fix(anthropic_messages): forward named params into MessagesIntercepto…

[samagana]

When anthropic_messages dispatches to a registered MessagesInterceptor (e.g. AdvisorOrchestrationHandler), it currently splats only **kwargs plus a handful of explicit positional/named args. Top-level parameters that are bound as named arguments on the public anthropic_messages function — thinking, metadata, stop_sequences, system, temperature, tool_choice, top_k, top_p — are silently dropped, because they live in local variables, not in kwargs.

This loses request fields on every interceptor sub-call. The most visible breakage: thinking={"type": "adaptive"} sent by clients (Claude Code, Anthropic SDK callers, etc.) is dropped on the executor sub-call, so downstream providers whose validation depends on thinking reject the request. Concretely, Vertex AI returns:

invalid_request_error: clear_thinking_20251015 strategy requires thinking to be enabled or adaptive

even though the caller correctly sent thinking: {type: adaptive}.

Fix: forward every named parameter explicitly into interceptor.handle, so the advisor (and any future interceptor) preserves the full request shape on its internal sub-calls.

Tests: added regression test_named_params_forwarded_into_advisor_executor_subcall that drives the full anthropic_messages → interceptor → executor path and asserts all named params arrive in the executor sub-call. Verified the test fails on master (None vs {type: adaptive}) and passes with this fix. All 4 tests in test_advisor_integration.py pass.

Relevant issues

None filed - discovered while debugging Claude Code + advisor against a litellm proxy fronting Vertex AI.

Linear ticket

N/A (external contributor)

Pre-Submission checklist

• I have added testing in the tests/test_litellm/ directory
• My PR passes all unit tests - make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

CI (LiteLLM team)

• Branch creation CI run — Link:
• CI run for the last commit — Link:
• Merge / cherry-pick CI run — Links:

Screenshots / Proof of Fix

Reproduction (against an unpatched litellm proxy fronting vertex_ai/claude-sonnet-4-6):

Send a request that mirrors what Claude Code with /advisor enabled produces — thinking: {type: adaptive} + context_management.clear_thinking_20251015 + an advisor tool registration:

curl http://localhost:4000/v1/messages \
  -H "x-api-key: sk-local-test" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: context-management-2025-06-27,advisor-tool-2026-03-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 8000,
    "thinking": {"type": "adaptive"},
    "context_management": {"edits":[{"type":"clear_thinking_20251015","keep":"all"}]},
    "tools": [{"name":"advisor","type":"advisor_20260301","model":"claude-opus-4-7"}],
    "messages": [{"role":"user","content":"hello"}]
  }'

Before fix — 400:

invalid_request_error: `clear_thinking_20251015` strategy requires `thinking` to be enabled or adaptive

After fix — 200, response carries a thinking block from the model.

Where the drop happens (root cause): litellm/llms/anthropic/experimental_pass_through/messages/handler.py — the interceptor dispatch splats **kwargs but never forwards the named parameters of anthropic_messages, so thinking (and 7 others) are silently lost on the executor sub-call.

Regression test (fails on master, passes with fix):

$ pytest tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_advisor_integration.py -v
test_anthropic_provider_bypasses_interceptor                            PASSED
test_full_dispatch_interceptor_fires_and_loop_completes                 PASSED
test_max_uses_enforced_through_full_handler                             PASSED
test_named_params_forwarded_into_advisor_executor_subcall               PASSED
4 passed

Screenshot

Type

🐛 Bug Fix

Changes

• litellm/llms/anthropic/experimental_pass_through/messages/handler.py — forward thinking, metadata, stop_sequences, system, temperature, tool_choice, top_k, top_p explicitly into interceptor.handle(...) so they're not dropped by the **kwargs splat.
• tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_advisor_integration.py — add test_named_params_forwarded_into_advisor_executor_subcall regression test.

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[samagana]

@greptileai

[greptile-apps]

Greptile Summary

This PR fixes a silent parameter drop in the anthropic_messages interceptor dispatch path: named params bound to the function signature (thinking, metadata, system, temperature, stop_sequences, tool_choice, top_k, top_p) were never forwarded to interceptor.handle, causing downstream providers (e.g. Vertex AI) to reject requests that depend on those fields.

