[feat] optimzie progressive update + adapt to new azure.yaml

[Zyysurely]

Optimize: progressive polling UX, azure.yaml agent detection, and --output json for status

Summary

Improves the azd ai agent optimize polling experience with live-updating candidate rows and phase-aware progress indicators, migrates agent kind detection to the new azure.yaml inline/config agent definition format, and adds --output json support to optimize status.

Changes

Agent detection from azure.yaml (eval.go, optimize.go)

• Replaced detectEvalAgentKind with projectpkg.LoadAgentDefinition, which supports the latest azure.yaml agent definition shapes:
  • Inline: agent kind defined directly in azure.yaml service config
  • Config reference: agent kind referenced via config pointer in azure.yaml
  • Legacy disk: standalone agent.yaml file on disk
• The agent kind source is now accurately reported (azure.yaml (inline), azure.yaml (config), agent.yaml) instead of just showing the file path.
• Emits a deprecation warning via WarnLegacyAgentShape when legacy config shapes are detected.
• Removed now-unused detectEvalAgentKind and relPathForYaml functions and their tests.

Optimize agent resolution refactor (optimize.go)

• Resolution order fixed: resolveOptimizeAgent now tries the azd project context first (passing --agent as a service name to resolveAgentService), then falls back to treating --agent as a raw Foundry agent name. Previously, a non-empty --agent flag would skip project resolution entirely.
• The --agent flag now also reads the agent version from the environment (via AGENT_{KEY}_VERSION), enabling version-aware optimization.
• Updated flag help text to clarify: "Agent service name from azure.yaml, or Foundry agent name outside a project".
• Same resolution fix applied to optimize apply and optimize deploy flag descriptions.

Progressive polling UX (optimize.go, optimize_helpers.go)

• Live candidate table: Completed candidates are printed as individual table rows as they finish, instead of waiting for the entire job to complete.
• Phase-aware status line: The spinner distinguishes "Generating" vs "Evaluating" using the new InProgressCandidate.CandidateGenerated field.
• Candidate progress counter: Shows Generating candidate 2/5 (from MaxCandidates) or Generating candidate 2 when max is unknown.
• Server-side elapsed time: Uses InProgressCandidate.CreatedAt (server timestamp) instead of a locally-tracked timer.
• Extracted table rendering helpers: candidateDisplayName, candidateTableHeader, formatCandidateRow, writeCandidateRow are now shared helpers, eliminating duplicated rendering between polling and final results.

optimize status --output json (optimize_status.go)

• Adds --output flag (json | table, default table) via azdext.RegisterFlagOptions.
• --output json: prints the full OptimizeJobStatus as indented JSON.
• --output json --watch: suppresses table output during polling, prints final status as JSON.
• Informational "Using last job: ..." line suppressed in JSON mode.

API model updates (optimize_api/models.go)

• Added InProgressCandidate struct with CreatedAt and CandidateGenerated fields.
• Added InProgressCandidate field to JobProgress.

Tests

• optimize_status_test.go: --output flag registration, default poll interval, printOptimizeStatusJSON, job summary duration display (present/missing/equal timestamps).
• optimize_helpers_test.go: candidateDisplayName, candidateTableHeader, formatCandidateRow (empty/non-empty eval & strategy, hidden columns).
• optimize_test.go: TestPrintOptimizeResults_StrategyDashForMixedMutations.
• Removed TestDetectEvalAgentKind, TestRelPathForYaml (functions deleted).
• Renamed TestOptimizeDeployCommand_AgentResolvedFromFlagOrYaml → …FromFlagOrProject.

[Copilot]

Pull request overview

This PR polishes the azd ai agent optimize command's live progress experience and adapts the eval/optimize commands to the newer azure.yaml inline agent-definition model. During polling it now renders a live-updating candidate results table (completed rows appended in place, with an overwritten in-progress status line) instead of a single spinner line, and it replaces the removed on-disk agent.yaml detection in eval with the shared projectpkg.LoadAgentDefinition, which understands inline, legacy-config, and disk sources.

Changes:

• Progressive polling UI: extracted shared table helpers (candidateTableHeader, formatCandidateRow, candidateDisplayName, writeCandidateRow), added a live candidate table with server-timestamp-based "Total time"/"Duration" output, softened the pre-run warning to a "Note", and bumped the default --poll-interval from 5s to 10s.
• azure.yaml adaptation: eval now resolves agent kind via LoadAgentDefinition (with legacy-shape warnings) and drops the local detectEvalAgentKind/relPathForYaml helpers; --agent help text is aligned across optimize/deploy/apply/eval to describe service-name vs Foundry-name semantics.
• Model additions: JobProgress.BaselineScore and a new InProgressCandidate (CreatedAt, CandidateGenerated) type for richer server progress reporting.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
internal/pkg/agents/optimize_api/models.go	Adds BaselineScore and InProgressCandidate to job progress payload.
internal/cmd/optimize.go	Live candidate table rendering during polling; softened warning; --poll-interval default 10; total-time output.
internal/cmd/optimize_helpers.go	New shared candidate-table rendering helpers; removes loadLastOptimizeJobID.
internal/cmd/optimize_status.go	Uses loadOptimizeJobIDForAgent; adds Updated/Duration to summary; poll default 10.
internal/cmd/optimize_deploy.go / optimize_apply.go	--agent help text and comments updated for azure.yaml semantics.
internal/cmd/eval.go	Replaces on-disk kind detection with LoadAgentDefinition; removes now-unused helpers.
internal/cmd/eval_generate.go	--agent help text updated.
*_test.go (optimize, status, helpers, deploy, eval)	Updates poll-interval assertions, adds table-helper and summary-duration tests, removes tests for deleted helpers.

[jongio]

Clean refactoring that consolidates agent kind detection into the shared LoadAgentDefinition utility and improves the optimize polling UX with a live-updating candidate table.

Key things I verified:

• loadOptimizeJobIDForAgent("", envName) correctly falls back to the global OPTIMIZE_LAST_OPERATION_ID key, matching the removed loadLastOptimizeJobID behavior.
• The resolveOptimizeAgent reordering (project-first, flag-as-Foundry-name fallback) is consistent with other commands and preserves backwards compatibility for users specifying a Foundry name directly.
• New table rendering helpers are well-tested and the extraction from the inline loop is a good separation of concerns.

One optional suggestion inline.

[jongio]

Incremental review of the new commit (1760b79, "support -o").

The JSON output support and the improved progress status line both look solid. One logic gap in the terminal-job path when --watch is combined with --output json.

[trangevi]

This looks good overall — just a couple of comments.