Agent Plugin

Load the portable project-context Skill and MCP configuration, verify Memory with a Citation, and do not pretend automatic recall exists.

Agent Plugin has no dedicated setup or doctor command. This package does not implement automatic recall, capture, or Runtime. The repository has package contract tests; a real host-process E2E is Not validated. Permission is Host-gated. Managed Skill export is Unsupported.

Agent Plugin is a portable package for hosts that can load Agent Plugin skills and MCP configuration. It is a client of a running PowerContext Server. It does not embed storage, start the Server, or add MCP tools. Memory, Handoff, and persistence stay with the Server’s existing MCP tools.

For the full loading procedure, use the Configure Agent Plugin how-to in the official docs. PowerContext 101 verifies the runnable boundary here.

Obtain the package and load it

Clone a checkout that contains the integration package, then start the Server separately:

git clone https://github.com/oceanbase/powercontext.git
cd powercontext

uv run powercontext server run

The package root is:

integrations/agent-plugin/powercontext/

It contains:

File	Role
plugin.json	Agent Plugins 1.0.0 metadata; name is powercontext, version is 0.1.0
mcp.json	Streamable HTTP configuration pointing at http://127.0.0.1:8000/mcp
skills/project-context/SKILL.md	Reusable instructions for explicit Memory and Handoff

VS Code loads local directories through chat.pluginLocations. Replace the path with your clone, and point at the package root that contains plugin.json:

{
  "chat.plugins.enabled": true,
  "chat.pluginLocations": {
    "/absolute/path/to/cloned/powercontext/integrations/agent-plugin/powercontext": true
  }
}

Reload VS Code. Confirm that PowerContext appears in the Agent Plugins view, and that chat can see the project-context skill and the powercontext MCP server.

Configuration boundary

mcp.json is the entire connection surface of this package:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json",
  "mcpServers": {
    "powercontext": {
      "type": "streamable-http",
      "url": "http://127.0.0.1:8000/mcp"
    }
  }
}

Agent Plugins 1.0.0 does not define a portable credential-reference field for remote MCP servers. The checked-in mcp.json therefore contains no static credentials, token placeholders, or POWERCONTEXT_* variables. When the Server requires auth, configure it in the loading host or client. Do not put a Bearer token in this file.

This package has no host-specific environment-variable table and no owned paths written by setup.

Minimal verification

Check the Server first:

powercontext doctor
powercontext ready
powercontext capabilities

Agent Plugin has no dedicated doctor command. A successful load only means the host can see the Skill and MCP entry. It does not prove Server reachability or recall.

In a chat that has loaded the package, save the learning-path answer explicitly:

Use PowerContext to remember this decision:
after changing the OpenAPI contract, regenerate the client and inspect the diff.
Return the Memory Citation. Do not store secrets.

This is MCP remember_memory and does not need a generation model. The host should return an exact Citation.

Ask again in a new request that reuses the same scope_id:

What should happen after the OpenAPI contract changes?

There is no automatic injection channel. The model must call search_memory or get_memory_entry to read the Memory you just stored. Recalled content includes a Citation and remains untrusted historical evidence.

Scope resolution

This Skill does not read .git/powercontext/codex-workspace.json and does not derive Scope from a Git remote. Before the first PowerContext tool call, choose one scope_id for the current task and reuse it for every Memory and Handoff call in that task.

Prefer a project identifier that stays stable across compatible hosts. When the GitHub repository is known:

git:github.com/owner/repository

If the user or host provides an explicit Scope, use that value. When Scope is ambiguous, ask before writing.

No automatic path

compatible host loads plugin.json + mcp.json + Skill
→ you start the Server separately
→ chat may use MCP tools
→ no prepare_context on prompt
→ no Source capture

Automatic recall, capture, and flush are Unsupported here. They are not fail-open helper paths. When MCP is unavailable, the Skill says to report it once and continue: do not retry in a loop, invent restored context, or claim that Memory or Handoff succeeded.

Explicit Memory and Handoff still go through Server MCP. The Skill uses model-free handoff_current_work, then has the receiver call continue_handoff with selection: "prepared". It does not call activate_handoff or finalize_handoff.

The bundled Skill is not a managed Skill export. Agent Plugin is not a current export target.

Negative checks

• Stop the Server, then call remember_memory. MCP must report failure and must not pretend the write succeeded.
• Stop the Server, then ask a normal question. There is no automatic recall to fail open; the host should continue and say MCP is unavailable when PowerContext is required.
• Open the checked-in mcp.json. It must have no headers, token, or POWERCONTEXT_*.
• Point chat.pluginLocations at the repository root instead of the package root. The host should not treat that as a valid Agent Plugin.

Current limits

• There is no setup or doctor command.
• A real VS Code or other Agent Plugin host-process E2E is Not validated. tests/agent_plugin/test_contract.py only checks the package and documentation contract.
• Automatic recall, capture, and flush are Unsupported.
• Managed Skill export is Unsupported.
• Authentication is not portable; the 1.0.0 schema has no credential-reference for remote MCP.

Completion checklist

• The package root contains plugin.json, mcp.json, and skills/project-context/SKILL.md.
• The host can see the project-context Skill and the powercontext MCP server.
• Server health, readiness, and capabilities were checked separately.
• Explicit remember_memory returned a Citation for the OpenAPI decision.
• The next question found it through MCP search or get, not through automatic injection.
• MCP failure stayed visible, and you did not treat the bundled Skill as a managed export.

Continue with Let the agent extend itself.