Command palette

The Command Palette is Team-X’s natural-language command surface. Instead of clicking through menus to hire an employee, assign a ticket, or call a meeting, type what you want and Team-X classifies the intent, resolves the entities, and executes, with a confirmation gate for destructive actions.

Overview

• One keystroke: Ctrl+K (Windows / Linux) or Cmd+K (macOS) opens the palette from any view.
• 15 intents: covers the full spectrum of org operations (hire, fire, promote, assign, create, close, reopen, project, goal, meeting, status, navigation, vault search) plus a fallback to the conversational agent.
• Local-first: classification runs through your configured LLM provider; with Ollama on the local privacy tier, nothing leaves your machine.
• Destructive-action gate: fire, close, end-meeting, and promote require an explicit confirm click before the IPC fires.
• History plus audit: every executed command is written to history (recallable with ArrowUp) and appended to the audit log as a command.executed event.

Opening the palette

Action	Shortcut
Open the palette	Ctrl+K (Win / Linux) or Cmd+K (macOS)
Close the palette	Esc
Submit / confirm	Enter
Cycle history (from empty input)	ArrowUp / ArrowDown

The keybinding is platform-aware and resolves at runtime in app-shell.tsx. There is no per-view variation; the palette is global.

What you can say

Team-X classifies your input into one of 15 intents. Each intent has required and optional entity slots; the palette will ask for clarification when an entity is ambiguous and for confirmation when an action is destructive.

Intent	Examples
hire_employee	”Hire a senior backend engineer”, “We need a CTO”
fire_employee	”Fire James”, “Let go of Sarah Chen”
promote_employee	”Promote Mike to VP Engineering”, “Move Sarah to tech lead”
assign_ticket	”Assign the auth bug to Sarah”, “Give ticket #42 to James”
create_ticket	”File a ticket for the login crash”, “New ticket: optimize the dashboard query”
close_ticket	”Close ticket #17”, “Mark the auth bug as done”
reopen_ticket	”Reopen ticket #17”, “The login bug is back, reopen it”
create_project	”Start a project called Onboarding Redesign”, “New project for Q2 launch”
create_goal	”Add a goal: ship MVP by end of quarter”, “Set a goal for 99.9% uptime”
call_meeting	”All-hands with the engineering team”, “Call a meeting with Sarah and Mike about the launch”
end_meeting	”End the meeting”, “Wrap up the all-hands”
check_status	”What is everyone working on?”, “Show me the team’s status”
show_view	”Take me to projects”, “Open the audit log”
search_vault	”Find the API spec”, “Search vault for onboarding docs”
complex_request	”Why is the frontend team behind schedule?”, “Plan the next sprint based on open tickets”

complex_request is the catch-all: anything ambiguous, multi-step, or below the confidence threshold gets routed to the conversational agent (the agentic loop). You will never get a hard “I do not understand” response; you will get a thoughtful one.

File creation is handled by employee execution tools, not by a dedicated palette intent. Ask an employee in chat or on a ticket for deliverables such as Markdown briefs, XLSX trackers, DOCX reports, or PPTX updates. Created files appear in the employee workspace and, when vault storage is available, in Files and Autonomy > Artifacts.

Destructive actions

Four intents are gated by an explicit confirmation step:

• fire_employee
• close_ticket
• end_meeting
• promote_employee

When the palette classifies one of these, you will see a summary line (“Fire Sarah Chen (Senior Frontend Engineer)?”) with a Confirm and Cancel button. Pressing Enter while focused on Confirm executes; Esc or Cancel aborts. Non-destructive actions (hire_employee, assign_ticket, create_*) execute on the first Enter and surface an undo toast.

The gate is enforced in main-process code (CommandService.execute rejects without confirmed: true on destructive intents). The renderer cannot bypass it.

Ambiguous matches

If the entity resolver finds 2 to 5 candidates for a name or reference, the palette asks you to pick:

• “Sarah” with two Sarahs in the org: palette shows both, you click one.
• “the auth bug” matching three open tickets: palette ranks by FTS5 relevance and shows the top three.
• “frontend engineer” matching multiple roles in the catalog: palette lists role names with level chips.

Use ArrowDown / ArrowUp to navigate the candidate list, Enter to select. If no match is found, the palette returns not_found and you can refine the input or use a more specific reference (e.g., #42 for tickets, full names for employees).

History

Every successful command is written to history (last 20, FIFO) and persisted in the settings key-value store. Recall:

• Open the palette with empty input.
• Press ArrowUp to cycle through your last 20 commands, newest first.
• ArrowDown to cycle forward.
• Enter to re-execute, or edit the recalled text first.

History also appears as a “Recent Commands” panel on the Dashboard subview.

Slash commands

For navigation without going through the classifier, use the structured /show syntax:

Command	Action
/show dashboard	Switch to Dashboard tab
/show tickets	Switch to Tickets tab
/show projects	Switch to Projects tab
/show meetings	Switch to Meetings tab
/show telemetry	Switch to Telemetry tab
/show files	Switch to Files tab
/show audit	Switch to Audit tab
/show settings	Switch to Settings tab

Slash commands bypass NLU entirely. No LLM call, instant navigation. Use them when you know exactly where you want to go.

Troubleshooting

“It did not understand me.” Below the 0.5 confidence threshold, the classifier falls back to complex_request and routes the input to the conversational agent. The agent will either complete the multi-step request or ask a clarifying question. If you want a deterministic intent resolution, rephrase using a verb the classifier recognizes (“hire”, “fire”, “assign”, “create”, “close”, “call”, “end”, “show”, “find”).

“I want more intents.” The 15-intent set is intentionally tight; it covers the deterministic operations Team-X exposes today. New intents land alongside new IPC handlers. For org-specific commands, contribute to the role pack.

“The classifier is slow.” Classification cost depends on the configured provider. Local Ollama on a modest CPU can take 1 to 3 seconds for the first call, faster on subsequent calls (model stays warm). Cloud providers (Anthropic, OpenAI) typically respond in 200 to 500ms. If latency matters, configure a fast model in the balanced tier in Settings > Providers.

“I get malformed JSON errors.” Smaller local models (Ollama 8B parameter range) occasionally produce invalid JSON. The classifier retries once with a “your previous output was not valid JSON” nudge, then falls back to complex_request. Switch to a larger model or a cloud provider for higher reliability.

Privacy

The classifier runs through your configured LLM provider via the provider router, the same path as every other LLM call in Team-X. Privacy tier filtering applies: if your max tier is local, only Ollama is used. If you allow proprietary-cloud, the classifier may use Anthropic, OpenAI, or Google. Your text never leaves your machine unless you have explicitly enabled a cloud provider.

No phone-home. No analytics. No third-party telemetry. The command palette honors the same invariant as the rest of the app.