fix(bedrock/converse): drop toolSpec.strict for Opus 4.7/4.8 (#31582)

[ly-wang19]

What Problem This Solves

Closes #31582.

Sending a tool with strict: true to Claude Opus 4.7 or Opus 4.8 on Bedrock Converse fails with a 400:

litellm.BadRequestError: BedrockException - {"message":"The model returned the following errors: tools.0.custom.strict: Extra inputs are not permitted"}

The same request works on direct Anthropic and on Bedrock Sonnet 4.5/4.6 and Opus 4.5/4.6, so this is specific to Opus 4.7/4.8 on Bedrock Converse. It is a regression: working on 1.85.1, broken on 1.90.0/1.90.0-rc.1.

Why This Change Was Made

PR #29814 ("feat(bedrock): forward strict and additionalProperties to Converse toolSpec") started forwarding strict into the Bedrock Converse toolSpec, gated on get_bedrock_base_model(model).startswith("anthropic"). That gate is too broad: Bedrock Converse routes Opus 4.7/4.8 through an Anthropic-compatible validator that maps toolSpec to the native tool shape (the custom variant) and rejects the extra strict key, even though Anthropic's native API accepts strict as a top-level tool field for these models. Sonnet 4.5/4.6 and Opus ≤4.6 accept toolSpec.strict and keep the existing behavior.

The fix is a focused exclusion: replace the broad startswith("anthropic") gate in litellm_core_utils/prompt_templates/factory.py with a small helper bedrock_converse_supports_strict_tools in litellm/llms/bedrock/common_utils.py that:

• returns False for non-Anthropic Bedrock models (preserves the existing "drop strict for Nova/Llama/GPT-OSS" behavior from feat(bedrock): forward strict and additionalProperties to Converse toolSpec #29814),
• returns False for the Opus 4.7/4.8 family,
• returns True for every other Anthropic model on Bedrock (Sonnet 4.5/4.6, Opus ≤4.6, Haiku, etc.).

The alternative — always drop strict on Bedrock and re-add it elsewhere — would undo the user-visible enum-constraint fix from #29814 for Sonnet/Opus ≤4.6 callers. The targeted exclusion preserves the #29814 fix where it works and removes it only where Bedrock rejects the field.

User Impact

• Callers sending strict: true tools to bedrock/...claude-opus-4-7 or bedrock/...claude-opus-4-8 stop getting 400 BadRequestError: tools.N.custom.strict: Extra inputs are not permitted. The tool call succeeds (with strict dropped on the Bedrock hop, matching pre-1.86 behavior).
• Callers using Sonnet 4.5/4.6, Opus ≤4.6, or any other Anthropic model on Bedrock see no change — strict is still forwarded to toolSpec as before.
• Non-Anthropic Bedrock families (Nova, Llama, GPT-OSS) see no change.

Evidence

• New regression test test_bedrock_tools_pt_strict_dropped_for_opus_47_48 covers: Opus 4.7 (3 id spellings — bedrock/us.anthropic.claude-opus-4-7, anthropic.claude-opus-4-7-v1:0, mixed dash/underscore) and Opus 4.8 (same spellings) all drop strict; Sonnet 4.5 and Opus 4.6 still forward strict.
• Existing test_bedrock_tools_pt_strict_parameter (the feat(bedrock): forward strict and additionalProperties to Converse toolSpec #29814 regression test) still passes — confirms the Sonnet/Nova behavior is unchanged.
• Reproducer from the issue (bedrock/us.anthropic.claude-opus-4-7 with strict: true) is the case the new test locks in.
• Local run: python3 -m pytest tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py -k strict → 2 passed.
• One unrelated test (test_bedrock_converse_messages_pt_document_various_formats) fails in this file both with and without my changes (verified via git stash) — it's a pre-existing MIME-type environment issue, not caused by this PR.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing ly-wang19:fix/bedrock-opus-47-strict-tools-31582 (73658ad) with main (0ade44f)}

[greptile-apps]

Greptile Summary

This PR adds a targeted exclusion so that toolSpec.strict is dropped when calling Claude Opus 4.7/4.8 on Bedrock Converse, fixing a 400 error introduced in 1.86+ where Bedrock's validator rejects the strict key for those model variants. All other Anthropic models (Sonnet 4.5/4.6, Opus ≤4.6) continue to forward strict unchanged.

• Adds bedrock_converse_supports_strict_tools() in llms/bedrock/common_utils.py to replace the overly broad startswith("anthropic") gate in factory.py, returning False only for the Opus 4.7/4.8 family and non-Anthropic Bedrock models.
• Adds a new test file covering the Opus 4.7/4.8 exclusion (multiple model ID spellings), the "keep strict" path for Sonnet/Opus ≤4.6, and non-Anthropic families — all as pure unit tests with no network calls.

Confidence Score: 5/5

Safe to merge — the change is a narrow, well-tested exclusion that restores pre-regression behaviour for Opus 4.7/4.8 without touching any other code path.

The fix is a single-function swap in factory.py backed by a clearly scoped helper. The regression test covers all meaningful model-ID spellings for both the excluded (Opus 4.7/4.8) and preserved (Sonnet/Opus ≤4.6, non-Anthropic) paths. No existing tests were weakened, and the change has no impact on routing, auth, or any other critical path.

No files require special attention beyond common_utils.py, where the hardcoded model-name pattern list was already flagged in a prior review thread.

Important Files Changed

Filename	Overview
litellm/llms/bedrock/common_utils.py	Adds bedrock_converse_supports_strict_tools() helper with a hardcoded tuple of Opus 4.7/4.8 model name patterns to gate toolSpec.strict forwarding; logic is correct and well-documented.
litellm/litellm_core_utils/prompt_templates/factory.py	Single-line swap of get_bedrock_base_model(model).startswith("anthropic") for the new bedrock_converse_supports_strict_tools(model) helper — minimal, correct change.
tests/test_litellm/litellm_core_utils/prompt_templates/test_bedrock_converse_strict_tools_opus_47_48.py	New regression test file covering Opus 4.7/4.8 strict-drop, Sonnet/Opus ≤4.6 strict-keep, and non-Anthropic drop — pure unit tests with no network calls.

_{Reviews (2): Last reviewed commit: "fix(bedrock/converse): move strict-tools..." | Re-trigger Greptile}

[greptile-apps]

Move capability data
This adds another hardcoded model-specific capability list for Bedrock strict support. The repo rule asks these flags to live in model_prices_and_context_window.json and be read via model info; keeping claude-opus-4-7 / 4-8 in code means the next Bedrock validator exception requires a LiteLLM release instead of a model-data update.

Rule Used: What: Do not hardcode model-specific flags in the ... (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

Thanks for the PR! I'll make a quick copy branch to test this on internal e2e tests

[mateo-berri]

This has been merged. Thank you again for your contribution!