Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

fix(model-hub): one owner of a Source model's effort levels (#2583)

Call Whitece58c5d7fb0f
diff --git a/core/controller.py b/core/controller.py
index 6b7da394d..19d99f337 100644
--- a/core/controller.py
+++ b/core/controller.py
@@ -3000,6 +3000,13 @@ class Controller:
"Model Hub built-in reconciliation failed during startup",
exc_info=True,
)
+ try:
+ asyncio.run(self.model_hub_service.reconcile_source_reasoning_tiers())
+ except Exception:
+ logger.warning(
+ "Model Hub could not derive saved Source tiers again during startup",
+ exc_info=True,
+ )
self.model_hub_turn_gateway = ModelHubTurnGateway(
self.model_hub_service,
language_provider=lambda: self.config.language,
diff --git a/core/handlers/model_hub/pricing.py b/core/handlers/model_hub/pricing.py
index 7741dd55c..97aa72cb2 100644
--- a/core/handlers/model_hub/pricing.py
+++ b/core/handlers/model_hub/pricing.py
@@ -29,6 +29,8 @@ from datetime import date, datetime, timedelta
from pathlib import Path
from typing import Any, Callable, Final, Mapping, Optional
+from vibe.backend_model_catalog import catalog_model_id
+
logger = logging.getLogger(__name__)
PRICE_OVERRIDE_FILENAME: Final = "model_hub_prices.json"
@@ -102,10 +104,6 @@ _PLAN_KEYS: Final[Mapping[tuple[str, str], str]] = {
# resolves only through a fee key the override file names outright.
_PLAN_FAMILIES: Final[Mapping[str, str]] = {"anthropic": "anthropic", "openai": "openai", "codex": "openai"}
-_DATE_SUFFIX: Final = re.compile(r"(?:-\d{8}|-\d{4}-\d{2}-\d{2}|@[^/]*)$")
-_BRACKET_SUFFIX: Final = re.compile(r"\[[^\]]*\]$")
-
-
@dataclass(frozen=True)
class ModelPrice:
"""USD per million tokens."""
@@ -223,18 +221,10 @@ def _flat_price(cost: object) -> Optional[ModelPrice]:
def normalize_model_id(model_id: str) -> str:
- """Fold the spellings agents send onto the one a price table lists.
-
- `anthropic/claude-opus-5[1m]` and `claude-opus-4-5-20251101` become
- `claude-opus-5` and `claude-opus-4-5`: the context-window tag, the snapshot
- date, and a provider prefix do not change a list price.
- """
+ """Fold the spellings agents send onto the one a price table lists (``catalog_model_id``): the context-window
+ tag, the snapshot date, and a provider prefix do not change a list price."""
- text = model_id.strip().lower()
- text = text.rsplit("/", 1)[-1]
- text = _BRACKET_SUFFIX.sub("", text)
- text = _DATE_SUFFIX.sub("", text)
- return text.strip()
+ return catalog_model_id(model_id)
def _first_party(model_id: str, vendor_map: Mapping[str, Any]) -> Optional[str]:
diff --git a/core/handlers/model_hub/reasoning_tiers.py b/core/handlers/model_hub/reasoning_tiers.py
index d4e572cff..f280209a6 100644
--- a/core/handlers/model_hub/reasoning_tiers.py
+++ b/core/handlers/model_hub/reasoning_tiers.py
@@ -8,6 +8,7 @@ from typing import Literal, Mapping, Sequence
from vibe.backend_model_catalog import (
PROTOCOL_REASONING_EFFORT_DEFAULTS,
bundled_catalog_wire_efforts_by_model,
+ catalog_row_value,
without_client_presets,
)
@@ -33,23 +34,31 @@ def resolve_reasoning_tiers(
) -> ReasoningTierResolution:
"""Apply upstream, bundled-catalog, then user provenance in that order.
