feat(temporal): opt-in continue-as-new for long-lived agent workflows

[danielmillerp]

Why

Long-lived chat/session agents run as a single Temporal workflow that stays open indefinitely. Two things killed them:

1. Event history grows until it hits Temporal's ~50k-event / 50MB limit, then the workflow stalls.
2. The workflow execution timeout (24h SDK default) terminated the whole chain — a user returning to an old chat hit a dead workflow and their message silently vanished.

This adds an opt-in continue-as-new pattern to BaseWorkflow so a session can stay open forever, and defaults the execution timeout to infinite so a long-lived chat isn't capped at 24h.

Design

Opt-in by adoption, no flag, no patch gate. An agent gets recycling only by calling run_until_complete from its @workflow.run instead of a bare wait_condition(timeout=None). Agents that keep the old wait are untouched. (No workflow.patched() gate: we have no in-flight long-running workflows to preserve, and per Temporal guidance the right tool for evolving these "pure entity" workflows later is Worker Versioning + upgrade-on-continue-as-new — tracked in AGX1-420.)

BaseWorkflow helpers (src/agentex/lib/core/temporal/workflows/workflow.py):

• run_until_complete(*args, is_complete, timeout=None) — keeps the workflow open and recycles history when Temporal suggests it. Optional timeout caps the wait (default None = wait forever). Workflow-level lifetime cap is the execution timeout (WORKFLOW_EXECUTION_TIMEOUT_SECONDS, infinite by default).
• should_continue_as_new() — recycle when workflow.info().is_continue_as_new_suggested() (Temporal owns the threshold).
• drain_and_continue_as_new() — waits all_handlers_finished (so an in-flight turn isn't lost) and re-checks completion before workflow.continue_as_new.
• is_continued_run() — gate one-time @workflow.run prologue (welcome message, state rehydration) so it doesn't repeat on each recycle.

Execution timeout (environment_variables.py + temporal_task_service.py + temporal_client.py): WORKFLOW_EXECUTION_TIMEOUT_SECONDS now defaults to None = no execution timeout (None/0/negative → execution_timeout=None). It's chain-wide (continue-as-new does NOT reset it), so capping it would still kill a forever-chat.

Scope (deliberately concise)

Just the pattern + the timeout default. Restoring state after a recycle is framework-specific (rebuild from adk.messages, an adk.state snapshot, or a framework's own memory like a LangGraph checkpointer / Pydantic AI history) and is left to follow-up PRs, one per integration. The only example touched is 000_hello_acp, which keeps no cross-turn state — it adopts the pattern and gates its one-time welcome behind is_continued_run() so it isn't re-emitted on recycle.

Follow-ups (Temporal-team feedback)

• AGX1-420 — adopt Worker Versioning + upgrade-on-CAN for evolving long-running workflows (vs patching).
• AGX1-421 — durable idle-timeout to gracefully close idle workflows so they don't pile up.

Verification

• Unit tests for the decision helpers (should_continue_as_new, is_continued_run) — tests/lib/core/temporal/test_base_workflow_continue_as_new.py.
• py_compile + ruff + pyright clean.
• Follow-up: replay/integration test of drain_and_continue_as_new against a Temporal test server.

🤖 Generated with Claude Code

Greptile Summary

This PR adds opt-in Temporal continue-as-new support for long-lived workflows. The main changes are:

• New BaseWorkflow helpers for recycle decisions, handler draining, continued-run detection, and long waits.
• The 000_hello_acp Temporal example now opts into run_until_complete and skips its one-time welcome on recycled runs.
• Workflow execution timeouts now default to no timeout, with positive configured values still applied.
• Unit tests cover the pure continue-as-new decision helpers.

Confidence Score: 5/5

The changes are narrowly scoped to opt-in workflow recycling helpers and timeout configuration behavior, with no reported correctness or security issues.

The implementation is covered by focused unit tests for the pure helper behavior, and the touched example adopts the new pattern without adding persistent cross-turn state requirements.

T-Rex Logs

What T-Rex did

• T-Rex ran the requested verification, but its local artifact references were not uploaded.

_{Ran code and verified through T-Rex}

_{Reviews (18): Last reviewed commit: "feat(temporal): opt-in continue-as-new f..." | Re-trigger Greptile}