Harness invoke with --model-id drops model inference parameters (temperature, topP, maxTokens)

[jesseturner21]

Problem

When invoking a harness agent with --model-id to override the model at invoke time, all model inference parameters (temperature, topP, topK, maxTokens) configured in harness.json are dropped from the request. The server treats the invoke-time model override as a full replacement of the deployed model config, so omitting these fields results in unbounded/default inference behavior.

Example: A harness configured with maxTokens: 500 will ignore that limit when invoked with --model-id anthropic.claude-v3, producing responses far exceeding 500 tokens.

Root Cause

buildHarnessBaseOpts in the invoke command constructed the model override object with only the modelId (and apiKeyArn for non-Bedrock providers), without forwarding the inference parameters from the harness spec.

Impact

• Affects all provider types (Bedrock, OpenAI, Gemini)
• Only triggered when using --model-id at invoke time
• Deploy-time configuration is unaffected (the parameters are correctly stored on the server)

Fix

PR #1287 — forwards inference parameters (temperature, topP, topK, maxTokens) from the harness spec when constructing the model override, and uses CLI-provided values (e.g. --api-key-arn) as precedence overrides.

Affected Versions

Preview branch only (harness feature has not shipped to mainline).

[notgitika]

PR merged, closing issue