Fix Bedrock thinking effort mapping for Claude models (#851)

[Edilbek]

What this does

Fixes #851.

Extended thinking on Bedrock Claude models (Sonnet 5, and every other Claude on Bedrock) raised RubyLLM::BadRequestError: The model returned the following errors: reasoning_effort: Extra inputs are not permitted whenever with_thinking(effort:) was used.

The Converse protocol branched on reasoningSupported.embedded metadata to decide how to send thinking params. That flag is unrelated to how a model accepts reasoning input, and it's false for every Bedrock Claude, so effort-based thinking always fell into a { reasoning_effort: ... } branch. Bedrock forwards additionalModelRequestFields straight to Anthropic, which rejects the unknown reasoning_effort key.

Bedrock actually publishes each model's accepted fields in converse.additionalRequestFieldsSchema. Claude models expose a reasoningConfig.budgetTokens schema and only accept a token budget, never an effort level. This PR:

• Replaces the reasoning_embedded? check with reasoning_budget_schema, which reads that published schema.
• Maps effort onto a budget: enumerated schemas use the token value they name for each level (e.g. Sonnet 5 low → 1024, medium → 40000, high → 63999); integer-range schemas spread low/medium/high across the range; efforts a schema doesn't enumerate clamp to its maximum.
• Keeps sending reasoning_effort for models whose schema advertises no budget (e.g. openai.gpt-oss-120b).
• Prefers an explicit budget: over effort:, and clamps a derived budget below max_output_tokens when one is set so it can't consume the whole output allowance.
• Falls back to a sibling regional registry entry for the same foundation model, since Bedrock only stores the Converse schema on some regional entries — without this, eu./global./bare anthropic. ids still failed.

Type of change

• Bug fix
• New feature
• Breaking change
• Documentation
• Performance improvement

Scope check

• I read the Contributing Guide
• This aligns with RubyLLM's focus on LLM communication
• This isn't application-specific logic that belongs in user code
• This benefits most users, not just my specific use case

Required for new features

N/A — bug fix.

• I opened an issue before writing code and received maintainer approval
• Linked issue: [BUG] Sonnet 5 Extra inputs are not permitted #851

Quality check

• I ran overcommit --install and all hooks pass
• I tested my changes thoroughly
  • For provider changes: Re-recorded VCR cassettes with bundle exec rake vcr:record[provider_name]
  • All tests pass: bundle exec rspec
• I updated documentation if needed
• I didn't modify auto-generated files manually (models.json, aliases.json)

AI-generated code

• I used AI tools to help write this code
• I have reviewed and understand all generated code (required if above is checked)

API changes

• Breaking change
• New public methods/classes
• Changed method signatures
• No API changes

---

Notes for the maintainer:

• VCR cassettes not re-recorded. No Bedrock credentials were available to record a live effort-based thinking cassette. The change is covered by unit specs in spec/ruby_llm/protocols/converse/chat_spec.rb asserting the emitted payload, plus a WebMock-stubbed end-to-end reproduction. The derived payload shape matches what the existing budget-based Bedrock cassettes already send, but you may want to add { provider: :bedrock, model: 'claude-sonnet-5' } to the thinking models and record one.

[codecov]

Codecov Report

❌ Patch coverage is 82.00000% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.30%. Comparing base (091b16a) to head (1b7c9af).
⚠️ Report is 8 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/ruby_llm/protocols/converse/chat.rb	81.25%	1 Missing and 5 partials ⚠️
lib/ruby_llm/protocols/converse.rb	83.33%	1 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #855      +/-   ##
==========================================
+ Coverage   83.18%   83.30%   +0.12%     
==========================================
  Files         173      175       +2     
  Lines        8461     8798     +337     
  Branches     1373     1423      +50     
==========================================
+ Hits         7038     7329     +291     
- Misses        925      941      +16     
- Partials      498      528      +30     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[justwiebe]

Looks good, although I'm not a reviewer on the project