Agent Platform V1 flows ignore model_metadata from request headers, so model switching does not work for those flows

Summary

All Agent Platform V1 flows (fix_pipeline, resolve_sast_vulnerability, sast_fp_detection, code_review) do not respect the model_metadata passed in request headers (e.g., x-gitlab-agent-platform-model-metadata). When a user passes a specific model like gpt_5 via the header, these flows ignore it and use the default model from their feature setting instead.

Current Behavior

When a request is made to an Agent Platform V1 flow with model_identifier: "gpt_5" in the x-gitlab-agent-platform-model-metadata header:

1. ✅ ModelMetadataInterceptor correctly sets the model metadata in the context
2. ❌ AgentComponent.attach() calls prompt_registry.get() without passing model_metadata
3. ❌ The prompt registry falls back to the default model for the feature setting
4. ❌ The flow executes with the wrong model (e.g., claude_sonnet_4_20250514 instead of gpt_5)

Root Cause

The issue is in the Agent Platform V1 component architecture:

Affected Components

All three component types call prompt_registry.get() without model_metadata:

1. AgentComponent

   prompt = self.prompt_registry.get(
       self.prompt_id, self.prompt_version, tools=tools, tool_choice=tool_choice
   )  # ❌ No model_metadata parameter

2. OneOffComponent

   prompt = self.prompt_registry.get(
       self.prompt_id, self.prompt_version, tools=tools, tool_choice=tool_choice
   )  # ❌ No model_metadata parameter

3. HumanInputComponent

   prompt = self.prompt_registry.get(self.prompt_id, self.prompt_version)
   # ❌ No model_metadata parameter

Why This Matters

The BasePromptRegistry.get() method does NOT automatically retrieve from context:

# ai_gateway/prompts/registry.py:308-316
def get(
    self,
    prompt_id: str,
    prompt_version: str | None,
    model_metadata: Optional[TypeModelMetadata] = None,  # If None, uses default
    ...
) -> Prompt:

Affected Flows

All Agent Platform V1 flows are affected:

Flow	Components Affected	Prompts Affected
fix_pipeline	4 AgentComponent 1 OneOffComponent	fix_pipeline_context fix_pipeline_create_plan fix_pipeline_execution commit_changes fix_pipeline_push_changes
resolve_sast_vulnerability	4 AgentComponent 1 OneOffComponent	resolve_sast_vulnerability_context resolve_sast_vulnerability_execution resolve_sast_check_false_positive resolve_sast_evaluate_mr_readiness resolve_sast_vulnerability_commit_and_push
sast_fp_detection	3 AgentComponent 3 OneOffComponent	validate_sast_vulnerability_agent_prompt sast_vulnerability_source_file_agent_prompt sast_vulnerability_lines_agent_prompt sast_vulnerability_report_agent_prompt sast_fp_detection_agent_prompt sast_post_results_to_gitlab_agent_prompt
code_review	2 AgentComponent	code_review_prescan review_merge_request

Expected Behavior

When a model is specified in the request headers:

1. The flow should use that specific model
2. If a model-specific prompt exists (e.g., fix_pipeline_execution/gpt_5/1.0.0.yml), it should be loaded
3. The model provider should be set correctly (e.g., openai for GPT models, not anthropic)

Proposed Solution

Option 1: Retrieve from Context

Update all three component types to retrieve model_metadata from context:

from ai_gateway.model_metadata import current_model_metadata_context

def attach(self, graph: StateGraph, router: RouterProtocol) -> None:
    tools = self.toolset.bindable + [AgentFinalOutput]
    tool_choice = "any"

    prompt = self.prompt_registry.get(
        self.prompt_id, 
        self.prompt_version, 
        model_metadata=current_model_metadata_context.get(),  # ✅ Add this
        tools=tools, 
        tool_choice=tool_choice
    )
    # ... rest of method

Files to update:

• duo_workflow_service/agent_platform/v1/components/agent/component.py
• duo_workflow_service/agent_platform/v1/components/one_off/component.py
• duo_workflow_service/agent_platform/experimental/components/agent/component.py
• duo_workflow_service/agent_platform/experimental/components/one_off/component.py
• duo_workflow_service/agent_platform/experimental/components/human_input/component.py

✅ Flows That Work Correctly

These older workflows use get_on_behalf() and respect model metadata:

• convert_to_gitlab_ci
• issue_to_merge_request
• software_development
• chat

❌ Flows That Don't Work

All Agent Platform V1 flows using the component architecture.