v0.8.2

This release adds bounded worker continuation, typed worker tools and reliable live child activity,
alongside Python and documentation dependency updates and provider SDK compatibility fixes.

Added

• Agent(yield_after_steps=N) and RunHandoff for explicitly continued, checkpointed work slices.
  Durable runs preserve committed actions, model history, correction allowances and cumulative budgets.
  RunResult.handoff exposes structured progress; resume_task(root_run_id) grants another bounded slice.
  Hard budget exhaustion, cancellation, denial and uncertain effects remain stopping conditions.
• Agent.add_worker_tool(worker, tool=None, name=None, description=None) exposes an owned local worker
  with a small task: str schema or the selected child tool's typed arguments. Calls use native delegation,
  inherited policies, shared budgets and existing execution receipts.
• A detailed worker execution guide and runnable, offline
  example.

Changed

• Refresh Python and documentation dependency lockfiles, update Docusaurus to 3.10.2 and apply compatible
  security updates to the documentation worker pool and JavaScript serializer.
• Support Langfuse SDK 5 alongside SDK 4, and run real provider-SDK compatibility checks in CI for
  locked, minimum-direct and latest dependencies.

Fixed

• Route explicit no-change ToolRejectedError failures during tool preparation through the existing
  bounded correction loop, before approval or execution. Rejected preparation consumes no tool call.
• Forward owned local child events immediately to the public stream, including nested and externally
  spawned children, with stable correlation identities and one root terminal event.
• Support Anthropic SDK v1 across text, native-tool and streaming requests. Only max_tokens=1024 is
  supplied by default; explicit legacy sampling settings are preserved through extra_body for models
  that support them. See AnthropicLLM parameters.
• Propagate failed or incomplete OpenAI responses, honor Gemini's configured API URL, and reject
  multiple parallel native tool calls instead of silently selecting one.
• Isolate LangSmith trace state between tracker instances and restore parent runs after nested tasks.
• Count literal tokenizer control strings as ordinary prompt text instead of falling back to the
  character-based token estimate.