[THardy98]

Summary

Add an experimental sample showing how activity sessions keep a workflow's activities on one worker, so they can reuse worker-local conversation history, and how single-session workers bind workflows to stable identities.

• Each workflow is a conversation of sequential conversation_step activities. Turn n needs every earlier turn. Each worker caches the turns it holds and fetches only the turns it is missing, with simulated latency per turn (--fetch-seconds). The workflow, activity, worker, and fleet code is shared by two entrypoints.
• activity_sessions.affinity runs a batch of workflows per mode on a fresh worker fleet and compares turns fetched, time spent fetching, and the most workflows any one worker served:
  • none: ordinary task-queue routing; workers keep fetching turns they missed.
  • affinity: turns run inside workflow.activity_session(ActivitySessionOptions(placement=UseWorkflowSession())) on shared session workers (WorkerSessionConfig()); nothing is fetched.
  • single-session: the same on one single-session worker per workflow (SingleSessionAdmission(stable_identity=...)); nothing is fetched and each worker serves one workflow.
• activity_sessions.single_session shows both sides of single-session binding across the loss of the whole fleet, in four phases: N workflows and M workers start, each worker binds one workflow and the rest get none; the fleet is killed and every workflow stalls; N - M workers with new stable identities run only the unbound workflows while the bound ones stay stalled; the original M workers restart with their stable identities and resume their own workflows.
• Each run checks its expectations and exits non-zero on failure. A Rich live view shows a turn grid (who ran each turn and what it fetched), stat tiles for the comparison, and workflow and worker status tables with a phase banner for the lifecycle; --no-live prints only the summary.

Requirements and scope

Requires the experimental SDK from temporalio/sdk-python#1979 and a server built from the shivam/agentcore-single-session branch of https://github.com/temporalio/temporal, with sticky activity execution enabled for the namespace. The README explains running with the feature SDK interpreter.

Restarted single-session workers start with an empty cache and fetch the history once; persisting the cache across restarts is a follow-up.

Validation

• Both entrypoints run against the custom server: affinity comparison (no session fetched 27 turns vs 0 with affinity and single session; up to 5 workflows per worker vs 2 and 1), and all four single-session lifecycle checks pass.
• Unit tests and the opt-in integration tests (TEMPORAL_TEST_ACTIVITY_SESSION_SAMPLE=1): 13 passed against the custom server.
• ruff and mypy pass.