| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133 |
- """Model picker, /model switch application, runtime snapshot/restore, and codex runtime handling for the interactive CLI
- Mixin split out of ``cli.py``; bound onto ``HermesCLI`` via the MRO. cli.py-internal
- symbols are imported LAZILY inside each method (``from cli import ...``) — the mixin
- never imports ``cli`` at module load time (import cycle).
- Tests drive the switch paths with bare stubs lacking most mixin methods, so shared steps are
- module-level functions taking ``cli`` and siblings are called as ``HermesCLI.<name>(self, ...)``.
- """
- from __future__ import annotations
- import copy
- import logging
- import sys
- import threading
- from rich.markup import escape as _escape
- from agent.i18n import t
- from utils import base_url_host_matches
- from hermes_cli.cli_agent_setup_mixin import _retire_agent
- logger = logging.getLogger(__name__)
- # CLI-level fields describing the active model route; snapshotted before a switch / one-turn
- # override and restored wholesale on rollback. ``reasoning_config`` rides along because it is
- # resolved per model: a failed swap or `/model X --reasoning high --once` must not leave the
- # new model's effort behind with the old route.
- _RUNTIME_FIELDS = (
- "model", "provider", "requested_provider", "_explicit_api_key", "_explicit_base_url",
- "api_key", "base_url", "api_mode", "reasoning_config")
- def _runtime_fields(cli) -> dict:
- return {key: getattr(cli, key, None) for key in _RUNTIME_FIELDS}
- def _resolve_cli_reasoning(cli) -> None:
- """Re-resolve the CLI-level ``reasoning_config`` for ``cli.model`` through the shared chokepoint
- (per-model ``reasoning_overrides`` > global ``agent.reasoning_effort``). Startup resolves it once
- for the launch model; every path that moves ``cli.model`` must call this BEFORE the agent branch,
- because a lazily built agent inherits this field — an always-thinking model then goes out with
- the launch model's effort and 400s (#112921, #96012). ``agent.switch_model`` re-resolves its own
- copy for the live-agent path."""
- from cli import CLI_CONFIG
- from hermes_constants import resolve_reasoning_config
- # getattr: tests drive /new unbound on a SimpleNamespace without ``model`` (blank -> config default).
- cli.reasoning_config = resolve_reasoning_config(CLI_CONFIG, getattr(cli, "model", None) or "")
- cli._reasoning_user_set = False
- def stored_session_route(session_meta, *, current_model, current_provider):
- """The route a resumed session should run on, or ``None`` when the stored one is absent or
- already current. Returns ``(model, provider, base_url, api_mode, provider_changed)``; the
- canonical row reader is ``SessionDB.session_gateway_runtime`` (``model_config.gateway_runtime``,
- else the TUI's top-level keys). Bare ``custom`` is healed because the CLI resolve path
- hard-fails on it (the TUI gateway keeps it when a base_url exists)."""
- stored_model = str((session_meta or {}).get("model") or "").strip()
- if not stored_model:
- return None
- from hermes_state import SessionDB as _SessionDB
- runtime = _SessionDB.session_gateway_runtime(session_meta)
- base_url = runtime.get("base_url") or None
- api_mode = runtime.get("api_mode") or None
- stored_provider = runtime.get("provider") or None
- provider = _heal_bare_custom_provider(stored_provider, base_url=base_url, model=stored_model)
- if provider != stored_provider:
- # A healed entry resolves its own current endpoint. A route that didn't heal is gone with its
- # endpoint and wire, which must not ride on the ambient provider.
- base_url = None
- if provider is None:
- api_mode = None
- provider_changed = bool(provider) and provider != current_provider
- if stored_model == current_model and not provider_changed:
- return None
- from hermes_cli.runtime_provider import is_foreign_provider_endpoint
- if is_foreign_provider_endpoint(provider, base_url):
- # The endpoint and its wire belong to the provider this chat left; resolve the stored one's own.
- base_url = api_mode = None
- # A row's api_mode/base_url were written for whichever model the session last ran. Providers that
- # pick the wire per model (OpenCode Zen/Go, Nous, plugin profiles) re-derive both from the stored model, or a
- # resumed opencode-go session keeps a MiniMax-era anthropic_messages route for a chat_completions
- # model (#96066) — the CLI/oneshot twin of tui_gateway's _rederive_per_model_route.
- from hermes_cli.model_switch import model_derived_api_mode
- derived = model_derived_api_mode(provider or "", stored_model)
- if derived is not None:
- from hermes_cli.models import normalize_opencode_base_url
- api_mode = derived
- base_url = normalize_opencode_base_url(provider, api_mode, base_url) or None
- return stored_model, provider, base_url, api_mode, provider_changed
- def _heal_bare_custom_provider(provider, *, base_url, model):
- """Bare ``custom`` is a billing class, not a routable identity: persisting/restoring it makes a
- later resume hard-fail once the config default leaves the custom endpoint. Recover the durable
- ``custom:<name>`` menu key from the endpoint, else drop the provider (None)."""
- if str(provider or "").strip().lower() != "custom":
- return provider
- try:
- # Heal bare "custom" persisted by older builds / gateway turns: it's the resolved billing class, not
- # a routable identity. (Stricter than the TUI gateway's recovery, which keeps bare "custom" when a
- # base_url exists — the CLI's resolve path would hard-fail on it, #14676.)
- from hermes_cli.runtime_provider import canonical_custom_identity
- return canonical_custom_identity(base_url=base_url or None, model=model or None) or None
- except Exception:
- return None
- def _merge_preflight_warning(cli, result, custom_providers) -> None:
- """Fold the context-compression preflight warning into ``result`` (fail-soft)."""
- from cli import logger
- if cli.agent is None:
- return
- try:
- from hermes_cli.context_switch_guard import merge_preflight_compression_warning
- # Prefer the fresh inventory list (same source as switch_model / TUI); fall back
- # to the agent-init snapshot.
- merge_preflight_compression_warning(
- result,
- agent=cli.agent,
- messages=list(cli.conversation_history or []),
- custom_providers=custom_providers if custom_providers is not None
- else getattr(cli.agent, "_custom_providers", None),
- config_context_length=getattr(cli.agent, "_config_context_length", None))
- except Exception as exc:
- logger.debug("preflight-compression switch warning failed: %s", exc)
- def _print_switch_summary(cli, result, old_model, *, one_turn: bool, strict_context: bool) -> None:
- """Record the next-turn switch note and print the "Model switched" block.
- The note is prepended to the next user message (a mid-history system message would break
- providers and prompt caching). ``strict_context``: the typed /model path lets
- context-resolution errors propagate; the picker path swallows them.
