Make copy actually copy on Wayland (main-process clipboard + OSC 52)

[ymajoros]

Fixes #54.

Copy was a black hole on Linux/Wayland — paste fine, copy gone. Two reasons, both about the renderer's rocky relationship with the clipboard:

1. Ctrl+C went through the renderer. navigator.clipboard.writeText(...).catch(() => {}) rejects on Ozone/Wayland (Chromium wants window focus + a user gesture) and the empty .catch swallowed it whole. Textbook silent failure.

2. OSC 52 wasn't handled at all. That's the sequence Claude Code uses to copy. xterm doesn't register a handler for it on its own, so those copies went nowhere.

What this does

Adds a small clipboard-write-text IPC that calls the main-process clipboard.writeText. The main process has none of the renderer's focus/gesture hangups, so it just writes on Wayland. Then:

• Ctrl+C (with a selection) uses it instead of navigator.clipboard
• registered an OSC 52 handler that base64-decodes the payload and routes it through the same IPC (and ignores the ? read-back query, which we don't answer)

Paste was never broken, so it's untouched.

The diff (3 files, ~30 lines)

• main.js — import clipboard, add the IPC handler
• preload.js — expose writeClipboard
• public/terminal-manager.js — use it for Ctrl+C, add the OSC 52 handler

terminal.parser.registerOscHandler is stable xterm API and allowProposedApi: true is already set, so nothing new is required capability-wise.

Testing

Ubuntu / GNOME 49 / Wayland, native Ozone.

• Before: Claude Code copy, Ctrl+C, and right-click Copy all dead, no error.
• After: Claude Code's copy and Ctrl+C-with-selection both land in the system clipboard and paste cleanly into other apps.

I only have the Linux box, so I'd love a second pair of eyes confirming macOS/Windows still behave — though main-process clipboard is the boring-and-reliable option on every platform, so I'd be surprised if it regressed anything.

[ymajoros]

Confirmed working. Built this branch and ran it on Ubuntu / GNOME Shell 49 / native Wayland (Ozone), Electron 41.0.3 — the same box where stock 0.0.30 copies nothing.

Both paths now land in the system clipboard and paste cleanly into other apps:

• OSC 52 (Claude Code's copy) — tested with printf '\e]52;c;%s\a' "$(printf hi | base64)"
• Ctrl+C with a selection

Paste was unaffected throughout. No regressions spotted in normal terminal use.

[abasiri]

Reviewed this and I'm happy with the approach — pushed two small follow-ups to the branch (6368f2e) rather than sending you back and forth. Also merged current main in, since the branch was 7 behind and #75 has since added the module.exports block this now extends.

1. writeClipboard rejection. It's an ipcRenderer.invoke, so it returns a promise, and the navigator.clipboard.writeText(...) call it replaced ended in .catch(() => {}). Without that an IPC failure becomes an unhandled rejection. Restored.

2. Pinned the read-back refusal. This was the part I wanted to protect. if (!b64 || b64 === '?') return true; is doing real security work — answering a read-back query would write the user's clipboard back into the terminal, letting anything running in the session exfiltrate whatever they last copied. But it reads like an unimplemented case, and it's one well-meaning commit away from being "finished". I extracted the parse into decodeOsc52Payload() so the refusal is documented where it lives, and added test/clipboard-osc52.test.js covering write / query / empty / separator-less / multi-byte / malformed-base64, with the query test stating the reason.

Otherwise a pure refactor — I checked the extracted helper returns identical (handled, written) results to your inline logic across all twelve of those payload shapes before pushing. 12/12 tests green.

On the wider question of whether OSC 52 writes should sit behind a setting: I considered it and decided no. The capability is the feature here — it's how Claude Code copies — so a setting defaulting off just reintroduces the bug, and defaulting on is a knob nobody turns. The exfiltration direction is already refused, xterm.js bounds the payload at 10MB, and these terminals run agents with full shell access, so gating clipboard writes while handing out Bash wouldn't be buying anything real.

Thanks for the fix — the Wayland diagnosis and the main-process routing were both right. Merging once you've had a chance to look over my two commits.