v0.1.5

What's new since v0.1.4

• Added an opt-in Responses requester lane for provider bridge users while keeping chat completions as the default.
• Added Dynamic DAG preview support: TaskDAG-like plans can compile into runtime-owned DynamicWorkflowPlan nodes and execute through registered capabilities.
• Added richer workflow lifecycle evidence, including lifecycle state, execution id, state version, intervention metadata, close reason, workflow events, and UI event projection.
• Added neutral recall context packs and context record references so downstream code can consume recall evidence without depending on provider-native workspace objects.
• Added action artifact evidence references using artifact summary metadata without reading raw artifact bodies.
• Improved runtime bridge evidence for usage, provider transport, request id, model attribution, tool approvals, and streamed tool-call normalization.
• Added new runnable examples for Responses bridge, Dynamic DAG preview, TriggerFlow lifecycle, recall context packs, action artifact evidence, and a live runtime bridge showcase.

v0.1.4

Highlights

• Added multimodal support for precomposed_messages content parts.
• Supported stable v1 input parts now include text and OpenAI-compatible image_url parts.
• Multiple images and image-only prompt content lists are supported through the runtime boundary.
• Added prompt evidence summaries for multimodal requests without recording prompt text, media URLs, base64 payloads, or full messages.
• Documented multimodal usage, evidence behavior, and troubleshooting guidance in the public help docs.

Compatibility

• Existing text-only precomposed_messages remain compatible.
• Existing public runtime entry points and result/artifact contracts are unchanged.
• Multimodal binary outputs continue to use existing artifact locator fields.

Package

• PyPI: capability-runtime==0.1.4

v0.1.3.post1

What's Changed

• Treats OpenAI-compatible parser provider values as placeholders when supplemental gateway/provider metadata is available.
• Preserves the effective provider in usage reports for observability, tracing, and reconciliation.
• Keeps upstream token usage as the source of truth to avoid double-counting.
• Retains the upstream placeholder provider in usage events for auditability.

v0.1.3

What's Changed

• Preserves supplemental provider and gateway usage metadata when upstream LLM usage is already available.
• Avoids double-counting tokens by keeping upstream usage as the source of truth.
• Hardens runtime result boundaries to avoid mutating terminal reports across stream and structured output paths.
• Improves prompt-control handling and message preview safety, including stronger secret redaction and HTML escaping.
• Adds regression coverage for usage metadata merging, prompt hardening, result immutability, and event callback error tracking.

v0.1.2

What's Changed

• Exposed provider request metadata on NodeReport usage: optional request_id and provider fields are now projected for host observability, reconciliation, and tracing.
• Preserved provider request metadata from Agently/SSE completed usage payloads when available, while keeping usage generation compatible when providers omit it.
• Kept multi-call usage token aggregation unchanged; request metadata follows the latest non-empty provider metadata observed in the node.

v0.1.1

Highlights

• Added per-agent and per-run prompt rendering strategies for capability-runtime agents.
• Added direct_task_text for hosts that already compose final task text.
• Added precomposed_messages for hosts that need to provide final provider messages directly while still using Runtime events, NodeReport, WAL, and validation.
• Added prompt_profile pass-through for SDK prompt profiles such as generation_direct and structured_transform.
• Upgraded the pinned skills-runtime-sdk dependency to 0.1.12.

Install with:

pip install capability-runtime==0.1.1