Agently 4.1.3.2 Release Notes

Languages: English · 中文

Agently 4.1.3.2 completes the 4.1.3.2 AgentExecution task-step slice and adds the runtime visibility and bounded-diagnostics work needed to make those steps usable in real multi-step applications.

What Changed

• agent.create_execution(...) now supports the existing turn behavior and the new bounded mode="task_step" shape with explicit lineage, limits, and Workspace injection.
• AgentExecution metadata and stream items now carry execution id, execution mode, lineage, diagnostics, route, logs, close snapshots, and workspace references for step-to-step correlation.
• Model-request budgets are enforced across direct model calls, DynamicTask model tasks, and Skills model stages.
• Runtime stall control now covers whole AgentExecution deadlines, no-progress windows, provider stream idle waits, response materialization, ActionRuntime stages, and ActionFlow loop close.
• EventCenter remains the runtime event hub. It receives RuntimeEvent, normalizes and delivers it, and applies outlet-level delivery policies such as raw delivery, summary delivery, and async background dispatch.
• DevTools keeps its ObservationEvent projection while adapting to the new RuntimeEvent hierarchy and grouped rendering semantics.
• Release acceptance now requires a coverage-first argument before implementation artifacts are cited as proof.

Usage Shape

workspace = Agently.create_workspace("issue-intake")

execution = agent.create_execution(
    mode="task_step",
    lineage={"task_id": "issue-intake", "step_id": "collect-open-issues"},
    limits={"max_model_requests": 3, "max_seconds": 60, "max_no_progress_seconds": 15},
    workspace=workspace,
)

async for item in execution.async_stream():
    print(item.meta["execution_id"], item.meta["execution_mode"], item.meta["lineage"])

await execution.async_record_workspace(
    kind="checkpoint",
    data={"status": "collected", "source": "official site search"},
)
context = await workspace.async_build_context(query="latest unprocessed issues")

None is the preferred unlimited budget marker. -1 remains accepted for compatibility where numeric settings already use it.

Examples

• examples/agent_auto_orchestration/20_agent_execution_task_step_workspace_loop.py runs two task-step executions, passes lineage and limits explicitly, records observations and checkpoints into Workspace, rebuilds Workspace context between steps, and shows stream/meta correlation from a real model run.
• examples/agent_auto_orchestration/21_agent_execution_github_issue_intake.py demonstrates a DeepSeek-driven issue intake workflow where the agent receives shell capability, decides how to search, can use local gh when useful, and stores the intake result in Workspace without hard-coding a GitHub Issues API path.

Compatibility

• Package version: 4.1.3.2.
• Release manifest: compatibility/releases/4.1.3.2.json.
• Recommended agently-devtools: >=0.1.6,<0.2.0.
• Companion Skills guidance is aligned to the runtime, dynamic task, playbook, and debugging guidance used by this release.

Issue Scope

This release closes the runtime visibility and bounded-stall work behind issues #277 and #280. Broader provider adoption and long-run observation performance monitoring remain release-line follow-up work rather than 4.1.3.2 blockers.