#8639 Add Bedrock model profile for xAI Grok 4.6 and 4.7

• Model ids are us.xai.grok-4.6 and global.xai.grok-4.6 only; xai.grok-4.6 is served by Bedrock Mantle only.
• Effort levels low (default), medium, high, xhigh are accepted via additionalModelRequestFields: {"reasoning": {"effort": "<level>"}}; max returns a 400.
• inferenceConfig.temperature, topP and top_k are rejected with a ValidationException (This model doesn't support the temperature field. Remove temperature and try again.).
• toolConfig works with all three toolChoice modes (auto, any, tool).
• Explicit cachePoint blocks return AccessDeniedException; only implicit prompt caching applies, so bedrock_supports_prompt_caching stays False.

2026-10-22 · thread 1: forced tool choice with thinking on xai · iter -

• Decision: Exempt the xai variant in _thinking_blocks_tool_forcing so tool_choice='required'/single tool/ToolOutput work with thinking
• Why: Converse accepts all toolChoice modes for Grok while it reasons (Macroscope flagged UserError)
• Source: https://github.com/pydantic/pydantic-ai/pull/8639#discussion_r4105394626

2026-10-22 · add Grok 4.7 alongside 4.6 · iter -

• Decision: Add us.xai.grok-4.7 and global.xai.grok-4.7 to every list the PR adds the 4.6 ids to (LatestBedrockModelNames, KnownModelName incl. gateway/bedrock:, docs thinking table), and pin 4.7 profile resolution in the 4.6 test
• Why: Bedrock now serves Grok 4.7 (cross-region ids only, same reasoning levels as 4.6, default high)
• Source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-xai-grok-4-7.html

2026-10-22 · port xai forced-tool exemption onto main's guards · iter -

• Decision: Main (#8808, #8812) removed _thinking_blocks_tool_forcing; the xai exemption now lives in one predicate, _thinking_rejects_tool_forcing, used by _default_structured_output_mode, prepare_request and _support_tool_forcing. Dropped the PR's noqa: C901, which main's refactor made unused.
• Why: Keep the PR's intent (Grok accepts a forced toolChoice while reasoning) on main's API without scattering variant checks
• Source: https://github.com/pydantic/pydantic-ai/pull/8812
• Supersedes: thread 1: forced tool choice with thinking on xai

2026-10-22 · thinking=True on Grok 4.7 keeps the model default · iter -

• Decision: The xai Bedrock variant omits reasoning for thinking=True when the profile sets grok_default_effort_omitted (Grok 4.7), matching the direct xAI adapter
• Why: thinking=True means the provider's default effort; Grok 4.7's default is high, so sending medium would lower it
• Source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-xai-grok-4-7.html

2026-10-22 · Grok 4.7 structured output stays off · iter -

• Decision: supports_json_schema_output/supports_json_object_output stay False for Grok 4.7 even though its card lists structured outputs as supported
• Why: Not verifiable: the test account has no Grok access on Converse
• Source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-xai-grok-4-7.html

2026-10-22 · API gate waiver for LatestBedrockModelNames · iter -

• Decision: Add the exact waiver for the LatestBedrockModelNames widening (against v2.59.0) to .github/api-compatibility-allowlist.json; the advisory Protect .github check fails on this fork PR as expected
• Why: agent_docs/api-design.md requires an exact waiver for a permitted change the gate reports
• Source: https://github.com/pydantic/pydantic-ai/pull/8639

2026-10-22 · drop sampling settings for Grok on Bedrock · iter -

• Decision: Set bedrock_disallows_sampling_settings in bedrock_xai_model_profile; Grok 4.7 structured output stays off, documented as unverified
• Why: Grok 4.6 rejects temperature, topP and top_k with a ValidationException on Converse; #8668's flag covers this
• Source: https://github.com/pydantic/pydantic-ai/pull/8668