ouyangbo/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 9 Packages Projects Releases Wiki Activity

refactor(hermes_cli/web_routers): models — _main_model_fields/_preset_dict field table, nous default helper, compact docstrings (443->366 LOC)

Browse Source
This commit is contained in:
Teknium
2026-09-02 21:00:30 -07:00
parent 448e1fa50c
commit 8b26b07cdf
1 changed files with 127 additions and 204 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hermes_cli/web_routers/models.py
+127 -204
View File
@@ -1,18 +1,19 @@
"""Model assignment dashboard routes: model info/options/recommended default, auxiliary + MoA slots, /api/model/set.
Extracted from ``hermes_cli.web_server``; helpers/state that tests monkeypatch on
``web_server`` stay there and are imported lazily at call time (cycle-safe).
``web_server`` stay there and are resolved late at call time (cycle-safe).
"""
import logging
import asyncio
from fastapi import APIRouter
from hermes_cli.web_routers._common import http_failure
from hermes_cli.web_deps import late
from fastapi import HTTPException
from hermes_cli.web_models import ModelAssignment, MoaModelSlot, MoaPresetPayload, MoaConfigPayload
import logging
from typing import Optional
from fastapi import APIRouter, HTTPException
from hermes_cli.web_deps import LateState, late
from hermes_cli.web_models import ModelAssignment, MoaConfigPayload, MoaModelSlot
from hermes_cli.web_routers._common import http_failure
_log = logging.getLogger("hermes_cli.web_server")
router = APIRouter()
@@ -24,6 +25,7 @@ _profile_scope = late("_profile_scope")
load_config = late("load_config")
run_in_threadpool = late("run_in_threadpool")
save_config = late("save_config")
_AUX_TASK_SLOTS = LateState("_AUX_TASK_SLOTS")
_EMPTY_MODEL_INFO: dict = {
@@ -36,55 +38,45 @@ _EMPTY_MODEL_INFO: dict = {
}
def _main_model_fields(model_cfg) -> tuple[str, str]:
"""(model, provider) from config's ``model`` section, which may be a plain string."""
if isinstance(model_cfg, dict):
return model_cfg.get("default", model_cfg.get("name", "")), model_cfg.get("provider", "")
return (str(model_cfg) if model_cfg else ""), ""
def _load_config_scoped(profile: Optional[str]) -> dict:
with _profile_scope(profile):
return load_config()
@router.get("/api/model/info")
def get_model_info(profile: Optional[str] = None):
"""Return resolved model metadata for the currently configured model.
Calls the same context-length resolution chain the agent uses, so the
frontend can display "Auto-detected: 200K" alongside the override field.
Also returns model capabilities (vision, reasoning, tools) when available.
"""
"""Resolved metadata for the configured model: auto-detected vs configured
context length (so the UI can show "Auto-detected: 200K" beside the
override) plus models.dev capabilities when available."""
try:
with _profile_scope(profile):
cfg = load_config()
model_cfg = cfg.get("model", "")
# Extract model name and provider from the config
if isinstance(model_cfg, dict):
model_name = model_cfg.get("default", model_cfg.get("name", ""))
provider = model_cfg.get("provider", "")
base_url = model_cfg.get("base_url", "")
config_ctx = model_cfg.get("context_length")
else:
model_name = str(model_cfg) if model_cfg else ""
provider = ""
base_url = ""
config_ctx = None
model_cfg = _load_config_scoped(profile).get("model", "")
model_name, provider = _main_model_fields(model_cfg)
base_url = model_cfg.get("base_url", "") if isinstance(model_cfg, dict) else ""
config_ctx = model_cfg.get("context_length") if isinstance(model_cfg, dict) else None
if not model_name:
return dict(_EMPTY_MODEL_INFO, provider=provider)
# Resolve auto-detected context length (pass config_ctx=None to get
# purely auto-detected value, then separately report the override)
try:
from agent.model_metadata import get_model_context_length
auto_ctx = get_model_context_length(
model=model_name,
base_url=base_url,
provider=provider,
config_context_length=None, # ignore override — we want auto value
config_context_length=None, # ignore override — we want the auto value
)
except Exception:
auto_ctx = 0
config_ctx_int = 0
if isinstance(config_ctx, int) and config_ctx > 0:
config_ctx_int = config_ctx
config_ctx_int = config_ctx if isinstance(config_ctx, int) and config_ctx > 0 else 0
# Effective is what the agent actually uses
effective_ctx = config_ctx_int if config_ctx_int > 0 else auto_ctx
# Try to get model capabilities from models.dev
caps = {}
try:
from agent.models_dev import get_model_capabilities
@@ -106,7 +98,7 @@ def get_model_info(profile: Optional[str] = None):
"provider": provider,
"auto_context_length": auto_ctx,
"config_context_length": config_ctx_int,
"effective_context_length": effective_ctx,
"effective_context_length": config_ctx_int or auto_ctx, # what the agent actually uses
"capabilities": caps,
}
except HTTPException:
@@ -125,39 +117,28 @@ async def get_model_options(
include_unconfigured: bool = False,
explicit_only: bool = False,
):
"""Return authenticated providers + their curated model lists.
