| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115 |
- """Provider/model inventory — shared substrate for the dashboard ``/api/model/options``, the TUI
- ``model.options``/``model.save_key`` RPC handlers, and the interactive picker."""
- from __future__ import annotations
- import logging
- from contextvars import copy_context
- from dataclasses import dataclass, replace
- from threading import Lock, Thread, current_thread
- from typing import Any, Optional
- from datetime import UTC
- _pricing_prewarm_lock = Lock()
- _pricing_prewarm_threads: dict[tuple[str, tuple[tuple[str, str], ...]], Thread] = {}
- @dataclass(frozen=True)
- class ConfigContext:
- """Disk-config snapshot (``load_picker_context()``); the TUI overlays live agent state via
- ``with_overrides()``."""
- current_provider: str
- current_model: str
- current_base_url: str
- user_providers: dict
- custom_providers: list
- excluded_providers: list = None
- def with_overrides(
- self, *, current_provider: Optional[str] = None, current_model: Optional[str] = None,
- current_base_url: Optional[str] = None,
- ) -> ConfigContext:
- """Copy with TRUTHY overrides applied: the TUI reads agent attributes that may be empty strings
- before an agent is spawned — empties must not clobber the disk-config values."""
- overrides = (("current_provider", current_provider), ("current_model", current_model),
- ("current_base_url", current_base_url))
- kw = {k: v for k, v in overrides if v}
- return replace(self, **kw) if kw else self
- def load_picker_context() -> ConfigContext:
- """Load the disk-config snapshot every consumer needs."""
- from hermes_cli.config import (
- coerce_provider_id, get_compatible_custom_providers, load_config, stringify_provider_map,
- )
- cfg = load_config()
- model_cfg = cfg.get("model", {})
- if isinstance(model_cfg, dict):
- # YAML parses unquoted scalars as int (`provider: 2070`); keep strings so picker/options
- # paths never call `.strip()` on an int.
- current_model = str(model_cfg.get("default", model_cfg.get("name", "")) or "")
- current_provider = coerce_provider_id(model_cfg.get("provider", ""))
- current_base_url = str(model_cfg.get("base_url", "") or "")
- else: # config.model can be a bare string in older configs
- current_model, current_provider, current_base_url = (str(model_cfg) if model_cfg else ""), "", ""
- excluded = cfg.get("model_catalog", {}).get("excluded_providers") or []
- return ConfigContext(
- current_provider=current_provider, current_model=current_model, current_base_url=current_base_url,
- user_providers=stringify_provider_map(cfg.get("providers")),
- custom_providers=get_compatible_custom_providers(cfg),
- excluded_providers=excluded if isinstance(excluded, list) else [],
- )
- def _slug(row: dict) -> str:
- return str(row.get("slug") or "").strip().lower()
- def _without_slug(rows: list[dict], slug: str) -> list[dict]:
- return [r for r in rows if _slug(r) != slug]
- # ─── Public: payload builder ────────────────────────────────────────────
- def build_models_payload(
- ctx: ConfigContext, *, explicit_only: bool = False, include_unconfigured: bool = False,
- picker_hints: bool = False, canonical_order: bool = False, pricing: bool = False,
- pricing_cache_only: bool = False,
- capabilities: bool = False, featured: bool = False, force_fresh_nous_tier: bool = False,
- refresh: bool = False, probe_custom_providers: bool = True, probe_current_custom_provider: bool = False,
- for_picker: bool = False, max_models: int | None = None, non_blocking_catalogs: bool = False,
- fast_custom_probe: bool | None = None,
- ) -> dict:
- """Build the ``{providers, model, provider}`` shape every consumer needs. ``explicit_only`` keeps
- only providers the user explicitly configured — hides ambient/auto-seeded credentials from
- desktop chat pickers. ``pricing_cache_only``: with ``pricing``, use only values already resident
- in process caches (normal picker opens, while a background worker warms cold endpoints).
- ``non_blocking_catalogs``: provider catalogs come from the disk cache only — a degraded provider
- cannot stall the response (GUI picker opens). ``fast_custom_probe`` overrides the
- custom-endpoint discovery budget ``for_picker`` otherwise implies (1.5s vs 5s) — ``None`` keeps
- the coupling, ``False`` retains the full 5s budget (#103843)."""
- from hermes_cli.model_switch import list_authenticated_providers
- rows = list_authenticated_providers(
- current_provider=ctx.current_provider, current_base_url=ctx.current_base_url,
- current_model=ctx.current_model, user_providers=ctx.user_providers,
- custom_providers=ctx.custom_providers, force_fresh_nous_tier=force_fresh_nous_tier,
- max_models=max_models, refresh=refresh, probe_custom_providers=probe_custom_providers,
- probe_current_custom_provider=probe_current_custom_provider, for_picker=for_picker,
- excluded_providers=ctx.excluded_providers or [],
- non_blocking_catalogs=non_blocking_catalogs, fast_custom_probe=fast_custom_probe,
- )
- # Managed local runtime: staged GGUFs are selectable like any provider's models, but
- # list_authenticated_providers can't know about them (no credential — reachability is the
- # credential), so inject the row here where every picker surface inherits it.
- local_row = _local_runtime_row(ctx)
- if local_row is not None:
- rows = _without_slug(rows, "llamacpp") + [local_row]
- # A live session on the managed server reports provider "custom" (raw base_url label), which
- # would materialize a duplicate "Custom endpoint" row with the same staged models stealing the
- # checkmark. The Local row owns the managed server's identity — drop such custom rows.
- if local_row.get("is_current"):
- staged = set(local_row["models"])
- def _is_managed_custom(row: dict) -> bool:
- models = {str(m) for m in (row.get("models") or [])}
- return _slug(row) == "custom" and bool(models) and models <= staged
- rows = [r for r in rows if not _is_managed_custom(r)]
- moa_row = _moa_provider_row(ctx.current_provider)
- if moa_row is not None:
- rows = [moa_row] + _without_slug(rows, "moa")
- if explicit_only:
- rows = _filter_explicit_provider_rows(rows, ctx)
- # If the current provider lost its credential, list_authenticated_providers() omits it; keep
- # that one row so the UI shows the saved selection + a re-auth affordance instead of appearing
- # to jump providers. Exception: a "custom" current on the managed local server is already
- # represented by the Local row — the skeleton would resurrect the duplicate removed above.
