[staging-supernova-dx-appf-io]

What this does

Adds a new, opt-in :invoke_anthropic protocol to the Bedrock provider that talks to Claude via Bedrock's InvokeModel / InvokeModelWithResponseStream APIs using Anthropic's native Messages request format, instead of the Converse API. Supernova currently talks to Claude on Bedrock via Converse, but newer server-side context-management features (compaction compact-2026-09-04, tool-result clearing clear_tool_uses_20250919 / context-management-2025-06-27, and tool search with defer_loading) are only reachable through InvokeModel/InvokeModelWithResponseStream, not Converse. This PR is a spike that builds and tests the alternate wire protocol at parity with Converse, as a foundation for a later PR to actually turn on compaction/context-editing/tool-search. It does not enable any of those features yet.

Key pieces:

• lib/ruby_llm/providers/bedrock.rb: registers :invoke_anthropic alongside :converse and :mantle_responses. protocol_for never auto-selects it — it is opt-in only via protocol: :invoke_anthropic (or config.bedrock_protocol), so Converse remains the default with no behavior change unless explicitly requested. Adds normalize_invoke_params, which lifts everything under additionalModelRequestFields to the top level and strips model/stream (the single place intended for future anthropic_beta/context_management additions).
• lib/ruby_llm/protocols/invoke_anthropic.rb (new): protocol class subclassing Anthropic, mixing in Chat and Streaming, POSTing to /model/{id}/invoke, SigV4-signed via the existing Bedrock signer.
• lib/ruby_llm/protocols/invoke_anthropic/chat.rb (new): request rendering/response parsing matching Converse's replay guarantees (thinking blocks always replayed verbatim, consecutive same-role turns folded, tool results batched). Builds the InvokeModel envelope (anthropic_version: "bedrock-2023-05-31", merged/de-duped anthropic_beta, no model/stream). Images/PDFs are always inline base64 (supports_provider_file_references? returns false).
• lib/ruby_llm/protocols/invoke_anthropic/streaming.rb (new): decodes the AWS eventstream (via the extracted Converse::EventStream helper) whose chunk events wrap base64-encoded Anthropic SSE-shaped events, including thinking/redacted_thinking finalize/salvage logic mirroring Converse streaming.
• lib/ruby_llm/protocols/converse/event_stream.rb (new): extracted eventstream decoding/base64 unwrapping/error handling shared by both Converse and InvokeAnthropic streaming; converse/streaming.rb now includes it (net −100 lines), validated by re-running the full existing 50-spec Converse suite before building on top of it.
• lib/ruby_llm/protocols/converse/chat.rb: adds converse_thinking_block so thinking/redacted_thinking blocks captured under InvokeModel can be replayed correctly if a later message on the same chat falls back to Converse (cross-protocol replay for Supernova's endpoint-pin/kill-switch mechanism).
• Tests: spec/ruby_llm/protocols/invoke_anthropic_spec.rb (new, 358 lines) covering URL construction, request-body shape, tool formatting/folding, inline images, verbatim/cross-protocol thinking replay, SigV4-signed completion parsing, throttling→RateLimitError mapping, full synthetic streaming decode, thinking-block salvage rules, and streaming error mapping. spec/ruby_llm/providers/bedrock_spec.rb (+35 lines) covers #complete param normalization for the new protocol.

Reviewer notes: default behavior is unchanged by construction (no config flag to flip in this fork — Supernova gates it via its own feature flag in a separate PR). No beta/context-management values are set anywhere in this PR — that's left for a follow-up. Cross-protocol thinking replay (both directions) is implemented and unit-tested, but nothing has run against live Bedrock yet to confirm this end-to-end; that verification is planned as a staging probe on the Supernova side. The Supernova-side feature flag, admin page, and endpoint-pinning/kill-switch logic live entirely in the separate appfolio/supernova PR #7086 and are not part of this diff.

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

• I opened an issue before writing code and received maintainer approval
• Linked issue: #___

N/A — no approved GitHub issue for this feature was identified; this should be confirmed with the author before merge.

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)

Note: this fork's tests use WebMock/RSpec stubs and synthetic AWS-eventstream payloads rather than VCR cassettes, so no cassette re-recording applies. 106 targeted specs pass and rubocop reports no offenses.

