✦ Features 12
- breakwrite injected text into history as an InjectedBlock (typescript)contextmodel
- breakwrite injected text into history as an injected block (python)contextmemory
- organize sample guides by user goal
- server-side summary capability with Anthropic on-demand compactioncontextmodel
- accept pre-built vendor clients in the Anthropic and OpenAI providers (Python + TypeScript) and document Bedrock InvokeModel / Mantle through themmodel
- make the shell tool's default timeout configurabledevxtool
- give Model.summarize() a client-side base implementationcontextmodel
- give Model.summarize() a client-side base implementation (TS port of #5962)contextmodel
- let ClassifierStrategy route with any DecisionModelmultiagentdecisions
- add "auto" mode to HumanInTheLoophildecisions
- handle Bedrock InvokeModel in AnthropicModel (mid-stream errors, API version); add defaultEffort to the shared model info (py + ts)model
- async agent builders; report content-filtered and refused children as tool errorsasyncmultiagent
⊘ Fixes 25
- snapshot upstream content at deploy
- accumulate metrics for hook retriesotelhooks
- support anthropic SDK 1.x and 2.x in AnthropicModelmodel
- restore a pending tool execution only for an activated state, and validate it (Python)hilpersistence
- add GPT-6 Cyber context window (py + ts + cli)model
- apply the 2026-10-21 restrictions only from that API version on (Python)model
- convert SageMaker tool results to user messagesmodel
- report directories and missing paths distinctly in PosixShellSandbox read_file (Python + TypeScript)server
- summarizer truncation fallback also shrinks text blocks under 3 × preserve_charscontext
- warn when structured output drops non-text system prompt blocksdevxstructured-output
- reject None interrupt responses instead of silently re-interruptingdevxhil
- mirror anthropic 2.7.0 enum/const schema handling (python)modelstructured-output
- include shell releases in the changelog
- send tool result images and documents as mediamodel
- enforce tool_filters on call_tool (python)
- add Claude Fable 5.2 and Mythos 5.2 model info (py + ts + cli)model
- used strands' tracer and use span_attributesoteldecisions
- preserve same-index signed reasoning blocksmodel
- retry with drop_block when no thinking config is setmodel
- retry a 503 UNAVAILABLE as throttling (#5916)model
- show the output schema to Ollama and llama.cpp models in structured_outputmodelstructured-output
- send Claude Haiku 5.5 structured output to native outputmodelstructured-output
- suppress OpenSSH 10.1+ post-quantum warning on stderr
- bump mistral-large-latest context window to 1,048,576model
- Report tool usage for streamed calls finishing with stop reasonmodel
⋯ Other 13
- add glama.json to claim the MCP server on Glamamcp
- clarify that inputTokens may include cached tokens depending on providermodel
- cover RateLimitError mapping to ModelThrottledException
- parametrize format_chunk content_block_delta cases
- cover RateLimitError mapping to ModelThrottledException
- make the duplicate redacts_output test enable output redaction
- read the summary block from SummarizeResult in the Anthropic Bedrock testcontextmodel
- expect temporary messages for Claude Opus 5.5 in the model info lookup testmodel
- bump actions/setup-python from 7.0.0 to 7.1.0
- expect the new mistral-large-latest context windowmodel
- record the process list before a hung Windows test run is killedcli
- cap the test job at 30 minutes and tolerate artifact upload failures
- update mistral-large-latest context window assertionsmodel