- _local_owns_current = bool(local_row and local_row.get("is_current")
- and (ctx.current_provider or "").lower() == "custom")
- if not _local_owns_current:
- rows = list(rows) + _append_unconfigured_rows(rows, ctx, current_only=True)
- # A local proxy serving a model also in an aggregator's catalog would show under both, and picking
- # the aggregator row silently breaks the call — aggregators only list models no specific provider has.
- _strip_aggregator_overlaps(rows)
- if include_unconfigured:
- rows = list(rows) + _without_slug(_append_unconfigured_rows(rows, ctx), "moa")
- if picker_hints:
- _apply_picker_hints(rows)
- if canonical_order:
- rows = _reorder_canonical(rows)
- if pricing:
- _apply_pricing(rows, force_fresh_nous_tier=force_fresh_nous_tier, cached_only=pricing_cache_only)
- # Both metadata decorators consult ``model_overrides``. Snapshot the
- # read-only config once for this payload rather than letting each model
- # lookup reopen config.yaml through models_dev._cfg_get().
- metadata_config = None
- if capabilities or featured:
- try:
- from hermes_cli.config import load_config_readonly
- metadata_config = load_config_readonly()
- except Exception:
- metadata_config = None
- if capabilities:
- _apply_capabilities(rows, metadata_config=metadata_config)
- if featured:
- _apply_featured(rows, metadata_config=metadata_config)
- _apply_custom_aliases(rows)
- from hermes_cli.models_validate import drop_unofferable_model_ids
- drop_unofferable_model_ids(rows)
- return {"providers": rows, "model": ctx.current_model, "provider": ctx.current_provider}
- def _strip_aggregator_overlaps(rows: list[dict]) -> None:
- """Drop models from TRUE routing aggregators (OpenRouter, custom:* proxies) that a user-defined
- provider also serves. The is_user_defined guard matters: is_routing_aggregator() is True for every
- custom:* slug, so without it the dedup would empty a user's own custom row. Flat-namespace
- resellers (opencode-go/zen) serve every model first-party and keep shared names."""
- try:
- from hermes_cli.providers import is_routing_aggregator
- except Exception:
- return
- builtin_aggregators = {
- _slug(row) for row in rows
- if not row.get("is_user_defined") and is_routing_aggregator(_slug(row))
- }
- def _duplicates_builtin_aggregator(row: dict) -> bool:
- # A user row that IS the same upstream as a built-in aggregator (registered OpenRouter via
- # Settings → Providers, or a ``custom:openrouter`` slug) is that aggregator's twin, not a
- # rival: its catalog is a superset of the built-in row's, so counting it empties the
- # built-in row (openrouter → total=0 beside a live custom:openrouter row).
- row_slug = _slug(row)
- slug_suffix = (
- row_slug.split(":", 1)[1] if row_slug.startswith("custom:") else ""
- )
- if slug_suffix and slug_suffix in builtin_aggregators:
- return True
- from agent.model_metadata import _infer_provider_from_url
- inferred = _infer_provider_from_url(str(row.get("api_url") or ""))
- return inferred is not None and inferred in builtin_aggregators
- user_models: set[str] = set()
- for row in rows:
- if row.get("is_user_defined"):
- if builtin_aggregators and _duplicates_builtin_aggregator(row):
- continue # the twin IS that aggregator; it must not retro-strip it
- user_models.update(m.lower() for m in (row.get("models") or []))
- if not user_models:
- return
- for row in rows:
- if row.get("is_user_defined") or not is_routing_aggregator(row.get("slug", "")):
- continue
- # Only strip overlaps from TRUE routing aggregators (OpenRouter, custom:* proxies). Flat-namespace
- # resellers (opencode-go / opencode-zen) serve every listed model as a first-party model, so their
- # rows must keep models that a user's proxy happens to share a name with — otherwise a subscription
- # provider's own catalog (minimax-m3, glm-5, deepseek-v4-flash, ...) is silently gutted in the
- # picker. (#47077)
- original = row.get("models") or []
- filtered = [m for m in original if m.lower() not in user_models]
- if len(filtered) < len(original):
- row["models"] = filtered
- row["total_models"] = len(filtered)
- def build_model_options_payload(
- ctx: ConfigContext, *, explicit_only: bool = False, include_unconfigured: bool = False,
- refresh: bool = False,
- ) -> dict:
- """Shared API-server/dashboard/TUI payload. Normal open probes only the current custom provider so
- offline saved endpoints don't block the picker; explicit refresh probes all and busts the cache.
- A normal open (``refresh=False``) is a READ path: provider catalogs come from the disk cache
- only and stale/missing ones warm in the background, so a degraded provider (hanging endpoint,
- failed auth probe) delays neither the other providers' rows nor the response (#114215).
- ``for_picker=True`` keeps providers whose credential pool is entirely rate-limited visible:
- these are human-facing pickers, and hiding a temporarily exhausted pool makes providers vanish
- mid-session even though another model under the same provider may still work (same contract
- as ``/model`` and the aux pickers, #66584 / #66624). Visibility only: ``fast_custom_probe=False``
- keeps the live probe of the current custom endpoint on its full 5s discovery budget."""
- refresh = bool(refresh)
- payload = build_models_payload(
- ctx, explicit_only=bool(explicit_only), include_unconfigured=bool(include_unconfigured),
- picker_hints=True, canonical_order=True, pricing=True, pricing_cache_only=not refresh,
- capabilities=True, featured=True, for_picker=True, fast_custom_probe=False,
- refresh=refresh, probe_custom_providers=refresh, probe_current_custom_provider=not refresh,
- non_blocking_catalogs=not refresh,
- )
- _apply_limits(payload["providers"])
- _apply_usage(payload["providers"])
- if not refresh:
- _prewarm_pricing_async(payload["providers"], current_provider=ctx.current_provider,
- current_base_url=ctx.current_base_url)
- return payload
- def _apply_limits(rows: list[dict]) -> None:
- """Attach ``limit`` to rows whose credential pool is rate-limited, so a picker can say why and until
- when instead of the row just looking broken. Only providers with a persisted pool are read (no
- seeding), and a pool that fails to load says nothing rather than failing the whole catalog."""
- import logging
- from datetime import datetime, timezone
- from agent.credential_pool import load_pool
- from hermes_cli.auth import read_credential_pool
- def iso(epoch: float) -> str:
- return datetime.fromtimestamp(epoch, UTC).isoformat()
- pooled = {slug for slug, entries in read_credential_pool().items() if entries}
- for row in rows:
- slug = str(row.get("slug") or "")
- # Window plans say their limits in plan_usage (gauges + limit line); a credential cooldown chip
- # beside them would show a second, unrelated reset time.
