Gateway silence handling and empty-response retry fixes

Fixes on record

Merged to main. A stamp counts the issues a fix is linked to.

• #46080 fix(gateway): suppress exact silence tokens without mutating history

  merged to main

  files: gateway/response_filters.py, gateway/run.py, scripts/release.py, tests/gateway/test_gateway_silence_tokens.py, tests/gateway/test_response_filters.py

  Show the diff

  • gateway/response_filters.py

    @@ -0,0 +1,53 @@
    +"""Gateway response filtering helpers.
    +
    +These helpers operate at the gateway boundary: they decide whether a completed
    +agent turn should be delivered to the chat, not what should be persisted in the
    +conversation history.
    +"""
    +
    +from __future__ import annotations
    +
    +from typing import Any
    +
    +# Canonical model-emitted control token for intentional silence.
    +SILENT_REPLY_TOKEN = "[redacted:secret-assignment]"
    +
    +# Exact whole-response markers that mean "the agent intentionally chose not to
    +# reply".  Keep this list small and explicit; arbitrary empty output remains an
    +# error/empty-response path, not silence.
    +LIVE_GATEWAY_SILENT_MARKERS = frozenset({
    +    "[SILENT]",
    +    "SILENT",
    +    "NO_REPLY",
    +    "NO REPLY",
    +})
    +
    +
    +def _canonical_silence_candidate(text: str) -> str:
    +    return " ".join(text.strip().upper().split())
    +
    +
    +def is_intentional_silence_response(response: Any) -> bool:
    +    """Return True only when ``response`` is exactly a silence marker.
    +
    +    Substantive prose that merely mentions ``NO_REPLY`` or ``[SILENT]`` must be
    +    delivered normally.  A blank response is also not silence; blank output is
    +    handled by the empty-response failure path.
    +    """
    +    if not isinstance(response, str):
    +        return False
    +    stripped = response.strip()
    +    if not stripped:
    +        return False
    +    if len(stripped) > 64:
    +        return False
    +    return _canonical_silence_candidate(stripped) in LIVE_GATEWAY_SILENT_MARKERS
    +
    +
    +def is_intentional_silence_agent_result(agent_result: dict | None, response: Any) -> bool:
    +    """Silence markers suppress delivery only for successful agent turns."""
    +    if not isinstance(agent_result, dict):
    +        return False
    +    if agent_result.get("failed"):
    +        return False
    +    return is_intentional_silence_response(response)

  • gateway/run.py

    @@ -8642,13 +8642,20 @@ async def _handle_message_with_agent(self, event, source, _quick_key: str, run_g
                     return None
     
                 response = agent_result.get("final_response") or ""
    +            try:
    +                from gateway.response_filters import is_intentional_silence_agent_result
    +                _intentional_silence = is_intentional_silence_agent_result(
    +                    agent_result, response,
    +                )
    +            except Exception:
    +                _intentional_silence = False
     
                 # Convert the agent's internal "(empty)" sentinel into a
                 # user-friendly message.  "(empty)" means the model failed to
                 # produce visible content after exhausting all retries (nudge,
                 # prefill, empty-retry, fallback).  Sending the raw sentinel
                 # looks like a bug; a short explanation is more helpful.
    -            if response == "(empty)":
    +            if response == "(empty)" and not _intentional_silence:
                     response = (
                         "⚠️ The model returned no response after processing tool "
                         "results. This can happen with some models — try again or "
    @@ -8684,10 +8691,11 @@ async def _handle_message_with_agent(self, event, source, _quick_key: str, run_g
     
                 # Normalize empty responses: surface errors, partial failures, and
                 # the case where agent did work but returned no text. Fix for #18765.
    -            response = _normalize_empty_agent_response(
    -                agent_result, response, history_len=len(history),
    -            )
    -            response = _sanitize_gateway_final_response(source.platform, response)
    +            if not _intentional_silence:
    +                response = _normalize_empty_agent_response(
    +                    agent_result, response, history_len=len(history),
    +                )
    +                response = _sanitize_gateway_final_response(source.platform, response)
     
                 # Ordering contract: the agent thread already updated the contextvar
                 # in conversation_compression.py; propagate to SessionEntry + _save().
    @@ -8711,7 +8719,7 @@ async def _handle_message_with_agent(self, event, source, _quick_key: str, run_g
                     )
                 except Exception:
                     _show_reasoning_effective = getattr(self, "_show_reasoning", False)
    -            if _show_reasoning_effective and response:
    +            if _show_reasoning_effective and response and not _intentional_silence:
                     last_reasoning = agent_result.get("last_reasoning")
                     if last_reasoning:
                         # Collapse long reasoning to keep messages readable
    @@ -8741,7 +8749,7 @@ async def _handle_message_with_agent(self, event, source, _quick_key: str, run_g
                 except Exception as _footer_err:
                     logger.debug("runtime_footer build failed: %s", _footer_err)
                     _footer_line = ""
    -            if _footer_line and response and not agent_result.get("already_sent"):
    +            if _footer_line and response and not agent_result.get("already_sent") and not _intentional_silence:
                     response = f"{response}\n\n{_footer_line}"
     
