^{PR Author Agent⚡}

👋 I've got this PR — here's what I'll handle for you:

• Review feedback — implement suggestions, answer questions, fix what comes up
• CI failures — I get pinged when checks fail and try to fix them
• Merge conflicts — bring the PR back up to date when it falls behind
• Merge gates — watch CI, reviews, and verification
• Ping you the moment it's ready to merge

Marking it ready and picking reviewers are your call — I'll leave both alone.

Drop a comment anytime!

_{Note: opened under @mgmonteleone's identity — the org bot lacks PR write access on this repo, so attribution may differ from other automated PRs.}

^{PR Author Agent⚡}

✅ Ready for review on 07df2f8:

• No CI checks are configured on this repo, so none will run.

Reviewer pick is your call — I won't request anyone myself. I'll respond to any review feedback as it comes in (no merge — that stays a human gate).

Minor UX thought: if the user switches out of the float (e.g. <C-w>w) it's orphaned with no clear way back. Worth having :Augment chat-input refocus an existing float instead of opening a new one? Avoids losing typed content on accidental focus loss. Not blocking.

^{PR Author Agent⚡}

Good call on the orphaned-float UX. Implemented in f35f72a: augment#chat#OpenInputWindow now checks for an existing valid input window first and, if found, refocuses it (nvim_set_current_win + startinsert) and returns early instead of opening a new float. The existing buffer and its submit callback are preserved, so typed content survives an accidental focus change (e.g. <C-w>w).

Updated the README and :help doc to note the refocus behavior, and extended the headless smoke test to cover it (re-invoking after moving focus away keeps the same window and its content). All checks pass (SMOKE_OK).

^{PR Author Agent⚡}

Resolved the merge conflict with main (merge commit 78b1057). The conflict was a CHANGELOG.md overlap with #67's :Augment help entry — both Unreleased bullets are now kept. I also registered chat-input in the new :Augment help command list (added by #67) so it shows up in :Augment help and :Augment help chat-input.

Re-ran the headless Neovim smoke test → SMOKE_OK. PR is mergeable: clean again. No completion/suggestion surface touched.

Found during manual QA (CSS-902): whitespace guard misses pure-newline messages.

Repro: :Augment chat-input → press <CR> (or several) → <C-s>. A blank message gets sent instead of being treated as cancel.

Cause: the guard in s:ChatInputSubmit (and s:CommandChat) uses \s, which in Vim regex is [ \t] only — newlines aren't matched.

if a:message ==# '' || a:message =~# '^\s*$'

Fix: swap \s for \_s (whitespace incl. newlines) at both sites:

if a:message ==# '' || a:message =~# '^\_s*$'

The empty-string check is then redundant but harmless. Worth a sub-case in the smoke test (buffer = ["", ""] → expect cancel).

Found during manual QA (CSS-902): visual selection is not restored after submit, contrary to PR description.

Repro: open any file, V + jj to select 3 lines, :'<,'>Augment chat-input, type a message, <C-s>. After the float closes, the source buffer has no visual selection — gv was never invoked.

Cause: was_visual is computed from mode() inside s:CommandChatInput, but by then Vim has already left visual mode (the : opened the cmdline), so mode() returns 'n' and was_visual is always false. That false value is passed as the reselect arg to s:ChatInputSubmit, so the normal! gv branch never runs.

let was_visual = index(['v', 'V', "\<C-v>"], mode()) >= 0  " always false here
...
let Callback = function('s:ChatInputSubmit', [source_win, ranged, was_visual])

Fix: drive reselect off ranged instead — when invoked as :'<,'>... the '</'> marks are set and gv works:

let Callback = function('s:ChatInputSubmit', [source_win, ranged, ranged])

(Or simpler: drop the dead was_visual plumbing entirely and have s:ChatInputSubmit reselect whenever a:ranged.)

Worth a smoke-test sub-case: invoke from visual-line mode, submit, assert mode() ==# 'V' afterwards.

^{PR Author Agent⚡}

Thanks for the QA finds (CSS-902) — both fixed in 629bb0f:

1. Newline-only submissions now cancel. Swapped \s → \_s in the empty-message guard at both sites (s:ChatInputSubmit and s:CommandChat), so a buffer of only blank lines is treated as cancel instead of sending an empty message.
2. Visual selection restored on submit. Dropped the always-false was_visual plumbing and drive reselection off a:ranged. Since the command leaves visual mode (setting '</'>) before opening the float, gv now works whether invoked from visual mode or via an explicit :'<,'> range.

Extended the headless smoke test with an empty-guard sub-case (blank/newline-only → cancel) and a gv reselection sub-case (asserts mode() ==# 'V' after the visual-exit → gv sequence). Re-ran → SMOKE_OK. Still chat-only; no completion surface touched.

Re-tested on 629bb0f (CSS-902) — both fixes confirmed ✅

• Case 5 (blank-line cancel): empty buffer, spaces-only, and pure-newline submissions all close the float with Chat cancelled. No more blank messages.
• Case 4 (visual range): V+3j → :'<,'>Augment chat-input → <C-s> now leaves the source buffer back in -- VISUAL LINE -- with the original 4 lines highlighted. Yank-after-submit returns the full selection.

Rest of the matrix from the previous run was already green. No new issues. 👍

^{PR Author Agent⚡}

@mgmonteleone Ready to merge on 629bb0f:

• ✅ No CI configured on this repo
• ✅ Approved by johnny2211 (re-tested the CSS-902 fixes — confirmed)
• ✅ All review threads addressed
• ✅ mergeable: clean

Holding for the human merge gate — I won't merge automatically.