Offline mesh

Directory examples/python/features/offline_mesh

Host concierge and fares in one process with atfgraph.offline.OfflineMesh. It executes nested calls without a graph server. The offline-mesh test suite checks this local mode; it has different persistence and API coverage from the shared server.

Capabilities

Capability	Kind	Signature	Module	Summary
`plan`	traversal	`(origin: str, cities: List[str], nights: int = 2) -> dict`	`examples/python/features/offline_mesh/agent.py`	Quote each leg through the fares node, keep the itinerary in session memory, return the cost tree.
`quote`	skill	`(origin: str, destination: str, nights: int = 1) -> dict`	`examples/python/features/offline_mesh/agent.py`	Deterministic fixture fare; records usage as if a pricing model had been consulted.
`recall`	traversal	`() -> dict`	`examples/python/features/offline_mesh/agent.py`	Read the itinerary the session remembered (empty outside a session that planned one).

Commands

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

python -m atfgraph.worker --module examples.python.features.offline_mesh.agent:concierge
python -m atfgraph.worker --module examples.python.features.offline_mesh.agent:fares
python -m examples.python.features.offline_mesh.run
python -m examples.python.features.offline_mesh.run --serve 127.0.0.1:8765    # optional uvicorn
python -m examples.python.features.offline_mesh.run --serve 127.0.0.1:8765   # needs the optional uvicorn

Files

examples/python/features/offline_mesh/README.md
examples/python/features/offline_mesh/__init__.py
examples/python/features/offline_mesh/agent.py
examples/python/features/offline_mesh/agentnode.yaml
examples/python/features/offline_mesh/run.py

Note: More than one agent node is declared in this directory; capabilities are listed without a single label.

README

Rendered from examples/python/features/offline_mesh/README.md; the source file is not modified.

Piece	What it shows
`OfflineMesh([concierge, fares])`	each node mounted as an ASGI sub-app at `/{node_id}` of one root ASGI app (`/health`, `/nodes`, `/{node}/health`, `/{node}/definition`, `POST /{node}/capabilities/{name}`)
`mesh.invoke("concierge.plan", ..., session_id="trip-1")`	dispatch through the ASGI app; failures raise the usual `TraversalFailed` / `TraversalCancelled` / `TraversalTimedOut`
`await concierge.traverse("fares.quote", ...)` inside the handler	a nested child executed in-process by the other node, with its own task-local `ExecutionContext` (parent step, inherited root and session, real local lease)
`fares.memory.set(...)` / `concierge.memory.get(...)`	the mesh's in-memory store, scoped exactly as the SDK selects (session → workflow → global)
`record_usage` + `app.execution_cost()`	per-execution tracker, `usage` progress events, offline usage store, child usage propagated into the parent
`mesh.discovery`	the normal `AgentMesh` facade over the offline registry
`mesh.submit(...)` → `Handle`	poll, cancel, `path()`, `event_log()`, `events()` through `mesh.client()`

python -m examples.python.features.offline_mesh.run
python -m examples.python.features.offline_mesh.run --serve 127.0.0.1:8765    # optional uvicorn
curl -s -X POST localhost:8765/concierge/capabilities/plan -H 'x-atfgraph-session-id: t1' \
     -d '{"origin":"BNE","cities":["SYD"]}'

You can also serve the same Agent definitions as normal outbound workers, for example through examples.python.features.offline_mesh.agent:concierge. Server mode provides the shared runtime rather than the local in-memory graph.

What offline means here

Local execution never claims a remote lease. The offline graph issues and checks its own leases for nested calls, progress, usage and checkpoints; invalid leases return 409.
Unsupported routes, including approvals, knowledge, vector search, publications and retention, return 501 offline_unsupported. Inspect mesh.graph.unsupported to see which were attempted.
State exists only for the mesh lifetime and does not survive process loss.
Fares are deterministic fixtures; no supplier or model is contacted.