Provider management, Skill Runtime and governed interactions

Changes since v1.0.0

Added

Centralized provider management with configuration, testing, authentication,
model discovery, saved model selection and per-session runtime overrides.
Direct OpenAI Responses API support, including tool calls, multimodal input,
reasoning effort and integration with the Forge verifier.
Requesty, Cheaper Inference and API Route as OpenAI-compatible gateways.
directory_tar_gz Forge Node packages for executables with bundled runtime
dependencies and validated internal relative symbolic links.
An opt-in interaction foundation connecting AgentTask, Session, multi-step
decisions, independent root verification and one deduplicated experience
episode, backed by frozen bindings, a persistent Runner and a Supervisor.

Skill profiles support startup timeouts and required_tools overrides; startup
hooks receive profile, Python and Skill context for LIBERO 0.3.4 prerequisites.
Custom and Responses providers support configurable timeouts and retries,
with improved max_completion_tokens compatibility.
Failed model completions return explicit errors while retaining user input
and completed tool exchanges without storing error text as assistant history.
Ordinary and interactive tasks reject new Query, Action and Session intents
after cancellation is persisted or the task leaves the executing state.

Preserve valid cron jobs when individual records are malformed; quarantine
unreadable stores and refuse to overwrite them if quarantine fails.
Parse Dora JSON and NDJSON status output and inject profile environments into
rendered dataflow nodes to avoid stale flow and daemon-environment effects.
Improve archive path containment, symbolic-link validation, image-path cleanup
and compatibility across providers.
Pin Runtime identity for interaction dispatch, discard late decisions after
cancellation, prevent automatic retransmission of uncertain operations and
retain resource guards until outstanding execution is reconciled.
Expand regression coverage and build checks; improve bilingual documentation,
Piper and LIBERO-10 quickstarts, and benchmark scope and result reporting.

Built and verified by GitHub Actions in an Ubuntu 20.04 container:

Install the attached wheel with:

pip install ./phyagentos_ai-1.1.0-py3-none-any.whl