introspection-harbor

Harbor installed-agent adapter for local Introspection Recipes. Install it through introspection setup --target harbor, then run a task through:

introspection eval run --runner harbor --path evals/refund-task

Select an inherited Recipe agent variant by its declared YAML name:

introspection eval run --runner harbor --path evals/refund-task --agent agent2

For example, agents/agent2.yaml can declare name: agent2 and from: agent, then override only its model, tools, extensions, or other variant settings. The selected agent YAML is the only source of the evaluated model; the CLI and adapter do not apply a separate model override.

The CLI selects the Recipe and optional canonical replay context. The adapter copies that opaque context into Harbor's environment and forwards its path to introspection local, which creates the temporary native Pi session immediately before Pi starts. The adapter writes both native Pi events and trajectory.json (ATIF) under the agent log directory. Recipe execution preserves the Harbor task image's WORKDIR; the staged Recipe source is selected separately and does not replace the task's working directory.

E2B through scoped Data Plane egress

Operator can use E2B without Docker Compose and without receiving a real E2B, OpenAI, or Anthropic key. Configure E2B in the Data Plane's sandbox_providers; weight: 0 keeps it out of normal traffic while allowing an explicitly pinned Operator task to use it. Then select the Introspection environment class in the arguments forwarded to Harbor:

introspection eval run --runner harbor --path evals/refund-task -- \
  --env introspection_harbor.environment:IntrospectionE2BEnvironment

Harbor's native E2B implementation still owns preflight, template operations, sandbox creation, resources, and networking. The adapter only adds the public egress contract through Harbor's supported persistent_env constructor input. The Data Plane exchanges the scoped session locator for credentials after authorization. Model inference returns through that public egress path; model provider keys are never copied into the E2B environment. Adding a Kubernetes Agent Sandbox provider therefore does not require changing the CLI runner or task API.

The platform Operator image sets INTROSPECTION_HARBOR_ENVIRONMENT=introspection_harbor.environment:IntrospectionE2BEnvironment. When the variable is set, introspection eval run supplies that environment to Harbor unless the caller explicitly passes --env or -e. Ordinary developer installs leave it unset, so the same command keeps Harbor's local Docker default.

When none of the three Introspection egress variables are set, the same class adds no egress configuration and Harbor uses E2B's normal environment and credentials. A partial egress contract is rejected rather than silently mixing direct and routed traffic.

End-to-end MCP example

recipe-harbor-mcp-agent runs a Harbor trial in E2B while its Recipe calls a tiny MCP server on the developer's machine through introspection dev. It has no Docker Compose dependency. The Recipe declares a logical MCP id, so the local development bridge can later be replaced with an Introspection MCP endpoint binding without changing the Recipe or Harbor task.

From the example directory, run each command in a separate terminal. First start the included hello-world MCP server:

git clone git@github.com:introspection-org/recipe-harbor-mcp-agent.git
cd recipe-harbor-mcp-agent
introspection runtimes create
uv run --with 'fastmcp>=2.0' mcp/server.py

Before involving Harbor or E2B, verify the Recipe locally with Pi. In another terminal, configure the ignored local binding, provide the model credential selected by agents/agent.yaml, and run one prompt:

cd recipe-harbor-mcp-agent
introspection setup --yes
cp .pi/mcp.local.example.json .pi/mcp.local.json
export ANTHROPIC_API_KEY='...'
introspection local \
  --runtime hello-mcp \
  --agent agent \
  --print 'Call the hello MCP get_greeting tool and return exactly its result.'

This local mode runs the Recipe's Pi coding agent directly on the developer's machine. Pi resolves logical server hello from the ignored .pi/mcp.local.json, calls hello.get_greeting, and returns hello from the local MCP. The committed example binding contains only the loopback endpoint and no credential. This is a quick check of the Recipe, agent-level MCP authorization, and server before testing either Harbor or the development tunnel.

The Harbor task adds one more piece of work: the same Pi agent must call that tool and write its exact result to /app/greeting.txt; Harbor's deterministic verifier checks that file. Harbor and introspection_harbor run in the local evaluation process (or later in Operator), not inside the evaluated Recipe image.

Next attach that local server to the Recipe's development Runtime, leaving the command running while the evaluation executes:

cd recipe-harbor-mcp-agent
introspection dev --runtime hello-mcp \
  --mcp=hello=http://localhost:8010/mcp

Finally start the Harbor trial in E2B:

cd recipe-harbor-mcp-agent
introspection eval run \
  --runner harbor \
  --runtime hello-mcp \
  --path harbor-task \
  -- \
  --env introspection_harbor.environment:IntrospectionE2BEnvironment

The Recipe inside E2B calls hello.get_greeting through the development tunnel and writes the response for Harbor's verifier. See the standalone repository's README for prerequisites and the equivalent proxy configuration.