- A catalog row is a client's menu: its ``client_presets`` (Codex ``ultra``)
+ The catalog row is found by the model's id, or by the id a snapshot, tag or
+ provider prefix folds to (``catalog_model_id``), and it bounds the protocol
+ defaults an upstream signal stands for. A catalog row is a client's menu: its ``client_presets`` (Codex ``ultra``)
are names the client maps before sending, so a Source's wire tiers never
include them (#2539). ``catalog_efforts_by_model`` is therefore the wire
index, ``bundled_catalog_wire_efforts_by_model()``.
"""
+ catalog_efforts = catalog_row_value(
+ bundled_catalog_wire_efforts_by_model() if catalog_efforts_by_model is None else catalog_efforts_by_model,
+ model_id,
+ )
if supported_parameters is not None and any(
parameter.strip().lower() in _UPSTREAM_REASONING_PARAMETERS
for parameter in supported_parameters
):
defaults = PROTOCOL_REASONING_EFFORT_DEFAULTS.get(protocol)
if defaults is not None:
- return ReasoningTierResolution(defaults, "upstream")
+ # A generic reasoning parameter says the upstream takes the protocol's levels; the model's own row
+ # bounds them, so the hop declares exactly the levels its model has (Opus 4.6 has no xhigh).
+ return ReasoningTierResolution(
+ defaults if catalog_efforts is None else tuple(effort for effort in defaults if effort in catalog_efforts),
+ "upstream",
+ )
- catalog_efforts = (
- bundled_catalog_wire_efforts_by_model() if catalog_efforts_by_model is None else catalog_efforts_by_model
- ).get(model_id)
if catalog_efforts is not None:
return ReasoningTierResolution(catalog_efforts, "catalog")
@@ -66,3 +75,37 @@ def catalog_wire_efforts(model_id: str, efforts: Sequence[str]) -> tuple[str, ..
"""
return tuple(without_client_presets((model_id,), efforts))
+
+
+def rederive_persisted_tiers(
+ *,
+ protocol: str,
+ model_id: str,
+ efforts: Sequence[str],
+ source: ReasoningEffortsSource,
+ catalog_efforts_by_model: Mapping[str, tuple[str, ...]],
+) -> ReasoningTierResolution | None:
+ """A saved Source declaration derived again by this ladder, or None to keep it as saved.
+
+ An earlier build derived rows by its own rules: 3.2.0 matched only exact
+ ids, and gave every model behind a generic reasoning parameter the whole
+ protocol default list. A row the bundled catalog describes is derived again
+ when this ladder made it (``catalog``, ``upstream``) or found nothing for it
+ (no source and no tiers). An upstream row keeps its upstream signal. A
+ user's list waits for discovery, which reports when the catalog outranks it.
+ """
+
+ if catalog_row_value(catalog_efforts_by_model, model_id) is None:
+ return None
+ if source == "upstream":
+ return resolve_reasoning_tiers(
+ protocol=protocol,
+ model_id=model_id,
+ supported_parameters=("reasoning",),
+ catalog_efforts_by_model=catalog_efforts_by_model,
+ )
+ if source == "catalog" or (source is None and not efforts):
+ return resolve_reasoning_tiers(
+ protocol=protocol, model_id=model_id, catalog_efforts_by_model=catalog_efforts_by_model,
+ )
+ return None
diff --git a/core/handlers/model_hub/service.py b/core/handlers/model_hub/service.py
index bfc39267c..11946d5b0 100644
--- a/core/handlers/model_hub/service.py
+++ b/core/handlers/model_hub/service.py
@@ -181,7 +181,7 @@ from .provenance import (
produce_turn_outcome,
)
from .request import ModelHubRequest, without_opaque_history
-from .reasoning_tiers import catalog_wire_efforts, resolve_reasoning_tiers
+from .reasoning_tiers import catalog_wire_efforts, rederive_persisted_tiers, resolve_reasoning_tiers
from .stream_wire import ProtocolSSEState, ProtocolUsageReport
from .resolver import (
BackendName,
@@ -6546,6 +6546,43 @@ class ModelHubService:
return backend_builtin_models(backend)
+ async def reconcile_source_reasoning_tiers(self) -> list[tuple[str, str]]:
+ """Derive saved Source tiers again by the current ladder, once at startup; the (Source, model) pairs it changed.
