[PR #586] test: add cross-repo CWD attribution tests #591

New issue

Closed

opened 2026-03-02 04:14:03 +03:00 by kerem · 0 comments

kerem commented

2026-03-02 04:14:03 +03:00

Owner

Original Pull Request: https://github.com/git-ai-project/git-ai/pull/586

State: closed
Merged: Yes

test: add cross-repo CWD attribution tests

Summary

Adds 9 integration tests in a new tests/cross_repo_cwd_attribution.rs file verifying that AI attribution is correctly resolved when the agent's CWD (where the hook/checkpoint is called from) differs from the repo root(s) where files are being edited.

Scenarios covered (matching the 5 requested):

#	Scenario	Test(s)
1	CWD is an unrelated repo, single target repo	`test_cwd_different_from_repo_root_single_repo`
2	CWD is unrelated, edits across 3 different repos	`test_cwd_different_from_repo_root_multiple_repos`
3	CWD is one of the edited repos + edits in 2 other repos	`test_cwd_is_one_of_edited_repos_plus_others`
4	CWD is a parent dir above repos (~/projects style)	`test_cwd_is_parent_dir_above_repos_single_repo`, `test_cwd_is_parent_dir_above_multiple_repos`
5	CWD is parent dir, edits in deep subpaths of multiple repos	`test_cwd_parent_dir_edits_in_repo_subpaths`, `test_cwd_parent_dir_multiple_files_per_repo_subpaths`

Two additional blame-level verification tests confirm that git-ai blame correctly distinguishes human vs AI lines in cross-CWD scenarios.

No production code changes — test-only PR.

Review & Testing Checklist for Human

Verify that in multi-repo tests (scenarios 2, 3, 5), git_ai_from_working_dir is called on one TestRepo object but applies to files across all repos — confirm that the shared GIT_AI_TEST_DB_PATH env from the calling repo doesn't mask issues that would appear in production (where each repo has its own storage)
Most tests only assert !attestations.is_empty() — they don't verify which files the attestations are for. Consider whether a test that writes to repo_a and repo_b could pass even if attestations were swapped between repos
In scenarios 1 and 4, working log assertions are placed before the commit — confirm this ordering is correct (working logs are keyed by base commit SHA at checkpoint time, so they must be read before the commit advances HEAD)
Only 2 of 9 tests (test_cross_repo_cwd_blame_shows_correct_attribution, test_parent_cwd_blame_correct_across_repos) verify blame line-by-line; the other 7 only check attestation presence. Decide if blame-level checks should be added to more scenarios

Notes

All 9 tests passed locally and on CI (Ubuntu, macOS; Windows still in progress at time of PR)
One bug was caught and fixed during development: working log assertions were initially placed after commits, which would fail because the working log is keyed by the base commit SHA at checkpoint time. This was corrected to check working logs before committing.
Tests use git_ai_from_working_dir to simulate running checkpoints from different CWDs, matching the pattern in existing multi_repo_workspace.rs tests
Workspace cleanup is manual via fs::remove_dir_all at test end; if tests panic, temp dirs may leak (minor issue, they're in system temp)

Link to Devin run: https://app.devin.ai/sessions/9dc8a791659f4feea7cbbd57b5a1ce61
Requested by: @svarlamov

**Original Pull Request:** https://github.com/git-ai-project/git-ai/pull/586 **State:** closed **Merged:** Yes --- # test: add cross-repo CWD attribution tests ## Summary Adds 9 integration tests in a new `tests/cross_repo_cwd_attribution.rs` file verifying that AI attribution is correctly resolved when the agent's CWD (where the hook/checkpoint is called from) differs from the repo root(s) where files are being edited. **Scenarios covered (matching the 5 requested):** | # | Scenario | Test(s) | |---|----------|---------| | 1 | CWD is an unrelated repo, single target repo | `test_cwd_different_from_repo_root_single_repo` | | 2 | CWD is unrelated, edits across 3 different repos | `test_cwd_different_from_repo_root_multiple_repos` | | 3 | CWD is one of the edited repos + edits in 2 other repos | `test_cwd_is_one_of_edited_repos_plus_others` | | 4 | CWD is a parent dir above repos (~/projects style) | `test_cwd_is_parent_dir_above_repos_single_repo`, `test_cwd_is_parent_dir_above_multiple_repos` | | 5 | CWD is parent dir, edits in deep subpaths of multiple repos | `test_cwd_parent_dir_edits_in_repo_subpaths`, `test_cwd_parent_dir_multiple_files_per_repo_subpaths` | Two additional blame-level verification tests confirm that `git-ai blame` correctly distinguishes human vs AI lines in cross-CWD scenarios. No production code changes — test-only PR. ## Review & Testing Checklist for Human - [ ] Verify that in multi-repo tests (scenarios 2, 3, 5), `git_ai_from_working_dir` is called on one `TestRepo` object but applies to files across all repos — confirm that the shared `GIT_AI_TEST_DB_PATH` env from the calling repo doesn't mask issues that would appear in production (where each repo has its own storage) - [ ] Most tests only assert `!attestations.is_empty()` — they don't verify *which* files the attestations are for. Consider whether a test that writes to repo_a and repo_b could pass even if attestations were swapped between repos - [ ] In scenarios 1 and 4, working log assertions are placed **before** the commit — confirm this ordering is correct (working logs are keyed by base commit SHA at checkpoint time, so they must be read before the commit advances HEAD) - [ ] Only 2 of 9 tests (`test_cross_repo_cwd_blame_shows_correct_attribution`, `test_parent_cwd_blame_correct_across_repos`) verify blame line-by-line; the other 7 only check attestation presence. Decide if blame-level checks should be added to more scenarios ### Notes - All 9 tests passed locally and on CI (Ubuntu, macOS; Windows still in progress at time of PR) - One bug was caught and fixed during development: working log assertions were initially placed after commits, which would fail because the working log is keyed by the base commit SHA at checkpoint time. This was corrected to check working logs before committing. - Tests use `git_ai_from_working_dir` to simulate running checkpoints from different CWDs, matching the pattern in existing `multi_repo_workspace.rs` tests - Workspace cleanup is manual via `fs::remove_dir_all` at test end; if tests panic, temp dirs may leak (minor issue, they're in system temp) --- Link to Devin run: https://app.devin.ai/sessions/9dc8a791659f4feea7cbbd57b5a1ce61 Requested by: @svarlamov