fix(tmux): submit quick action prompts via Enter keycode after paste

[dhilgaertner]

Summary

Quick action prompts (Merge PR, Address Changes, Fix Checks, Fix Conflicts) end with a trailing `\n` so they auto-submit in Claude Code's input. The Ghostty backend honored that contract by splitting on `\n` and emitting a synthetic Return keycode (36) — those events fire outside any bracketed-paste bracket, so the TUI saw them as real Enter keys.

The tmux backend shipped the entire string through `tmux load-buffer` + `tmux paste-buffer`. Claude Code's TUI enables bracketed-paste mode, so tmux wrapped the buffer in `\e[200~ … \e[201~` and the trailing `\n` was delivered as literal text — typed as a newline character in the input field, never as Enter. Prompt visible, never submitted.

Changes

• `TmuxController.sendKeys(target:keys:)` — thin wrapper over `tmux send-keys -t <keys...>` reusing the existing `run` watchdog/timeout path.
• `TmuxBackend.sendText` — strip the single trailing `\n` from the paste payload, then send a separate `Enter` via `sendKeys` after `paste-buffer` completes (outside the bracket). Internal newlines are preserved as-is; an empty-after-strip payload skips the paste entirely (some tmux builds reject empty `load-buffer` input).

Ghostty path is untouched.

Test plan

• `make build` — green
• `swift test` for the CrowTerminal package — all 24 tests pass, including `TmuxBackendTests.sendTextRoundTripsThroughBuffer`
• Manual: with `experimentalTmuxBackend: true`, click Fix Conflicts / Merge PR on a PR session and confirm the prompt is submitted automatically
• Manual regression: same flow with `experimentalTmuxBackend: false` (Ghostty) — behavior unchanged

Closes #264

[dgershman]

Code & Security Review

Critical Issues

None.

Security Review

Strengths:

• No user-controlled data reaches shell execution — sendKeys passes ["Enter"] as a literal tmux key name, not raw shell input.
• Buffer names are derived from UUID, no injection vector via tmux load-buffer / paste-buffer.
• The existing run() watchdog/timeout path is correctly reused for the new sendKeys method, so a hung tmux server can't block the main actor.

Concerns:

• None identified. The change stays within the existing trust boundary (Crow → tmux IPC over a private socket).

Code Quality

Well done:

• The doc comment on sendText (TmuxBackend.swift:244-251) clearly explains the bracketed-paste root cause and references the Ghostty parity (keycode 36). Excellent for future maintainers.
• Empty-payload guard (if !payload.isEmpty) correctly avoids sending an empty load-buffer, which some tmux builds reject — good defensive coding.
• sendKeys in TmuxController is a thin, general-purpose wrapper (send-keys -t <target> <keys...>) that will be reusable for future key-event needs (e.g., C-c, Escape).
• The defer { ctrl.deleteBuffer(...) } stays inside the if !payload.isEmpty block, so the buffer is only cleaned up when it was actually created.

Observations (non-blocking):

• sendText only strips a single trailing \n (text.dropLast()). If a caller sends "hello\n\n", only one \n is stripped and the other is pasted as literal text inside the bracket. This matches the Ghostty backend's behavior (which splits on every \n and sends a keycode-36 for each), so multi-newline strings will behave differently between backends. Not a concern for the quick-action use case (always exactly one trailing \n), but worth noting if sendText is later used for multi-line submission. Current callers at registerTerminal (line 183: command + "\n") and crow send all append a single \n, so this is fine today.
• The existing test sendTextRoundTripsThroughBuffer (TmuxBackendTests.swift:103) sends a payload without a trailing newline, so it exercises the paste-only path but not the new Enter-after-paste path. A test case like sendText(id: id, text: payload + "\n") would cover the new branch. Low risk given the simplicity, but would strengthen regression coverage.

Summary Table

Priority	Issue
Green	Consider adding a test case for the trailing-newline + Enter path
Green	Document multi-newline behavior difference vs. Ghostty if sendText scope widens

Recommendation: Approve — clean, focused fix that correctly addresses the bracketed-paste issue (#264). The approach mirrors what the Ghostty backend already does, the code is well-commented, and there are no security concerns.

🤖 Reviewed by Crow via Claude Code