+
+ Rows an earlier build derived keep its rules until discovery revisits
+ them: 3.2.0 found no catalog row for a snapshot id and declared the whole
+ protocol default list for Opus 4.6. ``rederive_persisted_tiers`` decides
+ which rows are derived again. The result is saved once, so a later start
+ changes nothing.
+ """
+
+ async with self._mutation_lock:
+ previous = self.store.load()
+ config = self._clone_config(previous)
+ catalog_efforts_by_model = bundled_catalog_wire_efforts_by_model()
+ changed: list[tuple[str, str]] = []
+ for source in config.sources:
+ for model in source.models:
+ resolution = rederive_persisted_tiers(
+ protocol=source.protocol,
+ model_id=model.id,
+ efforts=model.reasoning_efforts,
+ source=model.reasoning_efforts_source,
+ catalog_efforts_by_model=catalog_efforts_by_model,
+ )
+ if resolution is None or (
+ list(resolution.efforts) == list(model.reasoning_efforts)
+ and resolution.source == model.reasoning_efforts_source
+ ):
+ continue
+ model.reasoning_efforts = list(resolution.efforts)
+ model.reasoning_efforts_source = resolution.source
+ changed.append((source.id, model.id))
+ if changed and self._reconcile_store_writable():
+ await self._commit_synced(previous, config)
+ return changed
+ return []
+
async def reconcile_builtin_models(
self,
backends: Iterable[BackendName] = ("claude", "codex"),
diff --git a/docs/plans/model-hub-capability-model.md b/docs/plans/model-hub-capability-model.md
index 5b1363058..9d4d54823 100644
--- a/docs/plans/model-hub-capability-model.md
+++ b/docs/plans/model-hub-capability-model.md
@@ -132,7 +132,7 @@ and is pruned at the next write.
| Consumer | Rule |
| --- | --- |
-| Effort menu for a model on an Agent | each hop's declared tiers, then the union across hops. A hop's declared tiers follow `model-hub-reasoning-tiers.md`'s ladder: an upstream reasoning signal (the protocol's ladder), else the bundled backend row's efforts minus that row's `client_presets` (provenance `catalog`; it outranks a user list, with a visible override event), else the user's list, else models.dev `reasoning_options` when the catalog row is built. A persisted `catalog` list is also read without `client_presets`, since released builds wrote them, and Vibey's own model rows always drop them at config load (3.2.0 saved rows with Codex's `ultra`; Vibey is never the client a preset belongs to, and no provider documents `ultra` as a wire effort). A user-authored Source list is never filtered. The overlay's `wire_efforts` slots in before the backend row once a vendor needs it; no shipped entry sets it yet. A tier is omitted only if **every** hop that can serve the model has negative evidence for it. A tier with negative evidence on some hops is shown with "not on: <Source>". Model Hub projects this per hop as the chain API's read-only `refused_efforts` (`agent-chain.schema.json`); the UI renders the omission and the hint from it. |
+| Effort menu for a model on an Agent | each hop's declared tiers, then the union across hops. A hop's declared tiers follow `model-hub-reasoning-tiers.md`'s ladder: an upstream reasoning signal (the protocol's ladder, bounded by the model's catalog row when it has one), else the bundled backend row's efforts minus that row's `client_presets` (provenance `catalog`; it outranks a user list, with a visible override event), else the user's list, else models.dev `reasoning_options` when the catalog row is built. A persisted `catalog` list is also read without `client_presets`, since released builds wrote them, and Vibey's own model rows always drop them at config load (3.2.0 saved rows with Codex's `ultra`; Vibey is never the client a preset belongs to, and no provider documents `ultra` as a wire effort). A user-authored Source list is never filtered. The overlay's `wire_efforts` slots in before the backend row once a vendor needs it; no shipped entry sets it yet. A tier is omitted only if **every** hop that can serve the model has negative evidence for it. A tier with negative evidence on some hops is shown with "not on: <Source>". Model Hub projects this per hop as the chain API's read-only `refused_efforts` (`agent-chain.schema.json`); the UI renders the omission and the hint from it. |