- """
- from cli import _cprint
- from hermes_cli.model_switch import format_model_for_display, resolve_display_context_length
- _display_old = format_model_for_display(old_model)
- _display_new = format_model_for_display(result.new_model)
- cli._pending_model_switch_note = (
- f"[Note: model was just switched from {_display_old} to {_display_new} "
- f"via {result.provider_label or result.target_provider}. "
- f"{'This override applies to the next turn only. ' if one_turn else ''}"
- f"Adjust your self-identification accordingly.]")
- _cprint(f" {t('cli.model.switched', model=_display_new)}")
- _cprint(f" {t('gateway.model.provider_label', provider=result.provider_label or result.target_provider)}")
- if result.target_provider == "moa":
- # The preset name hides who pays: the aggregator runs every tool-loop step (#112359).
- from hermes_cli.moa_config import normalize_moa_config
- moa_cfg = cli.config.get("moa") if isinstance(cli.config, dict) else {}
- agg = normalize_moa_config(moa_cfg)["presets"].get(result.new_model, {}).get("aggregator") or {}
- if agg:
- _acting = f"{agg.get('provider')}:{agg.get('model')}"
- _cprint(f" {t('cli.model.acting_model', model=_acting)}")
- # Provider-aware context chain: Codex OAuth / Nous / profile live-catalog caps win over the raw
- # models.dev entry (gpt-5.5 is 1.05M on openai but 272K on Codex OAuth).
- mi = result.model_info
- agent = cli.agent
- try:
- ctx = resolve_display_context_length(
- result.new_model, result.target_provider,
- base_url=result.base_url or cli.base_url or "",
- api_key=result.api_key or cli.api_key or "", model_info=mi,
- config_context_length=getattr(agent, "_config_context_length", None) if agent else None,
- custom_providers=getattr(agent, "_custom_providers", None) if agent else None)
- except Exception:
- if strict_context:
- raise
- ctx = None
- if ctx:
- from agent.context_pin import context_window_suffix
- _cprint(f" {t('gateway.model.context_label', tokens=f'{ctx:,}')}" + context_window_suffix(
- ctx, getattr(agent, '_config_context_length', None) if agent else None, model=result.new_model,
- provider=result.target_provider, base_url=result.base_url or cli.base_url or ""))
- if mi:
- if mi.max_output:
- _cprint(f" {t('gateway.model.max_output_label', tokens=f'{mi.max_output:,}')}")
- _cprint(f" {t('gateway.model.capabilities_label', capabilities=mi.format_capabilities())}")
- cache_enabled = (
- (base_url_host_matches(result.base_url or "", "openrouter.ai") and "claude" in result.new_model.lower())
- or result.api_mode == "anthropic_messages")
- if cache_enabled:
- _cprint(f" {t('gateway.model.prompt_caching_enabled')}")
- if result.warning_message:
- _cprint(f" ⚠ {result.warning_message}")
- if getattr(result, "usage_notice", ""):
- _cprint(f" ⚠ {result.usage_notice}")
- def _switch_model_from(
- cli, raw_input, *, is_global, explicit_provider, user_providers, custom_providers):
- """``switch_model`` seeded with this CLI's live route."""
- from hermes_cli.model_switch import switch_model
- return switch_model(
- raw_input=raw_input, current_provider=cli.provider or "", current_model=cli.model or "",
- current_base_url=cli.base_url or "", current_api_key=cli.api_key or "", is_global=is_global,
- explicit_provider=explicit_provider, user_providers=user_providers,
- custom_providers=custom_providers)
- def _run_confirm_and_apply(cli, target, *args) -> None:
- """Run a confirm+apply sequence off the UI thread when the TUI is live.
- The expensive-model modal blocks its thread on a response queue (_prompt_text_input_modal);
- on the prompt_toolkit main thread that freezes rendering, so the modal never appears and the
- switch silently cancels after the 120s timeout.
- """
- if getattr(cli, "_app", None):
- threading.Thread(target=target, args=args, daemon=True).start()
- else:
- target(*args)
- def _one_turn_selection_context(cli):
- """Context the owner's large-context guard would see, for a ``/model --once`` that does not mutate.
- A live compressor measurement wins. Otherwise refresh ``session.stats``, then the session row's
- persisted usage anchor (a resume, before this process has run a turn).
- """
- from hermes_cli.model_selection_guards import SelectionContext, selection_context_for_agent
- agent = getattr(cli, "agent", None)
- if agent is not None and hasattr(agent, "_refresh_stats"):
- agent._refresh_stats()
- live = selection_context_for_agent(agent)
- if live is not None:
- return live
- from agent.usage_anchor import persisted_anchor_tokens
- db = getattr(cli, "_session_db", None)
- sid = getattr(cli, "session_id", None)
- if db is None or not sid or not callable(getattr(db, "get_messages_as_conversation", None)):
- return None
- try:
- messages = db.get_messages_as_conversation(sid)
- except Exception:
- logger.debug("one-turn selection context unavailable", exc_info=True)
- return None
- tokens = persisted_anchor_tokens(db, sid, messages)
- if not tokens:
- return None
- return SelectionContext(context_tokens=tokens, current_model=getattr(cli, "model", None) or None)
- def _ask_switch_anyway(cli, title: str, detail: str) -> bool:
- """The model-guard prompt (switch anyway once, or cancel): True only on an explicit yes."""
- choices = [
- ("once", t("cli.model.choice_switch_anyway"), t("cli.model.desc_switch_anyway")),
- ("cancel", t("cli.model.choice_cancel"), t("cli.model.desc_keep_current_model"))]
- raw = cli._prompt_text_input_modal(title=f"!!! {title} !!!", detail=detail, choices=choices, timeout=120)
- return cli._normalize_slash_confirm_choice(raw, choices) == "once"
- # `/model` usage block: (invocation form, description key). Forms are command syntax (never
- # translated); descriptions are padded to one column at render time.
- _MODEL_USAGE_ROWS = (
- ("/model <name>", "cli.model.usage_switch"),
- ("/model <name> --global", "cli.model.usage_switch_global"),
- ("/model <name> --once", "cli.model.usage_switch_once"),
- ("/model <name> --session", "cli.model.usage_switch_session"),
- ("/model <name> --provider <slug>", "cli.model.usage_switch_provider_model"),
- ("/model <name> --reasoning <level>", "cli.model.usage_switch_reasoning"),
- ("/model --provider <slug>", "cli.model.usage_switch_provider"),
- ("/model --refresh", "cli.model.usage_refresh"))
- _MODEL_USAGE_COL = 36
- def _configured_default_route() -> tuple[str, str]:
- """``(model, provider)`` config.yaml names NOW (``/model reset``'s target). A fresh read, not the
- startup snapshot: the dashboard or another surface may have changed the default since launch."""
