ux(community): split Agent Supermarket into LLM and Open Agent shelves

dashboard/backend/domain/agents/marketplace.py

@@ -1,8 +1,9 @@
 """Open agent templates for the Agent Supermarket.
 
-Templates are defined in ``dashboard/config/marketplace.json`` so baseline open
-agents can be added without schema migrations. The listing is public; cloning
-creates a user-owned built-in agent with the template's pipeline copied in.
+Templates are defined in ``dashboard/config/marketplace.json`` so competition
+models and hosted agents can be added without schema migrations. The listing
+is public; cloning creates a user-owned built-in agent with the template's
+pipeline copied in.
 """
 
 from __future__ import annotations
@@ -11,14 +12,29 @@
 from functools import lru_cache
 from typing import Any, Dict, List, Optional
 
-from dashboard.backend.domain.agents.taxonomy import (
-    category_sort_rank,
-    normalize_category,
-)
+from dashboard.backend.domain.agents.taxonomy import normalize_category
 from dashboard.backend.paths import CONFIG_DIR
 
 _MARKETPLACE_PATH = CONFIG_DIR / "marketplace.json"
 
+# Community supermarket rows. Declared order is display order: LLMs first,
+# then Open Agents. Unknown / omitted values fall through ``_normalize_shelf``.
+MARKETPLACE_SHELVES = ("llms", "open")
+
+
+def _normalize_shelf(raw: Dict[str, Any]) -> str:
+    """Return ``llms`` or ``open``.
+
+    Explicit ``shelf`` on the catalog row wins. Otherwise a non-pipeline
+    runtime (today: AI Hedge Fund) is an open agent, so a future hosted
+    project does not have to remember the field to land on the right row.
+    """
+    explicit = str(raw.get("shelf") or "").strip().lower()
+    if explicit in MARKETPLACE_SHELVES:
+        return explicit
+    runtime_type = str(raw.get("runtime_type") or "pipeline")
+    return "open" if runtime_type != "pipeline" else "llms"
+
 
 def _public_template(raw: Dict[str, Any]) -> Dict[str, Any]:
     # This "category" and an agent's "category" used to be two different
@@ -47,6 +63,8 @@ def _public_template(raw: Dict[str, Any]) -> Dict[str, Any]:
         "author": raw.get("author") or "Community",
         "runtime_type": runtime_type,
         "step_count": step_count,
+        "shelf": _normalize_shelf(raw),
+        "card_subtitle": str(raw.get("card_subtitle") or "").strip() or None,
         "mode": (
             "runtime"
             if runtime_type != "pipeline"
@@ -82,18 +100,17 @@ def _load_catalog() -> Dict[str, Dict[str, Any]]:
 
 
 def list_marketplace_templates() -> List[Dict[str, Any]]:
-    """Return public marketplace cards grouped by shelf, then sorted by name.
+    """Return public marketplace cards grouped by supermarket shelf.
 
-    Ordered by ``category_sort_rank`` rather than by the slug itself: the slugs
-    are not alphabetical in shelf order, so a plain ``sorted`` on the raw value
-    leads the Community listing with the A-share shelf. Uncategorized templates
-    sort last.
+    ``MARKETPLACE_SHELVES`` declaration order is the page order (LLMs, then
+    Open Agents). Within a shelf the catalog's insertion order is preserved
+    (stable sort) so Community can list models in the leaderboard roster
+    order without a second sort key.
     """
     items = [_public_template(raw) for raw in _load_catalog().values()]
-    return sorted(
-        items,
-        key=lambda t: (category_sort_rank(t.get("category")), str(t.get("name") or "")),
-    )
+    # _public_template always sets "shelf" via _normalize_shelf, which only
+    # returns members of MARKETPLACE_SHELVES -- no fallback branch needed.
+    return sorted(items, key=lambda t: MARKETPLACE_SHELVES.index(t["shelf"]))
 
 
 def get_marketplace_template(template_id: str) -> Optional[Dict[str, Any]]:

dashboard/backend/tests/test_admin_analytics_frontend.py

@@ -190,8 +190,8 @@ def test_app_lifecycle_and_cache_versions_are_wired():
     assert "window.AdminAnalytics.syncAuth(user)" in APP_JS
     assert "window.AdminAnalytics.onEnter()" in APP_JS
     assert "window.AdminAnalytics.refresh()" in APP_JS
-    assert 'styles.css?v=130' in APP_HTML
-    assert 'app.js?v=124' in APP_HTML
+    assert 'styles.css?v=131' in APP_HTML
+    assert 'app.js?v=126' in APP_HTML
     assert 'js/admin-analytics.js?v=2' in APP_HTML
     assert 'js/admin-tabs.js?v=3' in APP_HTML
 

dashboard/backend/tests/test_agent_starter_defaults.py

@@ -175,10 +175,30 @@ def test_create_still_returns_the_one_time_api_key(client):
     assert body["session_id"]
 