| Effort at request time | Forwarded unchanged (reasoning-intent). A hop with **negative** evidence for that effort is skipped for this request, as if it were not a candidate. If no candidate remains, the request fails before invocation with `effort_unsupported` naming the effort and the Sources, instead of a provider 400. Never clamped silently. |
| Hosted tool mode on a hop | `native` only with positive `call` evidence taken on the current engine and probe shape, and the user's toggle on; otherwise the gateway proxy (`vibey-hosted-tools.md`). A wrong native declaration is a terminal 400 or a silent drop; the proxy is lossless. So this feature needs positive evidence, unlike efforts. |
| Native payload replay | only to the same Source, and only while its `replay` evidence is positive and taken on the current engine and probe shape. The decision is made at every request from `HopResolution`; the transcript never stores it. |
diff --git a/docs/plans/model-hub-reasoning-tiers.md b/docs/plans/model-hub-reasoning-tiers.md
index f767715f0..25d7ff455 100644
--- a/docs/plans/model-hub-reasoning-tiers.md
+++ b/docs/plans/model-hub-reasoning-tiers.md
@@ -51,7 +51,22 @@ shape and treats everything else as "not carried":
- OpenRouter-shape `supported_parameters` arrays: if the model entry contains
`"reasoning"` (or `"reasoning_effort"`), the model is reasoning-capable.
Upstream signals are capability booleans, not level enums, so the applied
- tier list is the UNIFIED VOCABULARY for the source's protocol family (below).
+ tier list is the UNIFIED VOCABULARY for the source's protocol family (below),
+ bounded by the model's own catalog row when it has one: a generic signal says
+ the upstream takes the protocol's levels, not that this model has all of
+ them (Opus 4.6 has `max` but no `xhigh`). A model without a row keeps the
+ whole list.
+- Every catalog lookup finds a model's row by its id, or by the id a snapshot
+ date, a context tag or a provider prefix folds to
+ (`backend_model_catalog.catalog_model_id`, the one folding owner, which
+ pricing also uses): `claude-opus-5-20260724` and `anthropic/claude-opus-5[1m]`
+ are `claude-opus-5`.
+- Rows a released build derived by older rules (3.2.0 matched only exact ids
+ and never bounded the protocol list) are derived again once at controller
+ startup (`ModelHubService.reconcile_source_reasoning_tiers`): every `catalog`
+ or `upstream` row, or a row with no source and no tiers, whose model has a
+ catalog row. A user's list waits for the next discovery, which records the
+ override event.
- Any other/unknown metadata shape → rung 1 does not apply. No guessing from
model-id patterns (family inference was explicitly deferred; wrong claims
turn into upstream 400s).
diff --git a/tests/fixtures/model_hub/released_3_2_0_source_tiers.json b/tests/fixtures/model_hub/released_3_2_0_source_tiers.json
new file mode 100644
index 000000000..943d4a343
--- /dev/null
+++ b/tests/fixtures/model_hub/released_3_2_0_source_tiers.json
@@ -0,0 +1,793 @@
+{
+ "enabled": true,
+ "runtime_default_applied": true,
+ "sources": [
+ {
+ "id": "src_tiers3201",
+ "created_at": "2026-10-01T00:00:00Z",
+ "last_discovered_at": null,
+ "kind": "api_key",
+ "vendor": "anthropic",
+ "display_name": "anthropic relay",
+ "protocol": "anthropic",
+ "base_url": null,
+ "supply_channel": "hub",
+ "billing": "metered",
+ "state": {
+ "status": "active",
+ "retry_at": null,
+ "detail_key": null
+ },
+ "models": [
+ {
+ "id": "claude-opus-4-6",
+ "display_name": null,
+ "origin": "discovered",