• handler.py: Each of the 8 named params is now pop-ped from request_kwargs immediately after the pre-request hook extraction step (mirroring the existing pattern for tools and stream), then passed explicitly into interceptor.handle. Popping before kwargs.update(request_kwargs) also prevents the TypeError: got multiple values for keyword argument that would fire when a hook returns one of these keys.
• test_advisor_integration.py: Adds test_named_params_forwarded_into_advisor_executor_subcall (asserts all 8 params reach the executor) and test_pre_request_hook_override_does_not_collide_with_explicit_kwargs (asserts hook-returned overrides propagate without collisions). Both are purely mock-based.

Confidence Score: 5/5

The change is narrowly scoped to the interceptor dispatch path and follows the existing extraction pattern for tools/stream; no existing behavior is altered for callers that do not use an interceptor.

The extraction-before-merge pattern is already established in the codebase for tools and stream; extending it to the remaining 8 named params is straightforward. The two new regression tests cover forwarding correctness and hook-collision prevention, and are mock-only with no external dependencies.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/messages/handler.py	Pops all 8 named params from request_kwargs before merging into kwargs, then forwards them explicitly to interceptor.handle — correctly fixes both the silent-drop bug and the potential duplicate-keyword TypeError.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_advisor_integration.py	Adds two regression tests: one asserting all 8 named params reach the executor sub-call, and one asserting that hook-returned overrides of named params don't cause a duplicate-keyword TypeError. All mocked — no real network calls.

_{Reviews (3): Last reviewed commit: "fix(anthropic_messages): forward named p..." | Re-trigger Greptile}

[greptile-apps]

Greptile Summary

Fixes silent dropping of 8 named parameters (thinking, metadata, stop_sequences, system, temperature, tool_choice, top_k, top_p) when anthropic_messages dispatches to a MessagesInterceptor. Because these are bound as named params — not captured in **kwargs — they were never forwarded to interceptor.handle, causing providers like Vertex AI to reject requests that depended on them.

• handler.py: Adds 8 explicit keyword arguments to the interceptor.handle(...) call so they are preserved on every interceptor sub-call, matching the forwarding already done for api_key and api_base.
• test_advisor_integration.py: Adds test_named_params_forwarded_into_advisor_executor_subcall, a mock-only regression test that exercises the full anthropic_messages → interceptor → executor path and asserts 7 of the 8 newly-forwarded params arrive in the executor sub-call (metadata is captured but not asserted).

Confidence Score: 4/5

Safe to merge; the change is a targeted two-line-group addition with a passing regression test and no structural rewrites.

The fix is correct and well-scoped: forwarding named params to the interceptor mirrors the existing pattern already used for api_key and api_base. The regression test validates the critical path. The only gap is that metadata is captured in the test but never asserted, meaning a future regression that drops only metadata would go unnoticed by the new test.

The test file's metadata assertion gap is the only item worth a second look before merging.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/messages/handler.py	Adds explicit forwarding of 8 named params (thinking, metadata, stop_sequences, system, temperature, tool_choice, top_k, top_p) into interceptor.handle() so they aren't silently dropped on sub-calls. Change is minimal and correct.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_advisor_integration.py	Adds regression test for named-param forwarding; all params asserted except metadata, leaving a small coverage gap for that one forwarded field.

_{Reviews (2): Last reviewed commit: "fix(anthropic_messages): forward named p..." | Re-trigger Greptile}

[samagana]

@greptileai

[oss-pr-review-agent-shin]

🤖 litellm-agent: This PR is currently BLOCKED from merge.

Score: 4/5 ❌

Why blocked:

• 1 unresolved reviewer concern (greptile) (unresolved_concern, -1 pts)

Details: Score docked for: 1 unresolved reviewer concern (greptile).

Fix the issues above and push an update — the bot will re-review automatically.

Note: This bot is still in beta and might not always work as expected. Please share any feedback via Slack.

[oss-pr-review-agent-shin]

🤖 litellm-agent: Auto-merge skipped — the staging branch shin_agent_oss_staging_05_13_2026 has 1 commit(s) not in your branch. Merging as-is would produce a confusing diff on the staging PR.

Please rebase your branch onto shin_agent_oss_staging_05_13_2026 and push; the agent will re-review automatically.