                 # Emit agent:end hook
    @@ -8975,6 +8983,18 @@ async def _handle_message_with_agent(self, event, source, _quick_key: str, run_g
                     last_prompt_tokens=[redacted:secret-assignment]"last_prompt_tokens", 0),
                 )
     
    +            # Intentional silence is a delivery decision, not a transcript
    +            # mutation.  The agent's [SILENT]/NO_REPLY assistant turn above is
    +            # still persisted in session history so later turns keep normal
    +            # user/assistant alternation; only the outbound chat delivery is
    +            # suppressed.
    +            if _intentional_silence:
    +                logger.info(
    +                    "Suppressing intentional silence marker for session %s",
    +                    session_entry.session_id,
    +                )
    +                response = ""
    +
                 # Auto voice reply: send TTS audio before the text response
                 _already_sent = bool(agent_result.get("already_sent"))
                 if self._should_send_voice_reply(event, response, agent_messages, already_sent=_already_sent):

  • scripts/release.py

    @@ -82,6 +82,7 @@
         "[redacted:email]": "synapsesx",
         "[redacted:email]": "itsflownium",
         "[redacted:email]": "dirtyren",
    +    "[redacted:email]": "aldoeliacim",
         "[redacted:email]": "MaxFreedomPollard",
         "[redacted:email]": "achaljhawar",
         "[redacted:email]": "claytonchew",