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ],
+ "reasoning_efforts_source": "upstream",
+ "discovered_at": null
+ },
+ {
+ "id": "claude-sonnet-4-6-20260301",
+ "display_name": null,
+ "origin": "discovered",
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ],
+ "reasoning_efforts_source": "upstream",
+ "discovered_at": null
+ },
+ {
+ "id": "claude-opus-5-20260724",
+ "display_name": null,
+ "origin": "discovered",
+ "reasoning_efforts": [],
+ "reasoning_efforts_source": null,
+ "discovered_at": null
+ },
+ {
+ "id": "relay-house-model",
+ "display_name": null,
+ "origin": "discovered",
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ],
+ "reasoning_efforts_source": "upstream",
+ "discovered_at": null
+ },
+ {
+ "id": "claude-haiku-4-5-20251001",
+ "display_name": null,
+ "origin": "discovered",
+ "reasoning_efforts": [
+ "low",
+ "high"
+ ],
+ "reasoning_efforts_source": "user",
+ "discovered_at": null
+ }
+ ],
+ "credential_ref": "cred_src_tiers3201",
+ "account_label": null,
+ "masked_credential": null
+ },
+ {
+ "id": "src_tiers3202",
+ "created_at": "2026-10-01T00:00:00Z",
+ "last_discovered_at": null,
+ "kind": "api_key",
+ "vendor": "openai",
+ "display_name": "openai relay",
+ "protocol": "openai_responses",
+ "base_url": null,
+ "supply_channel": "hub",
+ "billing": "metered",
+ "state": {
+ "status": "active",
+ "retry_at": null,
+ "detail_key": null
+ },
+ "models": [
+ {
+ "id": "gpt-6-astra-2026-09-01",
+ "display_name": null,
+ "origin": "discovered",
+ "reasoning_efforts": [],
+ "reasoning_efforts_source": null,
+ "discovered_at": null
+ }
+ ],
+ "credential_ref": "cred_src_tiers3202",
+ "account_label": null,
+ "masked_credential": null
+ }
+ ],
+ "agents": {
+ "opencode": {
+ "backend": "opencode",
+ "mode": "direct",
+ "menu_kind": "open",
+ "sources": {
+ "order": []
+ },
+ "routes": {},
+ "menu": {
+ "view": "featured",
+ "checked": []
+ },
+ "models": [],
+ "removed_model_ids": []
+ },
+ "claude": {
+ "backend": "claude",
+ "mode": "direct",
+ "menu_kind": "fixed",
+ "sources": {
+ "order": []
+ },
+ "routes": {},
+ "menu": null,
+ "models": [
+ {
+ "id": "claude-fable-5-1",
+ "display_name": "claude-fable-5-1 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-fable-5",
+ "display_name": "claude-fable-5 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-opus-5-5",
+ "display_name": "claude-opus-5-5 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-opus-5",
+ "display_name": "claude-opus-5 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-opus-4-8",
+ "display_name": "claude-opus-4-8 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-opus-4-7",
+ "display_name": "claude-opus-4-7 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-opus-4-6",
+ "display_name": "claude-opus-4-6 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-sonnet-5-5",
+ "display_name": "claude-sonnet-5-5 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-sonnet-5",
+ "display_name": "claude-sonnet-5 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-sonnet-4-6",
+ "display_name": "claude-sonnet-4-6 [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-haiku-4-5",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high"
+ ]
+ },
+ {
+ "id": "claude-opus-4-5",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high"
+ ]
+ },
+ {
+ "id": "claude-sonnet-4-5",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high"
+ ]
+ },
+ {
+ "id": "opus",
+ "display_name": "opus [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "sonnet",
+ "display_name": "sonnet [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "haiku",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high"
+ ]
+ },
+ {
+ "id": "opus[1m]",
+ "display_name": "opus[1m] [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "sonnet[1m]",
