Handoff — 2026-09-17 — terminal windows and tools on Windows

State

Done, tested where it can be, gate green, documented in the library and in the six editors, not committed, on main. git status in turbo-core lists the touched files; nothing is staged. The Windows path is compiled and vetted, never run.

Next steps

Commit and tag turbo-core v0.9.0 — minor, because tools.Shell (a constant) became tools.Shell() and tools.ShellArgs(). 01-release.tag.sh runs make check; in this sandbox release_test.go fails on the cp-NUL problem (documented in summary.md), not on the code — run it on the Mac.
Re-pin the six editors (go get codeberg.org/turbo-editors/turbo-core@v0.9.0 && go mod tidy && GOWORK=off make check), then tag each. Their documentation already describes Windows terminal windows; until the re-pin the binaries do not have them.
Run the five-step protocol on a Windows machine, the first anyone will have: (a) F8 — a cmd.exe prompt appears and dir lists the directory; (b) drag the window's corner — the prompt repaints at the new width (mode con reports it); (c) Alt-<toolchain> ▸ a tool with output = "terminal" (Run) — the program runs and, when it ends, Ctrl-W closes the window (the console-closing goroutine gave the reader its EOF); (d) a long command in a terminal window — Ctrl-C interrupts it; (e) a tool with output = "popup" — output arrives, and Escape while it runs stops the whole tree (the job object). Report what the window showed, in SGR if possible.
If (c) hangs — the window never says the command finished — the suspect is the order in windowsChild.Close / closeConsoleWhenTheShellExits: conhost may need the input pipe closed before ClosePseudoConsole returns. If the first screen is garbage, the suspect is conhost's start-up sequences reaching an emulator that has only ever read Unix shells; capture the raw bytes (TURBO_ACP_TRACE does not cover terminals — add a temporary io.TeeReader in View.read).

Watch out for

unsafe.Pointer(handle) is what go vet flags, which is why UpdateProcThreadAttribute is called through a LazyProc with uintptr arguments. Do not "simplify" it back to x/sys's Update.
cmd.exe's command line is composed verbatim ("cmd.exe" /S /C "<command>"), in two places that cannot share code without breaking the package layering: terminal/windows.go (windowsCommandLine) and tools/shell.go (cmdExeCommandLine). Change both or neither.
The job object is assigned after Start; a child cmd.exe spawns in those first milliseconds escapes it. Accepted; documented in tools/README.md.
GOOS=windows go test -c ./terminal compiles the Windows tests but cannot run them here. windows_test.go is untagged on purpose so its ten tests run on Linux.
The sandbox truncated one Write to zero bytes (tools/shell_unix.go); it was rewritten from Python with a byte comparison. Check file sizes after any write here.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20


# Handoff — 2026-09-17 — terminal windows and tools on Windows

## State

Done, tested where it can be, gate green, documented in the library and in the six editors, **not committed**, on `main`. `git status` in turbo-core lists the touched files; nothing is staged. The Windows path is compiled and vetted, never run.

## Next steps

1. **Commit and tag turbo-core `v0.9.0`** — minor, because `tools.Shell` (a constant) became `tools.Shell()` and `tools.ShellArgs()`. `01-release.tag.sh` runs `make check`; in this sandbox `release_test.go` fails on the `cp`-NUL problem (documented in `summary.md`), not on the code — run it on the Mac.
2. **Re-pin the six editors** (`go get codeberg.org/turbo-editors/turbo-core@v0.9.0 && go mod tidy && GOWORK=off make check`), then tag each. Their documentation already describes Windows terminal windows; until the re-pin the binaries do not have them.
3. **Run the five-step protocol on a Windows machine**, the first anyone will have: (a) `F8` — a cmd.exe prompt appears and `dir` lists the directory; (b) drag the window's corner — the prompt repaints at the new width (`mode con` reports it); (c) `Alt-<toolchain>` ▸ a tool with `output = "terminal"` (Run) — the program runs and, when it ends, `Ctrl-W` closes the window (the console-closing goroutine gave the reader its EOF); (d) a long command in a terminal window — `Ctrl-C` interrupts it; (e) a tool with `output = "popup"` — output arrives, and `Escape` while it runs stops the whole tree (the job object). Report what the window showed, in SGR if possible.
4. If (c) hangs — the window never says the command finished — the suspect is the order in `windowsChild.Close` / `closeConsoleWhenTheShellExits`: conhost may need the *input* pipe closed before `ClosePseudoConsole` returns. If the first screen is garbage, the suspect is conhost's start-up sequences reaching an emulator that has only ever read Unix shells; capture the raw bytes (`TURBO_ACP_TRACE` does not cover terminals — add a temporary `io.TeeReader` in `View.read`).

## Watch out for

- **`unsafe.Pointer(handle)` is what `go vet` flags**, which is why `UpdateProcThreadAttribute` is called through a `LazyProc` with `uintptr` arguments. Do not "simplify" it back to x/sys's `Update`.
- **cmd.exe's command line is composed verbatim** (`"cmd.exe" /S /C "<command>"`), in two places that cannot share code without breaking the package layering: `terminal/windows.go` (`windowsCommandLine`) and `tools/shell.go` (`cmdExeCommandLine`). Change both or neither.
- **The job object is assigned after `Start`**; a child cmd.exe spawns in those first milliseconds escapes it. Accepted; documented in `tools/README.md`.
- **`GOOS=windows go test -c ./terminal` compiles the Windows tests but cannot run them here.** `windows_test.go` is untagged on purpose so its ten tests run on Linux.
- **The sandbox truncated one `Write` to zero bytes** (`tools/shell_unix.go`); it was rewritten from Python with a byte comparison. Check file sizes after any write here.