github strands-agents/harness-sdk harness-cli/v0.4.0

latest release: harness-typescript/v0.5.0
one hour ago

harness-cli/v0.4.0

Auto-drafted from commits in harness-cli/v0.3.0..harness-cli/v0.4.0, grouped by conventional-commit type. Edit on the release page after publish if you want a polished writeup; the canonical release notes live on the website.

🚀 Features

  • feat(catalog): add strands-acgs (#5259) (9faf8a4)
  • feat(tools): allow sorting tool specs (#5933) (543affe)
  • feat(todos): add reminder interval (#6239) (2e33620)
  • feat(context)!: offload conditions decide granularity (python) (#6101) (84d814b)
  • feat(external-docs): mdx passthrough and nested sidebar groups (#6225) (f1f20bc)
  • feat(site): publish OpenAPI metadata for agents (#6309) (6e1d55b)
  • feat(interventions): add a Dogwood authorization handler (typescript) (#6087) (4df8537)
  • feat(harness): serve web_search from AgentCore InvokeWebSearch on Bedrock (py + ts) (#5997) (dd92a16)
  • feat(background-tasks): allow configuring the management tool name (#6178) (143b3b6)
  • feat(harness)!: default to Claude Opus 5.5 on Anthropic's Bedrock client (#4683) (5a3d25d)
  • feat(decisions): add SageMakerDecisionModel for System One models on SageMaker AI endpoints (Python) (#6085) (b6beefb)
  • feat(ts): add ModelInfo.merge; document model info in the provider guides (#6145) (46cbb8c)
  • feat: vend the todos plugin in the SDK and deprecate the harness built-in (#6084) (76f2f44)
  • feat(models): default Anthropic max_tokens to the model info output limit; raise when it has none (Python) (#5698) (9b1d9f3)
  • feat(mcp-py): add auto_reconnect to MCPClient for dropped sessions (#5297) (d5f4da6)
  • feat(multiagent): async agent builders; report content-filtered and refused children as tool errors (#6093) (3d741b4)
  • feat(site): add /frogs page (#6095) (9b38e2e)
  • feat(context)!: offload conditions decide granularity (typescript) (#5325) (50a391e)
  • feat(mcp): add find_integrations tool backed by the catalog feed (#5878) (085d4de)
  • feat(models): handle Bedrock InvokeModel in AnthropicModel (mid-stream errors, API version); add defaultEffort to the shared model info (py + ts) (#6064) (860c07c)
  • feat(injection)!: write injected text into history as an injected block (python) (#5977) (228990b)
  • feat(harness)!: drop todos from the default built-in plugins (#5974) (23de8a8)
  • feat(site): add Decider product navigation (#6036) (39de68e)
  • feat(hitl): add "auto" mode to HumanInTheLoop (#5975) (edfb509)
  • feat(decisions): add SageMakerDecisionModel for System One models on SageMaker AI endpoints (TypeScript) (#5890) (f8dda55)
  • feat(python): let ClassifierStrategy route with any DecisionModel (#5995) (7c35bcd)
  • feat(models): give Model.summarize() a client-side base implementation (TS port of #5962) (#6022) (c1a689e)