+ "display_name": "sonnet[1m] [1M]",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "claude-opus-4",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ },
+ {
+ "id": "claude-sonnet-4",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ },
+ {
+ "id": "claude-haiku-4",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ },
+ {
+ "id": "claude-sonnet-4-0",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ },
+ {
+ "id": "claude-sonnet-4-20250514",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ },
+ {
+ "id": "claude-sonnet-3-7",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ },
+ {
+ "id": "claude-haiku-3-5",
+ "display_name": null,
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": []
+ }
+ ],
+ "removed_model_ids": []
+ },
+ "codex": {
+ "backend": "codex",
+ "mode": "direct",
+ "menu_kind": "fixed",
+ "sources": {
+ "order": []
+ },
+ "routes": {},
+ "menu": null,
+ "models": [
+ {
+ "id": "gpt-6-astra",
+ "display_name": "GPT-6-Astra",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max",
+ "ultra"
+ ]
+ },
+ {
+ "id": "gpt-6.1-sol",
+ "display_name": "GPT-6.1-Sol",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "gpt-6-sol",
+ "display_name": "GPT-6-Sol",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "gpt-6-luna",
+ "display_name": "GPT-6-Luna",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "gpt-5.6-sol",
+ "display_name": "GPT-5.6-Sol",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max",
+ "ultra"
+ ]
+ },
+ {
+ "id": "gpt-5.6-terra",
+ "display_name": "GPT-5.6-Terra",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max",
+ "ultra"
+ ]
+ },
+ {
+ "id": "gpt-5.6-luna",
+ "display_name": "GPT-5.6-Luna",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh",
+ "max"
+ ]
+ },
+ {
+ "id": "gpt-5.5",
+ "display_name": "GPT-5.5",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh"
+ ]
+ },
+ {
+ "id": "gpt-5.4",
+ "display_name": "GPT-5.4",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh"
+ ]
+ },
+ {
+ "id": "gpt-5.4-mini",
+ "display_name": "GPT-5.4-Mini",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh"
+ ]
+ },
+ {
+ "id": "gpt-5.2",
+ "display_name": "GPT-5.2",
+ "origin": "builtin",
+ "models_dev_id": null,
+ "context_window": null,
+ "max_output_tokens": null,
+ "input_modalities": [],
+ "output_modalities": [],
+ "supports_tools": null,
+ "supports_reasoning": null,
+ "reasoning_efforts": [
+ "low",
+ "medium",
+ "high",
+ "xhigh"
+ ]
+ }
+ ],
+ "removed_model_ids": []
+ }
+ }
+}
diff --git a/tests/scenarios/model_hub/catalog.yaml b/tests/scenarios/model_hub/catalog.yaml
index 90714ecc9..34f10ee2c 100644
--- a/tests/scenarios/model_hub/catalog.yaml
+++ b/tests/scenarios/model_hub/catalog.yaml
@@ -263,6 +263,8 @@ scenarios:
- tests/test_model_hub_reasoning_tiers.py::test_a_persisted_catalog_list_offers_no_client_preset_to_vibey
- tests/test_model_hub_vibey_consumer.py::test_a_released_vibey_row_loads_without_its_client_preset
- tests/test_model_hub_evidence_rules.py::test_a_refused_effort_still_waits_for_a_capable_hop_that_is_cooling_down
+ - tests/test_model_hub_reasoning_tiers.py::test_the_hop_declares_exactly_its_models_levels
+ - tests/test_model_hub_reasoning_tiers.py::test_rows_saved_by_3_2_0_are_derived_again_once_at_startup
- id: MH-HOSTED-VIBEY-002
name: Every native hosted item becomes one hosted call with a provider-neutral action and its record; citations stay on the text that cites them; the response's counters and reported cost go to its summary
status: covered
diff --git a/tests/test_model_hub_api.py b/tests/test_model_hub_api.py
index 11bbd347b..ba022c6a7 100644
--- a/tests/test_model_hub_api.py
+++ b/tests/test_model_hub_api.py
@@ -7066,7 +7066,8 @@ def test_add_model_upsert_cannot_bypass_catalog_tier_lock(tmp_path):
id="gpt-5.6-sol",
supported_parameters=("reasoning_effort",),
),
- ["minimal", "low", "medium", "high", "xhigh"],
+ # The protocol defaults, bounded by the model's catalog row.
