feat(eval): expose maxTokens / temperature / topP flags for llm-as-a-judge evaluator (Bedrock) #1973

Summary

The evaluator llm-as-a-judge create handler only passes modelId to the Bedrock evaluator model config — there is no way to set inference parameters. CloudFormation's AWS::BedrockAgentCore::Evaluator InferenceConfiguration supports:

• MaxTokens — Integer, minimum 1
• Temperature — Number, 0–1
• TopP — Number, 0–1

These map to BedrockEvaluatorModelConfig.inferenceConfig and apply to both the Bedrock and Bedrock Mantle evaluator paths.

Current behavior

src/handlers/eval/evaluator/llm-as-a-judge/create/index.tsx builds the config with just the model id:

modelConfig: { bedrockEvaluatorModelConfig: { modelId: flags["model"] } },

There are no --max-tokens, --temperature, or --top-p flags, so users get whatever the service/construct defaults are and cannot tune scoring behavior.

For comparison, the harness handler already exposes these — see src/core/project/schema/harness.ts (temperature, topP, maxTokens) and src/handlers/harness/parameterHelp.tsx.

Requested change

Add flags to llm-as-a-judge create (and any update/edit path):

• --max-tokens <int> (min 1)
• --temperature <float> (0.0–1.0)
• --top-p <float> (0.0–1.0)

Thread them into bedrockEvaluatorModelConfig.inferenceConfig (maxTokens / temperature / topP), omitting any unset value so existing defaults are preserved. Enforce the CFN-documented ranges above.

Related

Companion construct/schema change: aws/agentcore-l3-cdk-constructs#315 (the L3 currently hardcodes { temperature: 0, maxTokens: 4096 } for Bedrock and rejects these fields unless the provider is OpenAI). This CLI change depends on that surface being available.