Pr/workspace two tier prompt

[Midway65]

feat(workspace): two-tier system prompt — slim header every turn, full load on first message

Replaces the per-turn full workspace JSON blob with a two-tier system: a ~100-token slim header on every turn, and a single full workspace load injected on the first message after a workspace is selected or restored. Also removes four dead per-turn async fetches.

Submit this PR before pr1-workspace-bugs — Bug 3 in that PR references pendingFullWorkspaceLoad introduced here.

Commits: 2f523359, f287fdc3, 3b9ee3ea, 9b5be688, 3c150070
Files changed: 3

• src/ui/chat/services/ModelAgentManager.ts
• src/ui/chat/services/SystemPromptBuilder.ts
• src/ui/chat/services/WorkspaceIntegrationService.ts

---

Problem

setWorkspaceContext() called loadWorkspace() (full MCP agent execution) at workspace-select time and stored the result in loadedWorkspaceData. SystemPromptBuilder.buildSelectedWorkspaceSection() then serialised the entire object — sessions, file tree, tasks, workflows, key files, preferences, recent activity — as a JSON blob inside <selected_workspace> on every message turn. Cost: 2,000–10,000+ tokens per turn, compounding with conversation history.

Four additional async calls ran inside buildSystemPromptWithWorkspace() on every send and were passed to SystemPromptBuilder.build(), where none were consumed:

• getVaultStructure() — reads vault root from disk
• listAvailableWorkspaces() — DB query
• getAvailablePromptSummaries() — DB query
• getToolAgentInfo() — agent registry lookup

All four were fetched, allocated, and immediately discarded on every turn.

---

Fix — two-tier workspace section

Slim header (every turn): A ~100-token <active_workspace> block containing only id, name, description, purpose, and rootFolder, plus a hint to call memoryManager.loadWorkspace for details:

<active_workspace id="ws-uuid" name="My Project">
  Description: Blog writing workspace
  Root folder: 02-Projects/my-blog
  For file structure, sessions, or task details, call memoryManager.loadWorkspace.
</active_workspace>

Full blob (first message only): On the first message after a workspace is selected or restored, the full loadWorkspace result is injected once into that turn's system prompt via a pendingFullWorkspaceLoad flag. Subsequent turns revert to the slim header. The full data remains available to SubagentController.getLoadedWorkspaceData() during tool-call processing for the current turn; it is cleared at the start of the next getMessageOptions() call (not at the end of the current one — this preserves subagent access during the current message cycle).

The four dead per-turn fetches are removed entirely.

---

New fields on ModelAgentManager

private pendingFullWorkspaceLoad = false;        // true from workspace selection until first message sent
private selectedWorkspaceSlimData: {             // always populated when workspace selected
  id: string; name: string;
  description?: string; purpose?: string; rootFolder?: string;
} | null = null;

---

SystemPromptOptions interface

// Added
selectedWorkspaceSlimData?: { id: string; name: string; ... } | null;

// Existing — kept for upstream interface compatibility, no longer consumed by section builder
workspaceContext?: WorkspaceContext | null;

---

WorkspaceIntegrationService — new method

// Direct SQLite query — no MCP agent, safe at any startup phase
async getWorkspaceBasic(workspaceId: string): Promise<{
  id: string; name: string; description?: string;
  rootFolder?: string; context?: Record<string, unknown>;
} | null>

Used by restoreWorkspace() (cheap identity restore) and setWorkspaceContext() (slim data population).

---

State management

All clear paths — clearWorkspaceContext(), initializeDefaultModel(), and the null-restore branch in restoreFromConversationMetadata() — now consistently clear all five workspace-related fields:

• selectedWorkspaceId
• workspaceContext
• loadedWorkspaceData
• pendingFullWorkspaceLoad
• selectedWorkspaceSlimData

Two code-review corrections applied in a follow-up commit:

1. pendingFullWorkspaceLoad is now set only inside the if (slim) branch of setWorkspaceContext() — not before the DB fetch. Prevents a wasted loadWorkspace call if the workspace does not exist.
2. Dead workspaceContext: this.workspaceContext pass removed from the build() call in buildSystemPromptWithWorkspace(). The field is kept in SystemPromptOptions for upstream interface compatibility.

---

WorkflowRunService — confirmed safe

WorkflowRunService is a second caller of SystemPromptBuilder.build(). It passes toolAgents, vaultStructure, availableWorkspaces — all still present in SystemPromptOptions. No breakage.

---

Testing

Verified in a live Obsidian vault (2026-04-06).

Two-tier prompt:

1. Selected the Nexus workspace and sent a first message asking for a full workspace resource inventory.
2. Model returned a complete inventory of 93 files across all subfolders, correctly navigating the workspace root folder (02-Projects/Nexus) and reading key state files — confirming full <selected_workspace> blob fired on the first message.
3. Token counter jumped from ~981 tokens (post-open baseline) to 12K/200K after the first exchange — consistent with full workspace data injection on first message only.
4. No errors in the developer console.

[ProfSynapse]

Hi @Midway65, thanks again for the continued effort here — you clearly have a solid understanding of the codebase and keep finding real things worth looking at.

I'm going to close this one in favor of #105 since it's a strict superset (same base + 3 additional bug fix commits). No point reviewing both separately.

See my comment on #105 for the full breakdown.