 
-def test_marketplace_clone_keeps_its_own_pipeline(client):
-    """A template with its own pipeline must not be overwritten by the seed."""
+def test_marketplace_clone_keeps_its_own_pipeline(client, monkeypatch):
+    """A template with its own pipeline must not be overwritten by the seed.
+
+    The shipped catalog is now one-step competition-model cards plus a hosted
+    runtime, so the three-step shape is faked — the guard is the copy, not
+    which template still ships it.
+    """
+    fake_template = {
+        "template_id": "three-step-template",
+        "name": "Three-Step Template",
+        "model_name": "anthropic/claude-haiku-4-5",
+        "pipeline": [
+            {"id": "sub_gather", "presetKey": "info_gather", "prompt": "gather"},
+            {"id": "sub_signal", "presetKey": "info_to_signal", "prompt": "signal"},
+            {"id": "sub_exec", "presetKey": "signal_to_execution", "prompt": "exec"},
+        ],
+    }
+    monkeypatch.setattr(
+        marketplace_module,
+        "get_marketplace_template",
+        lambda template_id: fake_template if template_id == "three-step-template" else None,
+    )
     cloned = client.post(
-        "/api/v1/agents/marketplace/pipeline-analyst/clone",
+        "/api/v1/agents/marketplace/three-step-template/clone",
         json={},
         headers={"X-Session-Id": str(uuid.uuid4())},
     )

dashboard/backend/tests/test_agents_api.py

@@ -735,7 +735,7 @@ def test_marketplace_listing_and_clone(client):
     assert listing.status_code == 200
     templates = listing.json()["templates"]
     assert templates
-    assert any(t["template_id"] == "balanced-starter" for t in templates)
+    assert any(t["template_id"] == "claude-haiku-4-5" for t in templates)
     hedge_fund_card = next(
         t for t in templates if t["template_id"] == "ai-hedge-fund"
     )
@@ -751,13 +751,13 @@ def test_marketplace_listing_and_clone(client):
     browser_session = str(uuid.uuid4())
     headers = {"X-Session-Id": browser_session}
     cloned = client.post(
-        "/api/v1/agents/marketplace/balanced-starter/clone",
+        "/api/v1/agents/marketplace/claude-haiku-4-5/clone",
         json={},
         headers=headers,
     )
     assert cloned.status_code == 200
     agent = cloned.json()["agent"]
-    assert agent["name"] == "Balanced Starter"
+    assert agent["name"] == "Claude Haiku 4.5"
     assert agent["agent_type"] == "builtin"
     assert agent.get("pipeline")
     assert agent["pipeline"][0]["presetKey"] == "simple_instruction"
@@ -825,43 +825,33 @@ def test_marketplace_catalog_shape():
         )
 
     # "Pipeline" is banned product-copy vocabulary (glossary: pipeline ->
-    # "multi-step strategy"); the template_id stays "pipeline-analyst" since
-    # it's an API identifier baked into clone URLs, but the display name --
-    # the card's largest text -- must not carry the word.
+    # "multi-step strategy"). The hosted card and the competition-model cards
+    # must not put that word in the display name -- the card's largest text.
     names = {t["template_id"]: t["name"] for t in templates}
-    assert names["pipeline-analyst"] == "Three-Step Analyst"
-    assert "Pipeline Analyst" not in names.values()
+    assert names["ai-hedge-fund"] == "AI Hedge Fund"
+    assert "Pipeline" not in " ".join(names.values())
 
 
 def test_marketplace_listing_is_ordered_by_shelf_not_by_slug():
-    """Community cards group by market in *declaration* order, not slug order.
-
-    The recategorization onto slugs quietly changed which card leads the page:
-    ``sorted`` on the raw value orders cn_ashares < us_stocks, so the A-share
-    template became card #1 on a U.S.-focused product. Nothing caught it because
-    no test asserted order. ``category_sort_rank`` keys on the AgentCategory
-    Literal's declaration order instead, which is also the order MARKET_LABELS
-    renders the market chips in, so the two surfaces agree.
+    """Community cards group by supermarket shelf, LLMs then Open Agents.
+
+    Within a shelf the catalog's insertion order is preserved so the LLM
+    row can follow the leaderboard roster instead of alphabetical names.
     """
     import dashboard.backend.domain.agents.marketplace as marketplace_mod
-    from dashboard.backend.domain.agents.taxonomy import (
-        AGENT_CATEGORY_ORDER,
-        category_sort_rank,
-    )
 
     marketplace_mod.reload_marketplace_catalog()
     templates = marketplace_mod.list_marketplace_templates()
 
