[kylehgc]

Closes #102.

What this adopts

Adoption of upstream rtk-ai/rtk#3646 by @ninearif (Arif Waram): keep +/-/@@ diff markers at column 0 in compact_diff.

The ticket named upstream #3269 (@iliaal). During Phase 2 a cluster check found rtk-ai#3646, opened 2026-08-23, which is a strict superset of rtk-ai#3269:

• Same core fix — all four format!(" {}", line) sites emit at column 0 (hunk header, added, removed, context).
• Additionally re-spells the per-file tally +N -M → (+N -M), matching the notation diff_cmd.rs already uses, so the tally can never be read as an added line even if a consumer strips indentation.
• Stronger regression test (test_compact_diff_keeps_diff_markers_at_column_zero): asserts @@ at column 0, context-line leading space, an exhaustive filter that only real hunk lines start with +/-, and the tally re-spell.

Adopting rtk-ai#3646 leaves no residual gap; rtk-ai#3269 is referenced as the passed-over alternative. Both fix the same defect; they conflict textually and cannot both be cherry-picked.

Commits

Commit	Author	Kind
fix(git): keep +/-/@@ diff markers at column 0 in compact_diff	Arif Waram (@ninearif)	Cherry-pick of upstream rtk-ai#3646, byte-identical, clean apply on post-sync develop

No fork-authored amendments were needed: all 129 existing git-module tests and the full suite pass unchanged — no fork test had encoded the old indent.

Bug

compact_diff indented every hunk header and hunk body line by two spaces, so any anchored consumer of rtk-filtered diff output (grep -E '^[+-]', grep '^-') silently matched nothing. Because the hook rewrites git → rtk git transparently, the empty result is indistinguishable from "no matching changes" — a convincing false negative, not an error. Affects all compact_diff callers: git diff, git show, git stash show, gh/glab diff views, rtk pipe.

Repro (real binaries, scratch repo with 3 changed lines)

Raw git:

$ git diff | grep -cE '^[+-]'
5        # 2 file headers + 3 real changed lines

Before (installed rtk.exe, pre-fix):

$ rtk git diff | grep -cE '^[+-]'
0
$ echo $?
1        # exit 1, empty stdout — looks exactly like "no changes"

Filtered output had every marker at column 2:

lorem.md
  @@ -1,6 +1,7 @@
  -Sed do eiusmod tempor.
  +Sed do eiusmod TEMPOR.
   Ut enim ad minim.
  +- new bullet
   Duis aute irure.
  +2 -1

After (this branch's build):

$ rtk git diff | grep -cE '^[+-]'
3        # exactly the real changed lines
$ rtk git diff | grep -E '^\+.*bullet'
+- new bullet

lorem.md
@@ -1,6 +1,7 @@
-Sed do eiusmod tempor.
+Sed do eiusmod TEMPOR.
 Ut enim ad minim.
+- new bullet
 Duis aute irure.
  (+2 -1)

(3 vs raw 5: compact mode drops the --- a/ / +++ b/ file headers by design; all real hunk lines are found and the tally no longer masquerades.)

Quality gate

cargo fmt --all && cargo clippy --all-targets && cargo test --all — clean on the branch (2717 unit tests + all integration suites, 0 failures, clippy zero warnings). The same gate also ran green on the merged develop right after today's upstream sync (13 commits), before branching.

Deliberately not fixed

• Compact output is still not an applyable patch (diff --git/---/+++ headers are dropped by design) — use rtk proxy git diff for a faithful patch. The indent was never what made that case fail.
• Upstream fix(git-diff): emit hunk lines at column 0 so ^- anchors again rtk-ai/rtk#3269 stays open upstream; no comment was posted there (declining/steering upstream PRs is contributor outreach, a separate decision).

[kylehgc]

Review round (pre-merge)

Three passes ran on the branch diff: upstream's code-reviewer.md prompt (uncontaminated, general-purpose agent), and /code-review's Standards + Spec axes.

Verdicts: upstream reviewer APPROVE, 0 critical / 2 important; Spec axis: no missing requirements, no scope creep; Standards axis: 2 hard findings, 2 judgement calls.

Acted on

• Stale doc example (Standards, hard — CONTRIBUTING §Documentation): docs/usage/FEATURES.md still showed the pre-fix indented rtk git diff format (and was already stale beyond that: a ~ marker and (+5/-2) tally the filter never emitted). Redrawn from real output of the new format in fork-authored amendment 22ee0d4. Docs-only change; the green gate on 75ec765 stands (no Rust code references the file).

Deliberately left byte-identical (cherry-pick integrity)

• Comment overstates patch as an anchored consumer (flagged by all three passes): compact output is never a valid patch regardless (file headers dropped, leading context dropped). Editing the contributor's commit would break byte-identical adoption and invite a conflict at the sync after upstream merges fix(git): keep +/-/@@ diff markers at column 0 in compact_diff rtk-ai/rtk#3646. Upstream's reviewer will see the same line on fix(git): keep +/-/@@ diff markers at column 0 in compact_diff rtk-ai/rtk#3646.
• Test fixture nit: @@ -1,4 +1,4 @@ header vs a 3-line old-file body — harmless (compact_diff ignores counts), same reasoning.

Follow-ups (out of scope for this adoption)

• condense_unified_diff in src/cmds/git/diff_cmd.rs:373,401 (the … | rtk diff stdin path) still indents changed lines two spaces — same silent-false-negative class, but a separate command with a deliberately different ultra-condensed format ([file] headers, no @@/context), outside compact_diff's call graph. Needs its own ticket and design decision (its output does not claim to look like git diff).
• No ≥60% count_tokens savings test exists for compact_diff — pre-existing gap; this change only shrinks output (−2 bytes per hunk line, +2 per file tally).

Cross-checked reviewer claims

• All 6 compact_diff call sites traced (git diff/show/stash-show, gh, glab, pipe); none re-parse or re-indent the output.
• CRLF: str::lines() keeps a trailing \r, which does not disturb the column-0 anchor — holds on Windows git output.