fix: guard unsupported Anthropic reasoning fields across providers

[amitksingh1490]

Summary

Guard Anthropic reasoning parameters by model capability across Anthropic, Bedrock, and OpenAI-compatible provider paths to prevent 400 errors like "This model does not support the effort parameter."

Context

The critical case is models that support thinking but not effort-style fields.

For example, Claude Haiku 4.5 supports reasoning/thinking budgets but rejects effort-style parameters. If effort is sent, Anthropic-compatible providers can return 400 invalid request errors.

Fixes #3030.

Changes

Implemented transformer-based compatibility guards per request pipeline so unsupported reasoning fields are removed before sending upstream requests.

Key Implementation Details

• Added Anthropic request transformer StripUnsupportedReasoning and wired it into the Anthropic provider pipeline.
• Added OpenAI-request transformer StripUnsupportedAnthropicReasoning and wired it into the OpenAI provider pipeline so proxy providers (OpenRouter, Requesty, GitHub Copilot, etc.) also sanitize Anthropic reasoning fields by model.
• Added Bedrock transformer StripUnsupportedReasoning to remove unsupported additional_model_request_fields thinking payloads for unsupported/non-Anthropic models.
• Updated model capability checks so Claude Haiku 4.5 keeps thinking while stripping effort.
• Registered new modules in transformer/provider mod declarations.
• Added focused unit tests for Haiku, Sonnet 3.7, Sonnet 4, non-Anthropic cases, and Bedrock path behavior.

Use Cases

• Running claude-haiku-4-5 with reasoning enabled no longer fails due to unsupported effort fields.
• Anthropic models routed via OpenAI-compatible APIs (OpenRouter/Requesty/GitHub Copilot) no longer receive unsupported effort fields.
• Bedrock requests avoid sending unsupported Anthropic thinking fields for incompatible model IDs.

Testing

cargo test -p forge_app -- dto::openai::transformers::strip_unsupported_anthropic_reasoning
cargo test -p forge_app -- dto::anthropic::transforms::strip_unsupported_reasoning
PROTOC="/tmp/protoc-bin/bin/protoc" cargo test -p forge_repo -- provider::bedrock_strip_unsupported_reasoning
PROTOC="/tmp/protoc-bin/bin/protoc" cargo check -p forge_repo
cargo check -p forge_app

Links

• Related issue: [Bug]: Guard Anthropic reasoning fields by model support #3030
• Branch: fix/anthropic-haiku-reasoning-effort

[tusharmath]

Too much of duplicate code across transformers. We can create one on Conversation to remove reasoning effort thing.

[Alexx999]

I'm afraid this doesn't fix the big Anthropic problem.
If I'm reading it correctly, this still sends legacy thinking: { "type": "enabled" } for Opus 4.7 which won't work. And simply stripping this will go into "no thinking" mode which makes the model unusable - I'm afraid something deeper is required to make it work across the board for both older and new models.

[amitksingh1490]

I'm afraid this doesn't fix the big Anthropic problem. If I'm reading it correctly, this still sends legacy thinking: { "type": "enabled" } for Opus 4.7 which won't work. And simply stripping this will go into "no thinking" mode which makes the model unusable - I'm afraid something deeper is required to make it work across the board for both older and new models.

This is for Haiku models getting 400 for Opus-4.7 we would build on top of this

[amitksingh1490]

closing in favour of #3067