mirror of
https://github.com/openswarm-ai/openswarm.git
synced 2026-09-03 08:18:43 +02:00
ciregenzandClaude Opus 5
0f2543db65
Every silent-stop fix so far was a detector for a shape somebody had already found in the field, which is why the class kept coming back wearing a new hat. turn_spoke.py moves the question down a tier: at the one exit every terminal path passes through, ask whether anything readable appeared since the user last spoke, and if not, say the honest line. Cause no longer has to be enumerated for the user to be answered. proactive_prune.py is the hermes trigger we were missing. Their own tests say our bug out loud: on a large window, a percentage-of-window compaction check almost never fires, so aged tool output rides in history and is re-sent verbatim every turn. Measured here, our shaping cut 0.0% at every session size; with a fixed 60K-token cost trigger it cuts 88% at 12 turns, 93% at 30, 94% at 60. The prompt-cache contract is load-bearing rather than optional, because our prune is a rebuild: it commits only when it reclaims enough to pay for the busted prefix, then disarms until history has regrown a full runway. Also: lane preflight now treats only 401/403 as a dead credential, since testStatus=="unavailable" conflated a throttled lane with a revoked one and told users to reconnect a merely rate-limited Gemini; and awaiting_reconnect is cleared when the retry budget is spent, so a stale flag can no longer muzzle the floor and end an ask in total silence. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014wtspwSFzZmjCx9UNPAorQ
89 lines
4.1 KiB
Python
89 lines
4.1 KiB
Python
"""The trigger that fires on cost instead of on a percentage of the window.
|
|
|
|
Measured bug this closes: our shaping cut 0.0% at every session size because the compaction
|
|
threshold is a fraction of the context window, and on a 1M lane that fraction is never reached, so
|
|
a 218K history shipped verbatim to a cliff the model chokes at first. Hermes hit the same wall and
|
|
solved it with a second, independent trigger (MIT, NousResearch/hermes-agent).
|
|
"""
|
|
|
|
from backend.apps.agents.core.models import AgentSession, Message
|
|
from backend.apps.agents.manager.session.proactive_prune import (
|
|
MIN_RECLAIM_TOKENS,
|
|
PROACTIVE_PRUNE_TOKENS,
|
|
arm_proactive_prune,
|
|
should_proactively_prune,
|
|
)
|
|
|
|
|
|
def p_session(input_tokens: int, bulky_msgs: int = 40, window: int = 1_000_000) -> AgentSession:
|
|
s = AgentSession(name="t", model="sonnet-5", dashboard_id="d")
|
|
s.context_window = window
|
|
s.tokens["input"] = input_tokens
|
|
s.messages.append(Message(role="user", content="go", branch_id=s.active_branch_id))
|
|
for i in range(bulky_msgs):
|
|
s.messages.append(Message(role="tool_call", content={"tool": "Read", "input": {"n": i}},
|
|
branch_id=s.active_branch_id))
|
|
s.messages.append(Message(role="tool_result", content={"text": "x" * 20_000},
|
|
branch_id=s.active_branch_id))
|
|
return s
|
|
|
|
|
|
def test_a_big_history_on_a_1m_window_is_finally_pruned():
|
|
"""The exact case that shipped 0.0%: far past any sane cost, nowhere near 50% of 1M."""
|
|
s = p_session(120_000)
|
|
assert should_proactively_prune(s) is True
|
|
|
|
|
|
def test_a_small_session_is_left_alone():
|
|
"""Negative control: pruning a cheap session spends a prompt cache for nothing."""
|
|
assert should_proactively_prune(p_session(5_000, bulky_msgs=2)) is False
|
|
|
|
|
|
def test_it_never_duplicates_the_work_the_real_threshold_is_about_to_do():
|
|
"""Above the compaction trigger the existing path owns it; two rebuilds would be one wasted."""
|
|
s = p_session(700_000)
|
|
assert should_proactively_prune(s) is False
|
|
|
|
|
|
def test_a_prune_that_reclaims_little_is_refused():
|
|
"""The prompt-cache contract: a rebuild rewrites bytes the provider cached, so it must earn it.
|
|
A long conversation of SHORT messages has nothing worth reclaiming."""
|
|
s = AgentSession(name="t", model="sonnet-5", dashboard_id="d")
|
|
s.context_window = 1_000_000
|
|
s.tokens["input"] = PROACTIVE_PRUNE_TOKENS + 10_000
|
|
for i in range(60):
|
|
s.messages.append(Message(role="user", content=f"q{i}", branch_id=s.active_branch_id))
|
|
s.messages.append(Message(role="assistant", content=f"a{i}", branch_id=s.active_branch_id))
|
|
assert should_proactively_prune(s) is False
|
|
|
|
|
|
def test_committing_disarms_until_history_regrows():
|
|
"""A session hovering at the line must not rebuild every single turn."""
|
|
s = p_session(120_000)
|
|
assert should_proactively_prune(s) is True
|
|
arm_proactive_prune(s)
|
|
assert s.needs_fresh_session is True, "the rebuild is what actually applies the aging"
|
|
assert s.proactive_prune_rearm_tokens > 0
|
|
|
|
# Same history again right after: disarmed.
|
|
assert should_proactively_prune(s) is False
|
|
|
|
# History genuinely regrown past the runway: armed again.
|
|
for i in range(40):
|
|
s.messages.append(Message(role="tool_call", content={"tool": "Read", "input": {"n": 900 + i}},
|
|
branch_id=s.active_branch_id))
|
|
s.messages.append(Message(role="tool_result", content={"text": "y" * 20_000},
|
|
branch_id=s.active_branch_id))
|
|
s.compacted_through_msg_id = None
|
|
assert should_proactively_prune(s) is True
|
|
|
|
|
|
def test_the_trigger_is_not_a_fraction_of_the_window():
|
|
"""The whole correction: identical history fires on a 200K lane and a 1M lane alike, because
|
|
the tokens cost the same money either way. Tying this to a percentage of the window is what
|
|
let a 218K history sail through untouched on the big lane."""
|
|
small = p_session(50_000, window=200_000)
|
|
big = p_session(50_000, window=1_000_000)
|
|
assert should_proactively_prune(small) is True
|
|
assert should_proactively_prune(big) is True
|