Add attribution test writing guidelines to AGENTS.md

[jwiegley]

Document patterns and conventions for writing integration tests
that exercise the attribution system, including test harness setup,
daemon interaction, and assertion helpers.

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

[jwiegley]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• Add wall-clock timeout and SSH hardening to authorship note fetch #1184
• Add attribution test writing guidelines to AGENTS.md #1033 👈 (View in Graphite)
• Add integration tests for Graphite operations and missing attribution #1032
• Replace silent error swallowing with debug logging #1031
• Refactor CheckpointKind API and improve attribution/storage performance #1030
• Harden daemon: socket permissions, env safety, telemetry, rewrite locking #1029
• Fix security: credential permissions, CAS validation, token redaction #1028
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🟡 Documentation example uses .human() assertions for lines captured by an untracked (checkpoint --) checkpoint

The new code example at lines 270-285 uses checkpoint -- code.rs (which creates a legacy/untracked CheckpointKind::Human checkpoint per src/commands/git_ai_handlers.rs:1155) but then asserts .human() (known human) attribution for those lines. According to the conventions established earlier in this same file (AGENTS.md:52), checkpoint -- corresponds to untracked/unattributed changes, while .human() represents known-human attribution (which requires checkpoint mock_known_human). The correct assertions should use .unattributed_human() instead of .human(), or the checkpoint call should be changed to mock_known_human. The existing example at AGENTS.md:108-197 correctly uses checkpoint mock_known_human → .human() and checkpoint human → .unattributed_human(). This won't cause test failures since assert_lines_and_blame at tests/integration/repos/test_file.rs:316 treats Human and UnattributedHuman identically, but it will mislead agents/developers into writing tests with incorrect semantic expectations.

Prompt for agents

The documentation example at AGENTS.md lines 266-286 has an inconsistency between the checkpoint type used and the attribution assertions. The example uses `checkpoint -- code.rs` (which creates a legacy/untracked Human checkpoint) but then asserts `.human()` for those lines. There are two valid fixes:

1. Change the checkpoint in step 1 (line 271) from `checkpoint -- code.rs` to `checkpoint mock_known_human -- code.rs` so that lines are truly attributed as known human, matching the `.human()` assertions.

2. Alternatively, keep the `checkpoint --` call but change the assertions at lines 281-284 from `.human()` to `.unattributed_human()` for the non-AI lines.

Option 1 is preferable because the section title says "Write human content" and the intent is clearly to show known-human attribution. This would also be consistent with the existing example at AGENTS.md lines 108-197 which correctly uses `mock_known_human` for human-attributed content.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[svarlamov]

Sorry I'm so late to reviewing this. I ended up writing up a lot re tests in agentsmd that's on main already