"""Authenticated providers + curated model lists — REST twin of the
``model.options`` JSON-RPC on tui_gateway, same response shape so
``ModelPickerDialog`` shares the types.
REST equivalent of the ``model.options`` JSON-RPC on tui_gateway, so the
dashboard Models page can render the picker without a live chat session.
The response shape matches ``model.options`` 1:1 so ``ModelPickerDialog``
can share the same types.
``profile`` scopes the picker context (current model/provider, custom
providers from config, per-profile .env auth state) so the Models page
reads the SAME profile /api/model/set writes.
``refresh`` busts the per-provider model-id disk cache so every row
re-fetches its live catalog — used by the picker's explicit "Refresh
Models" control. Normal opens leave it false to stay on the 1h cache.
``profile`` scopes the picker context so the Models page reads the SAME
profile /api/model/set writes. ``refresh`` busts the per-provider model-id
disk cache (picker's explicit "Refresh Models"); normal opens stay on the 1h cache.
"""
with http_failure("GET /api/model/options failed", 500, detail="Failed to list model options"):
skew_msg = _dashboard_code_skew_guard()
if skew_msg:
_log.warning("GET /api/model/options refused: %s", skew_msg)
raise HTTPException(
status_code=503, detail=f"Restart required: {skew_msg}"
)
raise HTTPException(status_code=503, detail=f"Restart required: {skew_msg}")
from hermes_cli.inventory import build_model_options_payload, load_picker_context
def _build_payload_scoped() -> dict:
# Keep the profile override inside the worker thread so the full
# sync picker build (config load, pricing, refresh probes) runs
# off the event loop under the requested profile.
# Use _config_profile_scope (contextvar only, no skill-module
# lock) — the payload build can block for 15s on a models.dev
# cache miss, and _profile_scope's RLock held across that block
# starves concurrent /api/config and freezes the server (#58576).
# The full sync picker build runs off the event loop under the
# requested profile. _config_profile_scope (contextvar only, no
# skill-module lock): the build can block 15s on a models.dev cache
# miss, and _profile_scope's RLock held across that starves
# concurrent /api/config and freezes the server.
with _config_profile_scope(profile):
return build_model_options_payload(
load_picker_context(),
@@ -169,82 +150,65 @@ async def get_model_options(
return await run_in_threadpool(_build_payload_scoped)
def _nous_recommended_default() -> dict:
from hermes_cli.models import (
get_curated_nous_model_ids,
get_pricing_for_provider,
check_nous_free_tier,
nous_policy_allowed_ids,
partition_nous_models_by_tier,
pick_silent_default_model,
restrict_to_nous_policy,
union_with_portal_free_recommendations,
union_with_portal_paid_recommendations,
)
from hermes_cli.auth import get_provider_auth_state
model_ids = get_curated_nous_model_ids()
pricing = get_pricing_for_provider("nous") or {}
free_tier = check_nous_free_tier(force_fresh=True)
try:
portal_url = (get_provider_auth_state("nous") or {}).get("portal_base_url", "") or ""
except Exception:
portal_url = ""
