Skip to content

Fix stale commit attribution in PR path overlap reports - #3104

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
Arham-Tahir64:codex/reviewable-result-01a1474b
Oct 17, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
Arham-Tahir64:codex/reviewable-result-01a1474b

Conversation

@Arham-Tahir64

Copy link
Copy Markdown
Contributor

Path overlap searches can currently attach commit B to a file list fetched at commit A: the file-pagination code checks the original head, then discards it before caching. A separately refreshed PR-body snapshot supplies the reported head.

Retain the file-list head, report the body head separately, and flag unverified joins. Coverage also names missing or mismatched inventories when no path matches, so an empty result cannot silently hide that gap. Online path queries refresh legacy headless file caches; offline or failed refreshes retain explicitly unbound evidence. Full file pagination, stale leads, body-only fetch costs and per-source atomic replacement remain covered.

Validation:

  • python3.12 -B -m unittest discover -s tests -p 'test_research_*.py' -v: 214 tests, OK with 18 native Linux/Windows skips on macOS.
  • Six new behavioral tests cover A/B-head disagreement, matched-head controls, legacy caches, no-hit warnings, missing inventories, body-plus-path queries, offline immutability, automatic migration and failed-refresh preservation. The original implementation fails the new counterexample controls.
  • Live PR3100 response: all eight paths retain head 72380f027199199ed4b4dbfc252ec39163acc56c; an intentionally different body head is flagged. No full live file-index refresh is claimed.
  • git diff --check passes.

Task and overlap: canonical main 18aacb215fc3717b06be8ceb9fca62aeca0d3e99, priority revision 6, chooser seed 5df01753e885c02e48a4482ad39d7a09 (adversarial audit). Scope changed from the supplier-boundary card to a concrete source-binding counterexample in its prescribed coordination tool. Final overlap refresh at 2026-10-17 00:48 UTC covered 254 open PR bodies and 1,861 Discussion records with complete pagination; targeted all-state searches inspected the predecessor tooling. Sparse local notes and unsearched diffs remain limitations, not a novelty guarantee.

Credit Douglas Colkitt's overlap/cache helper, Aaron Wang/navinor's #1623 path index, and ywz65's #2517 complete continuation paging. This patch adds the binding between file and body snapshots. Historical overlap-path-index and overlap-discussion-batching pins already differ from the base script; their frozen packages are unchanged, and their registered replays are not claimed. No mathematical supplier replay, new kappa or selected-frontier change.

Arham Tahir, with substantial OpenAI Codex assistance. Apache-2.0; no independent human or formal review claimed.

@Mister-Meeseeks
Mister-Meeseeks merged commit e25b980 into CrocSwap:main Oct 17, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants