ohadmo · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026
diff --git a/.github/scripts/prune-orphans.sh b/.github/scripts/prune-orphans.sh
@@ -5,8 +5,8 @@
 # FORK of NVIDIA/skills' prune-orphans.sh, adapted for this repo's GROUPED,
 # multi-root layout. Differences from upstream:
 #   - Upstream scans a single flat `skills/*/` root. We scan the SOURCED group
-#     roots in PRUNE_ROOTS (the wholly-native roots nim-skills/ and workflows/
-#     are never scanned, so they can't be pruned).
+#     roots in PRUNE_ROOTS (the wholly-native roots nim-skills/, workflows/,
+#     and workflow-skills/ are never scanned, so they can't be pruned).
 #   - catalog_dir values are full grouped paths (e.g. library-skills/nvmolkit-usage),
 #     kept at DEPTH 1 under their group root, so the scan and the expected-set
 #     comparison both use full relative paths.

diff --git a/.github/workflows/plugin-sync.yml b/.github/workflows/plugin-sync.yml
@@ -14,6 +14,7 @@ on:
       - "library-skills/**"
       - "open-models-skills/**"
       - "workflows/**"
+      - "workflow-skills/**"
       - "skills/**"
       - "plugins"
       - "skills.sh.json"

diff --git a/.github/workflows/skill-security.yml b/.github/workflows/skill-security.yml
@@ -11,6 +11,7 @@ on:
       - "library-skills/**"
       - "open-models-skills/**"
       - "workflows/**"
+      - "workflow-skills/**"
       - ".github/workflows/skill-security.yml"
       - "scripts/scan_skills.py"
       - "scripts/plugin_sync.py"

diff --git a/.github/workflows/sync-skills.yml b/.github/workflows/sync-skills.yml
@@ -4,7 +4,7 @@
 # Sync skills from source repos into the bionemo-agent-toolkit catalog.
 # Adapted from NVIDIA/skills' sync-skills.yml, STRIPPED to the aggregation
 # core (no signing / compliance / drift / contact-resolution / tracker steps —
-# NVCARPS is deferred; the review PR is the v1 gate). Adapted for GROUPED
+# automated validation is deferred; the review PR is the v1 gate). Adapted for GROUPED
 # catalog dirs: catalog_dir already includes its group prefix
 # (e.g. library-skills/nvmolkit-usage), so the rsync target is $catalog_dir, not
 # skills/$catalog_dir.

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -34,8 +34,8 @@ This repo is an **aggregator**. Most skills are authored in *other* GitHub repos
 and **vendored** in by the nightly sync (`.github/workflows/sync-skills.yml`),
 which reads `components.d/*.yml`, clones each source repo at its declared `ref`,
 and `rsync`s the declared skill folders into this repo's grouped catalog dirs
-(`nim-skills/`, `library-skills/`, `open-models-skills/`, `workflows/`), then
-opens a PR a maintainer reviews.
+(`nim-skills/`, `library-skills/`, `open-models-skills/`, `workflows/`,
+`workflow-skills/`), then opens a PR a maintainer reviews.
 
 A skill lives here in exactly one of two ways:
 
@@ -44,7 +44,7 @@ A skill lives here in exactly one of two ways:
   every run (`rsync --delete`), so **do not edit a sourced skill dir in this
   repo** — the change reverts on the next sync. Fix it upstream.
 - **Native** — authored directly here and listed in `catalog-exceptions.yml`
-  (e.g. the NIM skills and the `workflows/` meta-skills).
+  (e.g. the NIM skills and the `workflows/` and `workflow-skills/` meta-skills).
 
 Every skill directory must be claimed by **exactly one** registry
 (`components.d` or `catalog-exceptions.yml`), or the orphan pruner
@@ -56,6 +56,12 @@ copy directly; run `python scripts/plugin_sync.py --write` after changing a
 source skill. The root `plugins -> skills` symlink is a temporary compatibility
 path for existing plugin consumers.
 
