chore(deps): bump anthropic from 0.109.2 to 0.120.2

[dependabot]

Bumps anthropic from 0.109.2 to 0.120.2.

Release notes

Sourced from anthropic's releases.

v0.120.2

0.120.2 (2026-07-28)

Full Changelog: v0.120.1...v0.120.2

Bug Fixes

• mcp: support mcp sdk v2 alongside v1 (#300) (177f88c)

v0.120.1

0.120.1 (2026-07-28)

Full Changelog: v0.120.0...v0.120.1

Bug Fixes

• mcp: pin mcp extra to <2 (#1783) (fb66371)

v0.120.0

0.120.0 (2026-07-24)

Full Changelog: v0.119.0...v0.120.0

Features

• api: add claude-opus-5 model (bf4e31c)
• api: add tool addition/removal blocks and tool_change events (bf4e31c)
• api: expand client-side fallback credit token types and add server-side fallbacks default option (bf4e31c)

v0.119.0

0.119.0 (2026-07-23)

Full Changelog: v0.118.0...v0.119.0

Features

• api: add new stop reason 'model_context_window_exceeded' (d983cde)

Bug Fixes

• tools: handle binary files in agent toolset read/edit (#283) (417b76b)

v0.118.0

0.118.0 (2026-07-22)

Full Changelog: v0.117.1...v0.118.0

Features

... (truncated)

Changelog

Sourced from anthropic's changelog.

0.120.2 (2026-07-28)

Full Changelog: v0.120.1...v0.120.2

Bug Fixes

• mcp: support mcp sdk v2 alongside v1 (#300) (177f88c)

0.120.1 (2026-07-28)

Full Changelog: v0.120.0...v0.120.1

Bug Fixes

• mcp: pin mcp extra to <2 (#1783) (fb66371)

0.120.0 (2026-07-24)

Full Changelog: v0.119.0...v0.120.0

Features

• api: add claude-opus-5 model (bf4e31c)
• api: add tool addition/removal blocks and tool_change events (bf4e31c)
• api: expand client-side fallback credit token types and add server-side fallbacks default option (bf4e31c)

0.119.0 (2026-07-23)

Full Changelog: v0.118.0...v0.119.0

Features

• api: add new stop reason 'model_context_window_exceeded' (d983cde)

Bug Fixes

• tools: handle binary files in agent toolset read/edit (#283) (417b76b)

0.118.0 (2026-07-22)

Full Changelog: v0.117.1...v0.118.0

Features

• api: add support for Managed Agents model effort, initial session events, and threads delta streaming (712bc6f)

0.117.1 (2026-07-21)

Full Changelog: v0.117.0...v0.117.1

... (truncated)

Commits

• f5c30d0 release: 0.120.2
• 8b327c3 fix(mcp): support mcp sdk v2 alongside v1 (#300)
• 3b65b23 release: 0.120.1
• 6c127ae fix(mcp): pin mcp extra to <2 (#1783)
• 36863a0 codegen metadata
• 60c64fb release: 0.120.0
• 70c0a64 feat(api): add claude-opus-5 model
• eb178ec release: 0.119.0
• 6914718 feat(api): add new stop reason 'model_context_window_exceeded'
• d2bad4f fix(tools): handle binary files in agent toolset read/edit (#283)
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[claude]

🔍 Dependency Analysis Summary

Package	Old	New	Type	Ecosystem
anthropic	0.109.2	0.120.2	11 minor releases (0.x)	Python / uv

Scope: uv.lock only — a 3-line diff touching just the anthropic package block. pyproject.toml declares anthropic unpinned, so no constraint change is needed. No transitive dependencies changed.

Overall risk: 🟢 LOW

The reason risk is low despite spanning 11 minor releases: the direct use of the Anthropic SDK in OCS is four exception classes. Everything else goes through langchain-anthropic, which is unchanged in this PR.

---

📋 Detailed Changelog Review

anthropic-sdk-python 0.109.2 → 0.120.2 (full compare)

Version	Date	Highlights
0.120.2	2026-07-28	fix(mcp): support MCP SDK v2 alongside v1
0.120.1	2026-07-28	fix(mcp): pin mcp extra to <2
0.120.0	2026-07-24	claude-opus-5 model; tool addition/removal blocks + tool_change events; expanded client-side fallback credit token types
0.119.0	2026-07-23	New stop reason model_context_window_exceeded; binary-file handling in agent toolset
0.118.0	2026-07-22	Managed Agents model effort, initial session events, threads delta streaming
0.117.1	2026-07-21	AnthropicAWS.copy() credential fix; new refusal category
0.117.0	2026-07-16	Dreaming support; MCP Tunnels; credential material kept out of traceback frames via SecretStr
0.116.0	2026-07-02	agent-memory-2026-07-22 beta header
0.115.1	2026-07-01	chore: removed nonfunctional types from SDK
0.115.0	2026-06-30	Managed Agents: event delta streaming, overrides, reverse pagination, vault credential scoping, webhook events
0.114.0	2026-06-30	claude-sonnet-5 model; absolute path resolution in agent toolset
0.113.0	2026-06-29	20260318 web fetch/support tools; async count_tokens output_config merge fix
0.112.0	2026-06-24	system.message streaming events; memory tool directory permissions fix
0.111.0	2026-06-18	Refusal-fallback middleware request tagging
0.110.0	2026-06-18	code_execution_20260120 tool; header-merge and Bedrock stream-event fixes

Breaking changes: None declared in any release in the range. Nearly all additions are gated behind Managed Agents / Threads / MCP / agent-toolset APIs that OCS does not touch.