# This endpoint picks the model a user lands on without choosing it, so an
# unreachable one here is worse than in a picker. Narrow to policy before
# the tier split, so a rescued id still has to pass the free/paid predicate.
policy_allowed = nous_policy_allowed_ids()
union = union_with_portal_free_recommendations if free_tier else union_with_portal_paid_recommendations
model_ids, pricing = union(model_ids, pricing, portal_url)
model_ids = restrict_to_nous_policy(model_ids, policy_allowed, rescue_empty=True)
if free_tier:
model_ids, _unavailable = partition_nous_models_by_tier(model_ids, pricing, free_tier=True)
model = pick_silent_default_model(model_ids, provider="nous")
return {"provider": "nous", "model": model, "free_tier": bool(free_tier)}
@router.get("/api/model/recommended-default")
def get_recommended_default_model(provider: str = ""):
"""Return the recommended default model for a freshly-authenticated provider.
"""Recommended default model for a freshly-authenticated provider, mirroring
``hermes model``'s curation so GUI onboarding lands on a sensible default.
Mirrors the model-curation `hermes model` does so GUI onboarding lands on a
sensible default instead of blindly taking the first curated entry. For
Nous this honors the user's free/paid tier: free users get a free model,
paid users get the full curated default. For any other provider it falls
back to the first curated model (same as before).
Nous honors the user's free/paid tier. Any other provider gets the preferred
silent default when its curated list carries it, else the first curated model
— aggregator lists lead with the priciest Anthropic flagship, which must never
be the model a user lands on without explicitly picking it.
Response: {"provider": str, "model": str, "free_tier": bool | None}
where free_tier is True/False for Nous and None otherwise. `model` may be
empty if nothing could be resolved (caller degrades gracefully).
Response: {"provider", "model", "free_tier": bool | None} — free_tier only
for Nous; ``model`` may be empty (caller degrades gracefully).
"""
slug = (provider or "").strip().lower()
if slug == "nous":
try:
from hermes_cli.models import (
get_curated_nous_model_ids,
get_pricing_for_provider,
check_nous_free_tier,
nous_policy_allowed_ids,
partition_nous_models_by_tier,
pick_silent_default_model,
restrict_to_nous_policy,
union_with_portal_free_recommendations,
union_with_portal_paid_recommendations,
)
from hermes_cli.auth import get_provider_auth_state
model_ids = get_curated_nous_model_ids()
pricing = get_pricing_for_provider("nous") or {}
free_tier = check_nous_free_tier(force_fresh=True)
portal_url = ""
try:
state = get_provider_auth_state("nous") or {}
portal_url = state.get("portal_base_url", "") or ""
except Exception:
portal_url = ""
# This endpoint picks the model a user lands on without choosing it,
# so an unreachable one here is worse than in a picker. Narrow before
# the tier split, so a rescued id still has to pass the free/paid
# predicate.
_policy_allowed = nous_policy_allowed_ids()
if free_tier:
model_ids, pricing = union_with_portal_free_recommendations(
model_ids, pricing, portal_url
)
model_ids = restrict_to_nous_policy(
model_ids, _policy_allowed, rescue_empty=True,
)
model_ids, _unavailable = partition_nous_models_by_tier(
model_ids, pricing, free_tier=True
)
else:
model_ids, pricing = union_with_portal_paid_recommendations(
model_ids, pricing, portal_url
)
model_ids = restrict_to_nous_policy(
model_ids, _policy_allowed, rescue_empty=True,
)
model = pick_silent_default_model(model_ids, provider="nous")
return {"provider": "nous", "model": model, "free_tier": bool(free_tier)}
return _nous_recommended_default()
except Exception:
_log.exception("GET /api/model/recommended-default (nous) failed")
return {"provider": "nous", "model": "", "free_tier": None}
# Non-Nous: preferred silent default when the provider's curated list
# carries it, else the first curated model. Aggregator lists lead with the
# priciest Anthropic flagship (claude-fable-5), which must never be the
# model a user lands on without explicitly picking it.
try:
from hermes_cli.inventory import build_models_payload, load_picker_context
from hermes_cli.models import pick_silent_default_model
@@ -262,25 +226,14 @@ def get_recommended_default_model(provider: str = ""):
@router.get("/api/model/auxiliary")
def get_auxiliary_models(profile: Optional[str] = None):
"""Return current auxiliary task assignments.
