67275641f8
The gateway had a SEPARATE compression system ('session hygiene')
with hardcoded thresholds (100k tokens / 200 messages) that were
completely disconnected from the model's context length and the
user's compression config in config.yaml. This caused premature
auto-compression on Telegram/Discord — triggering at ~60k tokens
(from the 200-message threshold) or inconsistent token counts.
Changes:
- Gateway hygiene now reads model name from config.yaml and uses
get_model_context_length() to derive the actual context limit
- Compression threshold comes from compression.threshold in
config.yaml (default 0.85), same as the agent's ContextCompressor
- Removed the message-count-based trigger (was redundant and caused
false positives in tool-heavy sessions)
- Removed the undocumented session_hygiene config section — the
standard compression.* config now controls everything
- Env var overrides (CONTEXT_COMPRESSION_THRESHOLD,
CONTEXT_COMPRESSION_ENABLED) are respected
- Warn threshold is now 95% of model context (was hardcoded 200k)
- Updated tests to verify model-aware thresholds, scaling across
models, and that message count alone no longer triggers compression
For claude-opus-4.6 (200k context) at 85% threshold: gateway
hygiene now triggers at 170k tokens instead of the old 100k.
205 lines
8.3 KiB
Python
205 lines
8.3 KiB
Python
"""Tests for gateway session hygiene — auto-compression of large sessions.
|
||
|
||
Verifies that the gateway detects pathologically large transcripts and
|
||
triggers auto-compression before running the agent. (#628)
|
||
|
||
The hygiene system uses the SAME compression config as the agent:
|
||
compression.threshold × model context length
|
||
so CLI and messaging platforms behave identically.
|
||
"""
|
||
|
||
import pytest
|
||
from unittest.mock import patch, MagicMock, AsyncMock
|
||
from agent.model_metadata import estimate_messages_tokens_rough
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# Helpers
|
||
# ---------------------------------------------------------------------------
|
||
|
||
def _make_history(n_messages: int, content_size: int = 100) -> list:
|
||
"""Build a fake transcript with n_messages user/assistant pairs."""
|
||
history = []
|
||
content = "x" * content_size
|
||
for i in range(n_messages):
|
||
role = "user" if i % 2 == 0 else "assistant"
|
||
history.append({"role": role, "content": content, "timestamp": f"t{i}"})
|
||
return history
|
||
|
||
|
||
def _make_large_history_tokens(target_tokens: int) -> list:
|
||
"""Build a history that estimates to roughly target_tokens tokens."""
|
||
# estimate_messages_tokens_rough counts total chars in str(msg) // 4
|
||
# Each msg dict has ~60 chars of overhead + content chars
|
||
# So for N tokens we need roughly N * 4 total chars across all messages
|
||
target_chars = target_tokens * 4
|
||
# Each message as a dict string is roughly len(content) + 60 chars
|
||
msg_overhead = 60
|
||
# Use 50 messages with appropriately sized content
|
||
n_msgs = 50
|
||
content_size = max(10, (target_chars // n_msgs) - msg_overhead)
|
||
return _make_history(n_msgs, content_size=content_size)
|
||
|
||
|
||
# ---------------------------------------------------------------------------
|
||
# Detection threshold tests (model-aware, unified with compression config)
|
||
# ---------------------------------------------------------------------------
|
||
|
||
class TestSessionHygieneThresholds:
|
||
"""Test that the threshold logic correctly identifies large sessions.
|
||
|
||
Thresholds are derived from model context length × compression threshold,
|
||
matching what the agent's ContextCompressor uses.
|
||
"""
|
||
|
||
def test_small_session_below_thresholds(self):
|
||
"""A 10-message session should not trigger compression."""
|
||
history = _make_history(10)
|
||
approx_tokens = estimate_messages_tokens_rough(history)
|
||
|
||
# For a 200k-context model at 85% threshold = 170k
|
||
context_length = 200_000
|
||
threshold_pct = 0.85
|
||
compress_token_threshold = int(context_length * threshold_pct)
|
||
|
||
needs_compress = approx_tokens >= compress_token_threshold
|
||
assert not needs_compress
|
||
|
||
def test_large_token_count_triggers(self):
|
||
"""High token count should trigger compression when exceeding model threshold."""
|
||
# Build a history that exceeds 85% of a 200k model (170k tokens)
|
||
history = _make_large_history_tokens(180_000)
|
||
approx_tokens = estimate_messages_tokens_rough(history)
|
||
|
||
context_length = 200_000
|
||
threshold_pct = 0.85
|
||
compress_token_threshold = int(context_length * threshold_pct)
|
||
|
||
needs_compress = approx_tokens >= compress_token_threshold
|
||
assert needs_compress
|
||
|
||
def test_under_threshold_no_trigger(self):
|
||
"""Session under threshold should not trigger, even with many messages."""