-    ranks = [category_sort_rank(t.get("category")) for t in templates]
-    assert ranks == sorted(ranks), "templates are not grouped in shelf order"
-
-    # The U.S. market leads; uncategorized templates never do.
-    assert templates[0]["category"] == AGENT_CATEGORY_ORDER[0] == "us_stocks"
-    assert templates[-1]["category"] == "cn_ashares"
-
-    # Within a shelf, still by name.
-    us_stocks = [t["name"] for t in templates if t["category"] == "us_stocks"]
-    assert us_stocks == sorted(us_stocks)
+    shelves = [t.get("shelf") for t in templates]
+    assert set(shelves) <= set(marketplace_mod.MARKETPLACE_SHELVES)
+    llms = [t for t in templates if t["shelf"] == "llms"]
+    opens = [t for t in templates if t["shelf"] == "open"]
+    assert llms, "the LLM shelf is empty"
+    assert opens, "the Open Agents shelf is empty"
+    assert templates[0]["shelf"] == "llms"
+    assert templates[-1]["shelf"] == "open"
+    assert [t["name"] for t in opens] == ["AI Hedge Fund"]
 
 
 def test_uncategorized_templates_sort_last_and_carry_no_fake_shelf():
@@ -1416,7 +1406,7 @@ def test_builtin_listing_echoes_category(client):
 def test_clone_honours_a_model_name_override(client):
     """Community's "Choose model" affordance clones a template onto another model."""
     cloned = client.post(
-        "/api/v1/agents/marketplace/balanced-starter/clone",
+        "/api/v1/agents/marketplace/claude-haiku-4-5/clone",
         json={"model_name": "deepseek/deepseek-v4-pro"},
         headers={"X-Session-Id": str(uuid.uuid4())},
     )
@@ -1429,7 +1419,7 @@ def test_clone_falls_back_to_the_template_model(client, blank):
     """Omitted or blank means "use the template's model", not "use empty"."""
     body = {} if blank is None else {"model_name": blank}
     cloned = client.post(
-        "/api/v1/agents/marketplace/balanced-starter/clone",
+        "/api/v1/agents/marketplace/claude-haiku-4-5/clone",
         json=body,
         headers={"X-Session-Id": str(uuid.uuid4())},
     )
@@ -1441,7 +1431,7 @@ def test_clone_does_not_validate_the_model_name(client):
     """No whitelist here: POST /agents and PATCH /agents/{id} don't have one either,
     and a Literal would drag in the openapi enum deploy gate #313 discharged."""
     cloned = client.post(
-        "/api/v1/agents/marketplace/balanced-starter/clone",
+        "/api/v1/agents/marketplace/claude-haiku-4-5/clone",
         json={"model_name": "some/unreleased-model"},
         headers={"X-Session-Id": str(uuid.uuid4())},
     )

dashboard/backend/tests/test_analytics_frontend.py

@@ -41,7 +41,7 @@ def _function_body(source: str, signature: str) -> str:
 
 
 def test_analytics_script_loads_between_app_and_page_scripts():
-    app_at = APP_HTML.index('<script src="app.js?v=124" defer></script>')
+    app_at = APP_HTML.index('<script src="app.js?v=126" defer></script>')
     analytics_at = APP_HTML.index(
         '<script src="js/analytics.js?v=1" defer></script>'
     )

dashboard/backend/tests/test_app_copy_register.py

@@ -403,12 +403,12 @@ def test_marketplace_mode_chip_labels_avoid_banned_words():
     `renderMarketplaceGrid` so this doesn't over-match an unrelated string
     elsewhere in app.js.
     """
-    body = fn_body("function renderMarketplaceGrid")
+    body = fn_body("function buildMarketplaceCardHtml")
     assert "'Hosted runtime'" not in body
     assert "'Multi-step pipeline'" not in body
-    assert "'Hosted'" in body
-    assert "'Multi-step strategy'" in body
-    assert "'Simple instruction'" in body
+    assert "Open Source" in body
+    assert "'Simple instruction'" not in body
+    assert "'Multi-step strategy'" not in body
 
 
 def test_each_agent_gets_its_own_trading_session_is_gone():

dashboard/backend/tests/test_backtest_comparison_frontend.py

@@ -191,8 +191,8 @@ def test_exact_raw_ties_mark_every_tied_series_best():
 
 def test_comparison_script_and_semantic_table_ship_before_app():
     helper = '<script src="js/backtest-comparison.js?v=1" defer></script>'
-    app = '<script src="app.js?v=124" defer></script>'
-    assert 'href="styles.css?v=130"' in APP_HTML
+    app = '<script src="app.js?v=126" defer></script>'
+    assert 'href="styles.css?v=131"' in APP_HTML
     assert APP_HTML.index(helper) < APP_HTML.index(app)
     for element_id in (
         "performanceLegend",

dashboard/backend/tests/test_frontend_fast_boot.py

@@ -191,9 +191,9 @@ def test_cache_busters_bumped():
     # the next bump, so the exact one looks like the broken guard and gets
     # "fixed" by loosening it. That collision has already cost this repo one
     # round of follow-ups (#347/#348).
-    assert "app.js?v=124" in APP_HTML
+    assert "app.js?v=126" in APP_HTML
     assert "js/agent-editor.js?v=30" in APP_HTML
-    assert "styles.css?v=130" in APP_HTML
+    assert "styles.css?v=131" in APP_HTML
     assert "js/leaderboard.js?v=32" in APP_HTML
     assert "home-page.js?v=50" in APP_HTML
     assert "js/credit-format.js?v=1" in APP_HTML