fangliquanflq
c2088efe9e
Three mechanisms to detect and notify when gateway sessions stall silently: 1. Mid-turn activity heartbeats stamped to SessionDB so hermes sessions list and hermes status show progress during long turns without new message rows. 2. Stall watchdog: when a busy session has pending inbound and the shared activity clock is idle past agent.session_stall_timeout (default 300), log a WARNING and notify the user once to try /new. Notify-only; does not kill the turn. 3. Compaction timeout: fenceless compress_context callers get a progress-aware host budget (compression.context_timeout_seconds default 120 idle, compression.context_total_ceiling_seconds default 600 ceiling). On timeout, cancel via commit fence, skip compaction without dropping messages, and continue the turn. Closes #72016 (slices 1-3; slice 4 cumulative SSE stream-retry deadline remains a follow-up). Cherry-picked from PR #72424 by @fangliquanflq.
78 lines
3.2 KiB
Python
78 lines
3.2 KiB
Python
"""Regression tests for PR #48127: cached agent max_iterations refresh.
|
|
|
|
When a long-lived gateway reuses an agent from its cache, the agent must run
|
|
the *current* configured iteration budget — not the budget it was constructed
|
|
with on the first turn of that session. Two pieces make that true:
|
|
|
|
1. ``GatewayRunner._init_cached_agent_for_turn`` must NOT reset
|
|
``max_iterations`` itself (the gateway refreshes it explicitly right after,
|
|
from current config). If this helper ever started clobbering it, the
|
|
gateway's refresh would be silently undone.
|
|
2. The per-turn budget object is rebuilt from ``agent.max_iterations`` at the
|
|
start of every turn (``agent/turn_context.py`` -> ``IterationBudget``), so
|
|
refreshing ``max_iterations`` on the cached agent is sufficient to change
|
|
the operative cap the agent loop checks.
|
|
|
|
These tests exercise the real code paths rather than asserting a plain
|
|
assignment, so they fail if either contract regresses.
|
|
"""
|
|
|
|
import time
|
|
from types import SimpleNamespace
|
|
|
|
from agent.iteration_budget import IterationBudget
|
|
from agent.session_activity import ActivityProvenance
|
|
|
|
|
|
def _make_cached_agent(max_iterations: int) -> SimpleNamespace:
|
|
"""A minimal stand-in cached agent with the attributes the helpers touch."""
|
|
# The turn loop checks both api_call_count >= max_iterations AND
|
|
# iteration_budget.remaining <= 0 (turn_finalizer.py), so the budget must
|
|
# also reflect the new cap. Seed it with the stale value to prove the
|
|
# refresh propagates.
|
|
return SimpleNamespace(
|
|
_last_activity_ts=time.time() - 1000,
|
|
_last_activity_desc="previous turn",
|
|
_last_activity_provenance=ActivityProvenance.AGENT_COMPRESSION,
|
|
_api_call_count=42,
|
|
_last_flushed_db_idx=5,
|
|
max_iterations=max_iterations,
|
|
iteration_budget=IterationBudget(max_iterations),
|
|
)
|
|
|
|
|
|
def test_init_cached_agent_for_turn_does_not_touch_max_iterations():
|
|
"""The per-turn reset helper must leave max_iterations untouched.
|
|
|
|
The gateway refreshes max_iterations explicitly right after calling this
|
|
helper; if the helper ever reset it, that refresh would be undone.
|
|
"""
|
|
from gateway.run import GatewayRunner
|
|
|
|
agent = _make_cached_agent(90)
|
|
GatewayRunner._init_cached_agent_for_turn(agent, interrupt_depth=0)
|
|
|
|
# Per-turn state was reset...
|
|
assert agent._api_call_count == 0
|
|
assert agent._last_activity_desc == "starting new turn (cached)"
|
|
assert agent._last_activity_provenance is ActivityProvenance.UNKNOWN
|
|
assert agent._last_flushed_db_idx == 0
|
|
# ...but the iteration budget was NOT changed by the helper itself.
|
|
assert agent.max_iterations == 90
|
|
|
|
|
|
def test_init_cached_agent_preserves_max_iterations_on_interrupt_depth():
|
|
"""Interrupt-recursive turns must also leave max_iterations alone."""
|
|
from gateway.run import GatewayRunner
|
|
|
|
agent = _make_cached_agent(200)
|
|
GatewayRunner._init_cached_agent_for_turn(agent, interrupt_depth=1)
|
|
|
|
# Activity timestamps preserved for the inactivity watchdog (#15654)...
|
|
assert agent._last_activity_desc == "previous turn"
|
|
assert agent._last_activity_provenance is ActivityProvenance.AGENT_COMPRESSION
|
|
# ...and max_iterations untouched.
|
|
assert agent.max_iterations == 200
|
|
|
|
|