- if slug not in pooled or row.get("is_user_defined") or row.get("plan_usage"):
- continue
- try:
- state = load_pool(slug).limit_state(row.get("models") or [])
- except Exception: # an unreadable pool must not fail the whole catalog
- logging.getLogger(__name__).debug("Pool limit read failed for %s", slug, exc_info=True)
- continue
- if state is None:
- continue
- if state["scope"] == "account":
- row["limit"] = {"scope": "account", "resets_at": iso(state["resets_at"])}
- else:
- row["limit"] = {"scope": "models", "models": {m: iso(at) for m, at in state["models"].items()}}
- def _apply_usage(rows: list[dict]) -> None:
- """Attach ``usage`` (subscription windows: % spent + reset) to signed-in rows that can report it,
- from the cache only, and ask for a background refresh, so the picker never waits on a usage API
- and a chip can warn before the wall instead of at it.
- Single-account providers keep the legacy gauge (``usage.windows``). A provider with a
- MULTI-ENTRY credential pool reports ``usage.accounts`` instead — one row per account, each with
- its own windows/state/resets_at — because one account's quota says nothing about its siblings;
- a provider-wide percentage there would be a fabricated average. A snapshot older than the
- staleness bound renders ``unknown``/``unavailable``, never a confident stale gauge. Nothing here
- mutates the pool or disables routing: telemetry never benches a credential."""
- from agent.account_usage_cache import (
- cached_account_usage, has_account_usage, refresh_account_usage_async,
- refresh_account_usage_entries_async, snapshot_is_stale,
- )
- from hermes_cli.auth import read_credential_pool
- def _wire_windows(snapshot) -> list[dict]:
- return [
- {"label": w.label, "used_percent": float(w.used_percent),
- "resets_at": w.reset_at.isoformat() if w.reset_at else None,
- "scope": w.scope}
- for w in (snapshot.windows if snapshot else ()) if w.used_percent is not None
- ]
- def _account_resets_at(snapshot, live_cooldown_until: float | None) -> str | None:
- """A quota-exhausted account recovers at the LATEST of its exhausted account-scoped
- windows (all of them must reopen before the account is whole again); a live
- credential-wide cooldown wins when later. Model-scoped windows never reset an account.
- Unknown stays None — the frontend renders its own advisory (e.g. the earliest sibling)."""
- resets: list[str] = []
- if live_cooldown_until:
- resets.append(_iso_from_epoch(live_cooldown_until))
- for window in (snapshot.windows if snapshot else ()):
- if window.scope != "account" or window.used_percent is None or window.reset_at is None:
- continue
- if float(window.used_percent) >= _EXHAUSTED_WINDOW_PERCENT:
- resets.append(window.reset_at.isoformat())
- return max(resets) if resets else None
- wanted: list[str] = []
- entry_requests: list[dict] = []
- pooled_sizes = {slug: len(entries) for slug, entries in read_credential_pool().items() if entries}
- for row in rows:
- slug = str(row.get("slug") or "")
- if not slug or row.get("is_user_defined") or row.get("authenticated") is False:
- continue
- supports_usage = has_account_usage(slug)
- # A single-entry pool (or no pool) is the single-account case: the legacy gauge stays
- # (chip contract unchanged). Only a genuinely multi-entry pool gets per-account rows —
- # there, a provider-wide percentage would be a fabricated average across logins.
- if pooled_sizes.get(slug, 0) < 2:
- if not supports_usage:
- continue
- wanted.append(slug)
- snapshot = cached_account_usage(slug)
- windows = _wire_windows(snapshot)
- if windows and not snapshot_is_stale(snapshot):
- row["usage"] = {"windows": windows}
- continue
- if supports_usage:
- wanted.append(slug)
- accounts = _pool_usage_accounts(slug, _wire_windows, _account_resets_at, entry_requests)
- if accounts is not None:
- row["usage"] = {"accounts": accounts}
- refresh_account_usage_async(wanted)
- refresh_account_usage_entries_async(entry_requests)
- # A window this spent counts as exhausted when computing an account's recovery time.
- _EXHAUSTED_WINDOW_PERCENT = 100.0
- def _iso_from_epoch(epoch: float) -> str:
- from datetime import datetime, timezone
- return datetime.fromtimestamp(float(epoch), UTC).isoformat()
- def _pool_usage_accounts(slug: str, wire_windows, account_resets_at,
- entry_requests: list[dict]) -> Optional[list[dict]]:
- """Per-account usage rows for one pooled provider, or None when the pool cannot be read.
- Read-only: entries are inspected via ``load_pool(slug)`` properties (no select/rotate/refresh).
- Dedupe is scoped to provider+host+trusted decoded account id where supported (several
- credentials of one Codex account are ONE account), else per-credential token identity — never
- a provider-wide singleton. Providers without a usage fetcher still get one metadata row per
- account with state unknown/limited/unavailable, so the chip can count honestly without bogus
- fetches. DEAD auth rows stay visible as ``unavailable`` (never a quota row). No secrets, raw
- payloads or base URLs ride the wire."""
- import logging
- import time
- from agent.account_usage_cache import _identity_id_for, cached_account_usage, has_account_usage, snapshot_is_stale
- from agent.credential_pool import STATUS_DEAD, _exhausted_until, load_pool
- try:
- pool = load_pool(slug)
- entries = list(pool.entries())
- except Exception: # an unreadable pool must not fail the whole catalog
- logging.getLogger(__name__).debug("Pool usage read failed for %s", slug, exc_info=True)
- return None
- if not entries:
- return None
- live = [e for e in entries if e.last_status != STATUS_DEAD]
- sole = len(live) <= 1
- supports_usage = has_account_usage(slug)
- now = time.time()
- accounts: list[dict] = []
- seen: dict[tuple[str, str, str], dict] = {}
- for entry in entries:
- identity_id = _identity_id_for(slug, entry)
- host = _pool_entry_host(slug, entry)
- dedupe_key = (slug, host, identity_id)
- is_dead = entry.last_status == STATUS_DEAD
- snapshot = (cached_account_usage(slug, identity_id=identity_id)
- if supports_usage and not is_dead else None)
- stale = snapshot_is_stale(snapshot)
- windows = wire_windows(snapshot) if not stale else []
- # 'limited' only when a live credential-wide cooldown benches this account or its
- # account-scoped quota windows are exhausted — telemetry alone never benches anything.
