Advertise the full Chat model catalog, not just the SDK enum

README.md

@@ -50,7 +50,7 @@ from openai import OpenAI
 
 client = OpenAI()  # picks up the env vars above
 r = client.chat.completions.create(
-    model="claude-sonnet-4-6",
+    model="claude-sonnet-5",
     messages=[{"role": "user", "content": "Explain TEE attestation in one line."}],
 )
 print(r.choices[0].message.content)
@@ -75,7 +75,7 @@ With `og-veil` running, set a custom endpoint — either via the CLI:
 hermes config set model.provider custom
 hermes config set model.base_url http://127.0.0.1:11434/v1
 hermes config set OPENAI_API_KEY og-veil      # ignored; your Chat login authenticates
-hermes config set model.default claude-sonnet-4-6
+hermes config set model.default claude-sonnet-5
 ```
 
 …or by editing `~/.hermes/config.yaml` directly:
@@ -91,6 +91,43 @@ model:
 Now `hermes` runs against attested, end-to-end-encrypted inference with no other
 changes. Confirm it's flowing through the enclave with `og-veil status`.
 
+## Models
+
+The same lineup [OpenGradient Chat](https://chat.opengradient.ai) offers, reached
+through the same verified TEE path. `og-veil models` prints it (add `--all` for
+superseded ones), and `GET /v1/models` serves it to any agent that probes.
+
+| Provider | Models |
+|----------|--------|
+| Anthropic | `claude-haiku-4-5`, `claude-sonnet-5`, `claude-opus-5`, `claude-fable-5` |
+| OpenAI | `gpt-4.1-nano`, `gpt-5-mini`, `gpt-5`, `gpt-5.6-luna`, `gpt-5.6-terra`, `gpt-5.6-sol` |
+| Google | `gemini-3.6-flash`, `gemini-3.5-flash-lite`, `gemini-2.5-pro`, `gemini-2.5-flash-lite` |
+| xAI | `grok-4.5`, `grok-4.3`, `grok-4-fast`, `grok-4` |
+| DeepSeek | `deepseek-v4-flash`, `deepseek-v4-pro` |
+| ByteDance | `seed-2.0-lite`, `seed-1.8`, `seed-1.6` |
+| Nous Research | `hermes-4-405b`, `hermes-4-70b` |
+| Z.ai | `glm-5.2` |
+| Image generation | `gpt-image-2`, `gemini-3.1-flash-image`, `seedream-5.0-lite`, `seedance-4.5`, `glm-image` |
+
+Names are **bare** - send `claude-opus-5`, not `anthropic/claude-opus-5`. Earlier
+generations (`claude-opus-4-8`, `gpt-5.5`, `grok-4-1-fast`, `o3`, …) stay
+routable and stay listed, so an agent already configured for one keeps working.
+
+**Image models are called like any other chat completion** and the images come
+back on the response — but note what the guarantee covers: the request is
+OHTTP-sealed exactly as a text prompt is, so the privacy story is unchanged,
+while the enclave's signature covers the assistant text, *not* the generated
+image bytes. Those ride out-of-band and are unsigned (see the SDK's
+`response_content_for_hash`). So `X-OpenGradient-Verified: true` on an
+image-generation response attests the exchange, not the pixels.
+
+**No video models.** Chat's Video Studio (Veo, Kling, Seedance, Wan) doesn't run
+over this path at all — those requests go to chat-api, which holds the
+fal/BytePlus/Atlas keys and calls those providers directly, with no enclave, no
+OHTTP and no signature anywhere in the chain. Listing them here would advertise
+a guarantee that doesn't exist. It needs the gateway to proxy video generation
+in-enclave first.
+
 ---
 
 ## How it works
@@ -147,7 +184,8 @@ the local OpenAI-compatible server.
 | `og-veil stop` | Stop the background server. |
 | `og-veil restart` | Stop and start the background server — e.g. after `og-veil update`. |
 | `og-veil status` | Login + network config + whether the server is running. |
-| `og-veil test ["prompt"]` | Send a one-off prompt to the running server and print the verified reply. Automatically targets the port the server was started on; override with `--host`/`--port`. |
+| `og-veil test ["prompt"]` | Send a one-off prompt to the running server and print the verified reply. Automatically targets the port the server was started on; override with `--host`/`--port`. Pick a model with `--model`. |
+| `og-veil models [--all]` | List the models you can send, grouped by provider. |
 | `og-veil update` | Update og-veil to the latest version. |
 | `og-veil login` | Authorize / re-authorize this device. |
 | `og-veil setup` | Re-run the setup wizard. |

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "opengradient-veil"
-version = "0.2.9"
+version = "0.3.0"
 description = "OpenGradient Veil — a drop-in, self-verifying private-inference proxy for AI agents. Point your OpenAI SDK at it; it routes prompts through OpenGradient's decentralized network of attestable Nitro TEE gateways and cryptographically verifies every response before a single token reaches your code."
 readme = "README.md"
 requires-python = ">=3.11"

tests/test_cli.py

@@ -65,14 +65,29 @@ def test_env_prints_env_vars():
     assert result.exit_code == 0
 
 
-def test_models_lists_known_models():
+def test_models_lists_the_catalog_grouped_by_provider():
     result = CliRunner().invoke(cli.main, ["models"])
     assert result.exit_code == 0
-    # Should print at least one model name derived from the SDK's TEE_LLM enum.
