Two portable Agent Plugins: evidence-led research and durable sandboxed sessions

[denial123789]

SandBase now has two portable Agent Plugins 1.0 packages that passed the
Awesome Copilot external-plugin intake checks: spec compliance, install smoke
test, version matching, and (for the Skill package) Vally lint.

They cover two separate layers and can be installed independently or in the
same Copilot CLI environment.

1. Evidence-led research

copilot plugin install sandbaseai/sandbase-skills:agent-plugin

sandbase-research packages the multi-source-search Skill. It asks the agent
to distinguish discovery from validation, cross-check important claims against
independent sources, record disagreements, and emit a structured evidence
ledger. The included offline validator checks the ledger for duplicate or
unknown sources, unused evidence, confidence inflation, and undeclared source
conflicts.

It uses search and page-reading capabilities supplied by the host, so this
plugin does not require a SandBase account or API key.

Useful prompts:

• "Fact-check this release claim using independent primary sources and return
  a validated evidence ledger."
• "Compare the strongest evidence for and against this architecture decision;
  preserve source disagreements."

Source: https://github.com/sandbaseai/sandbase-skills/tree/0ff6bfd67332b620937d781d0cb81447cfd72a1d/agent-plugin

Intake: #2716

2. Durable agent sessions over MCP

copilot plugin install sandbaseai/sandbase-harness:agent-plugin

sandbase-harness starts a stdio MCP bridge to a running local or self-hosted
Harness runtime. It exposes six tools for listing agents, creating and running
sessions, reading results and artifacts, and stopping work. The session tool
waits for the streamed turn to reach a terminal state, so the client does not
need to reimplement the durable event loop.

The bridge stores no credentials. Its authority is limited by the optional
MANAGED_AGENTS_API_KEY, and session data remains in the configured Harness
workspace. The default local-process sandbox is for trusted development;
Docker or Kubernetes should be selected when stronger isolation is required.

Source: https://github.com/sandbaseai/sandbase-harness/tree/841ac14e1b5a37b52a97d89eebe19c030ee4f1c6/agent-plugin

Intake: #2717

The research plugin is intentionally usable on its own. Installing the Harness
plugin alongside it adds durable session and artifact operations without
changing the Skill's evidence rules or requiring it to call a hosted SandBase
service.

Feedback is welcome on the minimum plugin surface, evidence-ledger ergonomics,
and MCP session boundaries. Disclosure: I am affiliated with SandBase, and AI
assistance was used to prepare and verify this post.