- cooldown_until = None if is_dead else _exhausted_until(entry, sole_credential=sole)
- live_cooldown = cooldown_until if cooldown_until and cooldown_until > now else None
- quota_exhausted = bool(windows) and any(
- w.get("scope") == "account" and w.get("used_percent") is not None
- and float(w["used_percent"]) >= _EXHAUSTED_WINDOW_PERCENT for w in windows)
- if is_dead:
- state = "unavailable"
- elif live_cooldown is not None or quota_exhausted:
- state = "limited"
- elif snapshot is None or stale or not windows:
- # No numeric live windows (failed/empty fetch, stale snapshot, non-supporting
- # provider): unknown — including a "ready-looking" account with nothing to show.
- state = "unknown"
- else:
- state = "ready"
- account_row = {
- "id": identity_id, "label": str(entry.label or ""), "windows": windows,
- "state": state, "resets_at": account_resets_at(snapshot, live_cooldown),
- }
- if dedupe_key in seen:
- # Same account under a second credential: keep the row that says more (state rank
- # limited > ready > unknown > unavailable), but never duplicate the account.
- existing = seen[dedupe_key]
- rank = {"limited": 3, "ready": 2, "unknown": 1, "unavailable": 0}
- if rank.get(account_row["state"], 0) > rank.get(existing["state"], 0):
- accounts[accounts.index(existing)] = account_row
- seen[dedupe_key] = account_row
- continue
- seen[dedupe_key] = account_row
- accounts.append(account_row)
- if supports_usage and not is_dead:
- entry_requests.append({
- "provider": slug, "identity_id": identity_id,
- "base_url": _pool_entry_route_base_url(slug, entry), "api_key": entry.runtime_api_key,
- })
- return accounts or None
- def _pool_entry_host(slug: str, entry) -> str:
- """Normalized route host of one entry, for account dedupe scoping. Never serialized."""
- from agent.credential_pool import _norm_url
- if slug == "openai-codex":
- from hermes_cli.auth_codex import _codex_pool_route_base_url
- return _norm_url(_codex_pool_route_base_url(entry.runtime_base_url))
- return _norm_url(getattr(entry, "runtime_base_url", None))
- def _pool_entry_route_base_url(slug: str, entry) -> Optional[str]:
- """The base URL a usage fetch for *entry* must target (its own route, never a sibling's)."""
- if slug == "openai-codex":
- from hermes_cli.auth_codex import _codex_pool_route_base_url
- return _codex_pool_route_base_url(entry.runtime_base_url)
- return getattr(entry, "runtime_base_url", None) or None
- # ─── Public: auxiliary-task pickers ─────────────────────────────────────
- def build_aux_picker_rows(
- *, current_provider: str = "", current_model: str = "", current_base_url: str = "",
- max_models: int | None = None,
- ) -> list[dict]:
- """Provider rows for any auxiliary-task picker (vision, compression, …). Honours
- ``excluded_providers``; exhausted-pool providers stay visible (``for_picker``); only the active
- custom endpoint is probed. ``moa`` is excluded: auxiliary_client unwraps it to its aggregator
- anyway, so offering it would be a choice silently rewritten.
- Aux pickers kept re-deriving their own kwargs and each one silently dropped a different slice of the
- user's configuration. Two independent contributor PRs landed against the same two call sites for exactly
- this: 52642 (user ``providers:`` / ``custom_providers:`` entries never appeared) and #66624 (providers
- with an exhausted credential pool were hidden). Both were per-site kwarg patches, so the next aux picker
- would have reintroduced the same gap. Routing through one function makes the correct behaviour the
- default that a new caller cannot forget:
- """
- ctx = load_picker_context().with_overrides(
- current_provider=current_provider, current_model=current_model, current_base_url=current_base_url,
- )
- rows = build_models_payload(
- ctx, for_picker=True, probe_custom_providers=False, probe_current_custom_provider=True,
- max_models=max_models,
- )["providers"]
- return _without_slug(rows, "moa")
- def format_aux_picker_entries(
- rows: list[dict], *, current_provider: str = "", current_base_url: str = "",
- ) -> list[tuple[str, str, list[str]]]:
- """Render aux-picker rows as ``(slug, label, models)``. A raw ``base_url`` custom endpoint is
- "current" only through that URL, never a slug — so with ``current_base_url`` set no row is marked."""
- entries: list[tuple[str, str, list[str]]] = []
- current_slug = str(current_provider or "").strip().lower()
- has_base_url = bool(str(current_base_url or "").strip())
- for row in rows:
- slug = str(row.get("slug") or "")
- name = row.get("name") or slug
- total = row.get("total_models") or len(row.get("models") or [])
- model_hint = f" — {total} models" if total else ""
- marker = " ← current" if slug.lower() == current_slug and current_slug and not has_base_url else ""
- entries.append((slug, f"{name}{model_hint}{marker}", list(row.get("models") or [])))
- return entries
- def _reasoning_catalog_reader(slug: str):
- """Per-model reasoning-capability reader for aggregators that publish one. Cache-only — the picker
- must never block on HTTP; a cold cache warms in the background and reports no restriction until then."""
- try:
- from hermes_cli.models_reasoning_caps import (
- nous_model_reasoning_capabilities,
- openrouter_model_reasoning_capabilities,
- warm_nous_reasoning_caps_async,
- warm_openrouter_reasoning_caps_async,
- )
- except Exception:
- return None
- readers = {
- "nous": (warm_nous_reasoning_caps_async, nous_model_reasoning_capabilities),
- "openrouter": (warm_openrouter_reasoning_caps_async, openrouter_model_reasoning_capabilities),
- }
- if slug not in readers:
- return None
- warm, read = readers[slug]
- warm()
- return read
- def _apply_capabilities(rows: list[dict], *, metadata_config: dict | None = None) -> None:
- """Attach ``{model: {fast, reasoning, ...}}`` per row. ``reasoning`` defaults True when the catalog is
- silent (the dial is a no-op on models that ignore it; hiding it from a capable model is worse). A
- serving aggregator's detail overrides models.dev (adds ``can_disable_reasoning``). ``supported_efforts``
- is deliberately NOT forwarded — it under-reports levels that work."""
