| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164 |
- """Warn when an in-session model switch will trigger preflight compression on the next turn."""
- from __future__ import annotations
- from typing import Any, Callable, List, Optional
- from agent.model_metadata import MINIMUM_CONTEXT_LENGTH
- from hermes_cli.model_switch import ModelSwitchResult, resolve_display_context_length
- def _append_warning(result: ModelSwitchResult, text: str) -> None:
- if result.warning_message:
- result.warning_message = f"{result.warning_message} | {text}"
- else:
- result.warning_message = text
- def _threshold_tokens(compressor: Any, model: str, context_length: int, provider: str = "") -> int:
- """The trigger the compressor WILL use after the switch (cap, model_thresholds and small-window
- floor included), so the warning quotes the real number; duck-typed engines keep the plain ratio."""
- preview = getattr(compressor, "preview_threshold_tokens", None)
- if callable(preview):
- return int(preview(model, context_length, provider))
- return max(int(context_length * float(getattr(compressor, "threshold_percent", 0.5))), MINIMUM_CONTEXT_LENGTH)
- def _estimate_tokens(agent: Any, messages: Optional[list[dict]]) -> Optional[int]:
- cc = getattr(agent, "context_compressor", None)
- if cc is None:
- return None
- if messages is not None:
- protect = (
- int(getattr(cc, "protect_first_n", 3)) + int(getattr(cc, "protect_last_n", 20)) + 1)
- if len(messages) <= protect:
- return None
- try:
- from agent.model_metadata import estimate_request_tokens_rough
- system_prompt = getattr(agent, "_cached_system_prompt", None) or ""
- tools = getattr(agent, "tools", None)
- return int(
- estimate_request_tokens_rough(
- messages, system_prompt=system_prompt, tools=tools or None))
- except Exception:
- pass
- # session_prompt_tokens is a lifetime sum, not occupancy (#126343): never a fallback here.
- last = int(getattr(cc, "last_prompt_tokens", 0) or 0)
- return last if last > 0 else None
- def merge_preflight_compression_warning(
- result: ModelSwitchResult,
- *,
- agent: Any = None,
- messages: Optional[list[dict]] = None,
- custom_providers: list | None = None,
- config_context_length: int | None = None,
- configured_model: str | None = None,
- configured_provider: str | None = None,
- configured_base_url: str | None = None) -> None:
- """If the next user message will likely preflight-compress, append a warning."""
- if not result.success or agent is None:
- return
- if not getattr(agent, "compression_enabled", True):
- return
- cc = getattr(agent, "context_compressor", None)
- if cc is None:
- return
- # Fall back to the agent's custom providers: without them the shrink warning used the
- # hardcoded catalog (e.g. "qwen" → 131072) even when the provider declared 1M.
- if custom_providers is None:
- custom_providers = getattr(agent, "_custom_providers", None)
- def _or_agent(value, attr):
- return value if value is not None else getattr(agent, attr, None)
- old_ctx = int(getattr(cc, "context_length", 0) or 0)
- new_ctx = resolve_display_context_length(
- result.new_model,
- result.target_provider,
- base_url=result.base_url or getattr(agent, "base_url", "") or "",
- api_key=result.api_key or getattr(agent, "api_key", "") or "",
- model_info=result.model_info,
- custom_providers=custom_providers,
- config_context_length=config_context_length,
- configured_model=_or_agent(configured_model, "model"),
- configured_provider=_or_agent(configured_provider, "provider"),
- configured_base_url=_or_agent(configured_base_url, "base_url"))
- if not new_ctx:
- return
- estimate = _estimate_tokens(agent, messages)
- if estimate is None:
- return
- new_threshold = _threshold_tokens(cc, result.new_model, new_ctx, result.target_provider)
- if estimate < new_threshold:
- return
- if int(getattr(cc, "_ineffective_compression_count", 0) or 0) >= 2:
- return
- parts: list[str] = []
- if old_ctx and new_ctx < old_ctx:
- parts.append(f"Context window shrinks ({old_ctx:,} → {new_ctx:,}). ")
- parts.append(
- f"Session is ~{estimate:,} tokens; "
- f"{result.new_model} allows {new_ctx:,} "
- f"(auto-compress at ~{new_threshold:,}). "
- f"Your next message will run preflight compression before the model replies.")
- _append_warning(result, "".join(parts))
- def enrich_model_switch_warnings_for_gateway(
- result: ModelSwitchResult,
- runner: Any,
- *,
- session_key: str,
- source: Any,
- custom_providers: list | None = None,
- load_gateway_config: Callable[[], dict] | None = None) -> None:
- """Gateway helper: cached agent + session DB messages."""
- lock = getattr(runner, "_agent_cache_lock", None)
- cache = getattr(runner, "_agent_cache", None)
- agent = None
- if lock is not None and cache is not None:
- with lock:
- entry = cache.get(session_key)
- if entry and entry[0] is not None:
- agent = entry[0]
- if agent is None:
- return
- configured: dict = dict.fromkeys(
- ("config_context_length", "configured_model", "configured_provider", "configured_base_url"))
- if load_gateway_config is not None:
- try:
- cfg = load_gateway_config()
- model_cfg = cfg.get("model", {}) if isinstance(cfg, dict) else {}
- if isinstance(model_cfg, dict) and model_cfg.get("context_length") is not None:
- configured.update(
- config_context_length=int(model_cfg["context_length"]),
- configured_model=model_cfg.get("default") or model_cfg.get("model"),
- configured_provider=model_cfg.get("provider"),
- configured_base_url=model_cfg.get("base_url"))
- except Exception:
- pass
- messages = None
- db = getattr(runner, "_session_db", None)
- store = getattr(runner, "session_store", None)
- if db is not None and store is not None:
- try:
- entry = store.get_or_create_session(source)
- messages = db.get_messages_as_conversation(entry.session_id)
- except Exception:
- pass
- merge_preflight_compression_warning(
- result, agent=agent, messages=messages, custom_providers=custom_providers, **configured)
|