"""Current auxiliary task assignments: ``{"tasks": [{task, provider, model,
base_url}, ...], "main": {provider, model}}``.
Shape:
{
"tasks": [
{"task": "vision", "provider": "auto", "model": "", "base_url": ""},
...
],
"main": {"provider": "openrouter", "model": "anthropic/claude-opus-4.7"},
}
``profile`` scopes the read — without it, the Models page would show
the dashboard profile's auxiliary pins while /api/model/set wrote the
selected profile's (read/write asymmetry).
``profile`` scopes the read — without it the Models page would show the
dashboard profile's pins while /api/model/set wrote the selected profile's.
"""
from hermes_cli.web_server import _AUX_TASK_SLOTS
with http_failure("GET /api/model/auxiliary failed", 500, detail="Failed to read auxiliary config"):
with _profile_scope(profile):
cfg = load_config()
cfg = _load_config_scoped(profile)
aux_cfg = cfg.get("auxiliary", {})
if not isinstance(aux_cfg, dict):
aux_cfg = {}
@@ -295,16 +248,8 @@ def get_auxiliary_models(profile: Optional[str] = None):
"base_url": str(slot_cfg.get("base_url", "") or ""),
})
model_cfg = cfg.get("model", {})
if isinstance(model_cfg, dict):
main = {
"provider": str(model_cfg.get("provider", "") or ""),
"model": str(model_cfg.get("default", model_cfg.get("name", "")) or ""),
}
else:
main = {"provider": "", "model": str(model_cfg) if model_cfg else ""}
return {"tasks": tasks, "main": main}
model, provider = _main_model_fields(cfg.get("model", {}))
return {"tasks": tasks, "main": {"provider": str(provider or ""), "model": str(model or "")}}
@router.get("/api/model/moa")
@@ -318,30 +263,32 @@ def get_moa_models(profile: Optional[str] = None):
return normalize_moa_config(cfg.get("moa") if isinstance(cfg, dict) else {})
_MOA_PRESET_FIELDS = (
"reference_temperature", "aggregator_temperature", "reference_timeout",
"degraded_reference_policy", "max_tokens", "reference_max_tokens", "fanout", "enabled",
)
def _slot_dict(slot: MoaModelSlot) -> dict:
# Drop unset optionals so saved slots stay minimal ({provider, model}).
return {k: v for k, v in slot.dict().items() if v is not None}
def _preset_dict(preset) -> dict:
"""Raw preset dict from a MoaPresetPayload or the flat MoaConfigPayload fields."""
return {
"reference_models": [_slot_dict(slot) for slot in preset.reference_models],
"aggregator": _slot_dict(preset.aggregator),
**{name: getattr(preset, name) for name in _MOA_PRESET_FIELDS},
}
@router.put("/api/model/moa")
def set_moa_models(body: MoaConfigPayload, profile: Optional[str] = None):
"""Persist the Mixture-of-Agents provider/model slots."""
