From 73321dd034d9ce02c2d22e23a86cb1ee832900a0 Mon Sep 17 00:00:00 2001
From: Ralf Anton Beier <ralf_beier@me.com>
Date: Thu, 30 Apr 2026 07:25:50 +0200
Subject: [PATCH] perf: parallelise the three PR-data fetches in
 reviewPullRequest
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## Why
The hot path of `reviewPullRequest` (`src/ai-review.js:382-398`) issued three
strictly-sequential GitHub REST calls before any work began: PR metadata, the
diff via `mediaType: { format: 'diff' }`, and the files list. Each call costs
~200-500 ms of egress time on the netcup VPS; serially that's ~600-1500 ms
of wall-time waste on every review.

The three calls are independent — none depends on the other two's response —
so they parallelise trivially.

## What
`Promise.all([request, request, request])` wrapper in `reviewPullRequest`.
The downstream consumers (rivet oracle path, `buildReviewPrompt`,
`recordReview`) keep their `prData` / `diffResponse` / `filesResponse`
contract unchanged. octokit.request invocations are still dispatched in
source order (JS is single-threaded), so the existing
`mockResolvedValueOnce`-style test rigs work unchanged.

## Source
Wave-1 Performance engineer flagged this as Bug #19 in
`docs/agent-fleet/bugs.md`. Estimated savings: 0.5-1 s per review.

## Test plan
- [x] All 806 tests pass
- [x] eslint clean
- [ ] After deploy: PR-review wall-time drops by the GitHub-egress component;
      visible in `pr-review-timing` logs (when added per Performance agent's
      "one profile I'd run" recommendation).

## Risk & rollout
- Risk: low. Same data, same downstream consumers, just dispatched concurrently.
- Rollout: self-update on merge.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 src/ai-review.js | 41 ++++++++++++++++++++++++-----------------
 1 file changed, 24 insertions(+), 17 deletions(-)

diff --git a/src/ai-review.js b/src/ai-review.js
index 3307261..ea12e4f 100644
--- a/src/ai-review.js
+++ b/src/ai-review.js
@@ -379,23 +379,30 @@ async function reviewPullRequest(octokit, owner, repo, prNumber) {
   }
 
   try {
-    const prData = await octokit.request('GET /repos/{owner}/{repo}/pulls/{pull_number}', {
-      owner,
-      repo,
-      pull_number: prNumber
-    });
-
-    const diffResponse = await octokit.request('GET /repos/{owner}/{repo}/pulls/{pull_number}', {
-      owner,
-      repo,
-      pull_number: prNumber,
-      mediaType: { format: 'diff' }
-    });
-
-    const filesResponse = await octokit.request(
-      'GET /repos/{owner}/{repo}/pulls/{pull_number}/files',
-      { owner, repo, pull_number: prNumber }
-    );
+    // Three independent reads — fan them out instead of waiting serially.
+    // ~600-1500 ms before any work was wall-time waste; Promise.all halves
+    // it on the typical Probot path. octokit.request invocations are still
+    // dispatched in source order (JS is single-threaded), so the existing
+    // mockResolvedValueOnce-style test rigs and downstream consumers
+    // (rivet oracle path, buildReviewPrompt, recordReview) keep their
+    // prData / diffResponse / filesResponse contract.
+    const [prData, diffResponse, filesResponse] = await Promise.all([
+      octokit.request('GET /repos/{owner}/{repo}/pulls/{pull_number}', {
+        owner,
+        repo,
+        pull_number: prNumber
+      }),
+      octokit.request('GET /repos/{owner}/{repo}/pulls/{pull_number}', {
+        owner,
+        repo,
+        pull_number: prNumber,
+        mediaType: { format: 'diff' }
+      }),
+      octokit.request(
+        'GET /repos/{owner}/{repo}/pulls/{pull_number}/files',
+        { owner, repo, pull_number: prNumber }
+      )
+    ]);
 
     // Mechanical oracle: only runs for repos that ship rivet.yaml AND have
     // the binary configured. Failures are non-fatal — the model path still