Agently 4.1.4.6 Release Notes

Agently 4.1.4.6 is a focused compatibility patch over 4.1.4.5. It:

• exposes the package version through the standard __version__ spelling;
• normalizes generated reasoning across OpenAI-compatible Chat Completions, Anthropic-compatible Messages, and Responses adapters;
• makes the public compatible-request retry policy the sole owner of physical SSE reconnection across Chat Completions, Anthropic Messages, and Responses;
• keeps live runtime resources live when they cross a TriggerFlow sub-flow boundary.

The sandbox-provider drafts tracked in #342 are not part of this release.

Core changes

Area	What changed	Recommended usage	Compatibility / risk	Evidence
Package version	Added the standard agently.__version__ and Agently.__version__ surfaces; removed the mistaken 4.1.4.5 version attributes	Inspect either __version__ surface	Intentional breaking correction only for code that adopted the mistaken 4.1.4.5 attributes	Package/version tests, built wheel and clean-install smoke
Compatible model reasoning	Chat Completions, Anthropic Messages, and Responses now converge generated reasoning into reasoning_delta, reasoning_done, and provider-native original_done	Keep provider controls in request_options; consume the public reasoning lifecycle	Additive response normalization; provider request options remain unchanged	Adapter suites plus real deepseek-v4-flash Chat Completions and Anthropic Messages probes
Compatible SSE retry	Removed hidden transport reconnects and made request_retry / AttemptRunner the sole physical-request retry owner across all three adapters	Use request_retry=False or max_attempts=1 for one physical connection; use after_output=False when provisional output cannot be reset	Default bounded replay remains; newly observable boundaries prevent uncounted requests and mixed output	Physical-connection, retry-boundary, attempt-index, [DONE], and EDA TransportGuard regression tests
TriggerFlow sub-flow resources	Resource capture and isolated child templates preserve live runtime-resource object identity	Pass clients, locks, callbacks, and services through resources; re-inject them after restoring a saved root execution	Resource capture is shared by identity; ordinary input/value capture remains copy-isolated	TriggerFlow resource-identity tests and sub-flow foundation example
Sandbox provider contributions	No sandbox provider draft is merged in 4.1.4.6	Coordinate and consolidate the contributor work through issue #342	Explicitly deferred; no runtime or dependency change in this release	Issue #342 and unchanged optional-provider surface

Standard package version

Use either public version surface:

import agently
from agently import Agently

assert agently.__version__ == "4.1.4.6"
assert Agently.__version__ == agently.__version__

The mistakenly introduced agently.version and Agently.version attributes from 4.1.4.5 are removed rather than retained as aliases.

Unified reasoning lifecycle

Provider-specific request options remain under request_options and pass through unchanged. For example, an OpenAI-compatible endpoint can receive:

plugins:
  ModelRequester:
    OpenAICompatible:
      model: deepseek-v4-flash
      request_options:
        thinking:
          type: enabled

Generated reasoning is projected through the same public response lifecycle:

• reasoning_delta carries streamed reasoning increments;
• reasoning_done carries the reconciled reasoning value exactly once;
• original_done retains the provider-native completed response.

OpenAI-compatible Chat Completions now retains non-stream and streamed reasoning_content. Anthropic-compatible responses reconcile thinking blocks, signatures, tool-use continuations, and an empty or populated terminal reasoning event. Responses adapters map reasoning-summary delta/done events and completed-output reasoning items without exposing request configuration such as reasoning.effort or reasoning.summary as generated content.

This is adapter-level normalization, not a DeepSeek-only branch. Providers that use the corresponding compatible protocol inherit the same lifecycle behavior.

One observable SSE retry lifecycle

request_retry now owns every physical streaming request across OpenAICompatible, AnthropicCompatible, and OpenAIResponsesCompatible. The transports no longer perform an additional stamina reconnect inside one public attempt. Therefore request_retry=False and max_attempts=1 both make at most one physical SSE connection; after_output=False also prevents replay after partial output. When replay is enabled, each replacement connection has the public, incrementing attempt_index and retry boundary.

The removed inner reconnect also removes its synchronous SSE retry-delay sleep from the async event loop. [DONE] remains the logical terminal marker; a disconnect before [DONE] remains a transport failure governed by the public policy.

TriggerFlow sub-flow runtime resources

Direct resources -> resources capture forwards the live object by identity:

parent_flow.to_sub_flow(
    child_flow,
    capture={"resources": {"service": "resources.service"}},
)

execution = parent_flow.create_execution(
    auto_close=False,
    runtime_resources={"service": live_service},
)

An isolated child-flow template also inherits its flow-level runtime resources by identity. Clients, callbacks, locks, events, and other live handles are not deep-copied. Ordinary input/value capture remains isolated by copy, so this fix does not turn general sub-flow data into shared mutable state.

Saved executions still do not serialize live resources. The host must re-inject required resources when restoring a root execution.

Compatibility

• Python: >=3.10
• Agently-Stage: >=0.3.5,<0.4.0
• Recommended Agently DevTools: >=0.1.10,<0.2.0
• Skills authoring protocol: agently-skills.authoring.v2
• DevTools observation protocol: agently-devtools.observation-runtime.v1

The opt-in long-output continuation and Stage-backed TriggerFlow task ownership introduced in 4.1.4.5 remain unchanged.