hermes-agent/tests/gateway/test_telegram_text_batch_perf.py

"""Regression tests for the Telegram text-batch adaptive-delay fast-path
and _env_float_clamped helper introduced by PR #10388 (Telegram latency
tuning).

The fast-path lets short replies stream near-instantly while keeping the
configured cap as the upper bound, so an operator who tightens the cap
gets the lower number on every tier.

The env-clamped helper guarantees float env vars never produce NaN/Inf
or out-of-bounds values that could break asyncio.sleep().
"""

from __future__ import annotations

import math

import pytest

from gateway.platforms.telegram import TelegramAdapter


@pytest.fixture
def adapter():
    """Build a TelegramAdapter shell without going through __init__'s
    network-touching setup. Just need the class for static-method access
    and the instance for instance-method tests."""
    return TelegramAdapter.__new__(TelegramAdapter)


class TestEnvFloatClamped:
    """_env_float_clamped is the fence around every float env var the
    adapter reads — must reject NaN/Inf and honor min/max bounds."""

    def test_default_when_unset(self, monkeypatch):
        monkeypatch.delenv("HERMES_TEST_VAR", raising=False)
        assert TelegramAdapter._env_float_clamped("HERMES_TEST_VAR", 0.5) == 0.5

    def test_parses_valid_value(self, monkeypatch):
        monkeypatch.setenv("HERMES_TEST_VAR", "1.25")
        assert TelegramAdapter._env_float_clamped("HERMES_TEST_VAR", 0.5) == 1.25

    def test_falls_back_to_default_on_garbage(self, monkeypatch):
        monkeypatch.setenv("HERMES_TEST_VAR", "not-a-float")
        assert TelegramAdapter._env_float_clamped("HERMES_TEST_VAR", 0.5) == 0.5

    def test_rejects_nan(self, monkeypatch):
        monkeypatch.setenv("HERMES_TEST_VAR", "nan")
        result = TelegramAdapter._env_float_clamped("HERMES_TEST_VAR", 0.5)
        assert math.isfinite(result)
        assert result == 0.5

    def test_rejects_inf(self, monkeypatch):
        monkeypatch.setenv("HERMES_TEST_VAR", "inf")
        result = TelegramAdapter._env_float_clamped("HERMES_TEST_VAR", 0.5)
        assert math.isfinite(result)
        assert result == 0.5

    def test_clamps_below_min(self, monkeypatch):
        monkeypatch.setenv("HERMES_TEST_VAR", "0.01")
        assert TelegramAdapter._env_float_clamped(
            "HERMES_TEST_VAR", 0.5, min_value=0.1,
        ) == 0.1

    def test_clamps_above_max(self, monkeypatch):
        monkeypatch.setenv("HERMES_TEST_VAR", "10.0")
        assert TelegramAdapter._env_float_clamped(
            "HERMES_TEST_VAR", 0.5, max_value=2.0,
        ) == 2.0


class TestAdaptiveTextBatchTiers:
    """The fast-path tiers cap delay for short / medium messages.  Tier
    constants must compose with the configured cap (operators who set a
    lower cap get the lower number on every tier)."""

    def test_class_constants_are_sensible(self):
        """Sanity check that the tier constants form a non-overlapping
        ascending ladder."""
        assert TelegramAdapter._TEXT_BATCH_FAST_LEN < TelegramAdapter._TEXT_BATCH_SHORT_LEN
        assert TelegramAdapter._TEXT_BATCH_FAST_DELAY_S < TelegramAdapter._TEXT_BATCH_SHORT_DELAY_S
        assert TelegramAdapter._TEXT_BATCH_FAST_DELAY_S > 0
        assert TelegramAdapter._TEXT_BATCH_SHORT_DELAY_S > 0

    def test_fast_tier_uses_min_with_configured_cap(self, adapter):
        """A short message picks the lower of the fast-tier delay and
        the operator's configured cap."""
        # Operator set a generous cap (0.6s); fast tier should win.
        adapter._text_batch_delay_seconds = 0.6
        delay = min(
            adapter._text_batch_delay_seconds,
            TelegramAdapter._TEXT_BATCH_FAST_DELAY_S,
        )
        assert delay == TelegramAdapter._TEXT_BATCH_FAST_DELAY_S

        # Operator tightened the cap below the fast-tier delay; cap wins.
        adapter._text_batch_delay_seconds = 0.10
        delay = min(
            adapter._text_batch_delay_seconds,
            TelegramAdapter._TEXT_BATCH_FAST_DELAY_S,
        )
        assert delay == 0.10

    def test_short_tier_uses_min_with_configured_cap(self, adapter):
        """Same composition rule for the medium tier."""
        adapter._text_batch_delay_seconds = 0.6
        delay = min(
            adapter._text_batch_delay_seconds,
            TelegramAdapter._TEXT_BATCH_SHORT_DELAY_S,
        )
        assert delay == TelegramAdapter._TEXT_BATCH_SHORT_DELAY_S

    def test_long_message_uses_full_cap(self, adapter):
        """Messages above the medium threshold use the configured cap
        without the tier-clamp."""
        adapter._text_batch_delay_seconds = 0.5
        # Beyond _TEXT_BATCH_SHORT_LEN there's no tier-clamp; cap wins.
        delay = adapter._text_batch_delay_seconds
        assert delay == 0.5

    def test_split_threshold_takes_priority_over_fast_tier(self, adapter):
        """If the latest chunk hits the platform split threshold a
        continuation is almost certain — wait the longer split delay
        regardless of total length."""
        adapter._text_batch_delay_seconds = 0.3
        adapter._text_batch_split_delay_seconds = 1.0
        last_chunk_len = TelegramAdapter._SPLIT_THRESHOLD + 50
        # The flush path checks last_chunk_len first; assert the contract.
        assert last_chunk_len >= TelegramAdapter._SPLIT_THRESHOLD
        delay = adapter._text_batch_split_delay_seconds
        assert delay == 1.0
        assert delay > adapter._text_batch_delay_seconds