| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427 |
- """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 resolved late at call time (cycle-safe).
- """
- import asyncio
- import concurrent.futures
- import logging
- from typing import Optional
- from fastapi import APIRouter, HTTPException
- from hermes_cli.web_deps import LateState, late
- from hermes_cli.web_server_config import (
- _AUX_TASK_SLOTS, _UNSET, _apply_model_assignment_sync, _dashboard_code_skew_guard,
- _plugin_aux_tasks, _prepare_main_assignment,
- )
- from agent.model_metadata import is_local_endpoint
- from starlette.concurrency import run_in_threadpool
- from hermes_cli.web_models import ModelAssignment, MoaConfigPayload, MoaModelSlot
- from hermes_cli.web_routers._common import _CONFIG_MUTATION_LOCK, config_write_scope, http_failure
- _log = logging.getLogger("hermes_cli.web_server")
- router = APIRouter()
- # Late-bound so a test's monkeypatch on the owning module wins at call time.
- _config_profile_scope = late("_config_profile_scope", "hermes_cli.web_server_profiles")
- _profile_scope = late("_profile_scope", "hermes_cli.web_server_profiles")
- load_config = late("load_config", "hermes_cli.config")
- save_config = late("save_config", "hermes_cli.config")
- require_readable_config_before_write = late("require_readable_config_before_write", "hermes_cli.config")
- _EMPTY_MODEL_INFO: dict = {
- "model": "", "provider": "", "auto_context_length": 0, "config_context_length": 0,
- "effective_context_length": 0, "capabilities": {},
- }
- _CAPABILITY_FIELDS = ("supports_tools", "supports_vision", "supports_reasoning", "context_window",
- "max_output_tokens", "model_family")
- 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()
- # Blocking budget for /api/model/info's context-length resolution. The resolver
- # chain (agent.model_metadata.get_model_context_length) runs several sequential
- # provider probes, each with its own multi-second timeout, so an unreachable or
- # blackholed model.base_url can hold this response for tens of seconds — and the
- # Desktop Model Settings page waits on it (#63214).
- _MODEL_INFO_PROBE_BUDGET_S = 5.0
- def _bounded_context_length_probe(model: str, base_url: str, provider: str) -> int:
- """``get_model_context_length`` with the route's blocking budget.
- On timeout the abandoned probe keeps running in its worker thread (bounded
- by its own per-request timeouts) while the response degrades to
- ``auto_context_length = 0`` ("auto-detected: unknown").
- """
- from agent.model_metadata import get_model_context_length
- pool = concurrent.futures.ThreadPoolExecutor(max_workers=1, thread_name_prefix="model-info-probe")
- try:
- return pool.submit(
- get_model_context_length, model=model, base_url=base_url, provider=provider,
- config_context_length=None
- ).result(timeout=_MODEL_INFO_PROBE_BUDGET_S)
- except concurrent.futures.TimeoutError:
- _log.warning(
- "GET /api/model/info: context-length probe for %r at %s exceeded %.1fs — returning unknown",
- model, base_url or "<default>", _MODEL_INFO_PROBE_BUDGET_S,
- )
- return 0
- finally:
- # wait=False: never block the response (or interpreter exit) on the
- # abandoned probe.
- pool.shutdown(wait=False)
- @router.get("/api/model/info")
- def get_model_info(profile: Optional[str] = None):
- """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:
- 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)
- try:
- # config_context_length=None: ignore the override — we want the auto value.
- # Bounded: the resolver's provider probes can hang for tens of seconds
- # when model.base_url is unreachable (#63214).
- auto_ctx = _bounded_context_length_probe(model_name, base_url, provider)
- except Exception:
- auto_ctx = 0
- config_ctx_int = config_ctx if isinstance(config_ctx, int) and config_ctx > 0 else 0
- caps = {}
- try:
- from agent.models_dev import get_model_capabilities
- mc = get_model_capabilities(provider=provider, model=model_name)
- if mc is not None:
- caps = {name: getattr(mc, name) for name in _CAPABILITY_FIELDS}
- except Exception:
- pass
- return {
- "model": model_name, "provider": provider, "auto_context_length": auto_ctx,
- "config_context_length": config_ctx_int,
- "effective_context_length": config_ctx_int or auto_ctx, # what the agent actually uses
- "capabilities": caps,
- }
- except HTTPException:
- # Unknown/invalid profile must surface as 404, not degrade into a
- # 200 with empty model info (which would render as "no model set").
