Agently 4.1.3.4 Release Notes

Languages: English · 中文

Agently 4.1.3.4 is a release-line hardening slice. The primary release reason is structured output reliability across local and cloud OpenAI-compatible models. It also includes runtime capability policy hardening and the first bounded AgentTaskLoop slice for the 4.1.4 AgentTask target.

Structured Output Hardening

.output(..., format="auto") now selects formats by schema structure only. It does not inspect field names, business meaning, tokenization, keywords, or model outputs.

• flat string-only dict schemas resolve to xml_field;
• mixed string plus typed non-string schemas resolve to hybrid;
• all-complex, all-control, non-dict, and dense machine contracts stay json;
• flat_markdown remains explicit-only for compatibility;
• yaml_literal is explicit opt-in and remains outside auto.

New and revised parser formats:

• xml_field: an XML-like field envelope with a custom boundary parser, not a strict XML parser;
• hybrid: Markdown sections for text fields and fenced JSON values for typed fields;
• yaml_literal: YAML inside explicit Agently boundaries with literal scalars for long text;
• flat_markdown: kept for explicit legacy usage, no longer recommended as an auto/default path.

Reasoning normalization is handled before parser-specific logic. Provider-native reasoning fields and a leading outer <think>...</think> block before the payload flow into existing reasoning events; parser payload, code blocks, and ordinary text retain internal <think> content.

Request And Streaming Reliability

OpenAICompatible now retries transient transport failures before output starts. The default is one replay (request_retry.max_attempts = 2). The replay keeps the same model, prompt, and output format. Once output has started, Agently does not replay automatically because that could duplicate partial content.

Response materialization now propagates explicit stream/provider construction errors through get_text(), get_data(), and get_meta() instead of waiting for the materialization timeout.

Runtime Capability Policy

Skills capability execution now uses the framework-wide PolicyApproval surface. Skill capability needs are recorded in SkillExecutionPlan, host policy can configure auto-load surfaces, and high-risk capabilities stay behind approval or fail-closed behavior.

Built-in Search reports partial_success when fallback providers recover after earlier backend failures. partial_success remains continuable evidence rather than an Action failure.

AgentTaskLoop First Public Slice

agent.create_task(...) is available as a bounded single-Agent task loop: plan one step, execute through AgentExecution, write Workspace evidence, verify, replan if needed, and finish as completed, blocked, or partial after limits.

This is intentionally a first public slice, not the complete 4.1.4 AgentTask target. It does not provide multi-task coordination, background autonomy, distributed leases, or long-term memory management. Use it for bounded single-agent workflows where the host controls workspace, limits, and enabled capabilities.

Compatibility Notes

• Package version: 4.1.3.4.
• Release manifest: compatibility/releases/4.1.3.4.json.
• Agently recommends agently-devtools >=0.1.7,<0.2.0.
• Agently-Skills uses authoring protocol agently-skills.authoring.v2 and standard SKILL.md packages.
• The next development-line manifest is compatibility/in-development.json targeting 4.1.3.5.