feat(chat): chip recognised slash-commands anywhere in the message buffer

[ZeroPoint95]

Ready to review. All four open items from the first review are done, and a reviewer found a gap with waiting messages, which is now fixed too. CI is running on the newest commit, 68db4a0a2.

What changes

Every slash command codeaf recognises is now drawn as a chip, wherever it appears in the message: in the box, on home's line and in the transcript.

try running /senior-dev on this   →  /senior-dev is chipped (it used to be plain)
explain /compact this             →  /compact is chipped, and enter sends it as ordinary words
/standing tag this                →  no change, /standing still routes
say /standing  + backspace        →  plain in the box, and still plain after sending

What stays the same

Pressing enter does exactly what it did before:

1. A command at the start of the message runs.
2. /standing, /orders and /task still act anywhere in the message.
3. Any other command in the middle of a sentence is sent as text, even though it now has a chip.

So the chip used to mean "enter will act on this". Now it means "codeaf knows this command".

The three commits

1. 745fc854c: chips anywhere. commandSpans chips any recognised command, and the transcript draws the same wider chips.
2. 617ed0555: a tag backspaced to plain words stays plain once sent.
   • enterLine copies the backspaced ranges (plainTags) before the box is cleared, and passes them on to the transcript entry.
   • wrapWithOffsets records where each wrapped row starts, so transcriptCommandSpans can skip those ranges on every row. tabExpandedSegments adjusts the ranges when a tab is widened to spaces.
   • actedTags is gone, and plainTags replaces it.
   • Adds the change entry, and commands.md and screen.md now say a chip means "codeaf knows this command".
3. 68db4a0a2: the same holds for a message typed while an answer is still coming. Such a message waits in a queue until the answer finishes. That queue used to drop the backspaced ranges, so the tag came back chipped.
   • Each waiting message now keeps its ranges: park stores them, and both sendParked and the background-conversation send pass them on.
   • Pressing ↑ to pull a waiting message back into the box keeps the tag plain there too.
   • plainWithoutTag handles the /standing-tag case: that path queues the words without the tag, so the ranges have to be moved to match.

Tests

• The full go test ./internal/tui3 and ./internal/manual suites pass on 68db4a0a2, and go vet is clean.
• 4 new tests: a waiting message stays plain, a recalled message stays plain, a backspaced tag beside a live /standing stays plain, and a table test for plainWithoutTag. The first three fail when the fix is removed.

Try it

1. Run make build, then bin/codeaf.
2. Ask something long. While the answer streams, type say /standing, press backspace, then enter.
3. When the answer finishes and the message sends, /standing is plain, with no chip.

Trial binary: sha256 2744a5f6d27e6a8b1798be1b89ffadc17997c6a536a3f5b3f4da9dde55c00900, commit 68db4a0a2.

🤖 Generated with Claude Code

[AbirAbbas]

Taking over: @ZeroPoint95, checked by hand in the real binary on both roads (engine and chat --no-host, at 120 and 62 columns). Chips show in the box, on home's line and in the transcript. explain /compact this is sent as words, and /standing mid-sentence still routes. A backspaced tag stays plain through a direct send, ↑ recall and a cmd+enter waiting message. Paths, URLs, /modelx and (/model) stay plain.

Pushed:

• b469e7311 tui3: covers the places a message is drawn that the PR missed. A send with a picked harness dropped the demoted ranges, so say /standing made plain came back chipped. The guard's "send to conversation", a reopened conversation or a second window on it, a follow-up, and a question's replacement or clarification all drew a door word nobody acted on as a chip. On those roads a door word away from the head now stays plain, as it did before. Plain enter while an answer streams now steers the sentence into the turn (the └ row), and that row drew no chips at all, so /compact was chipped in the box and plain once sent. The waiting block had the same gap. Both now paint through the transcript's painter. The tab stop is now one constant, and the manual and the change entry are updated to match.
• 0597dff3d merges dev, which includes chat: the /model menu's current model wears no band and the picker owns the wheel #1696.
• 7ab7bc6e3 docs(chat): since chat: the /model menu's current model wears no band and the picker owns the wheel #1696, the model in use in /model has no band. The chip section and screen.md no longer say it shares the chosen row's tint.

Left for later:

• hi<TAB>/model is plain in the box but chipped once sent.
• Spaced inline code (` /model `) and fenced code blocks chip commands.
• A second window's replay draws /senior-dev plain, because the rows are cached before the program commands install.
• The rows are scanned one at a time, so a hard-wrapped fragment can chip, and a command split across narrow rows loses its chip.
• When chat: ctrl+enter queues a follow-up, and a click takes a queued message back #1714 lands, its recallQueuedAt should carry the demoted ranges, and its input.go conflict should keep plain (a.park(line, false, plain), a.submitShown(line, shownLine, plain)).

[AbirAbbas]

Verified by hand in the real binary on both roads (120 and 62 columns). Merging.