+    assert "Anthropic" in result.output
+    assert "claude-opus-5" in result.output
+    assert "hermes-4-405b" in result.output
+    # Superseded models are held back for --all.
+    assert "claude-opus-4-8" not in result.output
+
+
+def test_models_all_adds_superseded_and_sdk_only_models():
+    result = CliRunner().invoke(cli.main, ["models", "--all"])
+    assert result.exit_code == 0
+    assert "claude-opus-4-8" in result.output
+    # An id only the SDK's enum knows about still gets listed.
     from opengradient import TEE_LLM
 
-    sample = next(iter(TEE_LLM)).value.split("/", 1)[1]
-    assert sample in result.output
+    from veil import models as catalog
+
+    curated = {m.id for m in catalog.CATALOG}
+    sdk_only = sorted({m.value.split("/", 1)[1] for m in TEE_LLM} - curated)
+    if sdk_only:
+        assert sdk_only[0] in result.output
 
 
 def test_test_command_posts_prompt_to_localhost_and_prints_reply():

tests/test_models.py

@@ -0,0 +1,92 @@
+"""The advertised model catalog.
+
+The point of these is that the listing can only ever grow: the catalog mirrors
+what the Chat app offers, the SDK's ``TEE_LLM`` enum is folded in on top, and
+neither source can silently drop a model an agent is already configured for.
+"""
+
+from __future__ import annotations
+
+from opengradient import TEE_LLM
+
+from veil import models
+
+
+def test_catalog_ids_are_bare_wire_names():
+    # The gateway is sent `claude-opus-5`, never `anthropic/claude-opus-5` --
+    # both the SDK and the Chat app strip the provider prefix before the wire.
+    for model in models.CATALOG:
+        assert "/" not in model.id, model.id
+
+
+def test_catalog_ids_are_unique():
+    ids = [m.id for m in models.CATALOG]
+    assert len(ids) == len(set(ids))
+
+
+def test_every_provider_has_a_label():
+    for model in models.CATALOG:
+        assert model.provider in models.PROVIDER_LABELS, model.provider
+
+
+def test_catalog_covers_the_models_the_chat_app_offers():
+    # The ten the SDK enum lagged on, which is what this catalog exists to fix,
+    # plus a couple of long-standing ones to catch a wholesale regression.
+    expected = {
+        "gpt-5.6-luna",
+        "gpt-5.6-terra",
+        "gpt-5.6-sol",
+        "claude-sonnet-5",
+        "claude-opus-5",
+        "claude-fable-5",
+        "gemini-3.6-flash",
+        "gemini-3.5-flash-lite",
+        "grok-4.5",
+        "grok-4.3",
+        "hermes-4-405b",
+        "glm-5.2",
+        "deepseek-v4-pro",
+    }
+    assert expected <= {m.id for m in models.CATALOG}
+
+
+def test_known_models_is_a_superset_of_the_sdk_enum():
+    # Folding the enum in is what keeps a model the Chat app has retired (but
+    # the network still serves) listed, and picks up anything a future SDK adds.
+    sdk_ids = {m.value.split("/", 1)[1] for m in TEE_LLM}
+    assert sdk_ids <= set(models.known_model_ids())
+
+
+def test_known_models_includes_the_whole_catalog_first():
+    known = models.known_model_ids()
+    assert known[: len(models.CATALOG)] == [m.id for m in models.CATALOG]
+    assert len(known) == len(set(known))
+
+
+def test_default_model_is_in_the_catalog():
+    assert models.get(models.DEFAULT_MODEL) is not None
+
+
+def test_owned_by_reports_the_provider_for_curated_models():
+    assert models.owned_by("claude-opus-5") == "anthropic"
+    assert models.owned_by("deepseek-v4-pro") == "deepseek"
+    # An id we don't curate keeps the generic owner rather than guessing.
