refactor(eval-author): move the Eval Author agent into Experimentalist

packages/nemo_platform/pyproject.toml

@@ -295,12 +295,10 @@ nemo-deployments-plugin = [
 nemo-eval-author-plugin = [
   "pydantic>=2",
   "harbor>=0.18",
-  "nooa>=0.0.9",
   "nemo-experimentalist-plugin",
   "nemo-insights-plugin",
   "nemo-platform-plugin",
   "pyyaml>=6.0.3",
-  "tomlkit>=0.13.3",
 ]
 
 # Generated from [tool.bundle-package]; do not edit by hand.
@@ -342,7 +340,6 @@ nemo-experimentalist-plugin = [
   "protobuf>=6.0.0",
   "nooa>=0.0.9",
   "pyyaml>=6.0.3",
-  "nemo-eval-author-plugin",
   "nemo-insights-plugin",
   "nemo-platform-plugin",
   "tomlkit>=0.13.3",

plugins/nemo-eval-author/README.md

@@ -3,60 +3,24 @@
 
 # NeMo Eval Author Plugin
 
-Library-only plugin that owns the Eval Author agent (`eval_author/`).
+Owns the `nemo agents eval-author` command group, registered under `nemo.cli.agents` and
+mounted by the agents plugin. `discover` is implemented; `audit`, `propose`, `run`, and
+`doctor` are placeholders.
 
-## Direction of travel
-
-**Eval Author is meant to become standalone, with nothing imported from
-Experimentalist.** Prefer duplicating a helper over sharing one, even when sharing
-looks tidier.
+Use `discover` only with a trusted repository, because importing an agent runs
+module top-level code.
 
-Right now the two packages depend on each other:
+The Eval Author agent moved into the Experimentalist plugin, at
+[`nemo_experimentalist_plugin.eval_author`](../nemo-experimentalist/src/nemo_experimentalist_plugin/eval_author/README.md).
+Experimentalist insight mode is its only caller, so the agent sits beside the evaluator,
+staging, and trace helpers it depends on.
 
-| Arrow | Status | Why |
-| --- | --- | --- |
-| Experimentalist → Eval Author | permanent | insight mode imports `EvalAuthor` and `EvalAuthorConfig` at module scope |
-| Eval Author → Experimentalist | temporary | still borrows evaluator/Harbor, staging, trace, tools, cache, backend |
+## Direction of travel
 
-[`tests/test_plugin_boundary.py`](tests/test_plugin_boundary.py) pins the second list
-so it can only shrink, and names what each remaining import is still for. `uv`
-resolves the current cycle; install both packages with:
+The dependency is one arrow. `discovery/run.py` borrows `make_client` from Experimentalist,
+and Experimentalist imports nothing from here, so there is no package cycle for `uv` to
+resolve. Install both plugins with:
 
 ```bash
 uv sync --group experimentalist
 ```
-
-## Public API
-
-```python
-from nemo_eval_author_plugin.eval_author.agent import EvalAuthor, build_eval_author_agent
-from nemo_eval_author_plugin.eval_author.models import EvalAuthorConfig, EvalAuthorResult
-from nemo_eval_author_plugin.eval_author.run import run_eval_author
-
-# Still borrowed from Experimentalist, and on the way out. Treat these as Eval Author's
-# own types once they move; do not build new code on the Experimentalist paths.
-from nemo_experimentalist_plugin.entities import Dataset, DatasetRef
-from nemo_experimentalist_plugin.experimentalist.components.dataset_staging import stage_eval_author_inputs
-from nemo_experimentalist_plugin.experimentalist.components.trace_analyzer import TraceAnalyzer
-from nemo_experimentalist_plugin.experimentalist.components.trace_explorer import TraceExplorer
-```
-
-## Agent models
-
-Run `nemo setup` and select the default and fast models for the active Platform
-context. Eval Author uses the default model for authoring and the fast model for
-summarization. Press Enter at the fast-model prompt to reuse the default model.
-
-The selections are workspace-qualified Platform Model Entity IDs. The Platform
-routes each request to the provider registered for that entity and reads its
-credential from Platform Secrets; Eval Author does not accept separate provider,
-endpoint, key, or model environment variables.
-
-For non-interactive and isolated environments, `NEMO_DEFAULT_MODEL` and
-`NEMO_FAST_MODEL` can override the stored selections. Values must still use
-`workspace/model-name` and refer to Model Entities on the target Platform.
-
-A `nemo agents eval-author` CLI is registered under `nemo.cli.agents` and
-mounted by the agents plugin. `discover` is implemented; `audit`, `propose`,
-`run`, and `doctor` remain placeholders. The library runner already uses the
-configured Platform model pair.

