[tbolis-at-mulesoft]

Closes #1126
Closes #1127
Closes #1128
Closes #1129
Closes #1130
Closes #1132
Closes #1139
Closes #1140
Closes #1141
Closes #1142

Summary

• Ships the nightly lock refresh (bot/lock-refresh, "chore(deps): refresh the PR CI lock"). The workflow pushed the branch but couldn't open the PR. It moves every pending framework release in python/constraints/:

  • agent-framework 1.20.0 → 1.21.0, with agent-framework-openai 1.16.0 (framework-release: agent-framework 1.20.0 → 1.21.0 #1126)
  • anthropic 1.11.0 → 1.13.0 (framework-release: anthropic 1.11.0 → 1.13.0 #1127)
  • langchain-core 1.6.7 → 1.6.10 (framework-release: langchain-core 1.6.7 → 1.6.10 #1128), langchain-openai 1.6.7 → 1.7.1 (framework-release: langchain-openai 1.6.7 → 1.7.1 #1129) and langgraph 1.2.14 → 1.2.15 (framework-release: langgraph 1.2.14 → 1.2.15 #1130)
  • strands-agents 1.58.0 → 1.60.0 (framework-release: strands-agents 1.58.0 → 1.60.0 #1132)
  • google-adk 2.11.0 → 2.12.0 (framework-release: google-adk 2.11.0 → 2.12.0 #1139) and litellm 1.104.0 → 1.105.1 (framework-release: litellm 1.104.0 → 1.105.1 #1140)
  • crewai 1.15.27 → 1.15.28 (framework-release: crewai 1.15.27 → 1.15.28 #1141)
  • openai 3.24.0 → 3.31.0 in the openai-3 locks (framework-release: openai 3.24.0 → 3.31.0 #1142)
  • llama-index-llms-openai 0.8.2 → 0.8.3 and llama-index-workflows 2.25.0 → 2.26.0 (no issue)

  Every combo ships; none is held back. The co-bumps worth a look: pydantic 2.14.1 / pydantic-core 2.50.1, multidict 7.1.0 (a new major) in the adk and agent_framework locks, and opentelemetry-* 1.45.1 / 0.66b1 in the agent_framework locks, which add opentelemetry-exporter-http-transport and -otlp-common. agent-framework-anthropic 1.0.0b261008 moves the agent_framework locks from anthropic 0.116.0 to 1.13.0. The adapter doesn't import anthropic, and the anthropic tests pass on 1.13.0 in that venv. The regenerated docs/framework-support.md, compatibility page and llms artifacts come with the bot commit.

• The bot commit moved ruff to 0.17.0 in the locks but left the ruff-pre-commit rev in .pre-commit-config.yaml at v0.16.10. test_pre_commit_runs_the_lint_jobs_tool_versions requires the two to match, so every session that runs the whole suite failed on that one test. This PR bumps both ruff hooks to v0.17.0 (the tag resolves to 44b5042e). Under ruff 0.17.0, ruff check and ruff format --check pass on the tree, and the 11 python/**/*.md files the hook formats are already formatted. No code changes.

• docs/verified-apis.md §8 records an offline re-check for each framework at its new version: signature check, the flagged release-note items checked against the installed source, and wheel diffs where the issue's review asked for them. The VERIFIED (LIVE) rows keep their 2026-10-13 live dates and status, because no --live run was made. The emitted UNVERIFIED (signature confirmed) rows were not pasted over them. Other changes in that file:

  • The LangGraph row's constructor and default now say use_responses_api=False, matching connection_kwargs() since feat(adapters): default agent-framework and LangGraph to Chat Completions #1043. The row had still said True.
  • The agent-framework middleware row moves to 2026-10-14 and cites framework-release: agent-framework 1.20.0 → 1.21.0 #1126.
  • The §8.1 "framework extras together" row now notes that agent-framework-anthropic accepts anthropic>=1.11,<2, so its anthropic<1 conflict no longer applies to agent_framework. I did not re-run coinstall_matrix.py, so its status and date stay as they were.

• Two findings from the re-checks, recorded in the ledger and not changed here:

  • Through the bridge, the openai 3.30 HTTP/1.1 stream drain after [DONE] finds the forwarded stream already consumed. The response is closed, not pooled. This is the same as on 3.24, so there is no regression, but bridged streams still get no connection reuse.
  • Strands 1.59.0 removed streamablehttp_client from its MCP examples. The §9 row that names it stays UNVERIFIED.

Test plan

Run from python/, on this branch.

• nox -s "adapter-contract(langgraph)": 2994 passed, 229 skipped; smoke example passed
• nox -s "adapter-contract(adk)": 2992 passed, 191 skipped; smoke example passed
• nox -s "adapter-contract(strands)": 3010 passed, 200 skipped; smoke example passed
• nox -s "adapter-contract(agent_framework)": 3090 passed, 160 skipped; smoke example passed
• nox -s "adapter-contract(openai-agents)": 2966 passed, 242 skipped; smoke example passed
• nox -s "adapter-contract(anthropic)": 2973 passed, 244 skipped; smoke example passed
• nox -s "adapter-contract(crewai)": 3010 passed, 215 skipped; smoke example passed
• nox -s "adapter-contract(llamaindex)": 2953 passed, 242 skipped; smoke example passed
• nox -s agent-framework-middleware (96 passed), llamaindex-transport (68 passed), agents-strands-last-call (73 passed), langgraph-demo (conformance 4/4), openai-1x (2963 passed, local_gateway 9 passed)
• python scripts/verify_frameworks.py --only <framework> --require-installed in each adapter-contract venv: signature ok and expected class ok on all 11 rows (live not run)
• nox -s typecheck-and-lint test-3.11 new-dependencies commit-identities docs-llms-drift:
  • typecheck-and-lint: mypy clean on 82 files; ruff 0.17.0 check and format pass; 9 import contracts kept
  • test-3.11: 2998 passed, coverage 95%
  • new-dependencies: no new direct dependencies
  • commit-identities: passes
  • docs-llms-drift: website tests 22/22, llms artifacts up to date
• python scripts/generate_docs.py: no diff on top of the bot commit
• python scripts/compile_constraints.py --check: every combo still resolves, but six files are stale against releases that came out after the nightly. Those are langchain-openai 1.7.2, pydantic-settings 2.16.0, authlib 1.9.1 and filelock 4.1.2, in langgraph-demo, adapter-contract-langgraph, -adk, -agent_framework, -crewai and agent-framework-middleware. They are not in this PR. The next nightly and framework-release-watch pick them up.
• nox (the full gate is left to CI: the test-3.10/3.12–3.14 legs, base-only, the stacks sessions, benchmark, quickstart, docs-site, pip-audit)
• verify_frameworks.py --live was not run. The live evidence for each row is still the 2026-10-13 round-trip.

ADR needed?

• No: this PR changes none of those.

Post-deploy steps

None. The locks are CI-only, and no floor in pyproject.toml changes. Optional: run verify_frameworks.py --live on google-adk 2.12.0 / litellm 1.105.1 and crewai 1.15.28. Their live rows still rest on google-adk 2.11.0 and crewai 1.15.27.

🤖 Generated with Claude Code