mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-05-07 02:51:50 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 7

fix(agent): recompute compressor budgets on context updates

Browse source
This commit is contained in:
Tranquil-Flow 2026-04-25 10:08:47 +10:00
parent 89389f1894
commit 1484a559fe
2 changed files with 24 additions and 28 deletions
Download patch file Download diff file Expand all files Collapse all files

46
agent/context_compressor.py
View file

@ -288,6 +288,23 @@ class ContextCompressor(ContextEngine):
def name(self) -> str:
return "compressor"
def _recompute_thresholds_and_budgets(self) -> None:
"""Rebuild derived token thresholds after a context/model change."""
self.threshold_tokens = max(
int(self.context_length * self.threshold_percent),
MINIMUM_CONTEXT_LENGTH,
)
# Threshold must stay below context_length so compression can trigger
# before the provider rejects the request outright.
self.threshold_tokens = min(
self.threshold_tokens, int(self.context_length * 0.95)
)
target_tokens = int(self.threshold_tokens * self.summary_target_ratio)
self.tail_token_budget = target_tokens
self.max_summary_tokens = min(
int(self.context_length * 0.05), _SUMMARY_TOKENS_CEILING,
)
def on_session_reset(self) -> None:
"""Reset all per-session state for /new or /reset."""
super().on_session_reset()
@ -314,16 +331,7 @@ class ContextCompressor(ContextEngine):
self.provider = provider
self.api_mode = api_mode
self.context_length = context_length
self.threshold_tokens = max(
int(context_length * self.threshold_percent),
MINIMUM_CONTEXT_LENGTH,
)
# Clamp: threshold must stay below context_length so compression can
# actually trigger before the API rejects the request. The 95% cap
# gives 5% headroom — enough to catch the threshold in time.
self.threshold_tokens = min(
self.threshold_tokens, int(context_length * 0.95)
)
self._recompute_thresholds_and_budgets()
def __init__(
self,
@ -360,25 +368,9 @@ class ContextCompressor(ContextEngine):
# the percentage would suggest a lower value. This prevents premature
# compression on large-context models at 50% while keeping the % sane
# for models right at the minimum.
self.threshold_tokens = max(
int(self.context_length * threshold_percent),
MINIMUM_CONTEXT_LENGTH,
)
# Clamp: threshold must stay below context_length so compression can
# actually trigger before the API rejects the request. The 95% cap
# gives 5% headroom — enough to catch the threshold in time.
self.threshold_tokens = min(
self.threshold_tokens, int(self.context_length * 0.95)
)
self._recompute_thresholds_and_budgets()
self.compression_count = 0
# Derive token budgets: ratio is relative to the threshold, not total context
target_tokens = int(self.threshold_tokens * self.summary_target_ratio)
self.tail_token_budget = target_tokens
self.max_summary_tokens = min(
int(self.context_length * 0.05), _SUMMARY_TOKENS_CEILING,
)
if not quiet_mode:
logger.info(
"Context compressor initialized: model=%s context_length=%d "