plugins/nemo-eval-author/pyproject.toml

@@ -4,19 +4,17 @@
 [project]
 name = "nemo-eval-author-plugin"
 version = "0.1.0"
-description = "Eval Author agent for NeMo Platform (hard-depends on Experimentalist for evaluator/trace helpers)."
+description = "Eval Author commands for NeMo Platform (borrows the Experimentalist platform client)."
 requires-python = ">=3.12,<3.14"
 dependencies = [
     "pydantic>=2",
     # Harbor 0.18 provides the discovery APIs used by this plugin.
     "harbor>=0.18",
-    "nooa>=0.0.9",
     "nemo-experimentalist-plugin",
     "nemo-insights-plugin",
     "nemo-platform",
     "nemo-platform-plugin",
     "pyyaml>=6.0.3",
-    "tomlkit>=0.13.3",
 ]
 
 [project.entry-points."nemo.cli.agents"]

plugins/nemo-experimentalist/AGENTS.md

@@ -13,6 +13,27 @@ Inherited from the NeMo Platform monorepo that now hosts this plugin:
 
 ## Active migrations
 
+### 2026-08-19: Eval Author agent moved back into this plugin
+
+`nemo_experimentalist_plugin.eval_author` owns the Eval Author agent again, along with
+`eval_author/traces.py`. The 2026-07-28 extraction aimed at a standalone Eval Author
+plugin. That goal is retired: the customer-facing path is a skill rather than a plugin
+install, and insight mode was the agent's only caller.
+
+**The dependency is one arrow, and it points out of the Eval Author plugin.**
+Experimentalist imports nothing from `nemo-eval-author-plugin`, so the package cycle that
+`uv` had to resolve is gone.
+
+- The agent's borrows of the evaluator, staging, trace, tools, and reporting helpers are
+  intra-package imports. The Eval Author plugin's boundary ratchet existed to drive those
+  ten borrows to zero, so it went away with them.
+- `tests/test_contract_dependency.py` asserts that this plugin never declares
+  `nemo-eval-author-plugin` as a dependency, which is what keeps the cycle broken.
+- `plugins/nemo-eval-author/` keeps the `nemo agents eval-author` command group and its
+  `discovery/` package, whose one remaining borrow is `make_client`.
+- Agent tests live in `tests/eval_author/`. This plugin's `conftest.py` already covers the
+  isolation those tests need, so the Eval Author copy went away.
+
 ### 2026-07-31: Command group nested under `nemo agents`
 
 The only path is `nemo agents experimentalist <verb>`. `ExperimentalistCLI` is
@@ -25,28 +46,17 @@ Analyst and Eval Author follow the same rule: `nemo agents analyst run` (was
 `ctx.command_path` over a hardcoded path when a message quotes the command back
 to the user.
 
-### 2026-07-28: Eval Author extracted to its own plugin, heading for standalone
-
-`plugins/nemo-eval-author/` (`nemo-eval-author-plugin`) owns the Eval Author agent package
-(`eval_author/`).
+### 2026-07-28: Eval Author extracted to its own plugin, heading for standalone (superseded)
 
-**The target is one arrow: Experimentalist → Eval Author.** Eval Author is meant to stop
-depending on Experimentalist entirely, even where that means duplicating code. Today the
-arrow points both ways:
+Superseded by the 2026-08-19 entry, which moved the agent back.
 
-- Experimentalist → Eval Author is permanent. Insight mode imports `EvalAuthor` in
-  `components/loop.py` and `EvalAuthorConfig` in `resolve.py`, both at module scope, so
-  the dependency is declared in `pyproject.toml`.
-- Eval Author → Experimentalist is temporary. It still borrows the evaluator/Harbor
-  abstractions, dataset staging, trace analyzer/explorer, `GuardedShellTools`, the cache,
-  and the backend factory.
+`plugins/nemo-eval-author/` took ownership of the `eval_author/` package, targeting a
+standalone plugin that imported nothing from Experimentalist. Both arrows pointed at each
+other for the duration, and `uv` resolved the cycle.
 