+The `protein-binder-design` source skill uses `workflow-skills/` because CI treats
+the `workflows/` prefix as a whole-workflow validation target. This lets CI
+validate the generated skill individually while plugin-sync and security checks
+cover its source. Both roots remain registered so existing workflow sources can
+stay in `workflows/` until they are migrated.
+
 ### Onboarding a source repo
 
 Add `components.d/<slug>.yml`, meet the requirements below, and open a PR:
@@ -89,7 +95,7 @@ the same conventions `NVIDIA/skills` uses, so a compliant repo is ingestible by
 - **SRC-8 · No `AGENTS.md` / `CLAUDE.md` inside a skill dir** (auto-loaded by some harnesses).
 - **SRC-9 · No stray non-skill files** in the skills container.
 
-**Compliance — enforced by `NVIDIA/skills`; deferred here (NVCARPS):**
+**Compliance — enforced by `NVIDIA/skills`; deferred here:**
 - **SRC-10 · Co-located `evals/evals.json`** per skill + a `BENCHMARK.md` report.
   Eval definitions for a **sourced** skill must live under `evals/` **in the
   source component repo** — never authored into the catalog copy here. The sync

diff --git a/README.md b/README.md
@@ -40,7 +40,7 @@ flow as well. Skills are also discoverable by partner harnesses directly from th
 
 | Product | Description | Skills |
 |---------|-------------|--------|
-| **Protein Binder Design** | End-to-end de novo binder design workflows — a NIM route and a Proteina-Complexa route. | [`protein-binder-design`](workflows/generative-protein-binder-design/protein-binder-design), [`complexa-binder-design`](workflows/generative-protein-binder-design/complexa-binder-design) |
+| **Protein Binder Design** | End-to-end de novo binder design workflows — a NIM route and a Proteina-Complexa route. | [`protein-binder-design`](workflow-skills/generative-protein-binder-design/protein-binder-design), [`complexa-binder-design`](workflows/generative-protein-binder-design/complexa-binder-design) |
 | **Generative Virtual Screening workflow** | Generate candidate molecules, dock them to a target, and score binding affinity (GenMol → DiffDock → Boltz-2). | [`drug-discovery-pipeline`](nim-skills/meta-skills/drug-discovery-pipeline) |
 | **MSA-enabled protein structure prediction workflow** | Build a multiple sequence alignment, then predict structure (MSA-Search → OpenFold3). | [`msa-structure-prediction-pipeline`](nim-skills/meta-skills/msa-structure-prediction-pipeline) |
 | **Boltz-2** | Biomolecular structure prediction + binding affinity (NIM). | [`boltz2-nim`](nim-skills/boltz2-nim) |

diff --git a/catalog-exceptions.yml b/catalog-exceptions.yml
@@ -2,9 +2,9 @@
 # registration. The forked prune-orphans.sh scans the SOURCED roots
 # (library-skills/, open-models-skills/) and deletes any dir there that is
 # neither declared in components.d/*.yml nor listed below. The wholly-native
-# roots (nim-skills/, workflows/) are not scanned at all (see PRUNE_ROOTS in
-# .github/scripts/prune-orphans.sh), so they need no entry — but nim-skills and
-# workflows are listed here for documentation.
+# roots (nim-skills/, workflows/, workflow-skills/) are not scanned at all
+# (see PRUNE_ROOTS in .github/scripts/prune-orphans.sh), so they need no entry
+# — but they are listed here for documentation.
 #
 # Add an entry only with a documented reason and an owner.
 exceptions:
@@ -17,6 +17,10 @@ exceptions:
     reason: Native meta-skill workflows (generative-protein-binder-design, virtual_screening, ai_q) authored in-repo.
     owner: trvachov
     component: Workflows
+  - dir: workflow-skills
+    reason: Native workflow skills migrated out of workflows/ for individual skill validation in CI.
+    owner: trvachov
+    component: Workflows
 
   # --- native skills inside a SCANNED (sourced) root — MUST be listed or the pruner deletes them ---
   - dir: library-skills/parabricks

diff --git a/plugins.d/bionemo-agent-toolkit.yml b/plugins.d/bionemo-agent-toolkit.yml
@@ -40,6 +40,7 @@ include_skills:
   - open-models-skills/proteina-complexa/
   # Workflows
   - workflows/generative-protein-binder-design/
+  - workflow-skills/generative-protein-binder-design/protein-binder-design/
 
