ling/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-05-01 08:21:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: include cache tokens in dashboard analytics input totals

Browse Source 
The /api/analytics/usage endpoint summed only the raw input_tokens
column, which for Anthropic-direct sessions holds only the uncached
portion of the prompt.  cache_read_tokens and cache_write_tokens
(which complete the total prompt) were ignored.

This caused the dashboard to massively undercount token usage —
showing ~117M instead of ~345M over 30 days — since Anthropic
sessions with high cache hit rates stored almost all prompt tokens
in the cache columns.

Fix: fold COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)
into the input_tokens sum across all three SQL queries (daily, by-model,
totals).  This is correct for every provider because normalize_usage()
guarantees input_tokens + cache_read + cache_write = total prompt tokens
regardless of API shape (Anthropic / OpenAI / Codex).

Add a regression test that creates a session with Anthropic-style token
splits and asserts the endpoint returns the combined total.
This commit is contained in:
yoniebans
2026-04-27 19:14:30 +02:00
committed by morpheus
parent d5a89283b7
commit 18e6cd9938
2 changed files with 47 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
hermes_cli/web_server.py
View File

@@ -2212,7 +2212,7 @@ async def get_usage_analytics(days: int = 30):
cutoff = time.time() - (days * 86400)
cur = db._conn.execute("""
SELECT date(started_at, 'unixepoch') as day,
SUM(input_tokens) as input_tokens,
SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) as input_tokens,
SUM(output_tokens) as output_tokens,
SUM(cache_read_tokens) as cache_read_tokens,
SUM(reasoning_tokens) as reasoning_tokens,
@@ -2227,18 +2227,18 @@ async def get_usage_analytics(days: int = 30):
cur2 = db._conn.execute("""
SELECT model,
SUM(input_tokens) as input_tokens,
SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) as input_tokens,
SUM(output_tokens) as output_tokens,
COALESCE(SUM(estimated_cost_usd), 0) as estimated_cost,
COUNT(*) as sessions,
SUM(COALESCE(api_call_count, 0)) as api_calls
FROM sessions WHERE started_at > ? AND model IS NOT NULL
GROUP BY model ORDER BY SUM(input_tokens) + SUM(output_tokens) DESC
GROUP BY model ORDER BY SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) + SUM(output_tokens) DESC
""", (cutoff,))
by_model = [dict(r) for r in cur2.fetchall()]
cur3 = db._conn.execute("""
SELECT SUM(input_tokens) as total_input,
SELECT SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) as total_input,
SUM(output_tokens) as total_output,
SUM(cache_read_tokens) as total_cache_read,
SUM(reasoning_tokens) as total_reasoning,