-`plugins/nemo-eval-author/tests/test_plugin_boundary.py` pins that second list so it can
-only shrink. **Do not share Eval Author implementation helpers with Experimentalist.**
-The Platform-owned model client integration is intentionally different: both plugins use
-`nemo_platform_plugin.nooa_model_client` so provider routing, Platform authentication,
-and configured model selection have one owner outside either plugin. `uv` resolves the
-remaining package cycle; install both with `uv sync --group experimentalist`.
+One rule from that entry still holds: both plugins use
+`nemo_platform_plugin.nooa_model_client`, so provider routing, Platform authentication,
+and configured model selection have one owner outside either plugin.
 
 ### 2026-07-24: Optimizer renamed to Experimentalist
 
@@ -87,7 +97,7 @@ including components registered by a separately installed package.
 The Curator agent was renamed directly to Eval Author in ASE-643. This is a
 breaking rename with no compatibility aliases or migration layer:
 
-- `nemo_experimentalist_plugin.curator` → `nemo_eval_author_plugin.eval_author` (was `nemo_experimentalist_plugin.eval_author` before extraction)
+- `nemo_experimentalist_plugin.curator` → `nemo_experimentalist_plugin.eval_author` (briefly `nemo_eval_author_plugin.eval_author` while the agent had its own plugin)
 - `Curator`, `CuratorConfig`, and `CuratorResult` → `EvalAuthor`,
   `EvalAuthorConfig`, and `EvalAuthorResult`
 - `run_curator(...)` and `build_curator_agent(...)` →
@@ -112,7 +122,7 @@ instead of restoring Curator imports, configuration, or aliases. The obsolete
 - `nooa` comes from PyPI. This plugin's `pyproject.toml` declares the floor
   (`nooa>=0.0.9`), which is the first release carrying the callable
   `@strategy(llm=...)` support the components depend on. When raising the floor,
-  move it in this plugin, `nemo-insights`, `nemo-eval-author`, and
+  move it in this plugin, `nemo-insights`, and
   `examples/tau3-nooa-agent/pyproject.toml`, update the tag quoted in
   `framework-skills/nooa/SKILL.md`, and relock both lock files together. Keep the
   Platform-supplied Insights plugin separate.

plugins/nemo-experimentalist/README.md

@@ -76,8 +76,11 @@ single leader, so complementary strengths stay alive across rounds.
   failure pattern inferred from traces — that the Experimentalist optimizes
   against. Run `nemo agents analyst run` first. The Experimentalist does not
   analyze traces or host an Insight API.
-- [NeMo Eval Author](../nemo-eval-author/README.md) builds the
-  Insight-specific evaluation suite, invoked automatically in Insight mode.
+- [Eval Author](src/nemo_experimentalist_plugin/eval_author/README.md) builds the
+  Insight-specific evaluation suite, and Insight mode invokes it automatically. It
+  ships inside this plugin; the separate
+  [Eval Author plugin](../nemo-eval-author/README.md) owns the
+  `nemo agents eval-author` commands.
 - **Harbor** runs the task containers that score every candidate.
 - **NeMo Experiments** mirrors each run and its candidates as an experiment
   group, so the lineage is visible in Studio. Structure only — rewards and

plugins/nemo-experimentalist/pyproject.toml

@@ -15,7 +15,6 @@ dependencies = [
     "protobuf>=6.0.0",
     "nooa>=0.0.9",
     "pyyaml>=6.0.3",
-    "nemo-eval-author-plugin",
     "nemo-insights-plugin",
     "nemo-platform",
     "nemo-platform-plugin",

plugins/nemo-experimentalist/src/nemo_experimentalist_plugin/config.py

@@ -19,8 +19,8 @@
 from pathlib import Path
 from typing import Any, Self
 
-from nemo_eval_author_plugin.eval_author.models import EvalAuthorConfig
 from nemo_experimentalist_plugin.entities import MetricTarget
+from nemo_experimentalist_plugin.eval_author.models import EvalAuthorConfig
 from nemo_experimentalist_plugin.experimentalist.components.models import (  # noqa: F401 - re-exported
     has_metric_dimensions,
     pareto_objectives,