Stop the AI review failing on a large PR, and point it at the right files

[ako]

review failed on #1107 and #1106. The diagnosis is the one in the title, plus two things it uncovered.

The failure

could not find pull request diff: HTTP 406: Sorry, the diff exceeded the maximum number of lines (20000)
PullRequest.diff too_large

gh pr diff reads an endpoint capped at 20,000 lines. #1107 is 21,366 changed lines, 1,366 over. gh exits 1, the step runs under bash -e, the job dies — and it is the only hard-fail path in a workflow whose every other error (missing key, non-200, empty content) is a warning plus exit 0.

The per-file endpoint has no such cap. Measured on #1107 it returns all 278 files, 3 without a patch (the .bson fixtures). The diff is rebuilt from those, and a fetch problem now downgrades the review instead of failing the PR.

What the review was actually reading

head -c 80000 truncates by bytes across an alphabetically ordered diff. #1107's full diff is 1.29 MB, so the window covered 3 of 275 files — all of them .claude/skills/fix-issue/findings/*.jsonl, the append-only evidence logs. Zero Go code. #1095 was cut to 23% and #1090 to 19%, so this is what every large review has been looking at.

Selection is now file-aware (.github/scripts/build_review_diff.py):

• generated / fixture / append-only paths skipped (mdl/grammar/parser/, testdata/, findings/*.jsonl, binaries)
• whole file patches, most-changed first
• no file may exceed a tenth of the budget

That last clause is load-bearing. Ordering by size with no cap put three files in the budget and nothing else; ordering alphabetically missed mdl/backend/mpr/backend.go, the 1,163-line change #1107 is named after. With both, #1107 yields 11 source files including that one.

A manifest of every changed file goes into the prompt regardless of budget, and the note tells the model to say when a finding needs a file it cannot see rather than assume absence.

Normal PRs are unaffected — #1098 (752 lines) comes through as 13 of 14 files, nothing over budget.

Verification

Both rehearsed locally against the real API payloads:

PR	before	after
#1107 (21,366 lines)	job fails at step 1	11 source files, 278-file manifest
#1098 (752 lines)	14 files	13 files, unchanged behaviour

Note pull_request_target runs the workflow from the base branch, so this PR is reviewed by the old workflow — which is useful: it is the first run since the key rotation that can reach the model, so it tests that independently. The new path takes effect for PRs opened after merge.

Not fixed here

Every "successful" run since at least 2026-09-04 got HTTP 401 from OpenRouter and posted nothing — the green checks were not reviews. The key has been rotated, which is outside this repo.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BNDe35kDNsMX5cz4Ahn4rk

[github-actions]

AI Code Review

Critical Issues

None found.

Moderate Issues

None found.

Minor Issues

• In .github/scripts/build_review_diff.py, the SKIP tuple contains "mdl/grammar/parser/" which is correct to skip generated parser files, but the comment in CLAUDE.md says to "note but skip" these files. The implementation correctly skips them, so this is just a minor observation.
• The SKIP_SUFFIX includes .lockb (likely meant for LevelDB lock files) but this seems unrelated to the project. However, it's harmless to include.

What Looks Good

• The solution correctly addresses the root cause: the GitHub diff API's 20,000 line limit that was causing HTTP 406 errors on large PRs.
• The new script .github/scripts/build_review_diff.py implements a smart selection strategy:
  • Uses per-file API endpoint (no line limit)
  • Prioritizes code files (.go, .g4) over other types
  • Skips noise paths (generated code, fixtures, append-only logs)
  • Enforces per-file limits to prevent a few large files from dominating the budget
  • Includes a manifest of all changed files so the model knows what it's missing
• The workflow updates handle failure cases gracefully:
  • If fetching PR files fails, it downgrades the review instead of failing the PR
  • If no reviewable diff is available, it exits cleanly
  • Provides clear notes to the model about the selection process
• Local verification shows the fix works correctly:
  • Large PR Sync ako/mxcli: retire the legacy engine (sdk/mpr importers 27 → 0), workflow authoring, layouts, MCP fixes #1107 now produces 11 source files + manifest instead of failing
  • Normal PRs like feat(describe): Mode 2 — describe crossed branches faithfully by default #1098 remain unaffected
• The implementation follows the principle described in the PR: "a fetch problem now downgrades the review instead of failing the PR"
• The solution is minimal and focused - only touches the review workflow without changing core functionality

Recommendation

Approve. This PR effectively fixes the issue where the AI review was failing on large PRs due to the diff endpoint's line limit. The solution is well-designed, properly tested against real API payloads, and maintains review quality while avoiding hard failures on large PRs. The changes are scoped appropriately to the review workflow only.

---

Automated review via OpenRouter (Nemotron Super 120B) — workflow source

[github-actions]

AI Code Review

Review Summary

This PR modifies the AI review infrastructure to handle large PRs by switching from the GitHub diff API (which has a 20,000-line limit) to the per-file API endpoint, then intelligently selecting which files to include in the review diff based on a byte budget and file type priority.

Critical Issues

• None found. The PR successfully addresses the core issue where large PRs (like Sync ako/mxcli: retire the legacy engine (sdk/mpr importers 27 → 0), workflow authoring, layouts, MCP fixes #1107 with 21,366 lines) were causing the review job to fail due to HTTP 406 from the diff endpoint.

Moderate Issues

• None found.

Minor Issues

• Script clipping logic: In build_review_diff.py, when a patch lacks newlines within the first PER_FILE bytes, patch.rfind("\n", 0, PER_FILE) returns -1, resulting in patch[:0] (empty string). While this edge case didn't manifest in testing (PR Sync ako/mxcli: retire the legacy engine (sdk/mpr importers 27 → 0), workflow authoring, layouts, MCP fixes #1107 worked), it could produce confusing output for patches without newlines in the budget window. Consider improving the clipping logic to handle this case (e.g., take first PER_FILE bytes when no newline found).
• Missing script tests: As infrastructure code, the script lacks unit tests. While manually verified against real PRs, adding test coverage would improve robustness (e.g., testing edge cases like empty patches, noise filtering, budget allocation).

What Looks Good

• Effective solution: Switching to the per-file endpoint eliminates the hard line limit failure. The selection strategy (most-changed files first, skipping noise/generated files, per-file byte cap) ensures the review focuses on meaningful changes.
• Manifest inclusion: Providing a complete file manifest in the prompt allows the reviewer to note when findings depend on unseen files rather than assuming absence.
• Robust error handling: Fetch failures downgrade to an empty diff (warning) rather than failing the job, maintaining CI stability.
• Normal PRs unaffected: Verified that small PRs (like feat(describe): Mode 2 — describe crossed branches faithfully by default #1098) continue to work identically.
• Clean implementation: The Python script is well-commented, follows the budget logic described, and integrates cleanly with the workflow.

Recommendation

Approve. The PR fixes a critical blocker for reviewing large PRs with minimal risk. The minor issues (clipping edge case, missing script tests) are low-impact and can be addressed in follow-up work. The solution has been validated against the problematic PRs (#1107, #1106) and maintains existing behavior for normal PRs.

Note: The OpenRouter authentication issue mentioned in the PR description is external to this repository and correctly out of scope. This PR solely addresses the diff retrieval mechanism.

---

Automated review via OpenRouter (Nemotron Super 120B) — workflow source