mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-05-01 01:51:44 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 6

fix: improve context compression quality — named constants, tool tracking, degradation warning

Browse source 
Three targeted improvements to the compression system:

1. Replace hardcoded truncation limits with named class constants
   (_CONTENT_MAX=6000, _CONTENT_HEAD=4000, _CONTENT_TAIL=1500,
   _TOOL_ARGS_MAX=1500, _TOOL_ARGS_HEAD=1200). Previous limits
   (3000/500) heavily truncated the summarizer's input — a 200-line
   edit got cut to 3000 chars before the summarizer ever saw it.

2. Add '## Tools & Patterns' section to both compression prompt
   templates (first-pass and iterative). Preserves working tool
   invocations, preferred flags, and tool-specific discoveries
   across compaction boundaries.

3. Warn users on 2nd+ compression: 'Session compressed N times —
   accuracy may degrade. Consider /new to start fresh.'

Ref #499
This commit is contained in:
SHL0MS 2026-03-30 23:34:58 -04:00 • committed by Teknium
parent af4abd2f22
commit 8567031433
2 changed files with 36 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

39
agent/context_compressor.py
View file

@ -199,30 +199,39 @@ class ContextCompressor:
budget = int(content_tokens * _SUMMARY_RATIO) budget = int(content_tokens * _SUMMARY_RATIO)
return max(_MIN_SUMMARY_TOKENS, min(budget, self.max_summary_tokens)) return max(_MIN_SUMMARY_TOKENS, min(budget, self.max_summary_tokens))
# Truncation limits for the summarizer input. These bound how much of
# each message the summary model sees — the budget is the *summary*
# model's context window, not the main model's.
_CONTENT_MAX = 6000 # total chars per message body
_CONTENT_HEAD = 4000 # chars kept from the start
_CONTENT_TAIL = 1500 # chars kept from the end
_TOOL_ARGS_MAX = 1500 # tool call argument chars
_TOOL_ARGS_HEAD = 1200 # kept from the start of tool args
def _serialize_for_summary(self, turns: List[Dict[str, Any]]) -> str: def _serialize_for_summary(self, turns: List[Dict[str, Any]]) -> str:
"""Serialize conversation turns into labeled text for the summarizer. """Serialize conversation turns into labeled text for the summarizer.
Includes tool call arguments and result content (up to 3000 chars Includes tool call arguments and result content (up to
per message) so the summarizer can preserve specific details like ``_CONTENT_MAX`` chars per message) so the summarizer can preserve
file paths, commands, and outputs. specific details like file paths, commands, and outputs.
""" """
parts = [] parts = []
for msg in turns: for msg in turns:
role = msg.get("role", "unknown") role = msg.get("role", "unknown")
content = msg.get("content") or "" content = msg.get("content") or ""
# Tool results: keep more content than before (3000 chars) # Tool results: keep enough content for the summarizer
if role == "tool": if role == "tool":
tool_id = msg.get("tool_call_id", "") tool_id = msg.get("tool_call_id", "")
if len(content) > 3000: if len(content) > self._CONTENT_MAX:
content = content[:2000] + "\n...[truncated]...\n" + content[-800:] content = content[:self._CONTENT_HEAD] + "\n...[truncated]...\n" + content[-self._CONTENT_TAIL:]
parts.append(f"[TOOL RESULT {tool_id}]: {content}") parts.append(f"[TOOL RESULT {tool_id}]: {content}")
continue continue
# Assistant messages: include tool call names AND arguments # Assistant messages: include tool call names AND arguments
if role == "assistant": if role == "assistant":
if len(content) > 3000: if len(content) > self._CONTENT_MAX:
content = content[:2000] + "\n...[truncated]...\n" + content[-800:] content = content[:self._CONTENT_HEAD] + "\n...[truncated]...\n" + content[-self._CONTENT_TAIL:]
tool_calls = msg.get("tool_calls", []) tool_calls = msg.get("tool_calls", [])
if tool_calls: if tool_calls:
tc_parts = [] tc_parts = []
@ -232,8 +241,8 @@ class ContextCompressor:
name = fn.get("name", "?") name = fn.get("name", "?")
args = fn.get("arguments", "") args = fn.get("arguments", "")
# Truncate long arguments but keep enough for context # Truncate long arguments but keep enough for context
if len(args) > 500: if len(args) > self._TOOL_ARGS_MAX:
args = args[:400] + "..." args = args[:self._TOOL_ARGS_HEAD] + "..."
tc_parts.append(f" {name}({args})") tc_parts.append(f" {name}({args})")
else: else:
fn = getattr(tc, "function", None) fn = getattr(tc, "function", None)
@ -244,8 +253,8 @@ class ContextCompressor:
continue continue
# User and other roles # User and other roles
if len(content) > 3000: if len(content) > self._CONTENT_MAX:
content = content[:2000] + "\n...[truncated]...\n" + content[-800:] content = content[:self._CONTENT_HEAD] + "\n...[truncated]...\n" + content[-self._CONTENT_TAIL:]
parts.append(f"[{role.upper()}]: {content}") parts.append(f"[{role.upper()}]: {content}")
return "\n\n".join(parts) return "\n\n".join(parts)
@ -310,6 +319,9 @@ Update the summary using this exact structure. PRESERVE all existing information
## Critical Context ## Critical Context
[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation] [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
## Tools & Patterns
[Which tools were used, how they were used effectively, and any tool-specific discoveries. Accumulate across compactions.]
Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions.
Write only the summary body. Do not include any preamble or prefix.""" Write only the summary body. Do not include any preamble or prefix."""
@ -348,6 +360,9 @@ Use this exact structure:
## Critical Context ## Critical Context
[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation] [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
## Tools & Patterns
[Which tools were used, how they were used effectively, and any tool-specific discoveries (e.g., preferred flags, working invocations, successful command patterns)]
Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. The goal is to prevent the next assistant from repeating work or losing important details. Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. The goal is to prevent the next assistant from repeating work or losing important details.
Write only the summary body. Do not include any preamble or prefix.""" Write only the summary body. Do not include any preamble or prefix."""

9
run_agent.py
View file

@ -6013,6 +6013,15 @@ class AIAgent:
except Exception as e: except Exception as e:
logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e) logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e)
# Warn on repeated compressions (quality degrades with each pass)
_cc = self.context_compressor.compression_count
if _cc >= 2:
self._vprint(
f"{self.log_prefix}⚠️ Session compressed {_cc} times — "
f"accuracy may degrade. Consider /new to start fresh.",
force=True,
)
# Update token estimate after compaction so pressure calculations # Update token estimate after compaction so pressure calculations
# use the post-compression count, not the stale pre-compression one. # use the post-compression count, not the stale pre-compression one.
_compressed_est = ( _compressed_est = (