Pr/workspace bugs

[Midway65]

fix(workspace): workspace selection and restore reliability

Three targeted bug fixes in the workspace/chat-settings pipeline. No architectural changes — each fix is a correctness correction to existing behaviour.

Depends on: pr2-workspace-two-tier-prompt must be merged first (Bug 3 references pendingFullWorkspaceLoad introduced there).

Commits: 996c578f, 25ec9a62, 699bd008, 986eb384
Files changed: 2

• src/ui/chat/components/ChatSettingsModal.ts
• src/ui/chat/ChatView.ts

---

Bug 1 — Workspace selection silently skipped for 14 of 21 workspaces

ChatSettingsModal.handleSave() had a guard that called setWorkspaceContext() only when workspace.context was non-null. Workspaces without a contextJson column value — all user-created workspaces — hit this guard and were silently dropped. The workspace ID was never bound, the system prompt header was never populated, the session was never set. The UI appeared to select the workspace but state was not applied.

Fix: Remove the guard. Remove the redundant caller-side getWorkspace() fetch. setWorkspaceContext() now takes only workspaceId and derives context internally.

// Before
const workspace = await this.workspaceService.getWorkspace(settings.workspaceId);
if (workspace?.context) {
    await this.modelAgentManager.setWorkspaceContext(settings.workspaceId, workspace.context);
}

// After
await this.modelAgentManager.setWorkspaceContext(settings.workspaceId);

setWorkspaceContext signature changed from (workspaceId: string, context: WorkspaceContext) to (workspaceId: string). The caller no longer needs to fetch the workspace first; context is derived from getWorkspaceBasic() inside the method.

---

Bug 2 — Workspace restore wiped selectedWorkspaceId at startup

restoreWorkspace() called loadWorkspace() — a full MCP agent execution — to reconstruct workspace state on startup and on every conversation switch. If the agent was not ready (common at startup), it returned null and the catch/null branch cleared selectedWorkspaceId. The workspace ID was saved in conversation metadata but lost from in-memory state. On modal reopen the dropdown showed empty.

Fix: restoreWorkspace() now uses getWorkspaceBasic() — a direct SQLite query with no agent dependency, added to WorkspaceIntegrationService. selectedWorkspaceId is set tentatively before the await and cleared only if the workspace is genuinely absent from the DB, not on transient errors.

// New method in WorkspaceIntegrationService — cheap DB lookup, safe at any startup phase
async getWorkspaceBasic(workspaceId: string): Promise<{
  id: string; name: string; description?: string;
  rootFolder?: string; context?: Record<string, unknown>;
} | null>

---

Bug 3 — Full-load flag consumed during initialization, not on first user message

performFullInitialization() in ChatView.ts called getMessageOptions() at line 298 to read the current provider for a WebLLM pre-load hint. This fired the G-W3 block as a side effect: it loaded full workspace data, set pendingFullWorkspaceLoad = false, and cleared loadedWorkspaceData — all before the user sent any message. The user's first actual message then found the flag already consumed and received a slim header only. The full-load-on-first-message feature worked correctly for workspace selections (where setWorkspaceContext() fires after initialization) but was silently broken for workspace restores.

Fix: Replace the getMessageOptions() call with getSelectedModel()?.providerId — sync, no side effects, same data, already exists on ModelAgentManager.

// Before — triggered G-W3 block as a side effect
const currentProvider = (await this.modelAgentManager.getMessageOptions()).provider;

// After — reads provider directly, no side effects
const currentProvider = this.modelAgentManager.getSelectedModel()?.providerId;

---

Testing

All three bugs verified in a live Obsidian vault (2026-04-06).

Bug 1: ✅ PASS

1. Created a new workspace (no contextJson).
2. Opened chat settings, selected the workspace and a prompt, saved.
3. Model responded correctly with workspace context applied. Reopened settings — all selections retained.

Bug 2: ✅ PASS

1. Selected a workspace in a conversation and saved.
2. Restarted Obsidian.
3. Reopened the conversation — chat settings dropdown showed the correct workspace immediately.

