[joelteply]

A resident's edits are checkpointed into wip(...) commits every act (workspace_transfer), so a working-tree diff is always empty. Aiko read that as "my previous receipts were phantom" and re-read the same files for 80 minutes (2026-11-02, 16:35-17:57Z) while 16 commits and 38 lines of her work sat on her branch. The tool told her nothing existed; this is a tool defect, not her slip.

• DiffScope replaces staged: bool. It has three cases: YourWork (the default), Unstaged and Staged.
• YourWork diffs from work_base: the newest merge-base of HEAD with the origin's integration refs (origin/HEAD, canary, main, master). With none, it diffs the working tree against HEAD.
• The result names what it was taken against, so an empty diff says what it is empty against.
• Porcelain status is parsed by its fields (index, tree, space, path) instead of slicing [..2].
• Regression test: a checkpoint leaves the tree clean, Unstaged is empty, and YourWork shows the work.

Measured

Tests ran on the M5 with serving down, 2026-11-02 18:12Z:

• cargo check --workspace --tests is clean.
• continuum-hands: 295 passed, 0 failed.
• Ratchets against canary: exit 0.

After deploy, the check is to read whether residents still report "phantom" or "empty diff" and restart their reads.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

[joelteply]

Changes requested at cd36f0c. The default should expose committed checkpoint work, and naming the comparison is useful. Two boundary defects remain:

1. work_base converts every merge-base/log/parse failure into absence, then YourWork silently runs diff HEAD. A checkout with committed work but no conventional origin ref therefore reports an empty diff again. The comment that nothing commits into such a checkout is false (local repositories, nonstandard integration branches and missing remote refs are all possible). Return a typed/result failure for an unavailable work baseline, propagate actual git failures, and use the checkout owner's recorded integration/base identity rather than guessing. Explicit Unstaged/Staged remain usable independently.

2. Choosing a merge-base by commit timestamp does not select the nearest ancestor or the branch this work came from. Commit dates need not be monotonic; a newer-dated older ancestor can win over the correct descendant. Multiple integration refs also do not identify the card's actual base. Resolve the recorded work base through its owner; report that exact identity in the result. Add regressions for committed work with an unavailable baseline and conflicting candidate refs with skewed commit timestamps. The existing single-ref happy path cannot detect these cases.

Also replace SHOWN_SHA/manual slicing with the existing shared ID renderer or git's actual short-id operation; 12 characters is not git's computed unambiguous short form.

No tests executed locally for this review.

[joelteply]

Scoped source clearance at 4e15c71. The comparison now resolves the checkout owner's AllocationEvidence.start_ref, propagates failures, refuses absent allocation evidence, and uses git's computed short ID. The tests cover committed checkpoints with a competing integration ref and refusal without allocation evidence, while explicit Unstaged remains available. These address my review requests. The result describes a merge-base against the recorded integration ref as currently resolved, not an immutable historical snapshot; retain that distinction. 296 passing hands tests are author-reported; no local test execution claimed.

[joelteply]

Re-read 40210b3 against cleared 4e15c71: only schema documentation/comment reduction, no runtime comparison change. Scoped source clearance stands. Core 1185/0 and hands 296/0 are author-reported, not locally rerun. Keep wording precise: the baseline is the recorded integration merge-base, not necessarily immutable branch-creation history.