- raise
- except Exception:
- _log.exception("GET /api/model/info failed")
- return dict(_EMPTY_MODEL_INFO)
- @router.get("/api/model/options")
- async def get_model_options(
- profile: Optional[str] = None,
- refresh: bool = False,
- include_unconfigured: bool = False,
- explicit_only: bool = False,
- ):
- """Authenticated providers + curated model lists — REST twin of the ``model.options``
- JSON-RPC on tui_gateway, same response shape so ``ModelPickerDialog`` shares the types.
- ``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}")
- from hermes_cli.inventory import build_model_options_payload, load_picker_context
- def _build_payload_scoped() -> dict:
- # Full sync picker build 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(), explicit_only=bool(explicit_only),
- include_unconfigured=bool(include_unconfigured), refresh=bool(refresh))
- return await run_in_threadpool(_build_payload_scoped)
- def _nous_recommended_default() -> dict:
- from hermes_cli.models import recommended_nous_default_model
- return recommended_nous_default_model()
- @router.get("/api/model/recommended-default")
- def get_recommended_default_model(provider: str = "", profile: Optional[str] = None):
- """Recommended default model for a freshly-authenticated provider, mirroring
- ``hermes model``'s curation so GUI onboarding lands on a sensible default.
- 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", "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:
- # The tier, Portal URL and recommendation caches are all per profile home.
- with _config_profile_scope(profile):
- return _nous_recommended_default()
- except HTTPException:
- raise # an unknown ?profile= is the scope's 404, not an empty recommendation
- except Exception:
- _log.exception("GET /api/model/recommended-default (nous) failed")
- return {"provider": "nous", "model": "", "free_tier": None}
- try:
- from hermes_cli.inventory import build_models_payload, load_picker_context
- from hermes_cli.models import pick_silent_default_model
- # build_models_payload -> list_authenticated_providers -> _save_discovered_models_to_config:
- # this GET lazily PERSISTS discovered custom-provider models, so it needs the scope too.
- with _config_profile_scope(profile):
- payload = build_models_payload(load_picker_context())
- for row in payload.get("providers", []):
- if str(row.get("slug", "")).lower() == slug:
- models = [str(m) for m in (row.get("models") or [])]
- return {"provider": slug, "model": pick_silent_default_model(models, provider=slug), "free_tier": None}
- return {"provider": slug, "model": "", "free_tier": None}
- except HTTPException:
- raise # an unknown ?profile= is the scope's 404, not an empty recommendation
- except Exception:
- _log.exception("GET /api/model/recommended-default failed")
- return {"provider": slug, "model": "", "free_tier": None}
- @router.get("/api/model/auxiliary")
- def get_auxiliary_models(profile: Optional[str] = None):
- """Current auxiliary task assignments: ``{"tasks": [{task, provider, model,
- base_url}, ...], "main": {provider, model}}``. ``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.
- Built-in slots come first; plugin-registered tasks follow, each carrying the
- ``label``/``hint``/``plugin`` the plugin declared (built-ins are labelled client-side) and
- ``inherit_from`` (base slot key or null). An inheriting row also carries ``effective``: the
- route it resolves to right now, which is the base's while the row itself is unpinned.
- ``provider``/``model``/``base_url`` stay the slot's own stored values on every row."""
- from agent.auxiliary_client import _get_auxiliary_task_config
- with http_failure("GET /api/model/auxiliary failed", 500, detail="Failed to read auxiliary config"):
- with _profile_scope(profile):
- cfg = load_config()
- # Inside the scope on purpose: plugin discovery and the resolver key on the same
- # context-local home, so profile B's rows come from B's plugins and B's config.
