Review agent jq line-count formula cross-products additions and deletions, inflating PR size ~33x

[fullsend-ai-retro]

What happened

On PR #23, the review agent ran 5 times. On each run, the agent computed the PR's total changed lines using a jq expression against the GitHub API's pulls/{number}/files endpoint. The formula [.[].additions + .[].deletions] | add was used, which in jq generates a cross-product: for each element's .additions, it adds every element's .deletions, producing N-squared results instead of N. On run 4 (workflow 30523081668), this computed 91,080 lines; on run 5 (30524230162), it computed 93,908 lines. The actual PR size was ~2,700 lines — a ~33x inflation. The agent self-corrected each time by re-running with the correct formula [.[] | (.additions + .deletions)] | add, but this wasted one turn per review run.

What could go better

The agent wastes a turn (and associated tokens) on every review run to self-correct the line count. Over 5 reviews on this PR, that is 5 wasted turns. The inflated initial count may also influence early reasoning about the PR's complexity before the correction occurs — the agent briefly believes it is reviewing a 91K-line change. Confidence: high — the bug was observed on both run 4 and run 5 with consistent behavior, and the jq cross-product semantics are well-documented. The fix is straightforward.

Proposed change

Pre-compute the accurate PR line count in scripts/pre-review.sh and pass it to the agent as an environment variable (e.g., PR_TOTAL_CHANGED_LINES). This removes the need for the agent to construct jq expressions for basic PR metadata at all. Alternatively, if the line count must be computed by the agent, add guidance in the skills/pr-review/SKILL.md skill definition warning about the jq cross-product pitfall: 'When computing totals from arrays, use [.[] | (.additions + .deletions)] | add — do NOT use [.[].additions + .[].deletions] | add as the latter produces a cross-product in jq.' The pre-script approach is preferred because it eliminates the class of error entirely.

Validation criteria

On the next 3 review agent runs against any repo in per-repo install mode, the agent should not compute an inflated line count or need a self-correction turn. If using the pre-script approach, the PR_TOTAL_CHANGED_LINES env var should be set before the agent starts. If using the skill-guidance approach, the agent should use the correct jq formula on the first attempt.

---

Generated by retro agent from redhat-community-ai-tools/code-to-docs#23

[fullsend-ai-prioritize]

RICE Priority Score: 3.6

Score breakdown

Dimension	Score	Reasoning
Reach	2	This affects every PR review agent run across all repos in per-repo install mode. The bug was consistently reproduced across multiple runs (5 runs on a single PR), indicating it hits all users who trigger the review agent, not just a single org or edge case.
Impact	1	Medium — the agent self-corrects on every run, so the final output is still correct. However, it wastes one turn (and associated tokens/latency) per review run and briefly reasons about an inflated line count (~91K instead of ~2.7K), which could subtly influence early complexity assessments before correction occurs.
Confidence	0.9	Very high — the issue is exceptionally well-documented with specific workflow run IDs, exact jq expressions, quantified inflation (~33x), and a clear root cause (jq cross-product semantics). The proposed fix is concrete and the validation criteria are precise.
Effort	0.5	Simple — the preferred fix is adding a few lines to scripts/pre-review.sh to pre-compute the line count and export it as an environment variable. The alternative (adding guidance text to skills/pr-review/SKILL.md) is equally small. Either approach is a well-scoped, low-risk change.

Formula: (2 x 1 x 0.9) / 0.5 = 3.6