🐛 Fixes

  • fix(models): cache the system prompt on GPT-6.2 through the Responses API (#6396) (d2c4e70)
  • fix(models): place the Vercel conversation cache point on tool-result turns in a tool loop (#6199) (4d5b88e)
  • fix(decisions): reject binary media nested in TypeScript content (#6306) (794d37c)
  • fix(decisions): preserve prototype-named TypeScript result keys (#6305) (685797f)
  • fix(models): stream Ollama thinking and skip reasoningContent in requests (#6295) (0d22b9f)
  • fix(telemetry): export reasoning text as a reasoning part under the latest GenAI conventions (python) (#6378) (450976a)
  • fix(models): add Gemini 4 Neon context window (py + ts) (#6364) (8508b49)
  • fix(decisions): reject invalid Python choice containers and elements (#6304) (54d74bf)
  • fix(decisions): omit nonfinite TypeScript provider token usage (#6308) (ccf608d)
  • fix(decisions): normalize TypeSafe context overflow in both SDKs (#6307) (1b0cdc5)
  • fix(telemetry): record tool definitions on the agent span in TypeScript (#5439) (09fb8fd)
  • fix(web-fetch): escape body text, the page title, and link titles and hrefs in extracted markdown (#6310) (b47020e)
  • fix(models): let Nova read the prompt cache across user turns (Python and TypeScript) (#6292) (e1ddaf8)
  • fix(google): reference tool-result media parts from the function response output (typescript) (#6017) (b458767)
  • fix(bash): preserve UTF-8 across output chunk boundaries (#5067) (7fe5bc1)
  • fix(openai): let chat params omit stream_options with null (#5018) (5d742a0)
  • fix(models): validate Bedrock guardrail configuration (#4223) (0943408)
  • fix(strands-ts): keep reasoning text in trace output (#5575) (0918cae)
  • fix(openai): make tools work on Chat Completions for GPT-5.6 and GPT-6 models (py + ts) (#6291) (e912a00)
  • fix(multiagent): forward structured output as a JSON TextBlock to downstream graph nodes (#6070) (e7913fd)
  • fix(web-fetch): escape image alt text and src in extracted markdown (#6282) (26c8eed)
  • fix(hooks): await callable object callbacks (#6220) (281b4c3)
  • fix(models): add GPT-6.2 Astra and Sol model info (py + ts) (#6265) (a4c8c9c)
  • fix(models): send injected text to Claude Haiku 5.5 as plain text (#6284) (92ebb74)
  • fix(web-fetch): strip script and data: URLs the same way in Python and TypeScript (#6253) (c8dd322)
  • fix(interventions): track CedarAuthorization call counts per agent (Python and TypeScript) (#6279) (a34db20)
  • fix: redact the last user message when a guardrail blocks input (Python and TypeScript) (#6261) (f956b37)
  • fix(tools): skip tool start when cancel arrives during BeforeToolCall hooks (#6255) (81cd69e)
  • fix(session): read session files with O_NOFOLLOW and warn on a wide storage directory (#6254) (2bedc19)
  • fix(event_loop): retry a tool_use stop with no toolUse block, then raise (#6218) (d2c9f98)
  • fix(decisions): read decisionapi.net's enveloped response (#6191) (216e4b1)
  • fix(context): count the whole request in utilization gates; report retrieval failures as errors (#6030) (81e6d5c)
  • fix(mcp-py): state that auto_reconnect does not restart stdio servers (#6198) (c06c95f)
  • fix(structured_output): keep every branch of union fields in tool specs (#6144) (f2b4dd6)
  • fix(models): resolve Bedrock auto caching for Nova, OpenAI GPT, and xAI (Python) (#6149) (f8ef261)
  • fix(models): resolve Bedrock auto caching for Nova, OpenAI GPT, and xAI (#6082) (42c4e82)
  • fix(models): add Gemini 4 Hydrogen context window (py + ts) (#6112) (df3d0c7)
  • fix(background-tasks): persist results that contain bytes (#6045) (c2e3e28)
  • fix(ts): replace lone surrogates before sending model requests (#5939) (c17c933)
  • fix(harness): give the sandbox shell probe 30 seconds (#6128) (7f860e0)
  • fix(models): bump mistral-large-latest context window to 1,048,576 (#6104) (ad1ae00)
  • fix(sandbox): suppress OpenSSH 10.1+ post-quantum warning on stderr (#6069) (b48c73c)
  • fix(models): mirror anthropic 1.7.0 enum/const schema handling (typescript) (#6016) (5f88066)
  • fix(models): send Claude Haiku 5.5 structured output to native output (#6073) (b512a53)
  • fix(models): show the output schema to Ollama and llama.cpp models in structured_output (#6000) (3bee385)
  • fix(agent): keep other sources' tools when an MCP server refreshes its list (#6055) (4330e22)
  • fix(gemini): retry a 503 UNAVAILABLE as throttling (#5916) (#5947) (bd00d2d)
  • fix(models): retry with drop_block when no thinking config is set (#6061) (97bceac)
  • fix(streaming): preserve same-index signed reasoning blocks (#5708) (b0a1740)
  • fix(decision): used strands' tracer and use span_attributes (#6059) (dc6a377)