- from hermes_cli.config import load_config_readonly
- model_cfg = (load_config_readonly() or {}).get("model")
- if not isinstance(model_cfg, dict):
- return (model_cfg.strip() if isinstance(model_cfg, str) else ""), ""
- provider = str(model_cfg.get("provider") or "").strip()
- return (str(model_cfg.get("default") or model_cfg.get("model") or "").strip(),
- "" if provider.lower() == "auto" else provider)
- def _picker_reasoning_rows() -> list[tuple[str, str]]:
- """``(value, label)`` rows for the picker's effort step: the canonical ladder, the off state,
- then a keep-current row (empty value = leave the effort alone)."""
- from hermes_constants import VALID_REASONING_EFFORTS
- rows = [(lvl, lvl) for lvl in VALID_REASONING_EFFORTS]
- rows.append(("none", t("cli.model.effort_none")))
- rows.append(("", t("cli.model.effort_keep_current")))
- return rows
- def _picker_offers_reasoning(provider_data: dict, model: str) -> bool:
- """False only when the inventory's capability map says the picked model has no reasoning
- control; unknown capabilities keep the step (a no-op dial beats hiding a real one)."""
- caps = (provider_data or {}).get("capabilities")
- entry = caps.get(model) if isinstance(caps, dict) else None
- return not (isinstance(entry, dict) and entry.get("reasoning") is False)
- def _apply_reasoning_after_switch(cli, effort: str, *, persist_global: bool, one_turn: bool = False) -> None:
- """Apply a ``--reasoning <level>`` that rode along with a model pick. Runs AFTER the swap: the
- agent's ``switch_model`` re-resolves ``reasoning_config`` from config.yaml, so an earlier write
- would be clobbered. Session-scoped unless the pick itself persists (``--global``).
- ``one_turn`` (``--once``): the level rides the one-turn admission (``model_override.reasoning``)
- and the restore snapshot puts the CLI's own level back. It is never a session pin."""
- from cli import CLI_CONFIG, _cprint, _parse_reasoning_config, save_config_value
- parsed = _parse_reasoning_config(effort)
- if parsed is None:
- return
- cli.reasoning_config = parsed
- if one_turn:
- from hermes_cli.gateway_agent import GatewayAgent
- pending = getattr(cli, "_pending_one_turn_model_restore", None)
- if isinstance(pending, dict):
- pending["one_turn_reasoning"] = effort # the agent chat rebuilds carries it too
- if isinstance(cli.agent, GatewayAgent):
- cli.agent.stage_turn_reasoning(parsed)
- elif cli.agent is not None:
- cli.agent.reasoning_config = parsed
- _cprint(f" {t('cli.model.reasoning_effort', effort=effort)}")
- return
- cli._reasoning_user_set = True
- if cli.agent is not None:
- # The gateway session's agent commits an explicit level even when it equals the inherited one.
- pin = getattr(cli.agent, "pin_reasoning", None)
- if pin is not None:
- pin(parsed)
- else:
- cli.agent.reasoning_config = parsed
- saved = persist_global and save_config_value("agent.reasoning_effort", effort)
- if saved:
- CLI_CONFIG.setdefault("agent", {})["reasoning_effort"] = effort
- _cprint(f" {t('cli.model.reasoning_effort_saved' if saved else 'cli.model.reasoning_effort', effort=effort)}")
- def _commit_model_switch(
- cli, result, *, persist_global: bool, one_turn: bool = False, picker: bool = False,
- reasoning_effort: str = "") -> None:
- """Stage + swap, print the summary, persist (session row unless --once; config on --global).
- ``picker``: tolerate context-resolution errors and label the config write "(--global)"; the
- typed path additionally records the one-turn restore snapshot. ``reasoning_effort`` (from
- ``--reasoning`` or the picker's effort step) is applied after the swap."""
- from cli import HermesCLI, _cprint
- old_model, old_provider = cli.model, getattr(cli, "provider", None)
- # A second ``--once`` before the turn restores to the route before the first one.
- snapshot = (getattr(cli, "_pending_one_turn_model_restore", None)
- or cli._snapshot_model_runtime()) if one_turn else None
- if snapshot is not None:
- # The level belongs to this --once pick only: a second pick without --reasoning drops it.
- # So does the alias it resolved through.
- snapshot = {key: value for key, value in snapshot.items()
- if key not in ("one_turn_reasoning", "one_turn_alias")}
- # ``--once`` must not session.mutate.
- cli._model_switch_one_turn = bool(one_turn)
- try:
- staged = cli._stage_and_swap_model(result, old_model)
- finally:
- cli._model_switch_one_turn = False
- if not staged:
- return
- # A TUI slash worker replays /model on its shadow CLI; the tui_gateway mirror counts the real switch.
- if not getattr(cli, "is_slash_worker", False):
- from hermes_cli.observability.shared_metrics_events import record_model_switch
- record_model_switch(
- from_provider=old_provider, to_provider=result.target_provider, surface="cli", from_model=old_model,
- session_id=getattr(getattr(cli, "agent", None), "session_id", None) or getattr(cli, "session_id", None))
- if not picker:
- cli._pending_one_turn_model_restore = snapshot
- if snapshot is not None and getattr(result, "resolved_via_alias", ""):
- snapshot["one_turn_alias"] = result.resolved_via_alias # the agent chat rebuilds carries it
- _print_switch_summary(cli, result, old_model, one_turn=one_turn, strict_context=not picker)
- if reasoning_effort:
- _apply_reasoning_after_switch(cli, reasoning_effort, persist_global=persist_global and not one_turn,
- one_turn=one_turn)
- if persist_global:
- from hermes_cli.model_switch import persist_model_selection
- persist_model_selection(result)
- _cprint(f" {t('cli.model.saved_global' if picker else 'cli.model.saved')}")
- elif one_turn:
- _cprint(f" {t('cli.model.next_turn_only')}")
- else:
- _cprint(f" {t('cli.model.session_only')}")
- # The row records what THIS session runs even on --global (else a later resume restores the
- # stale creation-time model); --once is restored after one turn and never touches the row.
- if not one_turn:
- HermesCLI._persist_model_switch_to_session(cli, result)
- def _show_model_picker(cli, ctx, force_refresh: bool) -> None:
- """``/model`` with no args: open the picker, or print usage when nothing is authed."""
