fix optimize for managed agent- #9954

[Jessie Li (YoYoJa)]

Summary

Fixes prompt-agent optimization behavior for both job submission and candidate application.

Prompt agents use the definition already deployed to the Foundry service, so azd ai agent optimize no longer requires local copies of the agent model, instructions, tools, or skills. Applying a candidate now persists its supported mutations to the inline prompt-agent definition in azure.yaml.

Hosted and voice-agent behavior remains unchanged.

Changes

Optimize job submission

For kind: prompt agents:

• Detect the agent kind using the shared agentkind.Kind resolver.
• Skip local baseline discovery and creation.
• Skip prompts for:
  • Agent instructions
  • Skills directory
  • Tools file
  • Target agent model
• Avoid reading local instruction, skill, and tool files.
• Omit service-owned fields from optimization_config:
  • model
  • system_prompt
  • skills
  • tools
• Preserve unrelated settings such as model_search_space.
• Build the request from a sanitized copy without modifying the original configuration.

Dataset, evaluator, evaluation-model, and optimization-model handling is unchanged.

Apply an optimization candidate

For prompt agents, azd ai agent optimize apply now:

• Loads the selected candidate’s mutation metadata.
• Finds candidates by candidate ID or name.
• Uses mutation keys to identify fields changed by the candidate.
• Reads final values from the candidate configuration.
• Writes supported mutations to the inline agent definition in azure.yaml.

Supported mappings:

Candidate configuration	azure.yaml
system_prompt, systemPrompt, or instructions	instructions
model	model
tools	tools

Mutation metadata is used only to determine which fields changed because its values may have different response shapes. Final values come from the candidate configuration endpoint.

Hosted agents do not enter this new inline-definition update path.

Tests

Added regression coverage for:

• Starting prompt-agent optimization without local instructions, skills, tools, or model.
• Omitting service-owned fields from prompt-agent optimization requests.
• Preserving unrelated optimization settings and the original configuration.
• Preserving hosted-agent request behavior.
• Candidate lookup and missing-candidate errors.
• Applying prompt-agent instruction and tool mutations.
• Updating only fields listed in the candidate mutation metadata.
• Explicitly excluding voice agents from prompt-agent mutation persistence.

Validation

• go test ./internal/cmd
• go build ./...

Fixes #10003

[azure-pipelines]

Azure Pipelines:
2 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

Unresolved compatibility, candidate lookup, environment persistence, and test-capture issues remain.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Updates managed prompt-agent optimization to use deployed definitions and persist candidate mutations to inline azure.yaml definitions.

Changes:

• Sanitizes prompt-agent optimization requests and skips unnecessary local inputs.
• Applies instruction, model, and tool mutations to inline definitions.
• Adds regression coverage for optimization and candidate application behavior.

File summaries

File	Review notes
cli/azd/extensions/azure.ai.agents/internal/cmd/optimize.go	Critical (2 votes): Legacy file-based prompt agents are treated like inline agents, so optimization and application do not update the deployed definition.
cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_test.go	No findings.
cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply.go	Moderate (3 votes): Candidate names are sent to ID-only API paths before lookup. Moderate (1 vote): Legacy config fields are updated at the wrong YAML path. Moderate (1 vote): Prompt-agent bookkeeping is persisted as deployed runtime environment variables.
cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply_test.go	Nit (2 votes): The test recorder drops array values, causing tool persistence assertions to use zero values and fail.

Review details

Suppressed comments (2)

cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply.go:360

• This writes instructions, model, and tools at the service root even when the prompt definition was found under the legacy config: block. SetServiceConfigValue will therefore create unrelated top-level fields, while PromptAgentFromResolvedService continues reading the unchanged config.* values on the next deploy, so applying a candidate silently has no effect for legacy prompt-agent projects. Use the definition source to write config.<field> for the legacy shape (or explicitly migrate it before updating).

		if _, err := azdClient.Project().SetServiceConfigValue(
			ctx,
			&azdext.SetServiceConfigValueRequest{
				ServiceName: svc.Name,
				Path:        path,
				Value:       value,

cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply.go:224

• envUpdates is still persisted for prompt agents. That writes .agent_configs and the candidate ID into the service's env: block; deployPromptAgent passes all service env values into the managed agent API as runtime environment variables. Since prompt deployments now use the inline mutations and do not consume local config paths, this leaks local apply bookkeeping into the deployed prompt agent; skip this local-env update for isPromptAgent while retaining the azd environment tracking.

	// Step 4: Persist OPTIMIZATION_LOCAL_DIR and OPTIMIZATION_CANDIDATE_ID onto the
	// agent definition so the deploy pipeline knows which local optimization
	// config to use. New projects carry the definition inline in azure.yaml;
	// older projects still keep it in an on-disk agent.yaml.

• Files reviewed: 4/4 changed files
• Comments generated: 3
• Review effort level: Lite

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Jessie Li (YoYoJa)]

recorder fixed

[Copilot]

🔵 Needs a closer look

One or more issues must be addressed before approval.

Review details

Suppressed comments (3)

Previously missed (1) — in code that hasn't changed since the last review.

cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply.go:208

• Prompt-agent apply now also persists supported instructions, model, and tools mutations into the inline azure.yaml service, but the command help and file-level description still only explain writing files under .agent_configs/ and updating environment variables. Update the user-facing help to disclose this additional inline-definition change so users understand what optimize apply modifies before running azd deploy.

This issue also appears on line 339 of the same file.

cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply.go:342

• The prompt-agent path advertises model as a supported mutation and has a dedicated branch here, but the added persistence test deliberately omits model from the mutation metadata and asserts that it is not written. Add a regression case with a model mutation and verify the candidate model is persisted to azure.yaml, otherwise this new update path can regress without detection.

	if hasCandidateMutation(mutations, "model") {
		if value, found := candidateConfigValue(config, "model"); found {
			updates["model"] = value
		}

cli/azd/extensions/azure.ai.agents/internal/cmd/optimize_apply.go:360

• When the prompt definition is in the supported deprecated service.config shape, PromptAgentFromResolvedService still returns found=true, so this branch runs, but these paths are written at the service root. The deploy resolver continues to read config.* for that shape, so the mutation is ignored and azd deploy uses the old instructions/model/tools. Select the definition source and prefix these paths with config. for AgentDefinitionSourceLegacyConfig (and cover that shape with a test).

		if _, err := azdClient.Project().SetServiceConfigValue(
			ctx,
			&azdext.SetServiceConfigValueRequest{
				ServiceName: svc.Name,
				Path:        path,
				Value:       value,

• Files reviewed: 5/5 changed files
• Comments generated: 0 new
• Review effort level: Lite