📚 Documentation

  • docs(catalog): add ClickHouse Loom memory store to the integrations catalog (#6114) (e1185fd)
  • docs: say when the structured output retry isn't forced; scope forcedToolChoice to what Strands does (#6086) (ac6852a)
  • docs(changelog): sync release notes (#6235) (af90026)
  • docs(multiagent): say the default builder does not inherit the parent's hooks or plugins (#6281) (9cba751)
  • docs(harness): say session files are trusted input and add .gitignore guidance (#6280) (ca4b18f)
  • docs(mongodb): add memory launch post and catalog entries (#5823) (73fb693)
  • docs(decision-models): record questions and answers as span attributes (#6123) (2ea1255)
  • docs(changelog): sync release notes (#6188) (1e43962)
  • docs(routing): route with a decision model and use routing_policy (#6124) (0e38c45)
  • docs(changelog): sync release notes (#6106) (d982a0d)
  • docs(site): clarify multi-agent pattern decision process (#5553) (259619a)
  • docs(changelog): sync release notes (#6031) (770a002)
  • docs(streaming): add browser streaming guide with SSE (#5544) (7b651c3)
  • docs(models): add multi-tenant model routing guide (#5842) (687ce5d)
  • docs: add a standalone Strands FAQ (#6023) (d0563b8)
  • docs: add decision models section (#6002) (de8ae22)
  • docs: explain which context strategies keep thinking on models that preserve thinking (#6026) (7296cab)
  • docs(hitl): fix decision model risk classifier examples (#6001) (045697b)

✅ Tests

  • test(cli): read the region from Anthropic's Bedrock client in the Python worker tests (#6366) (078cbc2)
  • test(interventions): cover CedarAuthorization call counts through the subagent tool (#6290) (69c3314)
  • test(ts): load typesafe and jev keys for integ tests (#6192) (048ef9d)
  • test(bedrock): fold redaction event generation tests into it.each (TS) (#6183) (094386a)
  • test(openai): fold chat usage metadata tests into it.each (TS) (#6184) (d201ef3)
  • test(openai): fold chat error classification tests into it.each (TS) (#6185) (80f6b06)
  • test(anthropic): cover max_tokens and stop_sequence stop reasons (TS) (#6182) (cb475e9)
  • test(mistral): cover streamed tool calls (#6181) (462512f)
  • test(models): update mistral-large-latest context window assertions (#6133) (74e9953)
  • test(cli): give the subagent follow-up test 20 seconds (#6129) (4bb096b)
  • test(models): expect the new mistral-large-latest context window (#6122) (acbcbf3)
  • test(harness-ts): give real-shell tests room for a slow PowerShell start (#6127) (edd8e42)
  • test(models): expect temporary messages for Claude Opus 5.5 in the model info lookup test (#6092) (07b8a0b)
  • test(integ): read the summary block from SummarizeResult in the Anthropic Bedrock test (#6032) (b17ae40)

👷 CI

  • ci: run the SDK and harness suites when shared/ changes (#6135) (c8e1351)
  • ci(cli): split the Windows test leg into two shards (#6132) (c7838ee)
  • ci(typescript): cap the test job at 30 minutes and tolerate artifact upload failures (#6130) (7733955)
  • ci(cli): record the process list before a hung Windows test run is killed (#6131) (7f75a32)
  • ci: bump actions/setup-python from 7.0.0 to 7.1.0 (#6047) (1060cce)

🔧 Chores

  • chore(model-info): record that Magistral Small cannot call tools on Bedrock (#6328) (60d1d1d)
  • chore(site): refresh catalog stats (#6168) (09aa5ba)

🔖 Other

  • design(site): make Harness the primary toolkit path (#6148) (0951db9)
  • Report tool usage for streamed calls finishing with stop reason (#6049) (bea59d9)

Don't miss a new harness-sdk release

NewReleases is sending notifications on new releases.