- from cli import _cprint
- from hermes_cli.inventory import build_models_payload
- from hermes_cli.providers import get_label
- try:
- if ctx is None:
- raise RuntimeError("inventory context unavailable")
- providers = build_models_payload(
- ctx, probe_custom_providers=force_refresh,
- probe_current_custom_provider=not force_refresh,
- capabilities=True, # the effort step hides itself on reasoning-free routes
- # Keep providers whose pool is entirely in cooldown visible: limits are
- # per-model for many providers, so another model may work (#103829) —
- # same contract as the gateway picker (#66584) and aux pickers (#66624).
- for_picker=True,
- # Visibility only, not a faster probe: this call live-probes the current custom
- # endpoint, which historically enjoyed the full 5s discovery budget (#103843).
- fast_custom_probe=False,
- )["providers"]
- except Exception:
- providers = []
- if not providers:
- _cprint(f" {t('cli.model.no_authenticated_providers')}")
- _cprint("")
- for form, desc_key in _MODEL_USAGE_ROWS:
- _cprint(f" {form.ljust(_MODEL_USAGE_COL)} {t(desc_key)}")
- return
- cli._open_model_picker(
- providers, cli.model or "unknown", get_label(cli.provider) if cli.provider else "unknown",
- user_provs=ctx.user_providers if ctx is not None else None,
- custom_provs=ctx.custom_providers if ctx is not None else None)
- class CLIModelSwitchMixin:
- """Model picker, /model switch application, runtime snapshot/restore, and codex runtime handling for the interactive CLI"""
- def _normalize_model_for_provider(self, resolved_provider: str) -> bool:
- """Normalize provider-specific model IDs and routing."""
- from cli import _split_model_config_default
- current_model = str(self.model or "").strip()
- if isinstance(self.model, dict):
- current_model, _ = _split_model_config_default(self.model)
- changed = False
- def _adopt(canonical, notice) -> None:
- """Adopt ``canonical`` when it differs; ``notice(new)`` builds the warning text."""
- nonlocal current_model, changed
- if canonical and canonical != current_model:
- if not self._model_is_default:
- self._console_print(f"[yellow]⚠️ {notice(canonical)}[/]")
- self.model = canonical
- current_model = canonical
- changed = True
- def _adopt_with_mode(normalize, api_mode_of, notice) -> bool:
- """Provider families that also own the wire protocol: adopt id, then sync api_mode."""
- nonlocal changed
- try:
- _adopt(normalize(current_model), notice)
- resolved_mode = api_mode_of(current_model)
- if resolved_mode != self.api_mode:
- self.api_mode = resolved_mode
- changed = True
- except Exception:
- pass
- return changed
- try:
- from hermes_cli.model_normalize import (
- _AGGREGATOR_PROVIDERS, normalize_model_for_provider)
- if resolved_provider not in _AGGREGATOR_PROVIDERS:
- _adopt(
- normalize_model_for_provider(current_model, resolved_provider),
- lambda new: t("cli.model.normalized_model", old=current_model, new=new, provider=resolved_provider))
- except Exception:
- pass
- from providers import get_provider_profile
- profile = get_provider_profile(resolved_provider)
- try:
- native_id = profile.normalize_model_id(current_model, api_key=self.api_key or "") if profile else None
- except Exception: # a plugin hook failing must not block startup: generic repair continues
- logger.debug("normalize_model_id(%s) failed", resolved_provider, exc_info=True)
- native_id = None
- if native_id is not None: # a profile-owned id space also re-derives the wire for the new id
- from hermes_cli.model_switch import model_derived_api_mode
- return _adopt_with_mode(
- lambda m: native_id,
- lambda m: model_derived_api_mode(resolved_provider, m, self.api_key or "") or self.api_mode,
- lambda new: t("cli.model.normalized_model", old=current_model, new=new, provider=resolved_provider))
- from hermes_cli.models import opencode_provider_family
- if opencode_provider_family(resolved_provider) is not None:
- from hermes_cli.models import normalize_opencode_model_id, opencode_model_api_mode
- return _adopt_with_mode(
- lambda m: normalize_opencode_model_id(resolved_provider, m),
- lambda m: opencode_model_api_mode(resolved_provider, m),
- lambda new: t("cli.model.stripped_provider_prefix", old=current_model, new=new, provider=resolved_provider))
- if resolved_provider != "openai-codex":
- return changed
- # 1. Strip provider prefix ("openai/gpt-5.4" → "gpt-5.4")
- if "/" in current_model:
- slug = current_model.split("/", 1)[1]
- if not self._model_is_default:
- self._console_print(
- "[yellow]⚠️ " + _escape(t("cli.model.stripped_provider_prefix", old=current_model, new=slug, provider="OpenAI Codex")) + "[/]")
- self.model = slug
- current_model = slug
- changed = True
- # 2. Replace untouched default with a Codex model
- if self._model_is_default:
- from hermes_cli.codex_models import DEFAULT_CODEX_MODELS, get_codex_model_ids
- fallback_model = DEFAULT_CODEX_MODELS[0]
- try:
- # self.base_url is the route resolved with self.api_key (#121486).
- available = get_codex_model_ids(
- access_token=self.api_key if self.api_key else None, base_url=self.base_url or None)
- if available:
- fallback_model = available[0]
- except Exception:
- pass
- if current_model != fallback_model:
- self.model = fallback_model
- changed = True
- return changed
- def _persist_model_switch_to_session(self, result) -> None:
- """Persist a session-scoped /model switch to the session DB row.
- Writes the model column plus the route in both shapes readers use — nested
- ``gateway_runtime`` (CLI --resume) and top-level keys (TUI session.resume) — from one
- or-None dict so stale keys are DELETED (``_merge_model_config_json`` only deletes on
- explicit None) and the shapes never diverge.
- Writes the model column plus the runtime route so ``--resume`` (CLI, reads ``gateway_runtime``) and
- ``session.resume`` (TUI/desktop, reads top-level ``model_config`` keys via
- ``_stored_session_runtime_overrides``) both restore the switched provider instead of recombining the
- model with the ambient default (#79536). Mirrors the gateway's ``update_session_model()`` call.
- getattr: tests drive the switch paths with ``object.__new__`` stubs.
- """
- from cli import logger
- db = getattr(self, "_session_db", None)
- sid = getattr(self, "session_id", None)
- if not db or not sid:
- return
- route = {
- "provider": _heal_bare_custom_provider(
- result.target_provider, base_url=result.base_url, model=result.new_model,
- ) or None,
- # Both shapes use the same or-None discipline so stale keys from a previous switch are deleted
- # (not merely omitted) in BOTH the nested gateway_runtime dict (CLI reader) and the top-level
- # keys (TUI gateway reader). _merge_model_config_json only deletes on explicit None, so falsy
- # values must be converted, not filtered. Deriving the top-level from **route guarantees the two
- # shapes can never diverge — the asymmetry that caused the original stale-key bug (#85261
- # simplify-code review).