- from agent.codex_service_tiers import daybreak_offered, ultrafast_offered
- from hermes_cli.models import model_supports_ultrafast, resolve_fast_mode_overrides
- try:
- from agent.models_dev import get_model_capabilities
- except Exception:
- get_model_capabilities = None # type: ignore[assignment]
- # Daybreak rides the Responses request; the opt-in codex app-server runtime never sends it.
- _model_cfg = (metadata_config or {}).get("model")
- _app_server = isinstance(_model_cfg, dict) and str(
- _model_cfg.get("openai_runtime") or "").strip().lower() == "codex_app_server"
- for row in rows:
- slug = row.get("slug") or ""
- caps: dict[str, dict[str, Any]] = {}
- read_reasoning_catalog = _reasoning_catalog_reader(slug.lower())
- for model in row.get("models") or []:
- reasoning = True
- if get_model_capabilities is not None and slug:
- try:
- meta = get_model_capabilities(slug, model, config=metadata_config)
- if meta is not None and meta.supports_reasoning is not None:
- reasoning = meta.supports_reasoning
- except Exception:
- reasoning = True
- fast = resolve_fast_mode_overrides(
- model, provider=slug, base_url=row.get("api_url")) is not None
- entry: dict[str, Any] = {"fast": fast, "reasoning": reasoning}
- if fast and model_supports_ultrafast(model) and ultrafast_offered(
- model, provider=slug, base_url=row.get("api_url")):
- entry["ultrafast"] = True
- if not _app_server and daybreak_offered(model, provider=slug, base_url=row.get("api_url")):
- entry["daybreak"] = True
- if reasoning and read_reasoning_catalog is not None:
- try:
- detail = read_reasoning_catalog(model)
- except Exception:
- detail = None
- if detail and not detail.get("supports_reasoning"):
- # Aggregator catalog beats models.dev for a route it serves: no reasoning param
- # means no reasoning controls, so no disable to describe either.
- entry["reasoning"] = False
- elif detail:
- entry["can_disable_reasoning"] = not detail.get("mandatory")
- caps[model] = entry
- row["capabilities"] = caps
- # Newest N models per lab an aggregator row features by default (older tail behind search/show-all);
- # 5 keeps a lab's headliners without letting a prolific vendor flood the view.
- _FEATURED_PER_LAB = 5
- def _apply_featured(rows: list[dict], *, metadata_config: dict | None = None) -> None:
- """Attach a ``featured_models`` shortlist to each routing-aggregator row: newest
- ``_FEATURED_PER_LAB`` per vendor by models.dev ``release_date`` (ranked within the row, never vs.
- today, so it is stable); ties keep curated order. Non-aggregators — including every user-defined
- row, whose ``models:`` list is an explicit allow-list — get an empty list and keep top-N
- behaviour (#120217)."""
- try:
- from agent.models_dev import get_model_info
- except Exception:
- get_model_info = None # type: ignore[assignment]
- # "Is this row an aggregator?" is answered canonically by is_routing_aggregator() — the same
- # predicate _strip_aggregator_overlaps() uses. Deriving it from model-id spelling (does any id
- # contain "/" and span >= 2 prefixes?) misread every Org/Model-shaped user provider as a
- # multi-lab aggregator and hid the models its owner configured by hand (#120217).
- try:
- from hermes_cli.providers import is_routing_aggregator
- except Exception:
- is_routing_aggregator = None # type: ignore[assignment]
- for row in rows:
- slug = str(row.get("slug") or "").strip().lower()
- models = row.get("models") or []
- if row.get("is_user_defined") or not (is_routing_aggregator and is_routing_aggregator(slug)):
- row["featured_models"] = []
- continue
- by_lab: dict[str, list[tuple[int, str, str]]] = {} # only multi-lab aggregators get a shortlist
- for pos, model in enumerate(models):
- lab = model.split("/", 1)[0] if "/" in model else ""
- if not lab: # no vendor prefix → single-namespace provider, not an aggregator
- by_lab = {}
- break
- date = ""
- if get_model_info is not None:
- info = (get_model_info(slug, model, config=metadata_config)
- or get_model_info("openrouter", model, config=metadata_config))
- date = getattr(info, "release_date", "") if info else ""
- by_lab.setdefault(lab, []).append((pos, date, model))
- if len(by_lab) < 2:
- row["featured_models"] = []
- continue
- featured: list[str] = []
- for entries in by_lab.values():
- # Newest release_date first; earlier list position breaks ties (sole key when undated).
- ranked = sorted(entries, key=lambda e: (e[1], -e[0]), reverse=True)
- featured.extend(model for _pos, _date, model in ranked[:_FEATURED_PER_LAB])
- order = {m: i for i, m in enumerate(models)} # keep the row's model order for stable rendering
- row["featured_models"] = sorted(featured, key=lambda m: order[m])
- def _apply_custom_aliases(rows: list[dict]) -> None:
- """Attach the accepted identity set to each user-defined row: ``model.options`` reports the canonical
- ``custom:<key>`` while rows carry the bare key as ``slug``, so GUI exact-match never finds the row.
- GUI pickers compare the two to decide which row is active; exact equality never matches for custom
- providers (#87035). Exposing ``aliases`` — every current and legacy spelling from
- :func:`hermes_cli.providers.custom_provider_aliases` — lets the frontend do a membership check instead.
- """
- from hermes_cli.providers import custom_provider_aliases
- for row in rows:
- if not row.get("is_user_defined"):
- continue
- try:
- row["aliases"] = sorted(
- custom_provider_aliases(str(row.get("name", "")), str(row.get("slug", ""))))
- except Exception:
- continue
- # ─── Internal: row post-processing ──────────────────────────────────────
- def _provider_auth_hint(slug: str) -> tuple[str, str]:
- """``(auth_type, key_env)`` for a canonical provider (``("api_key", "")`` when unregistered)."""
- from hermes_cli.auth import PROVIDER_REGISTRY
- cfg = PROVIDER_REGISTRY.get(slug)
- auth_type = cfg.auth_type if cfg else "api_key"
- key_env = cfg.api_key_env_vars[0] if (cfg and cfg.api_key_env_vars) else ""
- return auth_type, key_env
- def _row(slug: str, name: str, is_current: bool, **extra: Any) -> dict:
- return {"slug": slug, "name": name, "is_current": is_current, "is_user_defined": False, **extra}
- def _canonical_row(entry, cur: str, **extra: Any) -> dict:
- from hermes_cli.models import _PROVIDER_LABELS
- return _row(entry.slug, _PROVIDER_LABELS.get(entry.slug, entry.label), entry.slug.lower() == cur, **extra)
- def _append_unconfigured_rows(
- rows: list[dict], ctx: ConfigContext, *, current_only: bool = False,
- ) -> list[dict]:
- """Empty setup skeletons for canonical providers missing from ``rows`` — except the *current* one:
- if config.yaml still points at it but credentials are gone, keep a row carrying the saved model so
- GUI pickers don't silently snap to another provider."""