- plugin_tasks = _plugin_aux_tasks()
- effective = {entry["key"]: _get_auxiliary_task_config(entry["key"])
- for entry in plugin_tasks if entry.get("inherit_from")}
- aux_cfg = cfg.get("auxiliary", {})
- if not isinstance(aux_cfg, dict):
- aux_cfg = {}
- def _row(slot: str) -> dict:
- slot_cfg = aux_cfg.get(slot, {}) if isinstance(aux_cfg.get(slot), dict) else {}
- base_url = str(slot_cfg.get("base_url", "") or "")
- return {
- "task": slot, "provider": str(slot_cfg.get("provider", "auto") or "auto"),
- "model": str(slot_cfg.get("model", "") or ""), "base_url": base_url,
- "reasoning_effort": str(slot_cfg.get("reasoning_effort") or "") or None,
- # Lets the UI tell a free local/LAN pin from a forgotten paid-provider pin.
- "local_endpoint": is_local_endpoint(base_url),
- }
- tasks = [_row(slot) for slot in _AUX_TASK_SLOTS]
- tasks.extend({
- **_row(entry["key"]),
- "label": str(entry.get("display_name") or entry["key"]),
- "hint": str(entry.get("description") or ""),
- "plugin": str(entry.get("plugin") or ""),
- "inherit_from": entry.get("inherit_from") or None,
- **({"effective": {field: str(effective[entry["key"]].get(field) or "")
- for field in ("provider", "model", "base_url")}}
- if entry["key"] in effective else {}),
- } for entry in plugin_tasks)
- 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")
- def get_moa_models(profile: Optional[str] = None):
- """Return the configured Mixture-of-Agents provider/model slots."""
- with http_failure("GET /api/model/moa failed", 500, detail="Failed to read MoA config"):
- from hermes_cli.moa_config import normalize_moa_config
- with _profile_scope(profile):
- cfg = load_config()
- return normalize_moa_config(cfg.get("moa") if isinstance(cfg, dict) else {})
- _MOA_PRESET_FIELDS = (
- "reference_temperature", "aggregator_temperature", "reference_timeout",
- "degraded_reference_policy", "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 import config as config_mod
- from hermes_cli.moa_config import normalize_moa_config, validate_moa_payload
- # Serialize the full RMW with dashboard autosaves and with non-HTTP
- # save_config callers, which only hold the config module's reentrant lock.
- with config_write_scope(body.profile or profile), config_mod._CONFIG_LOCK:
- legacy_flat_update = not body.presets
- if not legacy_flat_update:
- raw = {
- "default_preset": body.default_preset,
- "active_preset": body.active_preset,
- "presets": {name: _preset_dict(preset) for name, preset in body.presets.items()},
- }
- else:
- raw = _preset_dict(body) # legacy flat payload from older clients
- # Reject-don't-repair: normalize_moa_config() silently swaps any preset with
- # incomplete slots for the hardcoded defaults — 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.
- # See #64156.
- problems = validate_moa_payload(raw)
- if problems:
- raise HTTPException(status_code=422, detail="Invalid MoA config: " + "; ".join(problems))
- # Read the raw profile file because the effective config contains defaults and
- # ``merge_existing=True`` deep-merges dicts, which resurrects omitted preset names.
- # Keep unknown MoA keys and sibling sections by rebuilding only the MoA section on
- # the raw document, while treating the submitted named map as authoritative.
- # Strict read: the merge below builds a new dict and would discard the
- # failed-read marker, allowing a transient fallback to replace the file.
- raw_cfg = require_readable_config_before_write()
- existing_moa = raw_cfg.get("moa")
- if isinstance(existing_moa, dict) and "privacy_filter" in existing_moa:
- # ``privacy_filter`` is emitted by the normalizer but is not declared by
- # MoaConfigPayload; an editor round-trip must not reset it.
