mirror of
https://github.com/langgenius/dify.git
synced 2026-05-28 12:53:23 +08:00
71db7d6c1c
Implements the dify-agent run lifecycle contract for workflow Agent nodes so a single workflow run can fan out into multiple Agent backend runs that share an Agent session via persisted snapshots, and so a HITL pause routes through a workflow-level pause instead of being modeled as agent-run pause. Highlights: * New ``workflow_agent_runtime_sessions`` table (migration a7b9c8d1e2f3) keyed by (tenant_id, workflow_run_id, node_id, binding_id, agent_id). Stores the CompositorSessionSnapshot plus the replayable ``composition_layer_specs`` so a future cleanup-only run can match the agenton compositor's snapshot-vs-composition name-order validator. * ``WorkflowAgentRuntimeRequestBuilder`` now threads the prior snapshot back into ``CreateRunRequest.session_snapshot`` and defaults to ``on_exit.default = suspend`` so layers survive between runs in the same Agent session. A pydantic-ai history layer is added by default to carry the Agent's chat history across runs. * ``DifyAgentNode`` saves a fresh snapshot on success / paused terminals, reuses the prior snapshot on the next entry, and retires the local row on failed / cancelled terminals so a workflow loop cannot resume from stale state. A paused agent run is reported as a workflow-level ``PauseRequestedEvent`` rather than an Agent backend pause. * New ``WorkflowAgentSessionCleanupLayer`` (GraphEngineLayer) attached on Workflow and AdvancedChat app runners. On terminal workflow events it retires every ACTIVE session row for that workflow run. The HTTP cleanup path is fully built (composition spec replay + snapshot filtering for excluded plugin layers + wait_run gating before mark_cleaned) but kept behind ``_HTTP_CLEANUP_SUPPORTED = False`` because the current dify-agent ``runner._run_agent`` always requires an LLM layer and has no ``purpose=cleanup`` branch yet; flipping the flag enables the contract once the backend supports cleanup-only runs. * Adds a compositor integration test that runs the real ``agenton.Compositor.from_config`` + ``_create_run`` against the built cleanup request so the snapshot-vs-composition mismatch trap (which is invisible to FakeAgentBackendRunClient) is caught at unit-test time. Tests: 130 passing in agent_v2 + agent_backend; verified end-to-end against real agent backend, gunicorn + celery, and the new ``workflow_agent_runtime_sessions`` table. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>