fix: gateway token double-counting — use absolute set instead of increment (#3317)

The gateway's update_session() used += for token counts, but the cached
agent's session_prompt_tokens / session_completion_tokens are cumulative
totals that grow across messages. Each update_session call re-added the
running total, inflating usage stats with every message (1.7x after 3
messages, worse over longer conversations).

Fix: change += to = for in-memory entry fields, add set_token_counts()
to SessionDB that uses direct assignment instead of SQL increment, and
switch the gateway to call it.

CLI mode continues using update_token_counts() (increment) since it
tracks per-API-call deltas — that path is unchanged.

Based on analysis from PR #3222 by @zaycruz (closed).

Co-authored-by: zaycruz <zay@users.noreply.github.com>
This commit is contained in:
Teknium 2026-03-26 19:13:07 -07:00 committed by GitHub
parent 867eefdd9f
commit 22cfad157b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 68 additions and 2 deletions

View file

@ -785,7 +785,7 @@ class SessionStore:
if self._db and db_session_id:
try:
self._db.update_token_counts(
self._db.set_token_counts(
db_session_id,
input_tokens=input_tokens,
output_tokens=output_tokens,

View file

@ -415,6 +415,72 @@ class SessionDB:
)
self._conn.commit()
def set_token_counts(
self,
session_id: str,
input_tokens: int = 0,
output_tokens: int = 0,
model: str = None,
cache_read_tokens: int = 0,
cache_write_tokens: int = 0,
reasoning_tokens: int = 0,
estimated_cost_usd: Optional[float] = None,
actual_cost_usd: Optional[float] = None,
cost_status: Optional[str] = None,
cost_source: Optional[str] = None,
pricing_version: Optional[str] = None,
billing_provider: Optional[str] = None,
billing_base_url: Optional[str] = None,
billing_mode: Optional[str] = None,
) -> None:
"""Set token counters to absolute values (not increment).
Use this when the caller provides cumulative totals from a completed
conversation run (e.g. the gateway, where the cached agent's
session_prompt_tokens already reflects the running total).
"""
with self._lock:
self._conn.execute(
"""UPDATE sessions SET
input_tokens = ?,
output_tokens = ?,
cache_read_tokens = ?,
cache_write_tokens = ?,
reasoning_tokens = ?,
estimated_cost_usd = ?,
actual_cost_usd = CASE
WHEN ? IS NULL THEN actual_cost_usd
ELSE ?
END,
cost_status = COALESCE(?, cost_status),
cost_source = COALESCE(?, cost_source),
pricing_version = COALESCE(?, pricing_version),
billing_provider = COALESCE(billing_provider, ?),
billing_base_url = COALESCE(billing_base_url, ?),
billing_mode = COALESCE(billing_mode, ?),
model = COALESCE(model, ?)
WHERE id = ?""",
(
input_tokens,
output_tokens,
cache_read_tokens,
cache_write_tokens,
reasoning_tokens,
estimated_cost_usd,
actual_cost_usd,
actual_cost_usd,
cost_status,
cost_source,
pricing_version,
billing_provider,
billing_base_url,
billing_mode,
model,
session_id,
),
)
self._conn.commit()
def get_session(self, session_id: str) -> Optional[Dict[str, Any]]:
"""Get a session by ID."""
with self._lock:

View file

@ -846,7 +846,7 @@ class TestLastPromptTokens:
store.update_session("k1", model="openai/gpt-5.4")
store._db.update_token_counts.assert_called_once_with(
store._db.set_token_counts.assert_called_once_with(
"s1",
input_tokens=0,
output_tokens=0,