fix(review): PR summary "Changes Overview" counts diverge from GitHub's file/line totals

[devops-thiago]

The PR summary "Changes Overview" reports file/line counts derived from the bot's own reviewable-file list, not GitHub's authoritative PR totals, so it undercounts whenever files are removed by the ignore-glob. See #39 (walkthrough / file-by-file origin) and #234 (truncation disclosure). Not #239 (that is per-model tokenizer for token counts).

Problem

PrSummaryGenerator.generate renders "Changes Overview" straight from its filesChanged/additions/deletions int params (PrSummaryGenerator.java:102–105):

sb.append("### Changes Overview\n");
sb.append("- **Files changed:** ").append(filesChanged).append("\n");
sb.append("- **Lines added:** ").append(signed('+', additions)).append("\n");
sb.append("- **Lines removed:** ").append(signed('-', deletions)).append("\n\n");

Those params come from DiffStats.fromFiles(ctx.reviewableFiles(), ctx.omittedFiles()) (VerdictBuilder.java:60), passed into summaryGenerator.generate(...) at VerdictBuilder.java:229–232. DiffStats.fromFiles sums per-file additions()/deletions() over reviewableFiles and uses files.size() for the file count (VerdictBuilder.java:165–172); the omittedFiles argument is stored only as the 4th record field to drive the truncated() flag and does not reduce the counts.

reviewableFiles is the ignore-glob-filtered subset of the PR's files (ReviewContextLoader.java:120 → ReviewDiffFormatter.reviewableFiles, which filters out isIgnored(...) paths). So any file removed by the ignore-glob contributes nothing to the count. The result is a "Changes Overview" derived from the filtered file list rather than GitHub's authoritative changedFiles/additions/deletions, and it silently undercounts.

Note: line-budget truncation (omittedFiles, ReviewContextLoader.java:127) does not currently undercount this section — DiffStats.fromFiles still sums over the full reviewableFiles list regardless of how many files the rendered diff dropped. It is the ignore-glob that causes the divergence today. (The changedFiles.size() used for the walkthrough rollup — PrSummaryGenerator.java:276 — is the same ignore-glob-filtered source.)

Evidence

Dogfood on devops-thiago/MongOCOM#46. Posted summary "Changes Overview":

### Changes Overview
- Files changed: 26
- Lines added: +958
- Lines removed: -186

GitHub's own totals (gh pr view 46, verified live: changedFiles=27 additions=975 deletions=196):

changedFiles  27
additions     975
deletions     196

The 26 matches the walkthrough table exactly — 20 rows + "…and 6 more file(s)" = 26 — confirming the count is the ignore-glob-filtered file list, not GitHub's total (27). The line counts are short by the same reasoning (+958/-186 vs +975/-196), i.e. one file was dropped by the ignore-glob.

Secondary (minor, optional): "Key Findings" cited the HIGH finding at ci.yml:4 while the inline comment anchored ci.yml:5 — a one-line drift, likely a separate anchor issue.

Where

• PrSummaryGenerator.generate — the "Changes Overview" rendering (PrSummaryGenerator.java:88–105).
• VerdictBuilder.DiffStats.fromFiles — the counts fed in, computed from reviewableFiles (VerdictBuilder.java:155–172).
• VerdictBuilder.build — the DiffStats.fromFiles(ctx.reviewableFiles(), ctx.omittedFiles()) call site (VerdictBuilder.java:60); the summaryGenerator.generate(...) call is in VerdictBuilder.buildResult (VerdictBuilder.java:229).
• ReviewContextLoader.ReviewContext — already carries the full unfiltered files list (ReviewContextLoader.java:82), but no PR-level totals.
• GitHubPullRequestClient.PullRequestDetails — currently only models title, body, head, base (GitHubPullRequestClient.java:98); it does not carry changed_files/additions/deletions, so the fix must add those fields.

Proposed fix

1. Surface GitHub's authoritative PR totals — changed_files, additions, deletions — by adding them to PullRequestDetails (the same fields gh pr view reads), plumbed through ReviewContext alongside the existing file list.
2. Populate the "Changes Overview" filesChanged/additions/deletions from those totals rather than from DiffStats.fromFiles(reviewableFiles).
3. Keep the 20-row walkthrough cap (MAX_FILE_ROWS) and the "…and N more file(s)" note as-is — but derive the "N more" from the authoritative file total, not from the reviewable-list size, so the rollup count matches too.
4. On a truncated PR, "Changes Overview" should reflect the full GitHub totals while the walkthrough table and truncation banner disclose the omission (ties to fix(review): disclose diff truncation and don't auto-APPROVE a truncated review #234). Interim: if PR-level totals cannot be fetched, fall back to the current diff-derived counts rather than failing the summary.

Acceptance criteria

• "Changes Overview" Files changed / Lines added / Lines removed equal GitHub's changed_files / additions / deletions for the PR.
• On MongOCOM#46's shape (27 files, +975/-196, 20-row cap), the overview reads 27 / +975 / −196 and the walkthrough still shows 20 rows + "…and 7 more file(s)".
• Non-truncated PR with no ignore-globbed files: overview counts unchanged.
• Ignore-globbed PR: overview shows full GitHub totals; the ignored files are still excluded from the walkthrough table.
• Truncated PR: overview shows full GitHub totals while the walkthrough rollup + truncation banner disclose the omission.
• PrSummaryGeneratorTest covers a case where reviewable-file counts diverge from the passed-in authoritative totals.

Related

• feat(review): add a walkthrough and file-by-file summary to the PR summary #39 — walkthrough / file-by-file summary origin (same reviewableFiles source; the walkthrough rollup count must move to authoritative totals too).

• fix(review): disclose diff truncation and don't auto-APPROVE a truncated review #234 — truncation disclosure (the truncated-PR case of the fix must reflect full totals while disclosing omission).

• feat(review): pluggable per-model tokenizer for exact token counts #239 — NOT related: per-model tokenizer for token counts, unrelated to file/line counts.

• Sibling findings from the same v0.3.0 dogfood: fix(review): /describe, /changelog, /add-docs run on a truncated diff with no partial-coverage disclosure #296 (on-demand-command truncation disclosure) and fix(webhook): /summary reports "a summary already exists" from persistence state, not the live comment #297 (/summary "already exists" wording).

[devops-thiago]

Dogfood follow-up (audit of PRs #305–#309): this bug has a second facet beyond the ignore-globbed dropped file — the summary's line counts can diverge from GitHub even when the file set matches.

• Line divergence, no file dropped — PR fix(review): include nested review-thread replies in MaintainerReply context (#303) #306 (2 files, none ignored): summary reported +57 / −5 while GitHub's authoritative totals are +58 / −7.
• Dropped file (the original facet) — PR fix(security): document/suppress non-applicable jackson GHSA-5jmj-h7xm-6q6v (no 2.x fix) #308: pom.xml is ignore-globbed → summary 1 file / +22 / −0 vs GitHub 2 / +32 / −5, and the summary prose even mentions the pom.xml edit while the Changed Files table omits it.

Both are covered by the authoritative-totals approach in #309 (populate Files changed / additions / deletions from GitHub's changed_files / additions / deletions). Please confirm the line-count facet (the #306 shape) is fixed too, not just the dropped-file facet.

[devops-thiago]

Implementation ready for review in PR #309 — close when merged to main.