Show the terminal's live working directory in the pane header

[Zeus-Deus]

Problem

The terminal pane header rendered a bare Terminal — the same word the tab directly above it already shows. Two labels stacked 28px apart, zero information between them.

What this does

Appends the session's live working directory to the header, only when it differs from the workspace root.

Shell location	Header
Workspace root (default)	Terminal — unchanged
cd src-tauri	Terminal src-tauri
Deep	Terminal …/src/pty_daemon
Outside the workspace	Terminal ~/dotfiles
Outside $HOME	Terminal …/docker/volumes

Design notes

Signal only when it's surprising. ~80% of sessions sit at the workspace root, and the workspace identity is already on screen twice (sidebar + context bar). Staying quiet at the root means the header doesn't churn and doesn't add a third copy of the same string to every pane in a split.

Full absolute paths are deliberately not rendered. The header is a 28px strip that CSS-truncates on the right, so /home/user/.codemux/worktrees/codemux/some-branch would degrade to /home/user/.codemux/worktr… — dropping the only segment that mattered. Eliding the head keeps the meaningful tail and fits a narrow split; the untrimmed path is on the hover tooltip.

Foreground command was considered and rejected — showing the running process would flash the header on every ls.

Known tradeoff: absence is ambiguous — an empty hint means "at the root" or "cwd not known yet." This only bites on remote panes whose shell doesn't emit OSC 7. Judged the better trade against always-on noise; it's a one-line flip in formatCwdHint if we want to revisit.

Two sources, osc7 beats proc

	OSC 7	/proc/<pid>/cwd
Mechanism	Shell pushes on every prompt	Poll + readlink
Needs shell integration	Yes (fish, vte-bash ✅; bare zsh ❌)	No
Remote/SSH panes	Works	Can't (pid is on another host)
Latency	Instant	≤1 interval

Once a session reports via OSC 7 the poller drops it from its request set, so a fully shell-integrated setup costs zero IPC. A single app-level 2s interval batches one call for the active workspace's remaining sessions and pauses while the window is hidden.

The /proc read targets the session's shell pid, so cd moves the header but a long-running child (a build, an agent) does not. Sessions with no local pid are omitted rather than answered from an unrelated local /proc entry.

Constraints the codebase imposed

• State is frontend-only. TerminalSessionSnapshot.cwd stays the spawn directory — making it live would fan out a full state rebuild plus a persistence write on every cd for a cosmetic label.
• The cwd is appended, never substituted. PRESET_TITLE_TO_ICON keys the preset icon off the exact title string, so overwriting node.title with a path would silently kill the Claude/Codex icons. A regression test pins this.

Testing

• 42 new frontend tests — formatter rules (incl. Windows separators and shared-prefix siblings), the osc7-wins precedence, the OSC 7 payload parser, and the header's render states.
• 2 new Rust tests over the /proc reader, covering live pid, dead pid, and no-pid (remote) cases.
• Full suite: 3093 frontend tests pass, including pane-tree-rerender.test.tsx — the hoisted hooks resolve to stable primitives for non-terminal nodes, so there's no re-render fan-out.
• Verified end-to-end in the running UI by pushing real OSC 7 escape sequences through the pty stream (new __codemuxTerminalMock.emitOsc7 dev helper, which exercises the production handler rather than a mock shortcut). All five label cases confirmed, plus correct truncation at a 520px viewport, and the hint survived many poll ticks without being clobbered — confirming the precedence rule.

⚠️ The /proc fallback is unit-tested against real /proc but was not exercised end-to-end in the desktop app (tauri:dev opens a native window not visible to browser automation). The OSC 7 path is verified end-to-end.

[Zeus-Deus]

Review — Opus 5 (high effort)

The architecture is right: OSC 7 push beating a poll, the poller dropping OSC-7 sessions from its request set so an integrated shell costs zero IPC, setPolledCwds returning the identical state object when nothing changed, omitting no-pid sessions rather than answering them from an unrelated local /proc entry, the inFlight guard, and the document.hidden pause. parseOsc7 correctly ignores the authority and survives malformed percent-encoding. All good.

One real bug.

formatCwdHint drops the leading / for two-segment absolute paths

MAX_TAIL_SEGMENTS = 2, so trimTail only prepends …/ when parts.length > 2. At exactly two segments it emits the bare tail with no marker of any kind:

"/var/log"                -> "var/log"
"/opt/tools"              -> "opt/tools"
"/etc/nginx"              -> "etc/nginx"
"/var/lib/docker/volumes" -> "…/docker/volumes"   ✓
"/etc"                    -> "/etc"               ✓

workspace-relative /home/zeus/projects/codemux/var/log -> "var/log"

(Ran against formatCwdHint from this branch with ROOT = /home/zeus/projects/codemux, HOME = /home/zeus.)

The last two lines are the problem: cd /var/log and cd var/log render identically. The one-segment case explicitly keeps its slash (/etc) and the three-plus case is marked by …/, so the two-segment arity is the only gap — and it's the arity the tests skip. They cover /etc, /, and /var/lib/docker/volumes, jumping straight over the case that breaks.

Fix is a line in the final branch — parts.length <= MAX_TAIL_SEGMENTS should keep the absolute form, e.g. return { label: \/${parts.join("/")}`, full }— or havetrimTailtake a "this is absolute" flag. Either way, add/var/log` to the test table.

Two smaller notes

terminal_session_cwds is a sync command. It runs on the GTK main thread and does one read_link per requested session, every 2s. Each call is trivially cheap, so this is not a real stall — but docs/core/STATUS.md records the deliberate migration of blocking workspace.rs/files.rs commands onto spawn_blocking "so directory listing / ignore checks / workspace ops can't stall IPC", and this is a new blocking command on the polled path. Making it async fn costs nothing and keeps the invariant intact.

The GC loop writes once per dead session.

for (const id of Object.keys(cwdStore.cwds)) {
  if (!live.has(id)) cwdStore.clearCwd(id);
}

clearCwd is a set() per call, so closing a workspace with six terminals fires six store updates and six render passes on the next tick. setPolledCwds already demonstrates the bulk-write pattern; a pruneCwds(liveIds) would match it.

Also worth noting: this branch and #210 both touch src-tauri/src/lib.rs, src-tauri/src/terminal/mod.rs, and src/App.tsx.

[Zeus-Deus]

Review by GPT 5.6 Sol xhigh found a Windows OSC 7 parsing bug.

parseOsc7("file:///C:/Users/z/proj") currently returns /C:/Users/z/proj. The workspace root is represented as C:\Users\z\proj; after normalization the leading slash remains, so formatCwdHint treats the terminal as outside its workspace and renders a misleading hint even while it is at the root. I reproduced the current code path directly:

{"parsed":"/C:/Users/z/proj","hint":{"label":"…/z/proj","full":"/C:/Users/z/proj"}}

Windows has no /proc/<pid>/cwd fallback, so OSC 7 is the important source there. Please normalize file-URI drive paths (/C:/... → C:/...) and cover UNC authorities as well. The existing 42 focused tests all pass because the formatter has Windows cases but the parser only has Unix file-URI cases.