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

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

New public classes/modules: Protocols::InvokeAnthropic (plus its Chat and Streaming), Protocols::Converse::EventStream, and new Bedrock#complete/protocol_for branching logic for the :invoke_anthropic protocol.

---

In Supernova
Chat: open

[staging-supernova-dx-appf-io]

🤖 Supernova Code Review — View trace

🔄 Re-run this review

— via automation "Code Review Opus 5.5 Medium" (run), installed by Sam Boland (sam.boland@appfolio.com)

[staging-supernova-dx-appf-io]

🤖 Supernova adversarial review: 4 inline findings.

The biggest risks when turning on :invoke_anthropic:

• Routing: setting config.bedrock_protocol sends non-Claude Bedrock models to InvokeAnthropic.
• Output cap: the 4096 max_tokens fallback truncates output and rejects thinking budgets of 4096 or more.
• Param translation: top-level Converse params such as inferenceConfig are passed through untranslated.
• Render mismatch: render and complete produce different bodies.

The EventStream extraction and the thinking-block translation between Converse and InvokeAnthropic looked correct to me.

🔄 Re-run this review

— via automation "Code Review Opus 5.5 Medium" (run), installed by Sam Boland (sam.boland@appfolio.com)

[staging-supernova-dx-appf-io]

When the model has no max_tokens, this caps output at 4096. Converse doesn't do that, so the two protocols behave differently here.

Converse's format_inference_config never sets maxTokens, so Bedrock falls back to the model's own default. InvokeModel requires max_tokens, and this code fills it with model.max_tokens || 4096. Bedrock registry entries often have no max_tokens, so Claude requests would hit the 4096 fallback. That causes two problems:

• Long answers get cut off with stop_reason: max_tokens on requests that work fine over Converse.
• Any budget thinking request with budget_tokens >= 4096 gets a 400, because Anthropic requires max_tokens > thinking.budget_tokens.

Fix: use a Claude-appropriate default (for example the model family's max output). At minimum, when a thinking budget is set, bump max_tokens above it. Also honor a caller-supplied max_tokens.

[staging-supernova-dx-appf-io]

This only unwraps additionalModelRequestFields. Other top-level Converse params go into the InvokeModel body unchanged.

The docstring says callers build params in the Converse shape. But keys like inferenceConfig (maxTokens, topP, stopSequences), guardrailConfig, performanceConfig, requestMetadata and additionalModelResponseFieldPaths get deep-merged into the Anthropic Messages body as they are. Bedrock then rejects them with a 400 (extraneous key ... is not permitted). So a caller that works on Converse today breaks as soon as it opts into :invoke_anthropic.

Fix: translate inferenceConfig to max_tokens / top_p / stop_sequences. For the Converse-only keys, either strip them or raise a clear ArgumentError, the way strip_converse_only_params does for mantle.

[staging-supernova-dx-appf-io]

Setting config.bedrock_protocol = :invoke_anthropic sends every Bedrock model to InvokeAnthropic, not just Claude.

Provider#resolve_protocol lets an explicit or configured protocol win over protocol_for. With this config set, Llama, Nova, Mistral and openai.gpt-oss-* all get an Anthropic Messages body posted to /model/{id}/invoke. Mantle-only openai.gpt-5* ids get their params run through strip_converse_only_params, then still go to the InvokeAnthropic protocol on bedrock-runtime. The PR advertises this config key as one of the two opt-in paths, so any non-Anthropic Bedrock call would fail with 400s.

Fix: only honor :invoke_anthropic for Anthropic model ids (match anthropic. including region prefixes like us./global., plus inference-profile ARNs), and fall back to protocol_for otherwise. Or raise an ArgumentError naming the model.

[staging-supernova-dx-appf-io]

The InvokeModel param normalization only runs in complete, not in render, so render and complete produce different bodies.

Bedrock doesn't override Provider#render. A dry-run render with protocol: :invoke_anthropic keeps additionalModelRequestFields nested and never lifts thinking / anthropic_beta / context_management to the top level. The rendered body then doesn't match what's actually sent, which undermines using render to check parity. Fix: apply the same params = branching in a render override, or pull it into a shared helper both methods call.