|
||
# 250 short messages — lots of messages but well under token threshold
|
||
history = _make_history(250, content_size=10)
|
||
approx_tokens = estimate_messages_tokens_rough(history)
|
||
|
||
# 200k model at 85% = 170k token threshold
|
||
context_length = 200_000
|
||
threshold_pct = 0.85
|
||
compress_token_threshold = int(context_length * threshold_pct)
|
||
|
||
needs_compress = approx_tokens >= compress_token_threshold
|
||
assert not needs_compress, (
|
||
f"250 short messages (~{approx_tokens} tokens) should NOT trigger "
|
||
f"compression at {compress_token_threshold} token threshold"
|
||
)
|
||
|
||
def test_message_count_alone_does_not_trigger(self):
|
||
"""Message count alone should NOT trigger — only token count matters.
|
||
|
||
The old system used an OR of token-count and message-count thresholds,
|
||
which caused premature compression in tool-heavy sessions with 200+
|
||
messages but low total tokens.
|
||
"""
|
||
# 300 very short messages — old system would compress, new should not
|
||
history = _make_history(300, content_size=10)
|
||
approx_tokens = estimate_messages_tokens_rough(history)
|
||
|
||
context_length = 200_000
|
||
threshold_pct = 0.85
|
||
compress_token_threshold = int(context_length * threshold_pct)
|
||
|
||
# Token-based check only
|
||
needs_compress = approx_tokens >= compress_token_threshold
|
||
assert not needs_compress
|
||
|
||
def test_threshold_scales_with_model(self):
|
||
"""Different models should have different compression thresholds."""
|
||
# 128k model at 85% = 108,800 tokens
|
||
small_model_threshold = int(128_000 * 0.85)
|
||
# 200k model at 85% = 170,000 tokens
|
||
large_model_threshold = int(200_000 * 0.85)
|
||
# 1M model at 85% = 850,000 tokens
|
||
huge_model_threshold = int(1_000_000 * 0.85)
|
||
|
||
# A session at ~120k tokens:
|
||
history = _make_large_history_tokens(120_000)
|
||
approx_tokens = estimate_messages_tokens_rough(history)
|
||
|
||
# Should trigger for 128k model
|
||
assert approx_tokens >= small_model_threshold
|
||
# Should NOT trigger for 200k model
|
||
assert approx_tokens < large_model_threshold
|
||
# Should NOT trigger for 1M model
|
||
assert approx_tokens < huge_model_threshold
|
||
|
||
def test_custom_threshold_percentage(self):
|
||
"""Custom threshold percentage from config should be respected."""
|
||
context_length = 200_000
|
||
|
||
# At 50% threshold = 100k
|
||
low_threshold = int(context_length * 0.50)
|
||
# At 90% threshold = 180k
|
||
high_threshold = int(context_length * 0.90)
|
||
|
||
history = _make_large_history_tokens(150_000)
|
||
approx_tokens = estimate_messages_tokens_rough(history)
|
||
|
||
# Should trigger at 50% but not at 90%
|
||
assert approx_tokens >= low_threshold
|
||
assert approx_tokens < high_threshold
|
||
|
||
def test_minimum_message_guard(self):
|
||
"""Sessions with fewer than 4 messages should never trigger."""
|
||
history = _make_history(3, content_size=100_000)
|
||
# Even with enormous content, < 4 messages should be skipped
|
||
# (the gateway code checks `len(history) >= 4` before evaluating)
|
||
assert len(history) < 4
|
||
|
||
|
||
class TestSessionHygieneWarnThreshold:
|
||
"""Test the post-compression warning threshold (95% of context)."""
|
||
|
||
def test_warn_when_still_large(self):
|
||
"""If compressed result is still above 95% of context, should warn."""
|
||
context_length = 200_000
|
||
warn_threshold = int(context_length * 0.95) # 190k
|
||
post_compress_tokens = 195_000
|
||
assert post_compress_tokens >= warn_threshold
|
||
|
||
def test_no_warn_when_under(self):
|
||
"""If compressed result is under 95% of context, no warning."""
|
||
context_length = 200_000
|
||
warn_threshold = int(context_length * 0.95) # 190k
|
||
post_compress_tokens = 150_000
|
||
assert post_compress_tokens < warn_threshold
|
||
|
||
|
||
class TestTokenEstimation:
|
||
"""Verify rough token estimation works as expected for hygiene checks."""
|
||
|
||
def test_empty_history(self):
|
||
assert estimate_messages_tokens_rough([]) == 0
|
||
|
||
def test_proportional_to_content(self):
|
||
small = _make_history(10, content_size=100)
|
||
large = _make_history(10, content_size=10_000)
|
||
assert estimate_messages_tokens_rough(large) > estimate_messages_tokens_rough(small)
|
||
|
||
def test_proportional_to_count(self):
|
||
few = _make_history(10, content_size=1000)
|
||
many = _make_history(100, content_size=1000)
|
||
assert estimate_messages_tokens_rough(many) > estimate_messages_tokens_rough(few)
|
||
|
||
def test_pathological_session_detected(self):
|
||
"""The reported pathological case: 648 messages, ~299K tokens.
|
||
|
||
With a 200k model at 85% threshold (170k), this should trigger.
|
||
"""
|
||
history = _make_history(648, content_size=1800)
|
||
tokens = estimate_messages_tokens_rough(history)
|
||
# Should be well above the 170K threshold for a 200k model
|
||
threshold = int(200_000 * 0.85)
|
||
assert tokens > threshold
|