- raw["privacy_filter"] = existing_moa["privacy_filter"]
- normalized = normalize_moa_config(raw)
- if legacy_flat_update and isinstance(existing_moa, dict):
- existing_presets = existing_moa.get("presets")
- existing_default = existing_moa.get("default_preset")
- if isinstance(existing_presets, dict):
- if isinstance(existing_default, str) and existing_default in existing_presets:
- # Older clients update the flat default preset; keep the existing
- # default name and any valid active selection from the stored map.
- synthesized_default = normalized.get("default_preset")
- if (
- isinstance(synthesized_default, str)
- and synthesized_default in normalized["presets"]
- ):
- normalized["presets"][existing_default] = normalized["presets"].pop(
- synthesized_default
- )
- normalized["default_preset"] = existing_default
- existing_active = existing_moa.get("active_preset")
- if isinstance(existing_active, str) and existing_active in existing_presets:
- normalized["active_preset"] = existing_active
- updated = config_mod._merge_partial_save(raw_cfg, {"moa": normalized})
- if not legacy_flat_update:
- # ``_merge_partial_save`` retains undeclared metadata on surviving presets;
- # filter after that merge so omitted names are authoritative deletions.
- merged_presets = updated["moa"]["presets"]
- updated["moa"]["presets"] = {
- name: merged_presets[name] for name in normalized["presets"]
- }
- save_config(updated)
- return {"ok": True, **normalized}
- @router.post("/api/model/set")
- async def set_model_assignment(body: ModelAssignment, profile: Optional[str] = None):
- """Assign a model to the main slot or an auxiliary task slot. Writes
- ``~/.hermes/config.yaml`` — applies to **new** sessions only; a running chat
- PTY hot-swaps via the ``/model`` slash command instead."""
- scope, task = (body.scope or "").strip().lower(), (body.task or "").strip().lower()
- provider, model = (body.provider or "").strip(), (body.model or "").strip()
- base_url, api_key = (body.base_url or "").strip(), (body.api_key or "").strip()
- if scope not in {"main", "auxiliary"}:
- raise HTTPException(status_code=400, detail="scope must be 'main' or 'auxiliary'")
- with http_failure("POST /api/model/set failed", 500, detail="Failed to save model assignment"):
- # #99859 (R2): the options picker already refuses on code skew; the WRITE path
- # must too — a stale process persisting a post-update model string is the
- # invalid-model-serving failure the reporter hit.
- skew_msg = _dashboard_code_skew_guard()
- if skew_msg:
- _log.warning("POST /api/model/set refused: %s", skew_msg)
- raise HTTPException(status_code=503, detail=f"Restart required: {skew_msg}")
- # Expensive-model warning runs BEFORE the profile scope is entered: _profile_scope
- # must never be held across an await (the RLock is reentrant per-thread, so a second
- # coroutine interleaving on the event-loop thread could cross-restore module globals).
- if model and not body.confirm_expensive_model:
- try:
- from hermes_cli.model_selection_guards import combined_selection_warning
- # Pricing lookup can hit models.dev / a /models endpoint on a cache miss — off the loop.
- warning = await asyncio.to_thread(combined_selection_warning, model, provider=provider, base_url=base_url)
- except Exception:
- warning = None
- if warning is not None:
- return {"ok": False, "scope": scope, "provider": provider, "model": model,
- "confirm_required": True, "confirm_message": warning.message}
- reasoning_effort = body.reasoning_effort if "reasoning_effort" in body.model_fields_set else _UNSET
- def _apply_assignment():
- # Same RMW span as PUT /api/config: applyMainModel fires this while the
- # settings-page autosave is in flight — hold the mutation lock. switch_model's
- # catalog fetches / endpoint probes are network I/O, so they run BEFORE the lock;
- # only load→apply→save holds it.
- with _profile_scope(body.profile or profile):
- prepared = (_prepare_main_assignment(load_config(), provider, model, base_url, api_key)
- if scope == "main" else None)
- with _CONFIG_MUTATION_LOCK:
- return _apply_model_assignment_sync(
- scope, provider, model, task, base_url, api_key,
- reasoning_effort=reasoning_effort, prepared=prepared)
- return await asyncio.to_thread(_apply_assignment)
|