Co-consult MCP profile

Directory examples/atfhar/co-consult

This atfHar profile calls co-consult.chat through the graph server's MCP endpoint at http://127.0.0.1:18880/mcp. It enables the five tools below; the co-consult worker still runs separately and executes the travel workflow.

Commands

Taken from this example's own README and module docstrings.

python -m atfgraph.worker --module examples.python.co_consult.main:app --server http://127.0.0.1:18880

Files

examples/atfhar/co-consult/README.md
examples/atfhar/co-consult/extensions/session.ts
examples/atfhar/co-consult/harness.yml
examples/atfhar/co-consult/mcp.example.json
examples/atfhar/co-consult/prompts/coordinator.md

README

Rendered from examples/atfhar/co-consult/README.md; the source file is not modified.

Tool	Use
`discover_agents`	list nodes and their entry points
`get_reasoner_schema`	read the input/output schema of `co-consult.chat`
`execute_reasoner`	submit one chat turn; returns a run ID
`get_run`	inspect a run
`wait_run`	wait up to 120 s for a terminal state

In harness.yml, the tool names use the form mcp__atfgraph_<tool>.

Files

harness.yml defines the profile. ${ATFHAR_DEFAULT_MODEL} and an empty modelProviders.allow leave model selection to the operator's atfHar configuration. The example includes no credentials.
mcp.example.json defines the MCP connection. Copy it as .mcp.json in the installed profile directory. For secured servers, configure an environment-backed X-API-Key header using atfHar's MCP settings.
prompts/coordinator.md instructs the coordinator how to use vNext sessions.
extensions/session.ts retains one server-created session per harness conversation. It reads the first result's session_id, supplies it on later turns, and uses the tool-call ID as request_id to deduplicate retries.

Use

Copy the example to a profile directory you control; it does not install itself into ~/atf-harness/profiles. Point atfHar serve --profile-dir at the containing directory, start the graph server on 18880 and the co-consult worker, then open the harness UI:

python -m atfgraph.worker --module examples.python.co_consult.main:app --server http://127.0.0.1:18880
atfHar serve --workspace-root "$PWD" --profile-dir /path/to/profiles --ui-dir packages/ui/dist

Check your connection

Start a new conversation and request a travel plan. Check that the first wait_run result contains session_id. Ask about a saved offer and verify that the next call reuses it.

Verify the complete flow with your installed atfHar version and provider. Missing tools usually require checking MCP enablement, URL and API key. New sessions on every follow-up require checking extension loading and support for the tool_result hook.