mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-06-23 10:42:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 7

Merge pull request #50726 from NousResearch/salvage/compression-token-progress

Browse source 
fix(agent): count tokens, not just message rows, as compression progress (#23767, #39550)
This commit is contained in:
kshitij 2026-06-22 15:44:38 +05:30 • committed by GitHub
commit 33efff0d8c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 68 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

31
agent/conversation_loop.py
View file

@ -2983,6 +2983,7 @@ def run_conversation(
agent._buffer_status(f"⚠️ Request payload too large (413) — compression attempt {compression_attempts}/{max_compression_attempts}...")
original_len = len(messages)
original_tokens = estimate_messages_tokens_rough(messages)
messages, active_system_prompt = agent._compress_context(
messages, system_message, approx_tokens=approx_tokens,
task_id=effective_task_id,
@ -2992,8 +2993,18 @@ def run_conversation(
# messages to the new session, not skipping them.
conversation_history = None
if len(messages) < original_len:
agent._buffer_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...")
# Re-estimate tokens after compression. Same-message-count
# compression (tool-result pruning, in-place summarization)
# can materially reduce request size without reducing the
# message array. (#39550)
new_tokens = estimate_messages_tokens_rough(messages)
approx_tokens = new_tokens # update for downstream logging
if len(messages) < original_len or (new_tokens > 0 and new_tokens < original_tokens * 0.95):
if len(messages) < original_len:
agent._buffer_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...")
else:
agent._buffer_status(f"🗜️ Compressed ~{original_tokens:,} → ~{new_tokens:,} tokens, retrying...")
time.sleep(2) # Brief pause between compression retries
_retry.restart_with_compressed_messages = True
break
@ -3139,6 +3150,7 @@ def run_conversation(
agent._buffer_status(f"🗜️ Context too large (~{approx_tokens:,} tokens) — compressing ({compression_attempts}/{max_compression_attempts})...")
original_len = len(messages)
original_tokens = estimate_messages_tokens_rough(messages)
messages, active_system_prompt = agent._compress_context(
messages, system_message, approx_tokens=approx_tokens,
task_id=effective_task_id,
@ -3148,9 +3160,18 @@ def run_conversation(
# messages to the new session, not skipping them.
conversation_history = None
if len(messages) < original_len or new_ctx and new_ctx < old_ctx:
# Re-estimate tokens after compression. Same-message-count
# compression (tool-result pruning, in-place summarization)
# can materially reduce request size without reducing the
# message array. (#39550)
new_tokens = estimate_messages_tokens_rough(messages)
approx_tokens = new_tokens # update for downstream logging
if len(messages) < original_len or (new_tokens > 0 and new_tokens < original_tokens * 0.95) or (new_ctx and new_ctx < old_ctx):
if len(messages) < original_len:
agent._buffer_status(f"🗜️ Compressed {original_len} → {len(messages)} messages, retrying...")
elif new_tokens > 0 and new_tokens < original_tokens * 0.95:
agent._buffer_status(f"🗜️ Compressed ~{original_tokens:,} → ~{new_tokens:,} tokens, retrying...")
time.sleep(2) # Brief pause between compression retries
_retry.restart_with_compressed_messages = True
break
@ -3159,13 +3180,13 @@ def run_conversation(
agent._flush_status_buffer()
agent._vprint(f"{agent.log_prefix}❌ Context length exceeded and cannot compress further.", force=True)
agent._vprint(f"{agent.log_prefix} 💡 The conversation has accumulated too much content. Try /new to start fresh, or /compress to manually trigger compression.", force=True)
logger.error(f"{agent.log_prefix}Context length exceeded: {approx_tokens:,} tokens. Cannot compress further.")
logger.error(f"{agent.log_prefix}Context length exceeded: {new_tokens:,} tokens. Cannot compress further.")
agent._persist_session(messages, conversation_history)
return {
"messages": messages,
"completed": False,
"api_calls": api_call_count,
"error": f"Context length exceeded ({approx_tokens:,} tokens). Cannot compress further.",
"error": f"Context length exceeded ({new_tokens:,} tokens). Cannot compress further.",
"partial": True,
"failed": True,
"compression_exhausted": True,

42
tests/run_agent/test_413_compression.py
View file

@ -440,6 +440,48 @@ class TestHTTP413Compression:
assert result.get("partial") is True
assert "413" in result["error"]
def test_413_retries_on_token_only_compression(self, agent):
"""Same message COUNT but fewer TOKENS must count as progress and retry.
Regression for #39550/#23767: tool-result pruning / in-place
summarization can shrink request size without dropping the message
count. The old gate (len(messages) < original_len) treated that as
'cannot compress further' and aborted; the fix re-estimates tokens and
retries when they drop materially.
"""
err_413 = _make_413_error()
ok_resp = _mock_response(content="OK after token-only compaction", finish_reason="stop")
agent.client.chat.completions.create.side_effect = [err_413, ok_resp]
# 3 large messages in, 3 much smaller messages out (same count, far
# fewer tokens) — exactly the token-only-progress case.
prefill = [
{"role": "user", "content": "x" * 4000},
{"role": "assistant", "content": "y" * 4000},
{"role": "user", "content": "z" * 4000},
]
with (
patch.object(agent, "_compress_context") as mock_compress,
patch.object(agent, "_persist_session"),
patch.object(agent, "_save_trajectory"),
patch.object(agent, "_cleanup_task_resources"),
):
# Same message count (3) but ~10x smaller content → token drop.
mock_compress.return_value = (
[
{"role": "user", "content": "x" * 300},
{"role": "assistant", "content": "y" * 300},
{"role": "user", "content": "z" * 300},
],
"compressed prompt",
)
result = agent.run_conversation("hello", conversation_history=prefill)
mock_compress.assert_called_once()
assert result["completed"] is True
assert result["final_response"] == "OK after token-only compaction"
class TestPreflightCompression:
"""Preflight compression should compress history before the first API call."""