- "base_url": result.base_url or None,
- "api_mode": result.api_mode or None}
- try:
- db.update_session_model(sid, result.new_model)
- db.patch_session_model_config(sid, {"gateway_runtime": route, **route})
- except Exception:
- logger.debug("Failed to persist model switch to session DB", exc_info=True)
- def _restore_session_model(self, session_meta: dict, *, quiet: bool = False) -> None:
- """Restore model/provider from the session DB row on every resume path.
- Skips when no model is recorded or the CLI got an explicit ``-m`` (user intent wins).
- A different stored provider gets its credentials re-resolved — the ambient ``api_key``
- must not be sent to the session's endpoint; on failure the ambient credentials are kept
- so the session still opens (the first turn surfaces the auth error).
- """
- from cli import logger
- if not (session_meta or {}).get("model") or getattr(self, "_explicit_model_override", False):
- return
- route = stored_session_route(session_meta, current_model=self.model, current_provider=self.provider)
- if route is None:
- return
- stored_model, stored_provider, stored_base_url, stored_api_mode, provider_changed = route
- from hermes_cli.local_runtime.endpoint import LLAMACPP_ALIASES
- managed = str(stored_provider or "").strip().lower() in LLAMACPP_ALIASES
- self.model = stored_model
- if stored_provider:
- self.provider = stored_provider
- self.requested_provider = stored_provider
- if stored_base_url and not managed:
- self.base_url = stored_base_url
- if stored_api_mode:
- self.api_mode = stored_api_mode
- if managed and not (getattr(self, "_explicit_base_url", None) and not provider_changed):
- # The supervisor owns the live port: last boot's loopback URL (an ephemeral fallback when
- # 18434 was busy) must not pin the resume onto a dead endpoint. A launch-time --base-url
- # for this same provider is user intent and keeps winning.
- self._explicit_api_key = None
- self._explicit_base_url = None
- try:
- from hermes_cli.runtime_provider import resolve_runtime_provider
- resolved = resolve_runtime_provider(requested=stored_provider, target_model=self.model or None)
- if resolved.get("api_key"):
- self.api_key = resolved["api_key"]
- self._credential_pool = resolved.get("credential_pool")
- if resolved.get("base_url"):
- self.base_url = resolved["base_url"]
- if not stored_api_mode and resolved.get("api_mode"):
- self.api_mode = resolved["api_mode"]
- except Exception:
- if stored_base_url:
- self.base_url = stored_base_url
- logger.debug(
- "Credential re-resolution for resumed session provider "
- "%s failed; keeping ambient credentials",
- stored_provider, exc_info=True)
- elif provider_changed:
- # Launch-time explicit overrides belong to the AMBIENT provider and would poison
- # _ensure_runtime_credentials for the restored one. api_key is never persisted to
- # the session DB — runtime provider resolution owns credentials.
- self._explicit_api_key = None
- self._explicit_base_url = stored_base_url
- try:
- from hermes_cli.runtime_provider import resolve_runtime_provider
- resolved = resolve_runtime_provider(requested=stored_provider, target_model=self.model or None)
- if resolved.get("api_key"):
- self.api_key = resolved["api_key"]
- self._credential_pool = resolved.get("credential_pool")
- if not stored_base_url and resolved.get("base_url"):
- self.base_url = resolved["base_url"]
- if not stored_api_mode and resolved.get("api_mode"):
- self.api_mode = resolved["api_mode"]
- except Exception:
- logger.debug(
- "Credential re-resolution for resumed session provider "
- "%s failed; keeping ambient credentials",
- stored_provider, exc_info=True)
- _resolve_cli_reasoning(self)
- # Mid-chat /resume swaps the live agent; on startup --resume _init_agent picks up
- # self.model / self.provider / self.reasoning_config.
- if self.agent is not None:
- try:
- self.agent.switch_model(
- new_model=self.model, new_provider=self.provider, api_key=self.api_key or "",
- base_url=self.base_url or "", api_mode=self.api_mode or "")
- except Exception:
- logger.debug("In-place agent model swap on resume failed", exc_info=True)
- msg = t("cli.model.restored_from_session", model=stored_model)
- if stored_provider:
- msg += f" ({stored_provider})"
- if quiet:
- print(msg, file=sys.stderr)
- else:
- self._console_print(f"[dim]{_escape(msg)}[/dim]")
- def _open_model_picker(self, providers: list, current_model: str, current_provider: str, user_provs=None, custom_provs=None) -> None:
- """Open prompt_toolkit-native /model picker modal."""
- self._capture_modal_input_snapshot()
- self._model_picker_state = {
- "stage": "provider",
- "providers": providers,
- "selected": next((i for i, p in enumerate(providers) if p.get("is_current")), 0),
- "current_model": current_model,
- "current_provider": current_provider,
- "user_provs": user_provs,
- "custom_provs": custom_provs,
- "filter": ""}
- self._invalidate(min_interval=0.0)
- def _confirm_expensive_model_switch(self, result) -> bool:
- """Ask for explicit confirmation before applying costly model switches."""
- self._confirmed_switch_warning = None
- if not getattr(result, "success", False):
- return True
- try:
- from hermes_cli.model_selection_guards import (
- combined_selection_warning, selection_context_for_agent)
- primed = getattr(self, "_one_turn_selection_context", None)
- warning = combined_selection_warning(
- result.new_model, provider=result.target_provider,
- base_url=result.base_url or self.base_url or "",
- api_key=result.api_key or self.api_key or "", model_info=result.model_info,
- selection_context=primed if primed is not None else selection_context_for_agent(
- getattr(self, "agent", None)))
- except Exception:
- warning = None
- if warning is None:
- return True
- if not _ask_switch_anyway(self, warning.title, warning.message):
- return False
- self._confirmed_switch_warning = warning.message
- return True
- def _confirm_owner_model_switch(self, refusal) -> bool:
- """The gateway owner refused a guarded target (``confirmation_required``: cost, data policy,
- large context; Ink/Desktop/ACP show the same ``confirm_message``): ask in the same prompt as
- ``_confirm_expensive_model_switch``. Only a yes re-sends with the owner's token. The identical
- warning the user just confirmed for this switch is not asked twice."""