+ ["low", "medium", "high", "xhigh"],
"upstream",
"upstream_tiers",
),
diff --git a/tests/test_model_hub_reasoning_tiers.py b/tests/test_model_hub_reasoning_tiers.py
index 7d28bb05b..50626ffb7 100644
--- a/tests/test_model_hub_reasoning_tiers.py
+++ b/tests/test_model_hub_reasoning_tiers.py
@@ -16,9 +16,10 @@ def test_upstream_reasoning_signal_uses_only_the_protocol_default(
protocol: str,
expected: tuple[str, ...],
) -> None:
+ # A model no catalog row describes: nothing bounds what its protocol allows.
resolution = resolve_reasoning_tiers(
protocol=protocol,
- model_id="gpt-5.6-sol",
+ model_id="relay-reasoning-model",
supported_parameters=("temperature", "reasoning"),
existing_efforts=("user-only",),
existing_source="user",
@@ -28,6 +29,31 @@ def test_upstream_reasoning_signal_uses_only_the_protocol_default(
assert resolution.source == "upstream"
+ ("protocol", "model_id", "reported", "efforts", "source"),
+ (
+ # A model's own row bounds the protocol defaults: Opus and Sonnet 4.6 have max but no xhigh.
+ ("anthropic", "claude-opus-4-6", ("reasoning",), ("low", "medium", "high", "max"), "upstream"),
+ ("anthropic", "claude-sonnet-4-6-20260301", ("reasoning",), ("low", "medium", "high", "max"), "upstream"),
+ ("openai_responses", "gpt-5.6-sol", ("reasoning_effort",), ("low", "medium", "high", "xhigh"), "upstream"),
+ # Mythos 5 and 5.1 have every Anthropic level, which is exactly the protocol default list, so no row is
+ # needed for them.
+ ("anthropic", "claude-mythos-5-1", ("reasoning",), ("low", "medium", "high", "xhigh", "max"), "upstream"),
+ # A snapshot, a context tag or a provider prefix names the same row.
+ ("anthropic", "claude-opus-5-20260724", None, ("low", "medium", "high", "xhigh", "max"), "catalog"),
+ ("anthropic", "anthropic/claude-opus-5[1m]", None, ("low", "medium", "high", "xhigh", "max"), "catalog"),
+ # ... and the row's client presets with it: a snapshot of gpt-6-astra still never offers Codex's ultra.
+ ("openai_responses", "gpt-6-astra-2026-09-01", None, ("low", "medium", "high", "xhigh", "max"), "catalog"),
+ ),
+)
+def test_the_hop_declares_exactly_its_models_levels(protocol, model_id, reported, efforts, source) -> None:
+ """MH-HOSTED-006: one owner of effort levels. A snapshot or qualified id finds its model's catalog row, and that
+ row bounds the protocol defaults a generic reasoning parameter stands for."""
+ resolution = resolve_reasoning_tiers(protocol=protocol, model_id=model_id, supported_parameters=reported)
+
+ assert (resolution.efforts, resolution.source) == (efforts, source)
+
+
def test_catalog_row_outranks_user_and_drops_its_client_presets() -> None:
"""MH-HOSTED-006 (#2539): Codex's ``ultra`` is a client preset, never a Source's wire tier."""
resolution = resolve_reasoning_tiers(
@@ -157,3 +183,46 @@ def test_a_persisted_catalog_list_offers_no_client_preset_to_vibey(tmp_path, pro
[row] = [row for row in providers if row["id"] == "gpt-5.6-sol"]
assert row["reasoning_efforts"] == expected
+
+
+async def test_rows_saved_by_3_2_0_are_derived_again_once_at_startup(tmp_path) -> None:
+ """MH-HOSTED-006: 3.2.0 found no row for a snapshot id and declared xhigh for Opus and Sonnet 4.6. The fixture
+ was written by v3.2.0's own ladder; the startup reconcile derives those rows again, and only those."""