- from hermes_cli.models import _PROVIDER_ALIASES, _model_requires_account_discovery
- from hermes_cli.models_catalog_static import listed_canonical_providers
- seen = {r["slug"].lower() for r in rows}
- cur = (ctx.current_provider or "").lower()
- cur_model = str(ctx.current_model or "").strip()
- # Honor ``model_catalog.excluded_providers`` like ``list_authenticated_providers`` does — unconditionally,
- # even for the current provider — or ``include_unconfigured`` pickers (TUI ``/model``) resurrect every
- # excluded provider as a skeleton row (#68816). A slug is hidden when it or any alias is excluded, so an
- # alias exclusion (``google`` → ``gemini``) can't leak the canonical row back.
- excluded = {str(p).strip().lower() for p in (ctx.excluded_providers or []) if p}
- names_for: dict[str, set[str]] = {}
- for alias, canon in _PROVIDER_ALIASES.items():
- names_for.setdefault(canon.lower(), {canon.lower()}).add(alias.lower())
- extras: list[dict] = []
- for entry in listed_canonical_providers():
- slug = entry.slug.lower()
- if slug in seen:
- continue
- if names_for.get(slug, {slug}) & excluded:
- continue
- if current_only and entry.slug.lower() != cur:
- continue
- if entry.slug.lower() == cur:
- saved_model = "" if _model_requires_account_discovery(entry.slug, cur_model) else cur_model
- auth_type, key_env = _provider_auth_hint(entry.slug)
- tail = (
- "Astra requires successful account-scoped model discovery."
- if cur_model and not saved_model else "Showing the saved model only."
- )
- warning = (
- f"Configured provider missing usable credentials; paste {key_env} to reactivate. {tail}"
- if auth_type == "api_key" and key_env
- else f"Configured provider is not authenticated; run `hermes model` to reactivate. {tail}"
- )
- extras.append(_canonical_row(
- entry, cur, models=[saved_model] if saved_model else [], total_models=1 if saved_model else 0,
- source="configured-current", authenticated=False, auth_type=auth_type, key_env=key_env,
- warning=warning,
- ))
- continue
- extras.append(_canonical_row(entry, cur, models=[], total_models=0, source="canonical"))
- return extras
- def _anthropic_oauth_credentials_present() -> bool:
- """True when the user explicitly authenticated Anthropic via OAuth (Hermes device flow or Claude Code
- login) — those leave no trace in active_provider / model.provider / API-key env vars."""
- try:
- from agent.anthropic_credentials import read_claude_code_credentials, read_hermes_oauth_credentials
- readers = (read_hermes_oauth_credentials, read_claude_code_credentials)
- if any((read() or {}).get("accessToken") for read in readers):
- return True
- except Exception:
- return False
- # Pool-only OAuth entries (auth.json credential_pool.anthropic) are equally deliberate — discovery
- # accepts them via pool.has_credentials(), so the filter must too or those rows are built then
- # silently dropped. Read-only (no load_pool) so a picker open never mutates auth.json.
- try:
- from agent.credential_pool import AUTH_TYPE_OAUTH
- from hermes_cli.auth import read_credential_pool
- for entry in read_credential_pool("anthropic"):
- if (isinstance(entry, dict) and entry.get("auth_type") == AUTH_TYPE_OAUTH
- and str(entry.get("access_token") or "").strip()):
- return True
- except Exception:
- pass
- return False
- def _filter_explicit_provider_rows(rows: list[dict], ctx: ConfigContext) -> list[dict]:
- """Keep only rows backed by explicit user configuration — ``list_authenticated_providers`` also
- discovers ambient credentials (e.g. another CLI's login token) Desktop chat pickers must not show."""
- from hermes_cli.auth import is_provider_explicitly_configured
- current_slug = str(ctx.current_provider or "").strip().lower()
- def _is_explicit(row: dict, slug: str) -> bool:
- # Managed local models are explicit configuration by existence (gigabytes downloaded into the
- # machine-scoped dir); there is deliberately no config credential, so without the source clause
- # the row would only survive on the profile where Use was last clicked.
- if (row.get("is_user_defined") or (current_slug and slug == current_slug)
- or row.get("source") == "local-runtime"):
- return True
- if slug == "moa":
- # Virtual routing mode, not a configured provider: hide unless current (above) or the user
- # wrote an enabled preset into RAW config (the DEFAULT_CONFIG preset must not show MoA).
- return _raw_config_has_enabled_moa_preset()
- return (
- # Anthropic OAuth (device flow / Claude Code) and external-process CLIs (ACP agents) are
- # deliberate sign-ins that leave no trace in config/env; keep the rows discovery accepted.
- (slug == "anthropic" and _anthropic_oauth_credentials_present())
- or _external_process_signed_in(slug)
- or is_provider_explicitly_configured(slug)
- )
- return [row for row in rows
- if (slug := str(row.get("slug", "")).strip().lower()) and _is_explicit(row, slug)]
- def _external_process_signed_in(slug: str) -> bool:
- """True when an external-process provider has verified CLI credentials."""
- try:
- from hermes_cli.auth import PROVIDER_REGISTRY, get_external_process_provider_status
- pconfig = PROVIDER_REGISTRY.get(slug)
- return bool(pconfig and pconfig.auth_type == "external_process"
- and get_external_process_provider_status(slug).get("auth_verified"))
- except Exception:
- return False
- def _raw_config_has_enabled_moa_preset() -> bool:
- """True when the user's RAW config enables MoA: ``load_config()`` merges the DEFAULT_CONFIG preset for
- everyone, which is not a user choice; visible once one enabled preset (or legacy flat config) is saved."""
