OSC 633 shell integration for PowerShell

[nedtwigg]

What

Extends OSC 633 shell integration to PowerShell (pwsh) and Windows PowerShell (powershell.exe) — bash and zsh already had it. Dormouse now injects prompt/command boundaries so PowerShell sessions get real command history, exit codes, and cwd tracking instead of falling back to the keystroke heuristic.

How

• New emitter standalone/sidecar/shell-integration/pwsh/shellIntegration.ps1 wraps the user's prompt function to emit the A/B/C/D/E/P;Cwd sequences. PowerShell has no preexec hook, so — like VS Code's PowerShell integration — a command's E/C/D are reported from the next prompt render (command line from Get-History, exit code from $?/$LASTEXITCODE). Boundaries and exit codes are exact; the C/D markers land at the following prompt rather than bracketing output in real time. Escaping matches the bash/zsh emitters (\→\, ;→\x3b, etc.); Cwd is sent raw because the parser reads it verbatim.
• standalone/sidecar/pty-core.js: inject via -NoExit -Command ". '<script>'" for both PowerShell flavors (profile still loads, so the user's prompt is defined before we wrap it). Skipped when the caller passed explicit args, mirroring bash. Added a shellStem() helper that handles \ separators and strips .exe so Windows shell paths match; resolveLoginArg reuses it.
• Spec: docs/specs/terminal-escapes.md marks PowerShell implemented and documents the mechanism + the next-prompt limitation.

Testing

• New unit tests: pwsh injection, Windows PowerShell injection, explicit-args passthrough, missing-script fallback. All pass.
• Verified the .ps1 in a real PowerShell session: first prompt emits A/P;Cwd/B; a full command cycle emits E (escaped), C, D;<exit>; failing native command reports the true exit code (D;3); install guard prevents re-wrapping.

Note: the 3 pre-existing pty-core.test.js failures seen locally are Windows-path artifacts (those zsh/bash tests hardcode POSIX /); they pass on Linux CI and are unrelated to this change.

Packaging

No build change needed — the VS Code build already copies shell-integration recursively and standalone ships it via the tauri resources glob, so pwsh/ is bundled automatically. .ps1 isn't a dotfile, so the dotfile-stripping caveat doesn't apply.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	3ebbf5e
Status:	✅  Deploy successful!
Preview URL:	https://16f2df33.mouseterm.pages.dev
Branch Preview URL:	https://powershell-osc633.mouseterm.pages.dev

View logs

[dormouse-bot]

One correctness issue, verified by simulating the history-id state machine in pwsh: the first interactive command of every session never gets its E/C/D reported.

The -1 sentinel is meant to suppress only the genuine first prompt, but it lingers through that first render because $Global:__dormouse_633_last_history_id is assigned only when history is non-empty. At the very first prompt Get-History is empty, so the sentinel stays -1; when the next prompt renders after the user's first command, the guard ... -ne -1 is still false and the command's line/output-start/exit-code are dropped. From the second command onward it works. (This is easy to miss in manual testing — by the time you're inspecting output you've already run several commands.) VS Code's PowerShell integration avoids this by assigning the id unconditionally each render. Inline suggestion below; I confirmed it restores reporting for the first command while still suppressing empty-line / Ctrl+C renders.

Secondary, non-blocking: by the time the user's original prompt is invoked ($Global:__dormouse_633_original_prompt.Invoke()), $? has been clobbered by the intervening Get-History and assignments, so it is always $True there. $LASTEXITCODE survives (nothing reassigns it), so starship-style prompts are fine, but a prompt that colors itself off $? will always see success. Restoring $? before the invoke is awkward in PowerShell — worth a note rather than a required change.

[nedtwigg]

@dormouse-bot fix and commit

[dormouse-bot]

Done — fixed in ed788e3. The first interactive command's E/C/D are now reported (the -1 sentinel was lingering past the genuine first prompt because the history id was only assigned when Get-History was non-empty). Details in the inline thread. Verified in pwsh, all 53 sidecar unit tests still pass. Watching CI now.