Anchor OSC 8 links and OSC 133 marks to the buffer

[JohnCampionJr]

OSC 8 and OSC 133 both parse today, and both are stored as a current value plus an event: Terminal.CurrentHyperlink is the URL in force, ShellIntegrationState is the last mark seen. The events say something happened. Neither says where — and every use of either is a question about a position: hit-testing a click, jumping to the previous prompt, selecting a command's output, putting an exit status beside the command that produced it.

So the information arrives and is thrown away. This anchors both halves.

Where it goes, and why not in the cell

To the line, following what #34 did for pictures.

An id in the cell is the obvious alternative and it is closed off twice over. AttributeData.Extended is fully allocated — bits 0–8 flags, 9–11 underline style, 12–31 the interned underline colour from #36 took the last of it. And a new field on BufferCell takes it from 24 bytes to 32, measured at 22% on scroll-heavy output in #38.

The line costs nothing and already has the properties both need: it survives scroll and reflow by being moved rather than copied, and releases what it holds when it falls out of the scrollback. Null on every line that has neither, which is nearly all of them.

The cell is untouched — still 24 bytes, IsReferenceOrContainsReferences still false, and #42's layout test says so on every build.

Marks

Recorded on the line they were emitted on, with the column and, for D, the exit status. Kept as int? and null where none was reported, which is not the same as zero: only D carries a status and it is optional even there, since cmd.exe cannot read the previous command's status from its prompt and always sends a bare D.

Two decisions that would have failed silently:

Erasing does not take a mark with it. A mark records a position in the history, not anything about the content there. A shell redrawing its prompt with EL — which is most of them — would otherwise destroy the A mark it had just emitted, a moment before the prompt it marks is even printed. Intermittent-looking, not broken-looking.

Line reuse does. The ring hands back the object it is about to drop, so a mark left on it comes back as history that never happened. Same trap the placement work hit in #38.

TryFindPreviousPrompt / TryFindNextPrompt are here rather than in every host, because it is the same walk each time with one off-by-one worth getting right once: the search is strictly past the row given, so calling it from its own answer walks through the history instead of sticking on the prompt it just found.

Links

The half that touches the print path, so it is worth saying up front that it costs nothing measurable — checked rather than asserted, below.

The URL is a string on the line, not interned. Interning is free only when the set is bounded, and cluster text is: a terminal sees a handful of distinct emoji sequences in a session. URLs are not — ls --hyperlink over a large directory emits one per file — so a table nothing is ever released from would be a leak that grows with output. The line owns it, exactly as it owns its images.

Three writers keep the bookkeeping, and two are the ones that get missed. Print writes a cell at a time; both batched run writers bypass it entirely. Without those two, a link covers the text or does not depending on which writer took it. Removing them fails eight of the twelve tests, including one that writes the same input through both paths and compares, and one through the byte entry.

Guarded at the call as well as inside. NoteLinkRun is too big to inline, so an unguarded call would have every printed character pay one to be told there is no link and never has been — the shape that cost the alt-redraw corpus 12% earlier in this project.

Text written over a link is not part of it. A write takes those columns out of whatever covered them, splitting a link written through the middle — which a client can produce with nothing but a cursor move.

Measured

Against main, three runs of each alternating, 200M characters per corpus, through the harness from #42:

corpus	bytes/char	ns/char	Δ
scroll-ascii	0.00 → 0.00	1.68 → 1.68	+0.5%
sgr-churn	2.33 → 2.33	4.41 → 4.37	−0.8%
truecolor	1.72 → 1.72	4.50 → 4.58	+1.7%
alt-redraw	0.26 → 0.26	5.53 → 5.57	+0.7%
unicode	7.66 → 7.66	12.75 → 12.69	−0.5%
flood	0.00 → 0.00	41.46 → 42.09	+1.5%

Allocation is byte-identical on every corpus — the half that is a count rather than a measurement, and so immune to what else the machine was doing. Timing deltas are all inside the noise the job observed in itself.

Tests

1179 pass, twenty-two new. Each guard was checked by removing it and watching the right tests fail: the mark recycling one drops a single test, the batched link bookkeeping drops eight.

What this unblocks

Everything left on that roadmap row is now a consumer of this, and none of it belongs in the emulator: hit-testing a link is line.TryGetLinkAt(column); jump-to-prompt is the two helpers above; selecting a command's output and an exit status in the gutter are a host reading line.Marks.

Worth noting the OSC 133 half stopped being theoretical today — before the shell-integration work, stock bash and zsh emitted no marks at all, so there was nothing to anchor.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

There are confirmed correctness issues in hyperlink bookkeeping (ID reset, ordering invariants, and erase consistency) and integer-overflow edge cases in the new prompt-search APIs.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR anchors OSC 8 hyperlink spans and OSC 133 shell-integration marks onto BufferLine so consumers can query where links/marks occurred (hit-testing, prompt navigation, etc.) instead of relying on “current value + event” state only.

Changes:

• Add per-line storage for OSC 133 marks (LineMark) and OSC 8 hyperlink spans (LineHyperlink) on BufferLine.
• Update print paths to record hyperlink spans and OSC 133 handling to record marks at the current line/column.
• Add Terminal.TryFindPreviousPrompt / TryFindNextPrompt plus new unit tests for marks and links anchoring.

File summaries

File	Description
src/XTerm.NET/Terminal.cs	Adds prompt-navigation helpers that scan anchored marks.
src/XTerm.NET/InputHandler.cs	Records OSC 133 marks onto the current line and records hyperlink spans during printing.
src/XTerm.NET/Buffer/LineMark.cs	New value type representing an anchored OSC 133 mark (column/kind/exit).
src/XTerm.NET/Buffer/LineHyperlink.cs	New value type representing an anchored OSC 8 hyperlink span (column/len/url/id).
src/XTerm.NET/Buffer/BufferLine.cs	Adds per-line collections for marks/links and hyperlink span bookkeeping logic.
src/XTerm.NET.Tests/ShellIntegrationMarkAnchorTests.cs	New tests asserting OSC 133 marks are anchored and prompt navigation works.
src/XTerm.NET.Tests/HyperlinkAnchorTests.cs	New tests asserting OSC 8 hyperlinks are anchored/split/joined and paths agree.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 5
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.