6 milestone) |
8s, all 3 specialists ok, comment correct |
| #2 | demo_provider_swap.py (step 7) |
groq ok → gemini fails gracefully (real error) → groq restored |
| #3 | Real GitHub webhook delivery (quick tunnel + PATCH /app/hook/config JWT-updated URL + seed_demo_pr.py) |
8s PR-created → comment-appeared, all 3 specialists found real issues |
| #4 | First hosted run (Render + Supabase, 2026-08-07) — happy path, seed_demo_pr.py against the deployed service |
~9.2s PR-created → comment-appeared, real findings via groq; tickets created by the app's own first boot against a real Supabase project (see docs/2026-08-05-first-hosted-run-findings.md for full detail) |
| #5 | Hosted Segment B — LLM_PROVIDER=github_models redeploy (real 2026-07-30 retirement — see Provider history) → all 3 specialists fail visibly → groq redeploy → follow-up commit → same comment updates in place |
Redeploys 65.5s / 56.7s (not a 2s local restart); ticket survived both restarts intact |
| #6-#9 | Hosted Segment C — 4 new PRs + a follow-up commit fired in quick succession under groq |
No 429 observed (current Groq headroom exceeds the 2026-08-03 token-math measurement; not retried, per the testing-hygiene rules in Provider history) |
PR #3 is the definitive rehearsal. It's the first run of the actual webhook path end-to-end — GitHub → tunnel → HMAC verify → dedup → background task → orchestrator → comment — not a direct function call or a synthetic payload, and it's the run most often cited elsewhere as proof the 15-second review target holds: the comment appeared 8 seconds after PR creation, all three specialists found real issues, well under target.
PRs #6-#9 are the closest this project came to a sustained-volume test: four
new PRs plus a follow-up commit fired in quick succession, with no 429
observed. That's a useful data point, not a guarantee — it wasn't retried at
higher volume, deliberately, per the same testing-hygiene discipline that
governs live LLM calls (see Provider history).