Inject OSC 633 shell integration for zsh and bash

[nedtwigg]

What

Makes shells emit real OSC 633 shell-integration sequences instead of relying on the keystroke heuristic. The terminal parser already consumed the full OSC 633 family (prompt/command boundaries A/B/C/D, command line E, cwd P) — this wires up the emit side per shell, giving authoritative command boundaries, real exit codes, and reliable cwd. The keystroke heuristic stays as the automatic fallback when injection can't apply.

This lands zsh and bash from the 5-shell plan; fish and PowerShell are follow-ups, and cmd.exe stays on keystrokes (no per-command hook exists).

Shell	Mechanism	Channel	Status
zsh	ZDOTDIR → our .zshrc chains to the user's	env	✅ this PR
bash	--init-file; script replicates login profile	shellArgs	✅ this PR
fish	XDG_DATA_DIRS vendor conf.d	env	follow-up
PowerShell	-NoExit -Command dot-source	shellArgs	follow-up
cmd.exe	none	—	keystroke fallback

How

• zsh (shell-integration/zsh/): ZDOTDIR points at our dir; .zshrc hands ZDOTDIR back to the user before sourcing their rc (so .zcompdump/.zsh_history land in their dir, not our read-only shipped one), then installs precmd/preexec hooks. Our precmd runs first so $? is the command's real exit code before user hooks (e.g. oh-my-zsh) clobber it.
• bash (shell-integration/bash/shellIntegration.bash): injected via --init-file. Since that conflicts with -l, the script replicates login-profile startup (/etc/profile + first of .bash_profile/.bash_login/.profile) then installs OSC 633 hooks. Written for bash 3.2 (macOS system bash): DEBUG trap for preexec, string PROMPT_COMMAND for precmd. Disarms across the prompt hook so the trap doesn't trip on itself or the user's PROMPT_COMMAND.
• Injection point (pty-core.js): applyShellIntegration sets the env var (zsh) or shellArgs (bash); fail-safe to today's behavior if the scripts are missing, and skipped when explicit args are present. detectUnixShells surfaces $SHELL plus common shells on disk so the picker can offer bash even when the login shell is zsh.
• Keystroke gating (terminal-state-store.ts): once a pane sees a real OSC 633 boundary it's promoted to OSC-driven and the keystroke-synthesized commandStart path is suppressed, so we never double-count.
• Exit-status UI: a finished non-zero command shows a red ✗ after the idle title. The fail state flows as a structured DerivedHeader.lastCommandFailed flag (not re-parsed off the title string).
• VS Code packaging: scripts copied into dist/shell-integration; the extension passes DORMOUSE_SHELL_INTEGRATION_DIR. Verified the .vsix ships the scripts at the flat path.

Testing

• pnpm test — 545 lib tests + 68 sidecar tests green; lib typecheck clean.
• Real-shell e2e for both zsh and bash: confirmed genuine OSC 633;C/;D;<exit> (not source:'user_input'), correct exit codes, user profile/rc loads intact (PATH/Homebrew/asdf).
• .vsix packaged and inspected: dist/shell-integration/{bash/shellIntegration.bash,zsh/.zshrc,…} present at the flat path.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	1b6350f
Status:	✅  Deploy successful!
Preview URL:	https://e52efb7a.mouseterm.pages.dev
Branch Preview URL:	https://osc-633.mouseterm.pages.dev

View logs

[dormouse-bot]

The injection logic is carefully built — the per-pane OSC-driven promotion gating (and the keystrokeHeuristic flag that keeps the fallback's own synthesized prompt markers from self-retiring the path that emits them) is exactly right, and the zsh ZDOTDIR hand-back / bash login-profile replication both look correct. One spec-sync gap:

docs/specs/terminal-state.md is now out of date with the header changes. This PR adds the ✗ glyph and a lastCommandFailed field, but the Header Derivation section still documents the old shape:

• The DerivedHeader type block declares only { primary; secondary? } — no lastCommandFailed.
• The prose says exit codes are surfaced through the alert machinery and "the header itself stays peaceful", and lists "exit-code badges" among consumers that "read it from [pane.activity]". The new trailing ✗ on primary for a non-zero lastCommand is the header now surfacing exit status directly, which contradicts that text.

terminal-state.md is the canonical spec for header derivation (it's the one AGENTS.md points to for terminal-state.ts / terminal-state-store.ts), and AGENTS.md asks that the spec be updated in the same change. The terminal-escapes.md update covers the OSC-633 emit side and the keystroke fallback well; the header/DerivedHeader change just needs the matching edit here. Worth also noting the per-pane keystroke-retirement invariant in this spec's Command Input Fallback section, since that's the store behavior this spec owns.

Not a correctness issue — the code and tests look sound (sidecar suite green locally; the store gating matches the new tests). Flagging the doc drift rather than blocking.