- from hermes_cli.gateway_mutations import confirmation_title
- message = refusal.get("confirm_message") or ""
- confirmed, self._confirmed_switch_warning = getattr(self, "_confirmed_switch_warning", None), None
- if message and message == confirmed:
- return True
- return _ask_switch_anyway(self, confirmation_title(refusal), message)
- def _confirm_and_apply_model_switch_result(
- self, result, persist_global: bool, custom_providers=None, reasoning_effort: str = "") -> None:
- from cli import _cprint
- try:
- if result.success and not self._confirm_expensive_model_switch(result):
- _cprint(f" {t('cli.model.switch_cancelled')}")
- return
- self._apply_model_switch_result(
- result, persist_global, custom_providers=custom_providers, reasoning_effort=reasoning_effort)
- except Exception as exc:
- _cprint(f" {t('cli.model.selection_failed', error=exc)}")
- def _close_model_picker(self) -> None:
- self._model_picker_state = None
- self._restore_modal_input_snapshot()
- self._invalidate(min_interval=0.0)
- def _snapshot_model_runtime(self) -> dict:
- """Capture current CLI and agent model runtime for one-turn restore."""
- agent = getattr(self, "agent", None)
- # ``reasoning_config`` is a mutable dict: deepcopy it so a later in-place edit cannot alias the snapshot.
- return {
- **_runtime_fields(self),
- "reasoning_config": copy.deepcopy(getattr(self, "reasoning_config", None)),
- "agent_primary_runtime": copy.deepcopy(
- getattr(agent, "_primary_runtime", None)
- ) if agent is not None else None}
- def _restore_model_runtime_snapshot(self, snapshot: dict | None) -> None:
- """Restore a model runtime captured before a one-turn override."""
- from cli import _cprint, logger
- if not snapshot:
- return
- one_turn = _runtime_fields(self)
- for key in _RUNTIME_FIELDS:
- if key in snapshot:
- setattr(self, key, snapshot.get(key))
- agent = getattr(self, "agent", None)
- if agent is None:
- return
- from hermes_cli.gateway_agent import GatewayAgent
- # A gateway agent's one-turn level only rode the admission: put it back locally, write nothing.
- restore_reasoning = (agent.restore_reasoning if isinstance(agent, GatewayAgent)
- else lambda config: setattr(agent, "reasoning_config", config))
- if "reasoning_config" in snapshot:
- restore_reasoning(snapshot["reasoning_config"])
- primary = snapshot.get("agent_primary_runtime")
- if primary and hasattr(agent, "_restore_primary_runtime"):
- try:
- agent._primary_runtime = copy.deepcopy(primary)
- agent._fallback_activated = True
- agent._rate_limited_until = 0
- if agent._restore_primary_runtime():
- return
- except Exception:
- logger.debug("CLI one-turn model restore via primary runtime failed", exc_info=True)
- if hasattr(agent, "switch_model"):
- try:
- # The classic CLI only builds a GatewayAgent (``_init_agent``): its ``once`` / ``restore`` are
- # named, so a rename fails loudly instead of turning the restore into a session mutate.
- restored = agent.switch_model(
- new_model=snapshot.get("model", ""), new_provider=snapshot.get("provider", ""),
- api_key=snapshot.get("api_key", ""), base_url=snapshot.get("base_url", ""),
- api_mode=snapshot.get("api_mode", ""),
- capabilities=snapshot.get("capabilities"), once=True, restore=True)
- except Exception as exc: # health: allow BLE001 -- runs in the turn's finally; reported below
- logger.warning("CLI one-turn model restore failed: %s", exc)
- restored = exc
- if restored is False or isinstance(restored, Exception):
- # The previous route did not come back. Stay on the one-turn model instead of
- # claiming the restore.
- for key, value in one_turn.items():
- setattr(self, key, value)
- notice = t("cli.model.switch_cancelled") if restored is False else t(
- "cli.model.switch_failed", model=snapshot.get("model", ""), error=restored, old_model=agent.model)
- _cprint(f" {notice}")
- return
- if "reasoning_config" in snapshot:
- restore_reasoning(snapshot["reasoning_config"])
- @staticmethod
- def _filter_model_picker_entries(entries: list, query: str) -> list:
- """Return (original_index, label) pairs matching ``query`` (case-insensitive subsequence;
- empty matches all). The ORIGINAL index keeps a filtered selection resolving to exactly one
- concrete model — filtering never introduces fuzzy *resolution*."""
- pairs = list(enumerate(entries))
- q = (query or "").strip().lower()
- if not q:
- return pairs
- def _subseq(needle: str, hay: str) -> bool:
- it = iter(hay)
- return all(ch in it for ch in needle)
- return [(i, e) for (i, e) in pairs if _subseq(q, str(e).lower())]
- @staticmethod
- def _compute_model_picker_viewport(
- selected: int, scroll_offset: int, n: int, term_rows: int, reserved_below: int = 6,
- panel_chrome: int = 6, min_visible: int = 3) -> tuple[int, int]:
- """Resolve (scroll_offset, visible) for the /model picker viewport. ``reserved_below``
- matches the approval/clarify panels (input, status bar, separators); ``panel_chrome`` is
- borders + blanks + hint row. The offset slides to keep ``selected`` on screen."""
- max_visible = max(min_visible, term_rows - reserved_below - panel_chrome)
- if n <= max_visible:
- return 0, n
- visible = max_visible
- if selected < scroll_offset:
- scroll_offset = selected
- elif selected >= scroll_offset + visible:
- scroll_offset = selected - visible + 1
- return max(0, min(scroll_offset, n - visible)), visible
- def _stage_and_swap_model(self, result, old_model) -> bool:
- """Stage ``result`` onto the CLI fields, then swap the live agent in place.
- CLI fields are snapshotted first so a failed agent swap rolls the whole CLI back —
- otherwise the staged broken credentials leak into the next turn even though the agent
- rolled back. Returns False after printing the failure (a failed switch is a no-op).
- """
- from cli import _cprint
- if self.agent is None and getattr(self, "_gateway_session_open", False) and not self._init_agent():
- # Before the first turn: the stand-in built for that turn reads the gateway session's model,
- # so the switch must commit on the session now; _init_agent printed why it could not.
- self._confirmed_switch_warning = None
- return False
- _cli_snapshot = _runtime_fields(self)
- self.model = result.new_model
- self.provider = result.target_provider
- self.requested_provider = result.target_provider
- # Always overwrite explicit overrides so stale credentials from the previous provider
- # (e.g. Ollama api_key/base_url) don't leak into the next resolution.
