| name | workspace-copilot |
| description | Analyze the user's sanitized Hyprland, tmux, Agent Deck, and ActivityWatch workspace context; explain what agents and windows are doing; identify context-switching or organization inefficiencies; and propose permissioned workspace changes. Use when the user asks about workspaces, windows, tmux sessions, agent fleet state, workflow habits, focus, navigation, keybind ergonomics, or workspace organization. |
| metadata | {"author":"kvn","version":"1.0"} |
Workspace Copilot
Use the local workspace-copilot CLI as the only workspace-observation interface. It exposes sanitized aggregates from Hyprland, the personal tmux server, Agent Deck, and privacy-filtered ActivityWatch.
Non-negotiable privacy boundary
Never inspect or request raw ActivityWatch events, window titles, pane contents, shell history rows, clipboard data, screenshots, audio, command arguments, browser URLs, or harness transcripts for this workflow. Never install input hooks or read /dev/input.
The approved context contains only application identity, workspace/monitor placement, project basenames, hashed paths, tmux topology, command basenames, explicit fleet session metadata, aggregate durations, semantic harness states, and prior accept/reject decisions.
Run this before relying on the data:
workspace-copilot --json doctor
Stop and report the failing boundary if ok is not true or activitywatch_privacy_ok is not true.
Analyze the workspace
Retrieve current sanitized context:
workspace-copilot --json context
Explain conclusions as:
- Observation — only fields present in the output.
- Inference — a clearly labeled interpretation.
- Recommendation — a concrete, reversible improvement.
- Evidence — counts, durations, session states, or repeated transitions.
Do not infer an agent's task from its pane output. Prefer the Agent Deck title, group, tool, state, and project basename. If those are insufficient, say so and ask the user to name or group the session rather than inspecting private content.
Generate recommendations
Refresh deterministic suggestions:
workspace-copilot --json suggest
workspace-copilot --json suggestions
Present every proposed mutation with its suggestion ID, evidence, exact allowlisted action, and undo availability. Advisory suggestions have no executable action.
Decision and execution protocol
A recommendation, acceptance, application, and undo are distinct transitions.
- Never treat discussion, interest, or a request for explanation as acceptance.
- Record acceptance only after the user explicitly accepts a suggestion ID.
- Record rejection and the user's reason when explicitly rejected.
- Never apply an accepted suggestion until the user separately requests application, unless the same message unmistakably says both accept and apply.
- Never construct an arbitrary shell command as a substitute for the allowlisted apply operation.
Record a decision:
workspace-copilot --json decide <id> accept --note '<brief reason>'
workspace-copilot --json decide <id> reject --note '<brief reason>'
Apply only an accepted suggestion:
workspace-copilot --json apply <id>
Undo only when requested:
workspace-copilot --json undo <id>
After apply or undo, rerun workspace-copilot --json context and verify the observable state transition.
Preference learning
Treat stored preference confidence as ranking evidence, not authority. Repeated acceptance may increase recommendation priority, but must never grant automatic execution permission. Rejections constrain future recommendations; explain any later proposal that appears to conflict with a recorded rejection.
