l9-cognitive-runtime — private Agent Plugin

Packages the existing L9 Cognitive Runtime as one private Agent Plugin (Agent Plugins 1.0.0).

Why there is no mcp.json

The portable mcp.json would have to name either a stdio command with its required environment or a remote URL. Neither can be stated truthfully here:

• stdio — the server refuses to start without L9_PACK_ROOT (no working-directory fallback). A pack root is a machine-specific path to a verified pack; ${PLUGIN_ROOT} does not contain one and installed plugins are copied to a host cache, so any value would be an invented path.
• streamable-http — the hosted endpoint and its OAuth registration are operator-owned and not yet bound to this package.

So the connection is configured by the operator in the host, and the skill fails closed when it is absent. Adding mcp.json later is a manifest-only change once a real endpoint exists.

Connection prerequisites

One of:

1. Local stdio (desktop hosts that allow local MCP servers)

   • Python ≥ 3.11 and the package installed from this repository, which provides the l9-cognitive-runtime-mcp console script.

   • A verified runtime pack. Build one from a checkout:

     uv sync --extra dev --frozen
     uv run python -m l9_cognitive_runtime.deployment \
       --source-root . --destination <pack-dir> --source-revision "$(git rev-parse HEAD)"
     

   • Register the server in the host as l9-cognitive-runtime, command l9-cognitive-runtime-mcp, environment L9_PACK_ROOT=<pack-dir> (absolute path). L9_MCP_TRANSPORT may be omitted; any value other than stdio is refused.

2. Hosted — the operator-registered HTTPS deployment of the same server (l9-cognitive-runtime-http, path /v1/mcp, OAuth 2.0 bearer). Connection and account authorization are done in the host UI; this package stores no URL, client id, or credential.

Blocked behavior (by design)

Human UI installation (operator)

Follows https://developers.openai.com/plugins/build/plugins (local/private marketplace). Not performed or verified by this package:

1. Add a repo marketplace entry in .agents/plugins/marketplace.json with "source": {"source": "local", "path": "./plugins/l9-cognitive-runtime"}, or copy this directory to a personal marketplace.
2. Restart the desktop app and confirm l9-cognitive-runtime is listed.
3. Connect the MCP server (local stdio or hosted, above).
4. Run the first compilation smoke from a new conversation.

First compilation smoke

In the host, after connecting:

Use compile-runtime-bundle to compile the mission "Update greeting text." with no governed context.

Expected: runtime_capabilities reports read_only with writes, execution and shell false; plan_context_requirements returns a context-plan.sha256:… id; compile_runtime with that expected_context_plan_id returns the same id, a run_id, resource_uri l9://runs/<run_id>, and execution_packet.compiled_task_context_digest == digests.context.

The same sequence is reproducible without a host — tests/test_plugin_package.py drives the skill's workflow against the real server, and from a clean checkout:

uv sync --extra dev --frozen --no-install-project --no-build
PYTHONPATH=src uv run --no-sync --no-build python -m pytest -q tests/test_plugin_package.py