- self._explicit_api_key = result.api_key
- self._explicit_base_url = result.base_url
- if result.api_key:
- self.api_key = result.api_key
- if result.base_url:
- self.base_url = result.base_url
- if result.api_mode:
- self.api_mode = result.api_mode
- _resolve_cli_reasoning(self)
- if self.agent is not None:
- try:
- fields = dict(
- new_model=result.new_model, new_provider=result.target_provider,
- api_key=result.api_key, base_url=result.base_url, api_mode=result.api_mode,
- capabilities=getattr(result, "runtime_capabilities", None))
- if getattr(self, "_model_switch_one_turn", False):
- # The CLI and agent show the resolved model; the admission sends the alias.
- fields.update(once=True, alias=getattr(result, "resolved_via_alias", "") or "")
- if self.agent.switch_model(**fields) is False:
- # Declined at the gateway owner's guard prompt: nothing was committed.
- for _k, _v in _cli_snapshot.items():
- setattr(self, _k, _v)
- _cprint(f" {t('cli.model.switch_cancelled')}")
- return False
- except Exception as exc:
- # The agent rolled itself back to the old working model/client. Roll the CLI's own staged
- # fields back too and abort the rest of the commit (note + success print) so a failed switch
- # is a no-op rather than a dead session (#50163).
- # Agent rolled itself back; roll the CLI back too and abort so a failed switch is a no-op
- # rather than a dead session (#50163).
- for _k, _v in _cli_snapshot.items():
- setattr(self, _k, _v)
- if str(exc) == "launch_credentials_unavailable":
- # No model switch can fix a chat whose launch key changed; say what does.
- from hermes_cli.cli_chat_error_copy import refusal_text
- _cprint(f" {refusal_text(exc)}")
- return False
- _cprint(f" {t('cli.model.switch_failed', model=result.new_model, error=exc, old_model=old_model)}")
- return False
- finally:
- self._confirmed_switch_warning = None # this switch's confirmation only
- return True
- def _apply_model_switch_result(
- self, result, persist_global: bool, custom_providers=None, reasoning_effort: str = "") -> None:
- """Picker-path commit (see _commit_model_switch)."""
- from cli import _cprint
- if not result.success:
- _cprint(f" ✗ {result.error_message}")
- return
- _merge_preflight_warning(self, result, custom_providers)
- _commit_model_switch(self, result, persist_global=persist_global, picker=True,
- reasoning_effort=reasoning_effort)
- def _handle_model_picker_selection(self, persist_global: bool = False) -> None:
- state = self._model_picker_state
- if not state:
- return
- selected = state.get("selected", 0)
- stage = state.get("stage")
- if stage == "provider":
- providers = state.get("providers") or []
- if selected >= len(providers):
- self._close_model_picker()
- return
- provider_data = providers[selected]
- # Curated list (same as `hermes model` / gateway pickers); live catalog only when
- # it is empty (user-defined endpoints, per-resource providers such as azure-foundry).
- # Disk-cached like the gateway pickers: the live probe can walk several api-version
- # fallbacks with a 6 s timeout each, which must not block the REPL on every select.
- model_list = provider_data.get("models", [])
- if not model_list:
- try:
- from hermes_cli.models import cached_provider_model_ids
- model_list = cached_provider_model_ids(provider_data["slug"]) or model_list
- except Exception:
- pass
- from hermes_cli.models_validate import offered_model_ids
- model_list = offered_model_ids(
- model_list, provider_data.get("slug"), provider_data.get("api_url"))
- state.update(
- stage="model", provider_data=provider_data, model_list=model_list,
- selected=0, filter="", _filtered_pairs=None)
- self._invalidate(min_interval=0.0)
- return
- if stage == "model":
- provider_data = state.get("provider_data") or {}
- model_list = state.get("model_list") or []
- # Map the row through the active fuzzy filter; pairs carry the ORIGINAL index.
- filtered_pairs = state.get("_filtered_pairs")
- if filtered_pairs is None:
- filtered_pairs = list(enumerate(model_list))
- visible_labels = [e for (_i, e) in filtered_pairs]
- back_idx = len(visible_labels)
- if selected == back_idx:
- state.update(
- stage="provider", filter="", _filtered_pairs=None,
- selected=next((i for i, p in enumerate(state.get("providers") or [])
- if p.get("slug") == provider_data.get("slug")), 0))
- self._invalidate(min_interval=0.0)
- return
- if selected > back_idx: # cancel row (and anything past it)
- self._close_model_picker()
- return
- if 0 <= selected < back_idx:
- result = _switch_model_from(
- self, visible_labels[selected], is_global=persist_global,
- explicit_provider=provider_data.get("slug"),
- user_providers=state.get("user_provs"),
- custom_providers=state.get("custom_provs"))
- if result.success and _picker_offers_reasoning(provider_data, result.new_model):
- # Third step: effort for the picked model (skipped for routes the catalog
- # marks reasoning-free). Rows come from the canonical level set.
- state.update(stage="reasoning", switch_result=result, selected=0, _scroll_offset=0)
- self._invalidate(min_interval=0.0)
- return
- self._commit_picker_result(result, persist_global)
- return
- self._close_model_picker()
- if stage == "reasoning":
- rows = _picker_reasoning_rows()
- result = state.get("switch_result")
- if selected == len(rows): # ← Back to the model list
- state.update(stage="model", selected=0, _scroll_offset=0, switch_result=None)
- self._invalidate(min_interval=0.0)
- return
- if selected > len(rows) or result is None:
- self._close_model_picker()
- return
- self._commit_picker_result(result, persist_global, reasoning_effort=rows[selected][0])
- def _commit_picker_result(self, result, persist_global: bool, reasoning_effort: str = "") -> None:
- """Close the picker and run the confirm+apply sequence for ``result``."""
- state = self._model_picker_state or {}
- # Capture before close — picker state is cleared on close.
- _picker_custom_provs = state.get("custom_provs")
- self._close_model_picker()
- # The effort is appended only when picked: stubs/tests pin the historical arity.
- extra = (reasoning_effort,) if reasoning_effort else ()
- _run_confirm_and_apply(
- self, self._confirm_and_apply_model_switch_result,
- result, persist_global, _picker_custom_provs, *extra)
- def _handle_model_switch(self, cmd_original: str):
- """Handle /model command — switch model.
- Supports:
- /model — show current model + usage hints
- /model <name> — switch model (this session only)
- /model <name> --once — switch for the next turn only
- /model <name> --session — switch for this session only (explicit)
- /model <name> --global — switch and persist to config.yaml
- /model <name> --provider <provider> — switch provider + model
- /model <name> --reasoning <level> — switch and set reasoning effort in one step
- /model --provider <provider> — switch to provider, auto-detect model
- /model reset — back to config.yaml's model (keeps history)
- Switches are session-scoped unless ``model.persist_switch_by_default`` or ``--global``.