- try:
- from hermes_cli.config import read_raw_config
- raw = read_raw_config()
- except Exception:
- return False
- moa = raw.get("moa") if isinstance(raw, dict) else None
- if not isinstance(moa, dict):
- return False
- presets = moa.get("presets")
- if isinstance(presets, dict):
- return any(
- not isinstance(preset, dict) or preset.get("enabled", True)
- for name, preset in presets.items() if str(name or "").strip()
- )
- legacy_keys = {"reference_models", "aggregator", "reference_temperature", "aggregator_temperature",
- "max_tokens", "reference_max_tokens", "fanout"}
- return any(key in moa for key in legacy_keys) and bool(moa.get("enabled", True))
- def _apply_picker_hints(rows: list[dict]) -> None:
- """Add ``authenticated``/``auth_type``/``key_env``/``warning`` per row."""
- for row in rows:
- if "authenticated" in row:
- continue
- # Skeleton rows (_append_unconfigured_rows) have empty `models` AND source="canonical".
- is_skeleton = row.get("source") == "canonical" and not row.get("models")
- row["authenticated"] = not is_skeleton
- if not is_skeleton or row.get("is_user_defined"):
- continue
- auth_type, key_env = _provider_auth_hint(row["slug"])
- row["auth_type"] = auth_type
- row["key_env"] = key_env
- row["warning"] = (f"paste {key_env} to activate" if auth_type == "api_key" and key_env
- else f"run `hermes model` to configure ({auth_type})")
- def _reorder_canonical(rows: list[dict]) -> list[dict]:
- """Canonical slugs in ``CANONICAL_PROVIDERS`` order, truly-custom rows last. Keys on slug membership,
- NOT ``is_user_defined`` — ``providers:`` config rows carry that flag even for canonical slugs."""
- from hermes_cli.models import CANONICAL_PROVIDERS
- order = {e.slug: i for i, e in enumerate(CANONICAL_PROVIDERS)}
- canon = sorted((r for r in rows if r["slug"] in order), key=lambda r: order[r["slug"]])
- extras = [r for r in rows if r["slug"] not in order]
- return canon + extras
- def _attach_nous_plan_usage(row: dict, *, warm: bool):
- """Stamp ``plan_usage`` from the portal snapshot (window plans only) and return the PickerUsage (or
- None). A cache-only open reads the last snapshot at any age: the 60s account cache has usually lapsed
- by the time someone opens the composer menu. Accounts without a window plan get nothing."""
- try:
- from hermes_cli.model_picker_usage import nous_picker_usage
- usage = nous_picker_usage(warm=warm, any_age=not warm)
- except Exception:
- logging.getLogger(__name__).debug("Nous plan usage unavailable", exc_info=True)
- return None
- if usage is not None:
- row["plan_usage"] = usage.to_payload()
- # Main's subscription-usage chip (`usage.windows`): the same gauge for a window
- # plan, so it warns near a limit in the existing slot. Once an allowance is full the picker's limit
- # line says it (with what happens next), so the chip would only repeat it; paused plans say so in
- # the plan line.
- if not usage.paused_reason and not row.get("usage") and not any(g.left_fraction <= 0.0 for g in usage.gauges):
- windows = [_plan_usage_window(g) for g in usage.gauges if g.resets_at_ms]
- if windows:
- row["usage"] = {"windows": windows}
- return usage
- def _plan_usage_window(g) -> dict:
- """One window-plan allowance in main's ``ProviderUsageWindow`` shape. The label is Hermes's row title
- (``Today``), like the other providers' English window labels."""
- from datetime import datetime
- from agent.usage_gauges import WindowGauge, gauge_title
- return {
- "label": gauge_title(WindowGauge(g.label, g.scope, g.title or g.label, 0)),
- "used_percent": round(g.used_fraction * 100.0, 2),
- "resets_at": datetime.fromtimestamp(g.resets_at_ms / 1000, UTC).isoformat(),
- "scope": "account" if g.scope == "all" else "model",
- }
- # health: allow CC -- usage-window plans: one branch attaches plan usage to the Nous row (built in _attach_nous_plan_usage); over target on main
- def _apply_pricing(rows: list[dict], *, force_fresh_nous_tier: bool = False, cached_only: bool = False) -> None:
- """Set ``row["pricing"] = {model_id: {input, output, cache | None, free}}``; for Nous also
- ``free_tier`` (account is free-tier) and ``unavailable_models`` (paid models a free user can't pick).
- ``cached_only`` never hits the network: unknown Nous entitlement fails closed (``free_tier_pending``,
- all models locked) and missing pricing is marked ``pricing_pending``."""
- from hermes_cli.models_pricing import (
- _format_price_per_mtok,
- compute_sale_discount,
- get_pricing_for_provider,
- )
- from hermes_cli.models import (
- check_nous_free_tier,
- get_cached_nous_free_tier,
- partition_nous_models_by_tier,
- )
- nous_free_tier: Optional[bool] = None # resolved once (cached in models.py for the TTL window)
- for row in rows:
- slug = str(row.get("slug", "")).lower()
- models = row.get("models") or []
- if not models:
- continue
- if row.get("free_tier_row"):
- # The free tier's one model has no Portal pricing and no entitlement to read: pricing
- # it would lock the only row a free-tier install can select.
- continue
- try:
- pricing_kwargs = {"cached_only": True} if cached_only else {}
- if slug.startswith("custom:"):
- pricing_kwargs["base_url"] = str(row.get("api_url") or "")
- raw_pricing = get_pricing_for_provider(slug, **pricing_kwargs) or {}
- except Exception:
- raw_pricing = {}
- cached_nous_tier: Optional[bool] = None
- plan_usage = _attach_nous_plan_usage(row, warm=not cached_only) if slug == "nous" else None
- if slug == "nous" and cached_only:
- cached_nous_tier = get_cached_nous_free_tier()
- if cached_nous_tier is None:
- # Entitlement unknown: stay nonblocking but fail closed until the prewarm has populated
- # both caches, else a free account could briefly select paid models on first open.
- row["free_tier_pending"] = True
- row["unavailable_models"] = list(models)
- if not row.get("warning"): # say why every model renders locked
- row["warning"] = ("Checking Nous plan entitlement… models unlock on the "
- "next picker open or refresh.")