+    assert models.owned_by("some-future-model") == "opengradient"
+
+
+def test_by_provider_hides_superseded_models_by_default():
+    visible = {m.id for _label, entries in models.by_provider() for m in entries}
+    assert "claude-opus-5" in visible
+    assert "claude-opus-4-8" not in visible
+
+    everything = {
+        m.id for _label, entries in models.by_provider(include_legacy=True) for m in entries
+    }
+    assert "claude-opus-4-8" in everything
+
+
+def test_by_provider_separates_chat_from_image_models():
+    chat = {m.id for _label, entries in models.by_provider() for m in entries}
+    image = {m.id for _label, entries in models.by_provider(kind=models.IMAGE) for m in entries}
+    assert "gpt-image-2" in image
+    assert not chat & image

tests/test_server.py

@@ -11,6 +11,7 @@
 from opengradient import RelayError, VerificationError, VerifiedChatResponse
 from opengradient.client.tee_verify import TeeProof
 
+from veil import models
 from veil.server import create_app
 
 
@@ -139,3 +140,18 @@ def test_models_and_health():
     health = client.get("/health").get_json()
     assert health["status"] == "ok"
     assert health["active_tee_id"] == "0xabc"
+
+
+def test_models_lists_the_advertised_catalog_in_openai_shape():
+    client = _client(_StubGateway())
+    data = client.get("/v1/models").get_json()["data"]
+
+    assert [m["id"] for m in data] == models.known_model_ids()
+    assert all(m["object"] == "model" for m in data)
+    by_id = {m["id"]: m for m in data}
+    # Bare wire names, and the provider is reported as the owner.
+    assert "/" not in "".join(by_id)
+    assert by_id["claude-opus-5"]["owned_by"] == "anthropic"
+    # Superseded models stay listed: an agent already configured for one must
+    # keep working after it drops out of the Chat app's UI.
+    assert "claude-opus-4-8" in by_id

veil/cli.py

@@ -14,6 +14,7 @@
 import click
 
 from veil.config import DEFAULT_APP_URL, ServerConfig
+from veil.models import DEFAULT_MODEL
 from veil.session import AuthError, Session, login, login_manual
 
 
@@ -276,20 +277,54 @@ def env_cmd() -> None:
 
 
 @main.command()
-def models() -> None:
+@click.option(
+    "-a",
+    "--all",
+    "show_all",
+    is_flag=True,
+    help="Also show superseded models and any extras the SDK knows about.",
+)
+def models(show_all: bool) -> None:
     """List the models available through OpenGradient Veil.
 
-    Derived from the SDK's canonical ``TEE_LLM`` enum — the same source the
-    local server uses for ``GET /v1/models`` — so it stays in sync with the
-    network as models are added or retired.
+    The same lineup the OpenGradient Chat app offers, grouped by provider. Names
+    printed here are what an agent puts in its ``model`` field.
     """
-    from opengradient import TEE_LLM
+    from veil import models as catalog
+
+    def _print(entries: list[catalog.Model]) -> None:
+        width = max(len(m.id) for m in entries)
+        for model in entries:
+            suffix = "  (superseded)" if model.legacy else ""
+            click.echo(f"    {model.id.ljust(width)}  {model.description}{suffix}")
 
-    # Each enum value is ``provider/model``; agents send the bare model name.
-    names = sorted(m.value.split("/", 1)[1] for m in TEE_LLM)
     click.echo("Models available through OpenGradient Veil:")
-    for name in names:
-        click.echo(f"  {name}")
+    for label, entries in catalog.by_provider(include_legacy=show_all):
+        click.secho(f"\n  {label}", bold=True)
+        _print(entries)
+
+    image = [
+        model
+        for _label, entries in catalog.by_provider(include_legacy=show_all, kind=catalog.IMAGE)
+        for model in entries
+    ]
+    if image:
+        click.secho("\n  Image generation", bold=True)
+        click.echo("  Same /v1/chat/completions call. The images come back on the response,")
+        click.echo("  outside the signed output hash -- the exchange is verified, not the pixels.")
+        _print(image)
+
+    if show_all:
+        # Older generations the Chat app has retired from its UI. Still routable,
+        # and still valid on an agent config that already names one.
+        curated = {m.id for m in catalog.CATALOG}
+        extra = [m for m in catalog.known_model_ids() if m not in curated]
+        if extra:
+            click.secho("\n  Also routable", bold=True)
+            for name in extra:
+                click.echo(f"    {name}")
+    else:
+        click.echo("\nRun `og-veil models --all` to include superseded models.")
 
 
 @main.command(name="login")
@@ -315,7 +350,12 @@ def login_cmd(app_url: str, no_browser: bool, manual: bool) -> None:
 
 @main.command(name="test")
 @click.argument("prompt", nargs=-1)
-@click.option("--model", default="gpt-4.1", show_default=True, help="Model to send the prompt to.")
+@click.option(
+    "--model",
+    default=DEFAULT_MODEL,
+    show_default=True,
+    help="Model to send the prompt to (see `og-veil models`).",
+)
 @click.option(
     "--host", default=None, help="Override the host to reach (default: the running server's)."
 )