  • tests/gateway/test_gateway_silence_tokens.py

    @@ -0,0 +1,165 @@
    +"""Gateway intentional-silence token behavior."""
    +
    +from datetime import datetime
    +from unittest.mock import AsyncMock, MagicMock
    +
    +import pytest
    +
    +import gateway.run as gateway_run
    +from gateway.config import GatewayConfig, Platform
    +from gateway.platforms.base import MessageEvent
    +from gateway.session import SessionEntry, SessionSource
    +from gateway.response_filters import (
    +    is_intentional_silence_agent_result,
    +    is_intentional_silence_response,
    +)
    +
    +
    +def _source():
    +    return SessionSource(
    +        platform=Platform.TELEGRAM,
    +        chat_id="-1001",
    +        chat_type="group",
    +        user_id="12345",
    +    )
    +
    +
    +def _event():
    +    return MessageEvent(
    +        text="side chatter",
    +        source=_source(),
    +        message_id="msg-42",
    +    )
    +
    +
    +def _runner(monkeypatch, tmp_path):
    +    runner = gateway_run.GatewayRunner(GatewayConfig())
    +    runner.adapters = {}
    +    runner._running_agents = {}
    +    runner._running_agents_ts = {}
    +    runner._pending_messages = {}
    +    runner._pending_approvals = {}
    +    runner._is_user_authorized = lambda _source: True
    +    runner._set_session_env = lambda _context: None
    +    runner._handle_active_session_busy_message = AsyncMock(return_value=False)
    +    runner._session_db = MagicMock()
    +    runner._recover_telegram_topic_thread_id = lambda _source: None
    +    runner._cache_session_source = lambda _key, _source: None
    +    runner._is_session_run_current = lambda _key, _gen: True
    +    runner._reply_anchor_for_event = lambda _event: None
    +    runner._get_guild_id = lambda _event: None
    +    runner._should_send_voice_reply = lambda *_a, **_kw: False
    +    runner.hooks = MagicMock()
    +    runner.hooks.emit = AsyncMock()
    +
    +    runner.session_store = MagicMock()
    +    runner.session_store.get_or_create_session.return_value = SessionEntry(
    +        session_key="agent:main:telegram:group:-1001:12345",
    +        session_id="sess-silent",
    +        created_at=datetime.now(),
    +        updated_at=datetime.now(),
    +        platform=Platform.TELEGRAM,
    +        chat_type="group",
    +    )
    +    runner.session_store.load_transcript.return_value = []
    +    runner.session_store.append_to_transcript = MagicMock()
    +    runner.session_store.update_session = MagicMock()
    +
    +    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
    +    monkeypatch.setattr(
    +        gateway_run, "_resolve_runtime_agent_kwargs", lambda: {"api_key": "fake"}
    +    )
    +    monkeypatch.setattr(
    +        "agent.model_metadata.get_model_context_length",
    +        lambda *_args, **_kwargs: 100_000,
    +    )
    +    return runner
    +
    +
    +def test_exact_silence_tokens_are_intentional_silence():
    +    for token in ("[SILENT]", " SILENT ", "NO_REPLY", "no reply"):
    +        assert is_intentional_silence_response(token)
    +
    +
    +def test_blank_and_prose_mentions_are_not_silence():
    +    assert not is_intentional_silence_response("")
    +    assert not is_intentional_silence_response("Use NO_REPLY when no answer is needed.")
    +    assert not is_intentional_silence_response("The reply was [SILENT], intentionally.")
    +
    +
    +def test_failed_agent_result_never_counts_as_intentional_silence():
    +    assert is_intentional_silence_agent_result({"failed": False}, "NO_REPLY")
    +    assert not is_intentional_silence_agent_result({"failed": True}, "NO_REPLY")
    +
    +
    [redacted:email]
    +async def test_silence_token_suppresses_delivery_but_preserves_transcript(monkeypatch, tmp_path):
    +    runner = _runner(monkeypatch, tmp_path)
    +    runner._run_agent = AsyncMock(return_value={
    +        "final_response": "[SILENT]",
    +        "messages": [
    +            {"role": "user", "content": "side chatter"},
    +            {"role": "assistant", "content": "[SILENT]"},
    +        ],
    +        "tools": [],
    +        "history_offset": 0,
    +        "last_prompt_tokens": 0,
    +        "api_calls": 1,
    +        "failed": False,
    +    })
    +
    +    response = await runner._handle_message_with_agent(
    +        _event(), _source(), "agent:main:telegram:group:-1001:12345", 1
    +    )
    +
    +    assert response == ""
    +    appended = [call.args[1] for call in runner.session_store.append_to_transcript.call_args_list]
    +    assert {"role": "assistant", "content": "[SILENT]"}.items() <= appended[-1].items()
    +    assert [msg["role"] for msg in appended if msg.get("role") in {"user", "assistant"}] == ["user", "assistant"]
    +
    +
    [redacted:email]
    +async def test_empty_success_still_gets_empty_response_warning(monkeypatch, tmp_path):
    +    runner = _runner(monkeypatch, tmp_path)
    +    runner._run_agent = AsyncMock(return_value={
    +        "final_response": "",
    +        "messages": [
    +            {"role": "user", "content": "question"},
    +            {"role": "assistant", "content": ""},
    +        ],
    +        "tools": [],
    +        "history_offset": 0,
    +        "last_prompt_tokens": 0,
    +        "api_calls": 1,
    +        "failed": False,
    +    })
    +
    +    response = await runner._handle_message_with_agent(
    +        _event(), _source(), "agent:main:telegram:group:-1001:12345", 1
    +    )
    +
    +    assert "no response was generated" in response
    +
    +
    [redacted:email]
    +async def test_prose_mentioning_silence_token_is_delivered(monkeypatch, tmp_path):
    +    runner = _runner(monkeypatch, tmp_path)
    +    text = "Use [SILENT] when no answer is needed."
    +    runner._run_agent = AsyncMock(return_value={
    +        "final_response": text,
    +        "messages": [
    +            {"role": "user", "content": "question"},
    +            {"role": "assistant", "content": text},
    +        ],
    +        "tools": [],
    +        "history_offset": 0,
    +        "last_prompt_tokens": 0,
    +        "api_calls": 1,
    +        "failed": False,
    +    })
    +
    +    response = await runner._handle_message_with_agent(
    +        _event(), _source(), "agent:main:telegram:group:-1001:12345", 1
    +    )
    +
    +    assert response == text

  • tests/gateway/test_response_filters.py

    @@ -0,0 +1,20 @@
    +from gateway.response_filters import (
    +    is_intentional_silence_agent_result,
    +    is_intentional_silence_response,
    +)
    +
    +
    +def test_exact_silence_tokens_are_intentional_silence():
    +    for token in ("[SILENT]", " SILENT ", "NO_REPLY", "no reply"):
    +        assert is_intentional_silence_response(token)
    +
    +
    +def test_blank_and_prose_mentions_are_not_silence():
    +    assert not is_intentional_silence_response("")
    +    assert not is_intentional_silence_response("Use NO_REPLY when no answer is needed.")
    +    assert not is_intentional_silence_response("The reply was [SILENT], intentionally.")
    +
    +
    +def test_failed_agent_result_never_counts_as_intentional_silence():
    +    assert is_intentional_silence_agent_result({"failed": False}, "NO_REPLY")
    +    assert not is_intentional_silence_agent_result({"failed": True}, "NO_REPLY")