python/v1.62.0
Auto-drafted from commits in python/v1.61.1..python/v1.62.0, 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(decisions): add decision usage to the request's invocation context (#5304) (5adc6a6)
- feat(python): add experimental OpenAI Live provider (#5286) (420cedc)
- feat(skills): add docs-verifier skill (#5268) (0a99337)
- feat(agent): enforce invoke limits and usage accounting across the whole agent tree (#4667) (6992d10)
- feat(vended-tools)!: add swarm tool in TypeScript (#5204) (78847d0)
- feat(changelog): add harness, CLI, and box release streams (#5192) (e67dda5)
- feat(stop): use agent.cancel() to stop the event loop (Python) (#4833) (cf0c6ae)
- feat(vended-tools): add swarm tool in Python (#4922) (b85c373)
- feat(middleware): make the Python middleware system public (#4976) (1632430)
- feat(telemetry): port baggage propagation to TypeScript (#5073) (fbc6977)
- feat(catalog): add attenu-guard (community intervention) (#3998) (a2c4fb4)
🐛 Fixes
- fix(ts): prevent text duplication across streamed content blocks (#5308) (774351c)
- fix(context): render tool calls as text in summarizer input (#5292) (b4c2054)
- fix(python): render tool calls as text in summarizer input (#5293) (6a077ea)
- fix(memory): run Bedrock knowledge base calls off the event loop (#5053) (c6bfe94)
- fix(deps): cap litellm below 1.107.0 (#5307) (f32ff21)
- fix(memory): always append the JSON output format to ModelExtractor prompts (#5269) (f0fd304)
- fix(cli): add GPT-6 Sol, Luna and 6.1 Sol to the fallback context windows (#5215) (f760f3a)
- fix(models): add context window limits for GLM-5, 5.3 and 4.7 Flash (py + ts) (#5301) (3c1291f)
- fix(harness): inject memory on user turns only, matching the SDK default (#5298) (2602fb1)
- fix(site): serve readable content at legacy docs URLs (#5155) (b412adc)
- fix: fix test flakiness caused by set_tracer_provider pollution (#5270) (b429fa1)
- fix(anthropic): drop unsigned reasoning instead of raising KeyError (#5247) (3fcf7e6)
- fix(multiagent): inherit the parent's interventions in the default subagent builder (#5200) (5f630db)
- fix(python): let providers enforce structured output when a model cannot be forced (#5190) (79b2d33)
- fix(multiagent): add DISPATCHED status to prevent fan-in double-execution on resume (#3633) (#3724) (2ea678a)
- fix(sandbox): run the default local environment's commands on Windows (#5244) (04b7cf6)
- fix(bedrock-ts): omit empty citation source when replaying citations (#5266) (1f52310)
- fix(python): preserve tool result adjacency in mistral, writer and llamacpp (#5248) (cbec09b)
- fix(bedrock): drop reasoning blocks the target model rejects (TypeScript) (#5246) (28cc479)
- fix: replace retired Claude Sonnet 4 model IDs with Sonnet 4.6 (#5241) (6b0368f)
- fix(memory): delegate FileMemoryStore search, scope storage namespaces (#5117) (207b206)
- fix(models): retry with drop_block when a history edit invalidates replayed thinking (#5171) (25c2460)
- fix(ts): report correct paths for values after nested objects in deepCopyWithValidation (#4981) (bd0b6a9)
- fix(harness): enable Bedrock Mantle web search for GPT-6.1 Sol (#5188) (26700e6)
- fix(sagemaker): keep assistant text that sits beside tool calls (#4919) (8909ec4)
- fix(harness): make the file tools work on Windows and run Windows CI for harness-ts, harness-py and strands-cli (#4978) (4ab2303)
- fix(cli): trim comma-separated flag values (#5183) (c1487af)
- fix(sagemaker): invoke endpoints off the event loop (#5034) (e6cfa5c)
- fix(models): add context window limits for GPT-6 Sol, Luna and 6.1 Sol (py + ts) (#5147) (02a83fc)
📚 Documentation
- docs(memory): add file memory store docs (#4279) (2afa133)
- docs(agent-loop): recommend invocation limits for long-running agents (#5295) (79e0f89)
- docs(evals): update red-teaming guide for graduation and attacker-refusal detection (#5234) (9a125e8)
- docs(agents): clarify lifecycle controls and recovery (#5153) (95381d8)
- docs: clarify TypeScript integration in the README (#5154) (0788ccb)
- docs(blog): add co-authors to StructuredOutputSimilarity post (#5285) (24c814f)
- docs(designs): add 0021 cache-friendly tool management (#5161) (823eabf)
- docs(changelog): backfill harness, CLI, and box releases (#5193) (8a59681)
- docs(blog): add StructuredOutputSimilarity launch post (#5256) (6b6bc9c)
- docs: add agent foundation migration paths (#5152) (3bd98e6)
- docs(learning): clarify course updates and optional additions (#5095) (038f9d1)
- docs: split structured output into a build guide and a component page (#5089) (f63c143)
- docs(evals): add StructuredOutputSimilarity evaluator page (#4464) (6d34195)
- docs(evals): fix and refine getting-started examples and description and drop redundant Overview headings (#5195) (287ec07)
- docs: revise custom-tools page (#5088) (1f4b6ab)
- docs(catalog): add LangWatch observability integration (#5167) (d738205)
- docs: fix typos in comments and docstrings (#4997) (c91524e)
✅ Tests
- test(litellm): skip flaky multi-turn cache read test (#5314) (7a0f7ba)
- test(bedrock): skip audio integ tests when Bedrock rejects audio input (#5310) (5021853)
👷 CI
- ci(docs): bump @astrojs/markdown-remark from 7.3.1 to 7.3.3 in /site (#5229) (1adcaf9)
- ci(docs): bump satori from 0.47.1 to 0.52.0 in /site (#5189) (7939e67)