- """
- from cli import _cprint
- from hermes_cli.model_switch import parse_model_switch_args, resolve_persist_behavior
- parts = cmd_original.split(None, 1) # split off '/model'
- request = parse_model_switch_args(parts[1].strip() if len(parts) > 1 else "")
- if request.errors:
- # CLI decoration: " ✗ " prefix over the canonical error copy.
- _cprint(f" ✗ {request.error_messages()[0]}")
- return
- target, explicit_provider = request.target, request.explicit_provider
- if request.is_reset:
- # Session-scoped switch back to config.yaml's model; drops an armed --once restore,
- # which would otherwise re-plant the model being reset away from after the next turn.
- target, explicit_provider = _configured_default_route()
- if not target:
- _cprint(f" ✗ {t('cli.model.reset_no_default')}")
- return
- if target == self.model and explicit_provider in ("", self.provider):
- self._pending_one_turn_model_restore = None
- _cprint(f" {t('cli.model.reset_none', model=target)}")
- return
- one_turn = request.is_once
- persist_global = not request.is_reset and resolve_persist_behavior(
- request.is_global, request.is_session, is_once=one_turn,
- explicit_provider=request.explicit_provider)
- # --refresh: wipe the picker cache so every authed provider's /v1/models is re-fetched.
- if request.force_refresh:
- try:
- from hermes_cli.models import clear_provider_models_cache
- clear_provider_models_cache()
- _cprint(f" {t('cli.model.cleared_picker_cache')}")
- except Exception:
- pass
- # Live session state is overlaid truthy-only so empty self.* attrs don't clobber config.
- from hermes_cli.inventory import load_picker_context
- try:
- ctx = load_picker_context().with_overrides(
- current_provider=self.provider or "", current_model=self.model or "",
- current_base_url=self.base_url or "")
- except Exception:
- ctx = None
- # switch_model() + _open_model_picker still need the raw provider dicts.
- user_provs = ctx.user_providers if ctx is not None else None
- custom_provs = ctx.custom_providers if ctx is not None else None
- if not request.target and not request.explicit_provider:
- return _show_model_picker(self, ctx, request.force_refresh)
- result = _switch_model_from(
- self, target, is_global=persist_global, explicit_provider=explicit_provider,
- user_providers=user_provs, custom_providers=custom_provs)
- if not result.success:
- _cprint(f" ✗ {result.error_message}")
- return
- _merge_preflight_warning(self, result, custom_provs)
- extra = (request.reasoning_effort,) if request.reasoning_effort else ()
- _run_confirm_and_apply(
- self, self._confirm_and_apply_cli_model_switch,
- result, persist_global, one_turn, custom_provs, *extra)
- def _confirm_and_apply_cli_model_switch(
- self, result, persist_global: bool, one_turn: bool, custom_provs=None, reasoning_effort: str = "") -> None:
- """Confirm an expensive model switch and apply it (typed /model path). Runs on a worker
- thread when the TUI is active (see _run_confirm_and_apply) so the modal can render."""
- from cli import _cprint
- # ``--once`` does not reach the owner's mutate guard. A resumed session's compressor is
- # still cold; the persisted anchor is what that guard would have measured.
- self._one_turn_selection_context = _one_turn_selection_context(self) if one_turn else None
- try:
- if not self._confirm_expensive_model_switch(result):
- _cprint(f" {t('cli.model.switch_cancelled')}")
- return
- _commit_model_switch(self, result, persist_global=persist_global, one_turn=one_turn,
- reasoning_effort=reasoning_effort)
- finally:
- self._one_turn_selection_context = None
- def _handle_codex_runtime(self, cmd_original: str) -> None:
- """Handle /codex-runtime — toggle the codex app-server runtime opt-in.
- Usage:
- /codex-runtime — show current state
- /codex-runtime auto — Hermes default (chat_completions)
- /codex-runtime codex_app_server — hand turns to codex subprocess
- /codex-runtime on / off — synonyms for the above
- """
- from cli import _cprint
- from hermes_cli import codex_runtime_switch as crs
- parts = cmd_original.split(None, 1)
- new_value, errors = crs.parse_args(parts[1].strip() if len(parts) > 1 else "")
- if errors:
- for err in errors:
- _cprint(f"❌ {err}")
- return
- try:
- from hermes_cli.config import load_config, save_config
- except Exception as exc:
- _cprint(t("cli.model.could_not_load_config", error=exc))
- return
- result = crs.apply(
- load_config(), new_value,
- persist_callback=(save_config if new_value is not None else None))
- prefix = "✓" if result.success else "✗"
- for line in result.message.splitlines():
- _cprint(f" {prefix} {line}" if line.startswith("openai_runtime") else f" {line}")
- if result.success and result.requires_new_session:
- _cprint(f" {t('cli.model.tip_reset_new_session')}")
- def _should_handle_model_command_inline(self, text: str, has_images: bool = False) -> bool:
- """Return True when /model should be handled immediately on the UI thread."""
- from cli import _looks_like_slash_command
- if not text or has_images or not _looks_like_slash_command(text):
- return False
- try:
- from hermes_cli.commands import resolve_command
- cmd = resolve_command(text.split(None, 1)[0].lower().lstrip('/'))
- return bool(cmd and cmd.name == "model")
- except Exception:
- return False
- def _cmd_moa(self, cmd_original: str):
- """/moa one-shot: run one prompt through the default MoA preset, then restore the prior
- model (a session-long MoA switch goes through the picker's virtual MoA provider)."""
- from cli import _cprint, _slash_args
- from hermes_cli.moa_config import moa_usage, normalize_moa_config
- payload = _slash_args(cmd_original)
- if not payload:
- _cprint(f" {moa_usage()}")
- return True
- moa_cfg = self.config.get("moa") if isinstance(self.config, dict) else {}
- preset = normalize_moa_config(moa_cfg)["default_preset"]
- self._pending_moa_restore_model = {
- key: getattr(self, key, None)
- for key in (
- "requested_provider", "provider", "model", "api_key", "base_url", "api_mode")}
- self.requested_provider = "moa"
- self.provider = "moa"
- self.model = preset
- self.api_key = "moa-virtual-provider"
- self.base_url = "moa://local"
- self.api_mode = "chat_completions"
- _retire_agent(self)
- self._pending_moa_disable_after_turn = True
- self._pending_agent_seed = payload
- _cprint(f" {t('cli.model.moa_one_shot_queued', preset=preset)}")
|