 # Inherits version/author/homepage/repository/license/brand_color/capabilities
 # from plugins.d/_defaults.yml.
diff --git a/scripts/plugin_sync.py b/scripts/plugin_sync.py
@@ -12,7 +12,7 @@
 
   1. COVERAGE  — every distributable source skill is listed in ``skills.sh.json``.
   2. FRESHNESS — for every listed skill, the aggregate skill folder is an exact
-                 copy of the source skill folder (including ``evals/`` — NVCARPS
+                 copy of the source skill folder (including ``evals/`` —
                  Tier 3 requires the eval dataset in the validated payload).
 
 Modes:
@@ -42,15 +42,15 @@
 CONFIG = REPO / "skills.sh.json"
 
 # Source roots that hold distributable skills.
-SOURCE_ROOTS = ["nim-skills", "library-skills", "open-models-skills", "workflows"]
+SOURCE_ROOTS = ["nim-skills", "library-skills", "open-models-skills", "workflows", "workflow-skills"]
 # Path segments that never contain a distributable skill.
 EXCLUDE_SEGMENTS = {"plugins", "vendor", "evals", "node_modules", ".git"}
 # Files/dirs ignored when copying/comparing (never part of the payload).
 # `.skillsource.json` is skills.sh metadata that the generator strips from the payload.
 JUNK = {".DS_Store", "__pycache__", ".skillsource.json"}
 # Subdirectories of a skill that are stripped from the aggregate payload.
 # Empty: the aggregate is a full copy. ``evals/`` is intentionally kept because
-# NVCARPS Tier 3 (live agent eval) validates the aggregate and requires the
+# Tier 3 (live agent eval) validates the aggregate and requires the
 # eval dataset there — stripping it makes Tier 3 skip → coverage invalid → gate block.
 STRIP_FROM_AGGREGATE: set[str] = set()
 # Generated artifacts the NVSkills signing/eval pipeline writes into the aggregate

