agent-team
agent-team is a public, owner-maintained Agent Plugins 1.0 package for GitHub Copilot in VS Code, using the Copilot harness in Agent Host.
It bundles reusable skills, custom agents, scoped rules, and Model Context Protocol (MCP) server configuration for software development and data analysis.
Publication status
Version 0.3.0 is the package manifest version in the awieczork-tools marketplace catalog.
The versioned notes describe this version; GitHub releases record published tags.
The former 0.2.0 release includes the changelog, open-issues, and project-status rules; 0.3.0 excludes them.
The catalog serves the full-repository main branch independently of release-tag creation.
Agent Host loading of this package's namespaced agents and rules remains unverified; static package checks do not establish client compatibility.
The owner waived runtime verification on 2026-10-28, not the runtime risk.
VS Code 1.142.0 did not run namespaced plugin agents or list namespaced agents and rules in the editor (runtime issue, editor issue); this package has not been tested in a later client.
Maintenance
The owner maintains this plugin for personal use. Public access supports installation without repository credentials; this repository does not accept external contributions or provide public support.
Requirements
- VS Code with GitHub Copilot, using the Copilot harness in Agent Host.
uvwithuvxonPATHfor Python execution and the Redshift MCP server.- Node.js with
npxonPATHfor the context7 MCP server. - The SQL Server extension
ms-mssql.mssqlor the Redshift MCP server, locally configured access, and read-only database permissions for database work.
Installation
The following steps manage the marketplace package; runtime support remains unverified.
Run each /plugin command separately in a VS Code chat session with the Copilot harness selected.
Replace the original plugin
Uninstall the original copilot-workbench before installing agent-team, so both copies do not load together.
For an installation made through /plugin, run:
/plugin uninstall copilot-workbench@awieczork-tools
For an installation made through the Extensions view, the Agent Customizations editor, or Chat: Install Plugin From Source, uninstall it in the Agent Plugins - Installed view or the Agent Customizations editor.
The editor and /plugin use separate plugin inventories; manage an installation through the interface where you installed it.
Install from the marketplace
Register the marketplace if it is not already registered, then refresh it and install the package:
/plugin marketplace add awieczork/copilot-workbench
/plugin marketplace update awieczork-tools
/plugin install agent-team@awieczork-tools
/plugin list
To update, refresh the catalog and then update the plugin:
/plugin marketplace update awieczork-tools
/plugin update agent-team@awieczork-tools
Manage this installation through /plugin.
The VS Code plugin documentation describes plugin management and its client requirements.
Package contents
Skills
| Skill | Purpose |
|---|---|
brief-subagent | Supplies self-contained briefs for one-shot delegation and how to act on the status each worker returns. |
interview-me | Clarifies the user's intent before work continues; only the user invokes it. |
redshift | Runs read-only queries against user-provided Redshift targets, with an MCP-first workflow and a bounded Python fallback. |
run-python | Runs Python scripts and inline code through uv, with or without the project environment. |
seed-task | Turns the current conversation, such as a forked discussion, into a task folder that an independent session starts from; only the user invokes it. |
sql-server | Runs read-only SQL Server queries against user-provided targets through saved connections. |
task-context | Keeps one task's state in a Git-ignored task folder that the task holder maintains, with a result file for a delegated result that must survive the session. |
Agents
The intended runtime behavior is for entry agents to appear in the agent picker and workers to run as subagents of an allowed entry agent or producer. Agent Host loading of these packaged agents remains unverified.
| Agent | Kind | Purpose |
|---|---|---|
lead | Lead | Delivers or plans any work in the workspace by deciding the approach, splitting it into parts for workers, and judging results. |
data-lead | Lead | Handles work that touches or rests on data: frames the question, has workers run the queries, and owns the interpretation and the conclusion. |
discuss | Chat | Shapes an unfinished idea into a request for a lead, and edits the documentation the user asks for. |
explorer | Helper | Searches the workspace for one question and returns locations with findings. |
database-explorer | Helper | Finds and profiles the SQL Server and Redshift objects and join paths for one data question, drafts and runs read-only queries. |
data-explorer | Helper | Finds and profiles the local dataset files for one data question and runs the analyses a brief specifies without changing the files. |
researcher | Helper | Gathers sourced external evidence for one bounded question. |
planner | Producer | Produces an executable plan for a change or an analysis with dependent steps or open design decisions. |
developer | Producer | Implements a change to code, tests, configuration, documentation, build, dependencies, or CI. |
ui-developer | Producer | Implements an interface change, checks it in the running app in a browser. |
reviewer | Checker | Assesses a finished result against its goal without repairing it. |
challenger | Helper | Tests one decision or framing with the strongest objection, the cheapest settling check, and a recommendation. |
Rules
| Rule | Scope |
|---|---|
markdown | Markdown documents; customization files get only its syntax rules. |
python | User-run task scripts; reusable code follows its project conventions. |
sql | Read-only SQL Server or Redshift queries: schema reading, result bounding, and query text; connection handling and execution safety belong to the database skills. |
MCP servers
| Server | Launcher | Purpose |
|---|---|---|
awslabs.redshift-mcp-server | uvx | Provides Redshift query and catalog tools. |
context7 | npx | Provides current library documentation; uses CONTEXT7_API_KEY when set, or anonymous access with lower rate limits. |
The configuration is in mcp.json. The SQL Server tools and the Redshift fallback runner can accept writes; the skills and agents authorize read-only work, and read-only logins provide an additional boundary.
Context7 API key
Create an API key in the Context7 dashboard.
Set CONTEXT7_API_KEY in the environment on the machine that runs the Context7 MCP server, before VS Code and Agent Host start.
The packaged server reads this variable directly; no change to mcp.json is required.
Keep the real key out of chat and repository files.
Set it in the environment that Agent Host starts from, and fully quit and restart VS Code after a change; for a remote, Windows Subsystem for Linux (WSL), or container session, set it where the MCP server runs.
Database access
Supply the target server or Redshift cluster and the database in the request. Keep credentials in local configuration and never provide them in chat or commit them to the package.
For SQL Server, create a saved connection in the SQL Server extension. The SQL Server skill connects to the saved server, verifies the active target, and disconnects when the task ends or becomes blocked.
The Redshift MCP configuration uses the local AWS profile redshift-mcp.
Configure that profile's credentials and region on the machine that runs the server; no cluster or database target is encoded in the skills.
The Redshift skill uses MCP first and switches to its Python fallback only when MCP cannot reach the cluster.
The fallback requires a workspace-root .env with these locally supplied settings:
REDSHIFT_USER=
REDSHIFT_PASSWORD=
REDSHIFT_HOST=
REDSHIFT_PORT=
REDSHIFT_DATABASE=
The runner reports the host and database it connected to, and the skill discards a result from another target. The runner returns at most 100 rows and reports truncation.