with http_failure("PUT /api/model/moa failed", 500, detail="Failed to save MoA config"):
from hermes_cli.moa_config import normalize_moa_config, validate_moa_payload
def _slot_dict(slot: MoaModelSlot) -> dict:
# Drop unset optionals so saved slots stay minimal ({provider, model}).
return {k: v for k, v in slot.dict().items() if v is not None}
def _preset_dict(preset: MoaPresetPayload) -> dict:
return {
"reference_models": [_slot_dict(slot) for slot in preset.reference_models],
"aggregator": _slot_dict(preset.aggregator),
"reference_temperature": preset.reference_temperature,
"aggregator_temperature": preset.aggregator_temperature,
"reference_timeout": preset.reference_timeout,
"degraded_reference_policy": preset.degraded_reference_policy,
"max_tokens": preset.max_tokens,
"reference_max_tokens": preset.reference_max_tokens,
"fanout": preset.fanout,
"enabled": preset.enabled,
}
with _profile_scope(body.profile or profile):
cfg = load_config()
if body.presets:
@@ -351,37 +298,19 @@ def set_moa_models(body: MoaConfigPayload, profile: Optional[str] = None):
"presets": {name: _preset_dict(preset) for name, preset in body.presets.items()},
}
else:
raw = _preset_dict(
MoaPresetPayload(
reference_models=body.reference_models,
aggregator=body.aggregator,
reference_temperature=body.reference_temperature,
aggregator_temperature=body.aggregator_temperature,
reference_timeout=body.reference_timeout,
degraded_reference_policy=body.degraded_reference_policy,
max_tokens=body.max_tokens,
reference_max_tokens=body.reference_max_tokens,
fanout=body.fanout,
enabled=body.enabled,
)
)
raw = _preset_dict(body) # legacy flat payload from older clients
# Reject-don't-repair: normalize_moa_config() silently swaps any
# preset containing incomplete slots for the hardcoded defaults —
# correct tolerance for hand-edited configs at READ time, silent
# data loss at WRITE time (#64156: desktop autosave of a
# half-filled slot replaced the user's whole preset). Refuse the
# save loudly so no client can corrupt config through this route.
# correct tolerance at READ time, silent data loss at WRITE time
# (desktop autosave of a half-filled slot replaced the user's whole
# preset). Refuse loudly so no client can corrupt config here.
problems = validate_moa_payload(raw)
if problems:
raise HTTPException(
status_code=422,
detail="Invalid MoA config: " + "; ".join(problems),
)
raise HTTPException(status_code=422, detail="Invalid MoA config: " + "; ".join(problems))
normalized = normalize_moa_config(raw)
# Merge instead of overwrite so that hand-edited keys not declared
# in MoaConfigPayload (e.g. save_traces, trace_dir) survive a GUI
# save. See issue #58819.
# Merge instead of overwrite so hand-edited keys not declared in
# MoaConfigPayload (e.g. save_traces, trace_dir) survive a GUI save.
cfg.setdefault("moa", {}).update(normalized)
save_config(cfg)
return {"ok": True, **normalized}
@@ -391,9 +320,8 @@ def set_moa_models(body: MoaConfigPayload, profile: Optional[str] = None):
async def set_model_assignment(body: ModelAssignment, profile: Optional[str] = None):
"""Assign a model to the main slot or an auxiliary task slot.
Writes to ``~/.hermes/config.yaml`` — applies to **new** sessions only.
The currently running chat PTY (if any) is not affected; use the
``/model`` slash command inside a chat to hot-swap that specific session.
Writes ``~/.hermes/config.yaml`` — applies to **new** sessions only; a
running chat PTY hot-swaps via the ``/model`` slash command instead.
"""
scope = (body.scope or "").strip().lower()
provider = (body.provider or "").strip()
@@ -417,10 +345,7 @@ async def set_model_assignment(body: ModelAssignment, profile: Optional[str] = N
# Pricing lookup can hit models.dev / a /models endpoint on a
# cache miss — keep it off the event loop.
warning = await asyncio.to_thread(
combined_selection_warning,
model,
provider=provider,
base_url=base_url,
combined_selection_warning, model, provider=provider, base_url=base_url
)
except Exception:
warning = None
@@ -436,8 +361,6 @@ async def set_model_assignment(body: ModelAssignment, profile: Optional[str] = N
def _apply_assignment():
with _profile_scope(body.profile or profile):
return _apply_model_assignment_sync(
scope, provider, model, task, base_url, api_key
)
return _apply_model_assignment_sync(scope, provider, model, task, base_url, api_key)
return await asyncio.to_thread(_apply_assignment)