Review — end-to-end ✅ (clean approve, regression test pushed)

What this ships

Fixes a regression introduced by #1182's empty-session guard: when a user has the workspace panel open and reloads the page after clicking New Conversation (without sending a message), the panel gets stuck closed and the toggle button is disabled. Recovery requires clicking a past conversation.

Four targeted changes in static/boot.js:

1. syncWorkspacePanelState() (boot.js:62-75) — when S.session is null, only force-close if mode is 'preview' (which legitimately requires a session). For 'browse' mode, call syncWorkspacePanelUI() instead so the panel renders its no-workspace placeholder rather than vanishing.

2. Ephemeral-session boot path (boot.js:908-912) — read panelPref from localStorage and set _workspacePanelMode='browse' BEFORE the sync call, matching what the existing session-restore path at boot.js:921-925 already does.

3. No-saved-session boot path (boot.js:933-937) — same pref-restore.

4. canBrowse and openWorkspacePanel() — extended to include S._profileDefaultWorkspace so the toggle stays enabled and clickable when a profile workspace is configured.

Traced against upstream hermes-agent

Pure WebUI frontend behaviour. Zero agent coupling.

End-to-end trace

syncWorkspacePanelState() semantics

Pre-fix: if(!S.session){ _setWorkspacePanelMode('closed'); return; } — unconditional close when no session.

Post-fix:

if(!S.session){
  if(_workspacePanelMode==='preview') _setWorkspacePanelMode('closed');
  else syncWorkspacePanelUI();
  return;
}

The mode-state machine has three values: 'closed', 'browse', 'preview'. With no session:

• 'preview': needs file content from a session → close ✅
• 'browse': render panel chrome (file tree, no-workspace placeholder, profile-default tree if configured) → keep open ✅
• 'closed': render closed UI → no-op via syncWorkspacePanelUI() ✅

localStorage pref-key parity

The new ephemeral and no-session paths read both keys with ||:

localStorage.getItem('hermes-webui-workspace-panel-pref')==='open'
  || localStorage.getItem('hermes-webui-workspace-panel')==='open';

This matches the existing session-restore path at boot.js:921-922. The two keys are documented in panels.js:2408-2417:

• hermes-webui-workspace-panel-pref — user-preference checkbox in settings (priority key)
• hermes-webui-workspace-panel — runtime state from toolbar toggle

Pref takes priority so closing via toolbar X doesn't suppress the "keep open" setting.

Toggle-stays-enabled extension

canBrowse at boot.js:109:

const canBrowse=!!S.session||_hasWorkspacePreviewVisible()||!!(S._profileDefaultWorkspace);

openWorkspacePanel() early-return at boot.js:77:

if(mode==='browse'&&!S.session&&!_hasWorkspacePreviewVisible()&&!S._profileDefaultWorkspace)return;

Both gates now include S._profileDefaultWorkspace, set during the settings fetch at boot.js:822 which runs well before the IIFE.

Regression-cause confirmation

Bug timing: introduced by #1182's empty-session guard, exposed when a user had the panel open before reloading. The session-restore path at boot.js:921-925 already restored panelPref correctly, but #1182 added two new syncWorkspacePanelState() call sites (the ephemeral guard and the no-session path) that didn't.

Edge-case trace

What I pushed — 2fbbb8d

Six static-analysis tests in tests/test_workspace_panel_persists_on_empty_boot.py locking the four invariants:

1. syncWorkspacePanelState checks _workspacePanelMode==='preview' before force-closing
2. The non-preview branch calls syncWorkspacePanelUI() rather than force-closing
3. Both new boot paths read 'hermes-webui-workspace-panel-pref' and set _workspacePanelMode='browse' before sync
4. canBrowse and openWorkspacePanel early-return both include S._profileDefaultWorkspace

These would have caught the original regression — they're exactly the invariants #1182 inadvertently broke.

CI re-ran green on 2fbbb8d: 3.11 ✅, 3.12 ✅, 3.13 ✅.

Tests

• PR's targeted test update (test_boot_sets_profile_default_workspace_in_settings_block): correctly tightened from "find first occurrence" to "find specific assignment expression" — the new _profileDefaultWorkspace references in canBrowse/openWorkspacePanel were pushing the first-find away from the settings block.
• My 6 new tests: pass.
• Local full suite: 2643 passed, 47 skipped, 1 PR-unrelated pre-existing failure (test_sprint3.py::test_workspace_add_rejects_system_paths — the macOS-only quirk that PR #1186 fixes; this PR is on top of master before #1186 lands).
• CI on PR: ✅ test (3.11), ✅ test (3.12), ✅ test (3.13).

Other audit — confirmed correct

• JS syntax: node --check passes on boot.js.
• No agent-side coupling: pure UI state machine.
• No race against settings fetch: S._profileDefaultWorkspace is set at line 822 in the settings fetch IIFE, well before the boot path runs around line 880-940.
• loadSession doesn't mutate _workspacePanelMode: confirmed via grep — so the pre-sync mode set in the ephemeral path survives loadSession's call.

Minor observations (non-blocking)

• The panelPref read pattern is now duplicated three times (existing session-restore path + two new paths). A small DRY helper (_readPanelPref()) would tighten this. Cosmetic — easy refactor in a follow-up.
• The two pref-key fallback (-pref AND non-pref) is documented in panels.js but the rationale ("pref priority over runtime state") is implicit in the || order. A one-line comment near each read site noting the priority semantics would help future readers. The existing session-restore path has this comment; the two new sites could pick it up too.

Recommendation

Approved. Tight, well-scoped fix for a real regression. The PR description's scenario table accurately characterises the four affected paths. State-machine reasoning ('preview' close vs 'browse' keep) is correct. The _profileDefaultWorkspace extension in canBrowse/openWorkspacePanel keeps the UI consistent (toggle enabled when there's something to show). Pushed 6 regression tests locking the new invariants. CI green; no agent coupling. Parked at approval — ready for the release agent's merge/tag pipeline.