Bug 3: ✅ PASS (verified via first-message token count and model behaviour — see pr2-workspace-two-tier-prompt.md testing section, which covers the full-load path this fix enables.)

[ProfSynapse]

Hi @Midway65, thanks for the continued work here — you keep finding real things worth looking at.

What we took

We pulled out the dead per-turn fetch removal and landed it on fix/remove-dead-per-turn-fetches. You were right that vaultStructure, availableWorkspaces, availablePrompts, and toolAgents were being fetched every message send and never consumed by the builder — I'd removed the consumption side a while back but never cleaned up the fetch side. That's now gone (-110 lines).

We also fixed the workspace?.context guard you identified — that was a real upstream bug dating back to Sep 2025 where most user-created workspaces were silently skipped during selection. Landed in #107.

Why we can't merge this PR as-is

1. It carries the full PR #97 payload. This PR is 40 files / 66 commits / +2,250/-1,896 but the workspace feature only touches ~5 files across 8 commits. The rest is the PR #97 content we already went through — schema migrations v12-v19, the action bar feature, the JSONL orphan pruning, and ~1,500 lines of CRLF whitespace conversion. All of that is still bundled in here, which makes it impossible to review or merge cleanly.

2. The two-tier system prompt doesn't match our architecture. The system prompt gets rebuilt before each message send and is always sent at position zero of the completion. The full workspace blob goes in there so the LLM is oriented to the workspace on every turn — it needs the folder tree, sessions, and context to function properly.

The "slim header on turn 2+" approach (G-W2/G-W3) saves tokens, but the LLM loses the folder tree and workspace context after the first message. Telling it to "call memoryManager.loadWorkspace if you need details" adds a round-trip and defeats the purpose — the LLM should just know what workspace it's in without having to ask. We need the full blob every time.

This also means the "cheap restore" (G-W1) can't be separated out — on main, the restoreWorkspace call via loadWorkspace() is what populates loadedWorkspaceData, which is what the system prompt builder serializes. Replacing that with a basic DB lookup would remove the folder tree from the system prompt.

3. Two of the 3 bug fixes address bugs introduced by the refactor, not bugs on main. The G-W3 flag race condition only exists because the G-W3 flag was introduced in this PR. The redundant context parameter is only redundant after the setWorkspaceContext refactor. These are valid fixes within your branch, but they don't address issues in the upstream codebase.

The workspace?.context guard that blocked 14/21 workspaces was a good catch — that one was a real bug on main (dating back to Sep 2025). We've fixed it separately in #107.

Specific items not taken and why

Item	Why not
Schema migrations v12-v19	Fork-specific cleanup for your prior Nomic/768-dim branch. Burns 8 schema version numbers that don't help upstream users.
Action bar (Insert/Append/Create File)	Previously discussed in #97 — adds UI complexity we don't want right now.
JSONL orphan pruning at startup	Deletes JSONL files not found in SQLite, but JSONL is our source of truth. If SQLite is corrupt/incomplete, this permanently destroys conversation data.
Two-tier prompt (G-W2/G-W3)	LLM needs full workspace context every turn, not just on the first message.
Cheap restore (G-W1)	Breaks the system prompt — loadWorkspace() is what populates the data serialized into the prompt.
G-W3 flag + redundant context param fixes	Fix bugs introduced by this PR's refactor, not bugs on main.
workspace?.context guard fix	Real upstream bug — good catch. Fixed separately in #107.
ProviderHttpClient require() switch	Already fixed on main via the mobile compat work (desktopRequire()).
CRLF whitespace changes	~1,500 lines of noise across 10 files that inflate the diff.

Going forward

One thing that would really help — please pull main so your fork is up to date before submitting PRs. Right now your branch has diverged significantly and every PR carries months of accumulated fork-specific changes mixed in with the new work. That makes it very hard to isolate what's actually new.

For the kind of issues you're finding, opening issues might work better than PRs. You clearly have good instincts for spotting inefficiencies — the dead per-turn fetches were a real find. If you file those as issues with the detail you put in your commit messages, we can implement fixes that fit cleanly on main without the merge headaches.

Thanks again for the time you put into this.