diff --git a/skills/bionemo-agent-toolkit/skills/protein-binder-design/BENCHMARK.md b/skills/bionemo-agent-toolkit/skills/protein-binder-design/BENCHMARK.md
@@ -0,0 +1,117 @@
+# Skill Benchmark: protein-binder-design
+
+> **Overall verdict: NEUTRAL — One or more dimensions remain below PASS**
+
+Live evaluation did not show a material gain or regression. Collect more evidence or improve the skill before making a publication decision.
+
+## Evaluation Metadata
+
+- Skill: `protein-binder-design`
+- Evaluation date: 2026-10-02
+- Evaluator version: `1.5.6`
+- Agents: Claude Code (`aws/anthropic/bedrock-claude-opus-5`), Codex (`openai/openai/gpt-5.5`)
+- Tasks: 1 evaluation tasks (1 positive)
+- Dataset digest: `sha256:109a2793e5be03aca71b3dc7844f3adb21799632e13edcf2fc73cc66cdfd1303` (skill-evaluator-dataset-snapshot/1)
+- Attempts per task: 1
+- Environment: `k8s-sandbox`
+- Tier 2 evidence: required for publication
+- Tier 3 evidence: required for publication
+
+Each task attempt ran in its own isolated sandbox pod.
+
+## What This Report Answers
+
+The three-tier evaluation checks whether the skill:
+
+- is safe to use;
+- produces correct answers;
+- is discovered and activated when needed;
+- helps the agent complete the user's goal and expected workflow; and
+- avoids wasted skill and tool usage.
+
+## Results at a Glance
+
+| Measure | Claude Code (Baseline → Skill Uplift) | Codex (Baseline → Skill Uplift) |
+|---|---:|---:|
+| Overall | 83.8% — baseline ran, but no comparable score was available; uplift unavailable | 85.5% — baseline ran, but no comparable score was available; uplift unavailable |
+| Security | 100.0% → 100.0% (±0.0 points) | 100.0% → 100.0% (±0.0 points) |
+| Correctness | 100.0% → 100.0% (±0.0 points) | 100.0% → 100.0% (±0.0 points) |
+| Discoverability | 35.0% — baseline ran, but no comparable score was available; uplift unavailable | 85.0% — baseline ran, but no comparable score was available; uplift unavailable |
+| Effectiveness | 100.0% → 100.0% (±0.0 points) | 100.0% → 58.3% (-41.7 points) |
+| Efficiency | 83.9% — baseline ran, but no comparable score was available; uplift unavailable | 84.2% — baseline ran, but no comparable score was available; uplift unavailable |
+
+**How to read this table:** baseline is the same task attempted without the target skill. Scores are rounded to one decimal; threshold-adjacent values use additional precision so their displayed band matches the verdict. Uplift is derived from those displayed scores and shown in percentage points.
+
+Example: `47.0% → 92.0% (+45.0 points)` means the skill-assisted run scored 92.0%, 45.0 percentage points above its 47.0% no-skill baseline.
+
+## Token Usage
+
+Actual Tier 3 execution usage is reported for every observed agent/case pair and both conditions.
+
+| Agent | Dataset case | With skill | Without skill | Delta | Change | Coverage |
+|---|---|---:|---:|---:|---:|---|
+| claude-code | All cases | 387,458 | 733,717 | -346,259 | -47.19% | skill 1/1; base 1/1 |
+| claude-code | offline-manifest | 387,458 | 733,717 | -346,259 | -47.19% | skill 1/1; base 1/1 |
+| codex | All cases | 204,321 | 182,661 | +21,660 | +11.86% | skill 1/1; base 1/1 |
+| codex | offline-manifest | 204,321 | 182,661 | +21,660 | +11.86% | skill 1/1; base 1/1 |
+| ALL AGENTS | Dataset aggregate | 591,779 | 916,378 | -324,599 | -35.42% | skill 2/2; base 2/2 |
+
+Prompt tokens include cached reads, so total tokens are `prompt + completion` (cached is not added twice). The Efficiency score uses `(prompt - cached) + completion`. N/A means the relevant trajectory counters were not available; coverage is never estimated.
+
+## Tier Status
+
+| Tier | Purpose | Status | Evidence |
+|---|---|---|---|
+| Tier 1 | Static validation | **PASSED WITH OBSERVATIONS** | 11 validator(s); 41 finding(s) |
+| Tier 2 | Semantic deduplication | **PASSED** | 2 validator(s); 0 finding(s) |
+| Tier 3 | Live agent evaluation | **NEUTRAL** | 2 agent(s); 1 task(s) |
+
+## Findings and Observations
+
+<details>
+<summary>Show detailed findings and successful checks</summary>
+
+- **MEDIUM** QUALITY/quality_correctness: No documented scripts in table format (`skills/bionemo-agent-toolkit/skills/protein-binder-design/SKILL.md`)
+- **MEDIUM** QUALITY/quality_correctness: Instructions don't mention 'run_script' (`skills/bionemo-agent-toolkit/skills/protein-binder-design/SKILL.md`)
+- **MEDIUM** QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.author' (`skills/bionemo-agent-toolkit/skills/protein-binder-design/SKILL.md`)
+- **MEDIUM** QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (`skills/bionemo-agent-toolkit/skills/protein-binder-design/SKILL.md`)