+ import json
+ from pathlib import Path
+
+ from config.v2_config import ModelHubConfig
+ from tests.test_model_hub_api import _service
+
+ section = json.loads(
+ (Path(__file__).parent / "fixtures/model_hub/released_3_2_0_source_tiers.json").read_text(encoding="utf-8")
+ )
+ service, store, _adapter = _service(tmp_path)
+ store.config = ModelHubConfig.from_payload(section)
+
+ changed = await service.reconcile_source_reasoning_tiers()
+
+ rows = {
+ (source.id, model.id): (model.reasoning_efforts, model.reasoning_efforts_source)
+ for source in store.load().sources
+ for model in source.models
+ }
+ assert rows == {
+ ("src_tiers3201", "claude-opus-4-6"): (["low", "medium", "high", "max"], "upstream"),
+ ("src_tiers3201", "claude-sonnet-4-6-20260301"): (["low", "medium", "high", "max"], "upstream"),
+ ("src_tiers3201", "claude-opus-5-20260724"): (["low", "medium", "high", "xhigh", "max"], "catalog"),
+ # No catalog row describes it: its upstream list stays as saved.
+ ("src_tiers3201", "relay-house-model"): (["low", "medium", "high", "xhigh", "max"], "upstream"),
+ # A user's list waits for discovery, which reports when the catalog outranks it.
+ ("src_tiers3201", "claude-haiku-4-5-20251001"): (["low", "high"], "user"),
+ ("src_tiers3202", "gpt-6-astra-2026-09-01"): (["low", "medium", "high", "xhigh", "max"], "catalog"),
+ }
+ assert sorted(changed) == [
+ ("src_tiers3201", "claude-opus-4-6"),
+ ("src_tiers3201", "claude-opus-5-20260724"),
+ ("src_tiers3201", "claude-sonnet-4-6-20260301"),
+ ("src_tiers3202", "gpt-6-astra-2026-09-01"),
+ ]
+ # Saved once: the next start finds nothing to change.
+ assert await service.reconcile_source_reasoning_tiers() == []
diff --git a/vibe/backend_model_catalog.py b/vibe/backend_model_catalog.py
index 313f619ff..253efc925 100644
--- a/vibe/backend_model_catalog.py
+++ b/vibe/backend_model_catalog.py
@@ -542,6 +542,32 @@ def prepare_codex_hub_catalog(
)
+#: Spellings that name the same model: a snapshot date or version tag, and a context-window tag.
+_MODEL_SNAPSHOT_SUFFIX: Final = re.compile(r"(?:-\d{8}|-\d{4}-\d{2}-\d{2}|@[^/]*)$")
+_MODEL_CONTEXT_TAG: Final = re.compile(r"\[[^\]]*\]$")
+
+
+def catalog_model_id(model_id: str) -> str:
+ """Fold the spellings agents send onto the one catalogs list.
+
+ ``anthropic/claude-opus-5[1m]`` and ``claude-opus-4-5-20251101`` become ``claude-opus-5`` and ``claude-opus-4-5``:
+ a provider prefix, the context-window tag and the snapshot date name the same model, its levels and its price.
+ """
+
+ text = model_id.strip().lower()
+ text = text.rsplit("/", 1)[-1]
+ text = _MODEL_CONTEXT_TAG.sub("", text)
+ text = _MODEL_SNAPSHOT_SUFFIX.sub("", text)
+ return text.strip()
+
+
+def catalog_row_value(index: Mapping[str, tuple[str, ...]], model_id: str) -> tuple[str, ...] | None:
+ """A per-model catalog list for ``model_id``: its own row's, else the row's its spelling folds to."""
+
+ value = index.get(model_id)
+ return value if value is not None else index.get(catalog_model_id(model_id))
+
+
def load_bundled_catalog(path: Path | None = None) -> dict[str, Any]:
return _read_catalog(path or get_bundled_catalog_path()) or {}
@@ -621,7 +647,7 @@ def without_client_presets(model_ids: Sequence[str], efforts: Sequence[str]) ->
"""
presets = bundled_client_presets_once()
- dropped = {preset for model_id in model_ids for preset in presets.get(model_id, ())}
+ dropped = {preset for model_id in model_ids for preset in catalog_row_value(presets, model_id) or ()}
return [effort for effort in efforts if effort not in dropped]