- continue
- if not raw_pricing:
- if slug == "nous":
- row["free_tier"] = bool(cached_nous_tier)
- row["pricing_pending"] = True
- row["unavailable_models"] = list(models) if cached_nous_tier else []
- continue
- formatted: dict[str, dict] = {}
- for mid in models:
- p = raw_pricing.get(mid)
- if not p:
- continue
- inp_raw, out_raw = p.get("prompt", ""), p.get("completion", "")
- cache_raw = p.get("input_cache_read", "")
- inp = _format_price_per_mtok(inp_raw) if inp_raw != "" else ""
- out = _format_price_per_mtok(out_raw) if out_raw != "" else ""
- entry: dict = {
- "input": inp, "output": out,
- "cache": _format_price_per_mtok(cache_raw) if cache_raw else None,
- "free": inp == "free" and out in ("free", ""), # both input and output cost nothing
- }
- # Sale chrome is Nous Portal-only (other catalogs' nested pricing.original is ignored); free
- # models get flat -100% chrome, was_* only when the gateway served an original.
- if slug == "nous":
- sale = compute_sale_discount(inp_raw, out_raw, p.get("original"))
- if sale is not None:
- discount_percent, was_prompt_raw, was_out_raw = sale
- entry["discount_percent"] = discount_percent
- for key, was_raw in (("was_input", was_prompt_raw), ("was_output", was_out_raw)):
- if was_raw != "":
- entry[key] = _format_price_per_mtok(was_raw)
- formatted[mid] = entry
- if formatted:
- row["pricing"] = formatted
- if slug == "nous":
- try:
- if nous_free_tier is None:
- nous_free_tier = (cached_nous_tier if cached_only
- else check_nous_free_tier(force_fresh=force_fresh_nous_tier))
- row["free_tier"] = bool(nous_free_tier)
- row["unavailable_models"] = (
- partition_nous_models_by_tier(list(models), raw_pricing, free_tier=True)[1]
- if nous_free_tier else [])
- except Exception: # tier detection failed — fail open (no gating)
- row["free_tier"] = False
- row["unavailable_models"] = []
- if plan_usage is not None:
- from hermes_cli.model_picker_usage import usage_levels
- # The gateway's per-plan word per model (free / low / medium / top_tier).
- row["usage_levels"] = usage_levels(list(models), raw_pricing)
- # Top-tier marks from the gateway (it decides, not the client).
- row["plan_usage"] = plan_usage.with_pricing({m: raw_pricing.get(m) for m in models}).to_payload()
- # Window plans: the picker shows the usage word, not per-token prices.
- row.pop("pricing", None)
- def _local_runtime_row(ctx: ConfigContext) -> dict | None:
- """The ``llamacpp`` row from staged GGUFs (``None`` when none) — downloaded models must be selectable
- before the server runs (selection starts it via the runtime_provider seam). The row's id comes from
- the provider registry's own definition, never a local literal: a row the resolver can't resolve is
- the bug this row's offline-first contract depends on not having."""
- try:
- from hermes_cli.local_runtime.bootstrap import staged_model_ids
- from hermes_cli.providers import LLAMACPP_ALIASES, LLAMACPP_PROVIDER_ID
- staged = staged_model_ids()
- if not staged:
- return None
- current = (ctx.current_provider or "").strip().lower() in LLAMACPP_ALIASES
- if not current:
- # A LIVE session on the managed server reports provider "custom" with the managed base_url;
- # match on the endpoint so the session being chatted in still shows a selection.
- try:
- from hermes_cli.local_runtime.endpoint import _state_endpoint
- managed = _state_endpoint()
- current = bool(managed and (ctx.current_base_url or "").strip().rstrip("/")
- == managed["base_url"].rstrip("/"))
- except Exception:
- current = False
- # Bare "Local" user-facing (engine name is an implementation detail); authenticated = reachability.
- return _row(LLAMACPP_PROVIDER_ID, "Local", current, models=staged, total_models=len(staged),
- source="local-runtime", authenticated=True, auth_type="local", warning=None)
- except Exception:
- return None
- def _prewarm_pricing_async(
- rows: list[dict], *, current_provider: str = "", current_base_url: str = "",
- ) -> Optional[Thread]:
- """Warm picker pricing caches without delaying the current payload (one worker per
- profile + endpoint scope; a live worker is reused)."""
- from hermes_constants import hermes_home_key
- from hermes_cli.models_pricing import pricing_cache_scope
- slugs = {
- (
- str(row.get("slug") or "").lower(),
- str(row.get("api_url") or "") if str(row.get("slug") or "").lower().startswith("custom:") else "",
- )
- for row in rows if row.get("slug")
- }
- endpoint_scope = tuple(sorted(
- (slug, pricing_cache_scope(
- slug, base_url=base_url, current_provider=current_provider, current_base_url=current_base_url,
- ))
- for slug, base_url in slugs))
- prewarm_key = (hermes_home_key(), endpoint_scope)
- with _pricing_prewarm_lock:
- current = _pricing_prewarm_threads.get(prewarm_key)
- if current is not None and current.is_alive():
- return current
- # The worker mutates only private copies; the pricing helpers populate shared process caches.
- worker_rows = [{**row, "models": list(row.get("models") or [])} for row in rows]
- def _worker() -> None:
- try:
- _apply_pricing(worker_rows)
- finally:
- with _pricing_prewarm_lock:
- if _pricing_prewarm_threads.get(prewarm_key) is current_thread():
- _pricing_prewarm_threads.pop(prewarm_key, None)
- thread = Thread(target=copy_context().run, args=(_worker,),
- name="hermes-picker-pricing-prewarm", daemon=True)
- _pricing_prewarm_threads[prewarm_key] = thread
- thread.start()
- return thread
- def _moa_provider_row(current_provider: str = "") -> dict | None:
- """The virtual ``moa`` row shared by the CLI inventory and gateway picker; ``None`` without presets.
- Strictly opt-in (#63353): the row only appears when the user's raw config.yaml explicitly
- enables at least one MoA preset. The synthesized ``default`` preset from
- ``normalize_moa_config({})`` — which every user gets via DEFAULT_CONFIG defaults — must not
- be treated as a user choice."""
- if not _raw_config_has_enabled_moa_preset():
- return None
- try:
- from hermes_cli.config import load_config
- from hermes_cli.moa_config import normalize_moa_config
- cfg = normalize_moa_config(load_config().get("moa") or {})
- models = list(cfg.get("presets", {}).keys())
- if not models:
- return None
- return _row(
- "moa", "Mixture of Agents", (current_provider or "").lower() == "moa", models=models,
- total_models=len(models), source="virtual", authenticated=True, auth_type="virtual",
- warning="Aggregator is the acting model billed for the run; references only advise once per user turn by default.")
- except Exception:
- return None
|