+- **MEDIUM** SCHEMA/folder_hierarchy: Unexpected nesting depth for general skill (`skills/bionemo-agent-toolkit/skills/protein-binder-design`)
+- 36 additional finding(s) are available in the full evaluation artifacts.
+
+</details>
+
+## Scoring Methodology
+
+<details>
+<summary>Show dimension definitions, source signals, and thresholds</summary>
+
+| Dimension | Question | Scored signals |
+|---|---|---|
+| Security | Is it safe to use? | `security` (100%) |
+| Correctness | Is the answer correct? | `accuracy` (100%) |
+| Discoverability | Was the right skill loaded when needed? | `skill_execution` (100%) |
+| Effectiveness | Did the skill help complete the task? | `goal_accuracy` (50%) + `behavior_check` (50%) |
+| Efficiency | Did it avoid wasted tool calls and token usage? | `skill_efficiency` (50%) + `token_efficiency` (50%) |
+
+- Dimension bands: PASS at 50% or above; NEUTRAL from 40% to below 50%; FAIL below 40%.
+- Overall Tier 3 lift: PASS at +5 points or more; FAIL at -10 points or less; values between those bands are NEUTRAL.
+- Overall verdict: PASS only when every configured dimension passes for at least one supported agent. Lift is reported as diagnostic evidence and does not override this gate.
+- The 50% attempt pass threshold is a separate per-task gate; it is not the dimension pass threshold.
+- Effectiveness is the equal-weight mean of goal completion (`goal_accuracy`) and expected workflow adherence (`behavior_check`).
+- Efficiency is 50% tool-call productivity (the backward-compatible `skill_efficiency` wire id) and 50% `token_efficiency`. Positive-case skill routing is scored under Discoverability, not Efficiency; a negative case without a routing target is N/A. N/A sources are omitted, remaining weights are renormalized, and the dimension is marked partial.
+
+Signals present in this run:
+
+- `security` (Security): unsafe operations, secret leakage, and unauthorized access.
+- `skill_execution` (Skill Execution): whether the expected skill was selected, decoys were avoided, and the workflow executed.
+- `skill_efficiency` (Tool Productivity): tool-call productivity (legacy wire id; routing is scored under Discoverability).
+- `accuracy` (Accuracy): final-answer correctness against the reference answer.
+- `goal_accuracy` (Goal Accuracy): whether the user's goal was achieved.
+- `behavior_check` (Behavior Check): whether the expected workflow behavior was followed.
+- `token_efficiency` (Token Efficiency): actual uncached prompt plus completion usage (50% of Efficiency).
+
+</details>
+
+## Freshness
+
+Regenerate this benchmark when the skill, evaluation dataset, target agent/model, evaluator version, environment, or scoring policy changes.
diff --git a/skills/bionemo-agent-toolkit/skills/protein-binder-design/README.md b/skills/bionemo-agent-toolkit/skills/protein-binder-design/README.md
@@ -20,7 +20,7 @@ protein-binder-design/
 │   ├── metrics.py           # Kabsch CA-RMSD (self-consistency)
 │   ├── controls.py          # scrambled negative controls
 │   └── registry.py          # target-registry loader
-├── assets/targets.json      # EXAMPLE target registry (verify before real use)
+├── assets/targets.json      # empty target-registry template (populate before use)
 └── evals/                   # trigger + assertion evals
 ```
 
@@ -30,7 +30,25 @@ protein-binder-design/
 - Access to the BioNeMo NIMs you intend to use (RFdiffusion, ProteinMPNN, Boltz2
   and/or OpenFold3, optional MSA-Search) — hosted at
   [build.nvidia.com](https://build.nvidia.com) or self-hosted via NGC.
-- Python ≥ 3.10 with `numpy` (`pip install numpy`). Scripts are otherwise stdlib.
+- Python 3.10-3.13 for the pinned helper environment. Install dependencies with
+  `python3 -m pip install -r requirements.txt`; mmCIF parsing and cached refold
+  recovery require Biotite. Manifest and PDB-only helpers use the standard library.
+
+## Offline tests
+
+Install `requirements-dev.txt` in a virtual environment, then run
+`python3 -m pytest tests`. To inspect coverage, run
+`python3 -m pytest --cov=scripts --cov-report=term-missing tests`.
+These tests use synthetic fixtures and credentials; they make no model calls.
+
+## Tier 1 validation
+
+CI selects `config/skillspector-baseline.yml` before the repository-wide
+baseline. Include this file, the requirements files, and the tests when committing
+the skill; synchronize the aggregate copy with the repository's plugin sync tool.
+The baseline records reviewed false positives as exact source fingerprints for
+SkillSpector 2.11.2. Source edits or a scanner upgrade require another review;
+the environment-copy finding is fixed in code and is not suppressed.
 
 ## Configure