Assistant skills

Skills give Spear Assistant a reusable procedure. Type / in the composer to discover workflows, or open Skills in the assistant header to inspect them. Use Up/Down to select a suggestion and Enter or Tab to insert it; press Enter again to send. Escape dismisses the picker. /skills asks for the available list.

Built-in workflows

• /deploy gathers device types, sites, operating systems, scale and reachability to plan agent placement and enrollment. Installation steps may require manual work.
• /monitor helps create or update checks using supported schemas and existing agents.
• /alerts guides conditions, thresholds, evaluation windows and notification routing.
• /investigate compares incidents, test configuration, companion results and path history, then proposes a focused change when evidence supports one.
• /onboard plans account setup; /configure coordinates related changes.
• /dashboard builds saved views; /coverage reviews agent and test coverage.
• /access reviews roles and grants; /baselines configures learned baselines.
• /logs manages collection; /infrastructure explores devices and floor layouts.
• /verify checks outcomes; /report assembles evidence; /optimize reviews cost.
• /support prepares reviewed tickets and replies.

Related changes can appear as one ordered plan. Review each step, dependencies, current configuration and recovery advice, then approve the exact plan. Spear stops at the first failure and preserves completed changes. Ask it to inspect the saved receipt before continuing partial work; rollback is not automatic.

Runs continue through a browser disconnect. Reopening the conversation reattaches to saved progress; Stop cancels execution. After a server restart, safe analysis checkpoints resume within the original limits. An uncertain write pauses for verification instead of being retried.

Secret entry, authentication, account switching and uploads open the relevant platform form. Completing a handoff is a user action; Spear can continue independent work and verify configuration afterward.

The assistant may ask several related questions in a form. Answer the relevant choices and free-text fields, or reply in ordinary chat. Nothing is preselected. The submitted answers remain in the conversation.

Create a personal workflow

Open Skills to create a skill with a command name, description and instructions. Alternatively, ask “Save this process as a skill named branch-setup.” Chat-based writes use the normal approval flow. Skills belong to your user within the organization; they are not an organization-wide library.

Built-ins are read-only. Copy one to make a personal version. A saved command name stays fixed; copy the skill to use a different name. Built-in names and skills are reserved. Use in chat runs the saved version. Unsaved edits are protected when switching skills or returning to chat.

Reading and using skills requires ai:read. Creating, editing and deleting them requires ai:manage. Running a skill never grants permissions its user lacks.

Declare the tools a workflow needs

The editor’s Tools section accepts up to 16 tool references. Select a catalog tool, mark it required or optional, and explain its use in up to 500 characters. Unknown names, duplicates and empty usage instructions are rejected.

An explicit slash command saves a workflow snapshot with the conversation. Required tools load at the start; optional tools can be discovered as needed. The assistant also discovers relevant tools during ordinary chat. A declaration guides tool selection, but does not bypass access checks or change approvals.

Investigate before changing configuration

For recurring path alerts, /investigate can compare stored traces with same-agent destination connectivity. A silent intermediate router alone does not establish packet loss at the destination. Path-history counts describe sampled flow traces, not independent check runs, and may be truncated.

An offered configuration fix should identify the old and new values, supporting evidence and how to undo it. Review the exact proposal before approving. A saved change needs fresh observations before you can conclude it fixed the problem; it does not itself resolve an incident. Path-history tools read stored evidence and do not launch an on-demand probe.

External clients can discover skills through MCP prompts.