Unify terminal context menus and add helper terminals

[nedtwigg]

Right-clicking a terminal's header, alert indicator, or uncaptured terminal body now opens one terminal context panel. It shows the surface identifier, title provenance, directory actions, port actions, and source alert controls, with a real helper terminal filling the remaining space.

The helper starts lazily in the source directory and runs the global default command (git status initially). Untouched idle helpers refresh on reopening; any user input preserves the session until an explicit Reset. Promote moves the same PTY and xterm into a regular split, preserving scrollback and running editors. Running helper work blocks closing its source and is revealed for the user to stop. Helpers stay outside alerts, dor, and Pocket until promotion.

Port actions open the system browser, an iframe, or an embedded/popped-out agent browser. Browser targets are reused per source and port. Shared compact switches and subtle labeled actions match the approved Storybook baseline.

Includes the native host protocol, recovery and process inspection, lifecycle regression tests, and updated specs. The presentation is shared by the live menu and Storybook state gallery.

Validation:

• Full pnpm test and pnpm build.
• Standalone frontend build and Rust check.
• Real PTY checks through dor ab: autorun, preserved input, background-job close protection, idle parent/helper closure, and promotion with unsaved editor contents intact.

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	ca86651
Status:	✅  Deploy successful!
Preview URL:	https://2b508937.mouseterm.pages.dev
Branch Preview URL:	https://new-right-click-a.mouseterm.pages.dev

View logs

[dormouse-bot]

Read the whole diff plus the merged-tree versions of helper-terminal.ts, TerminalContext*.tsx, Wall.tsx, pty-core.js, pty-manager.ts and message-router.ts. The helper lifecycle itself reads well — the conservative-inspection rule is applied consistently (busy: null → work, missing PTY in the process table → work, promoting → work), the promotion/rollback ownership handoff is genuinely covered by helper-terminal.test.ts, and the sidecar's validHelperOwner and openDirectory validation are tight. Four things below.

The one I'd most want addressed is the helper status poll: it never stops when the context closes, so each retained helper keeps issuing a host process-table inspection every 2s for the rest of the session. On Windows that is a synchronous powershell.exe spawn inside the sidecar process, which is the same process pumping every PTY's output.

Two smaller notes that don't fit on a diff line:

• onKill in Wall.tsx no longer calls exitTerminalMode() synchronously — it moved inside confirmSource(), which for a helper-owning source runs only after helperBlocksClose resolves. On the standalone host that inspection can take up to the 10s pty_context timeout, so pressing x in passthrough now leaves the wall in passthrough for that whole window before anything visible happens.
• docs/specs/terminal-context.md → Global autorun setting says the command is "at most 4096 characters, excluding CR, LF, and NUL". The hosts enforce all three (validCommand), but the Modify input only sets maxLength={4096} — a pasted multi-line value reaches the host and comes back as a thrown error rather than being rejected at the field. Fine if that's the intent; worth a line in the spec if so, since the spec reads as one rule over both.

[nedtwigg]

Addressed the review in cd4b7e3 and replied on each thread. The two general notes are covered too: onKill now exits passthrough synchronously before inspection, and the spec explicitly assigns command validation to the host and requires rejected edits to retain the prior preference and show the error in Modify.

Also corrected chunk-boundary trimming in Standalone's replay buffer, with a regression test preserving the complete 200,000-character tail. The nine stale Storybook interactions were fixed in ae83694; its automated visual regression run passed.

Full pnpm test and pnpm build pass on cd4b7e3. The new CI run is in progress.

[dormouse-bot]

Re-reviewed the increment. All five earlier threads are genuinely addressed: setHelperVisible gates both the 100ms tick and the process inspections (with the hidden-invalidation and pending-creation regressions to pin it), the focus-trap hook and pointInConvexPolygon are gone with no dangling references, the VS Code router now posts pty:exit for a rejected helper, and the Explorer path acknowledges launch rather than exit. The replay-trim correction is right too — slice(session.chars - REPLAY_CHARS) can never over-slice, because the loop stops precisely when chars - REPLAY_CHARS < chunks[0].length.

One thing the pause introduced, inline below: the status recomputed on reopen is never published, so TerminalContext can render the status the helper had when it was hidden.

standalone/sidecar/helper-terminal.test.js passes locally (8/8); the vitest suites need an install this checkout doesn't have, so I'm relying on CI for those.

[dormouse-bot]

The two general notes from my review check out on cd4b7e3. onKill calls exitTerminalMode() at Wall.tsx:1294, ahead of the helper lookup and the helperBlocksClose branch, so passthrough exits on the keypress instead of after the inspection round-trip. And docs/specs/terminal-context.md now carries Must enforce command validity at the host with the rejected-edit behavior, which stops the 4096/CR/LF/NUL rule above it from reading as a field-level constraint.

The five diff-line threads and the increment are answered in the review above.

[dormouse-bot]

Reviewed the five commits since 445fb1c — the move into the Lath leaf overlay, the clip-path reveal/dismiss, the copy and "Opening…" feedback, and the unfocused-window highlight suppression. Moving the context out of TerminalPanel's clipping box is a clean win: it drops the onMouseDown stopPropagation the old placement needed, and the mounted / id split (identity for the leaf, input-owner for everyone else) keeps the exit from leaking into the wall's keyboard state. Re-read helper-terminal.ts too — splitting setHelperVisible into its own [id, closing] effect still arms watchHelper before openHelper's first await resolves, so the retained-helper path is unaffected.

Two things below.

The one I'd most want addressed is the reveal animation's Chromatic exposure. Every other keyframe animation in the wall is either triggered by a play function or has a cfg static substitute; this is the first one that runs on mount over Chromatic-snapshotted content, and it is gated only by prefers-reduced-motion — motionIsInstant() already covers the dismissal but not the entrance.

One smaller note that doesn't fit on a diff line: the detail dialog renders outside .terminal-context-content, so during the first 160ms of a reveal it paints at full opacity over content still at opacity 0. Only reachable through initialDetail, which is exactly what the TitleSources / ModifyCommand / ResetHelper stories set — so it compounds the first finding rather than standing on its own.

Happy to push either fix if you'd rather not hand-apply them.

[dormouse-bot]

ca86651 closes both threads and the general note. Verified the Chromatic path end to end (isChromatic() → cfg.layout.animate = false at module scope → motionIsInstant() → no terminal-context-enter, so neither the clip-path reveal nor the now-scoped content fade can match), and the focus fix keeps the launch button in the tab order while suppressing the repeat click. Nothing new in the increment.

Automated checks are green on this head: Build & Test, Visual Regression Tests, Standalone Smoketest, both Standalone Platform Checks, Webview Boot Smoketest, relay, Cloudflare Pages, Storybook Publish. UI Tests is the Chromatic baseline gate — 16 changes awaiting your accept, which only you can clear.