Security-relevant: 0.117.0 stops credential material from appearing in traceback frames (SecretStr). Small but real hardening for a Django app that ships tracebacks to error reporting.

The only removal-shaped item is 0.115.1 "removed nonfunctional types from SDKs" — verified below that it removed nothing OCS or langchain-anthropic references.

---

⚠️ Impact Assessment

Breaking changes found: No. Verified against the installed 0.120.2, not just inferred from the changelog:

1. The direct SDK surface is one file. apps/service_providers/llm_service/retry.py:15 is the only import anthropic in the repo. It uses exactly four symbols, all confirmed present on 0.120.2:

• anthropic.RateLimitError, anthropic.InternalServerError, anthropic.APIConnectionError, anthropic.APITimeoutError

2. The real consumer is langchain-anthropic 1.4.6, which is unchanged by this PR.

• Its constraint is anthropic>=0.96.0,<1.0.0 — 0.120.2 satisfies it.
• I enumerated every anthropic.* symbol it references and confirmed each resolves on 0.120.2: Anthropic, AsyncAnthropic, Client, AsyncClient, BadRequestError, DefaultHttpxClient, DefaultAsyncHttpxClient, types.RawMessageStreamEvent, and the top-level transform_schema.

3. uv lock --check passes — lockfile is consistent with pyproject.toml.

4. The MCP extra churn (0.120.1/0.120.2) does not apply here. OCS declares plain "anthropic" with no extras, so the anthropic[mcp] pin never enters resolution. The mcp package in uv.lock comes from a different dependency and is untouched by this diff.

Watch items (not blockers, no action required for merge)

These are behaviours the new SDK makes possible that existing code handles by silently ignoring. Worth knowing about; none of them break on merge.

• New content block types (0.120.0 tool addition/removal blocks). parse_output_for_anthropic in apps/service_providers/llm_service/parsers.py:104-115 falls through to LlmChatResponse(text="") for any unrecognised dict block. That is a safe degrade — no crash — but if these blocks ever reach the parser their content is dropped rather than surfaced.
• New stop reason model_context_window_exceeded (0.119.0). Nothing in OCS branches on stop_reason (apps/api/openai.py only declares a static OpenAI-compat finish_reason field). Practical effect: a context overflow now arrives as a normal completion with truncated content instead of an error, so it will not be distinguishable from a normal short answer.
• Expanded token-detail keys (0.120.0 fallback credit token types). apps/service_providers/tracing/metrics.py:24 hardcodes _CACHE_READ_KEYS = ("cache_read", "flex_cache_read", "priority_cache_read"). Any new cache-read alias would be billed at the full input rate in cost reporting. This is latent, not active — it would require langchain-anthropic to start surfacing a new key, and that package is not changing here.
• New models claude-sonnet-5 (0.114.0) and claude-opus-5 (0.120.0) are now known to the SDK. Model availability in OCS is DB-driven via LlmProviderModel, so exposing them is a separate data change, not part of this PR.

Affected files

None require changes. Files reviewed for impact:

• apps/service_providers/llm_service/retry.py — only direct SDK import ✅
• apps/service_providers/llm_service/main.py (AnthropicLlmService, prompt caching middleware, built-in tools) ✅
• apps/service_providers/llm_service/parsers.py ✅
• apps/service_providers/tracing/metrics.py ✅

Test impact

No test updates expected. The Anthropic-related tests operate on LangChain message shapes, not raw SDK objects: test_output_parsing.py, test_agent_tool_parsing.py, test_llm_credentials.py, apps/pipelines/tests/test_llm_node.py.

Configuration changes

None.

---

🛠️ Recommendations

Action required: None.

⚠️ Disclosure on verification depth: I could not execute the test suite in this session — the sandbox blocked both setting SECRET_KEY in the environment and creating a .env, and I could not read the CI check results for this PR. My verification was static analysis plus live symbol resolution against the installed 0.120.2 (imports, exception classes, the full langchain-anthropic symbol surface, uv lock --check). Please confirm CI is green before merging — I have not independently seen it pass.

Testing focus (if doing manual verification beyond CI):

1. An Anthropic-backed pipeline run end-to-end, checking Trace/Span records — confirms streaming and output parsing still assemble correctly.
2. A response with web-search citations — exercises the citation-append path in parse_output_for_anthropic.
3. Token/cost attribution on an Anthropic call with prompt caching enabled — confirms input_token_details keys still land in the expected buckets.

Follow-up tasks (separate PRs, not this one):

• Consider adding claude-sonnet-5 / claude-opus-5 to LlmProviderModel data if the team wants them available.
• Consider making _CACHE_READ_KEYS in metrics.py resilient to unknown cache-read aliases so future SDK token-detail additions do not silently inflate cost reports.

Merge recommendation: ✅ APPROVE (conditional on CI passing, which I was unable to verify myself)

---

📚 Useful Links

• anthropic-sdk-python CHANGELOG
• Full diff v0.109.2...v0.120.2
• anthropic on PyPI
• langchain-anthropic on PyPI (pinned 1.4.6, requires anthropic>=0.96.0,<1.0.0)

_{🤖 Automated Dependabot review by Claude}