feat(frontend): migrate evaluator run invocation

[mmabrouk]

Summary

• add workflow invoke helper for SimpleEvaluator execution
• switch evaluator playground DebugSection to call /preview/workflows/invoke
• update migration plan/status docs for PR2 run migration

Testing

• pnpm lint-fix (warns about in web/oss/src/components/Editor/plugins/code/nodes/Base64Node.tsx)

---

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Feb 13, 2026 8:04pm

[mmabrouk]

Decisions & Tradeoffs

• Use /preview/workflows/invoke from the evaluator playground instead of the legacy /evaluators/{key}/run. This matches the new backend path and keeps run logic aligned with SimpleEvaluator.
• Fallback URI when evaluator not saved: if no config exists yet, build via . This keeps Run Evaluator working for unsaved drafts.
• **Send parameters in both and **. The backend accepts both; this keeps compatibility with how workflow service requests are handled today.
• Minimal response typing: only surface and since UI currently only needs evaluator outputs.

Alternatives considered

• Keep legacy run endpoint (): lower change risk but blocks full migration; would keep old adapter path around.
• Require saved evaluator config before run: strict but would break current UX; reduces ambiguity around URI.
• **Only pass **: cleaner request but might miss expectations in existing workflow data handlers.
• Expose trace_id/span_id and surface in UI: more debugging power but requires UI changes.

If you prefer a different tradeoff (e.g., require saved configs or drop ), I can adjust this PR.

[devin-ai-integration]

Devin Review found 1 potential issue.

View 6 additional findings in Devin Review.

[mmabrouk]

@jp-agenta

We kept backend hydration on purpose, even after moving schema ownership to the frontend.
This PR makes the frontend send data.schemas.outputs when it can determine the schema at configure time. For fixed evaluators, the frontend reads outputs_schema from GET /evaluators. For auto_ai_critique, it derives the schema from parameters.json_schema.schema. For json_multi_field_match, it derives the schema from parameters.fields. If an evaluator has no known schema in the template, the frontend does not send one.
We still keep _ensure_builtin_evaluator_data in SimpleEvaluatorsService as a fallback. This protects non-UI callers and older payloads that still send only uri and parameters. It also keeps behavior compatible with the legacy flow, where builtin schemas were derived on the backend. In short, frontend schema sending is now the primary path, and backend hydration is a safety net.
This also means we did not remove existing behavior for old evaluators. We added a safer default while we complete the migration and verify all clients send schemas consistently.