Skip to content
Python

Carry reviewer cache tokens into the metering badge

Devika Verma36656407ee40
diff --git a/coworker/audit.py b/coworker/audit.py
index d05a6976..5199a7b7 100644
--- a/coworker/audit.py
+++ b/coworker/audit.py
@@ -47,7 +47,9 @@ class AuditStore:
resource TEXT,
call_id TEXT,
tokens_in INTEGER DEFAULT 0,
- tokens_out INTEGER DEFAULT 0
+ tokens_out INTEGER DEFAULT 0,
+ cache_read INTEGER DEFAULT 0,
+ cache_write INTEGER DEFAULT 0
)
""")
# Existing databases predate the reviewer columns (2026-08-12): call_id joins a
@@ -58,6 +60,13 @@ class AuditStore:
("call_id", "TEXT"),
("tokens_in", "INTEGER DEFAULT 0"),
("tokens_out", "INTEGER DEFAULT 0"),
+ # Cached-prefix share of a reviewer check (2026-08-22). Without these the
+ # metering badge could only ever see the FRESH tokens — ~75 of a ~1,500-token
+ # check once the provider caches the instruction prefix — so it under-reported
+ # cost by more the longer a session ran. Same defect class as OPE-101, one
+ # layer further out.
+ ("cache_read", "INTEGER DEFAULT 0"),
+ ("cache_write", "INTEGER DEFAULT 0"),
):
try:
self._conn.execute(
@@ -78,8 +87,8 @@ class AuditStore:
self._conn.execute(
"""
INSERT INTO audit_events
- (session_id, agent, workspace, connector, tool, stage, status, approval, args, result_preview, reason, resource, call_id, tokens_in, tokens_out)
- VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+ (session_id, agent, workspace, connector, tool, stage, status, approval, args, result_preview, reason, resource, call_id, tokens_in, tokens_out, cache_read, cache_write)
+ VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
""",
(
event.get("session_id") or "",
@@ -97,6 +106,8 @@ class AuditStore:
str(event.get("call_id") or ""),
int(event.get("tokens_in") or 0),
int(event.get("tokens_out") or 0),
+ int(event.get("cache_read") or 0),
+ int(event.get("cache_write") or 0),
),
)
self._conn.commit()
@@ -112,14 +123,19 @@ class AuditStore:
"""
SELECT status, COUNT(*) AS n,
COALESCE(SUM(tokens_in), 0) AS tin,
- COALESCE(SUM(tokens_out), 0) AS tout
+ COALESCE(SUM(tokens_out), 0) AS tout,
+ COALESCE(SUM(cache_read), 0) AS cread,
+ COALESCE(SUM(cache_write), 0) AS cwrite
FROM audit_events
WHERE session_id = ? AND stage = ?
GROUP BY status
""",
(session_id, stage),
).fetchall()
- out = {"checks": 0, "allow": 0, "deny": 0, "unsure": 0, "tokens_in": 0, "tokens_out": 0}
+ out = {
+ "checks": 0, "allow": 0, "deny": 0, "unsure": 0,
+ "tokens_in": 0, "tokens_out": 0, "cache_read": 0, "cache_write": 0,
+ }
for row in rows:
status = str(row["status"])
if status in ("allow", "deny", "unsure"):
@@ -127,6 +143,8 @@ class AuditStore:
out["checks"] += int(row["n"])
out["tokens_in"] += int(row["tin"])
out["tokens_out"] += int(row["tout"])
+ out["cache_read"] += int(row["cread"])
+ out["cache_write"] += int(row["cwrite"])
return out
return {"live": _bucket("reviewer_verdict"), "shadow": _bucket("reviewer_shadow")}
diff --git a/coworker/engine.py b/coworker/engine.py
index 09b65d8c..46b7eba7 100644
--- a/coworker/engine.py
+++ b/coworker/engine.py
@@ -1030,6 +1030,8 @@ class TurnEngine:
call_id=tool_call.id,
tokens_in=verdict.tokens_in,
tokens_out=verdict.tokens_out,
+ cache_read=verdict.cache_read,
+ cache_write=verdict.cache_write,
)
except Exception:
pass # shadow must never surface a failure
@@ -1117,6 +1119,8 @@ class TurnEngine:
reason=verdict.reason,
tokens_in=verdict.tokens_in,
tokens_out=verdict.tokens_out,
+ cache_read=verdict.cache_read,
+ cache_write=verdict.cache_write,
)
if verdict.verdict == "allow":
allowed = True
diff --git a/surfaces/gui/src/api.ts b/surfaces/gui/src/api.ts
index 9cf13a78..f355faad 100644
--- a/surfaces/gui/src/api.ts
+++ b/surfaces/gui/src/api.ts
@@ -1657,6 +1657,10 @@ export interface ReviewerBucket {
unsure: number;
tokens_in: number;
tokens_out: number;
+ // Cached-prefix share of the input, billed at ~10%. Without it the badge only ever
+ // showed the FRESH tokens — a fraction of what a check really processes.
+ cache_read: number;
+ cache_write: number;
}
export interface ReviewerStats {
live: ReviewerBucket; // the mode actually deciding (Mode.AUTO_APPROVE)
diff --git a/surfaces/gui/src/components/Composer.tsx b/surfaces/gui/src/components/Composer.tsx
index f0230f0c..26e52da3 100644
--- a/surfaces/gui/src/components/Composer.tsx
+++ b/surfaces/gui/src/components/Composer.tsx
@@ -109,7 +109,7 @@ interface Props {
onUnattendedChange?: (on: boolean) => void;
// Auto-Approve metering (§1.7): the "Auto-Approve · N checks" badge + mode-menu summary.
// Absent/zero hides both; only the LIVE bucket surfaces here (shadow is a Settings concern).
- reviewerStats?: { checks: number; allow: number; deny: number; unsure: number; tokens_in: number; tokens_out: number } | null;
+ reviewerStats?: { checks: number; allow: number; deny: number; unsure: number; tokens_in: number; tokens_out: number; cache_read?: number; cache_write?: number } | null;
approvalSlot?: ReactNode;
// Push text + attachments into the composer (e.g. a start-panel task card). The `nonce` makes
// repeated identical prefills re-apply; the user can still edit before sending.
@@ -882,7 +882,7 @@ function ModeMenu({
onModeChange: (mode: string) => void;
unattended?: boolean;
onUnattendedChange?: (on: boolean) => void;
- reviewerStats?: { checks: number; allow: number; deny: number; unsure: number; tokens_in: number; tokens_out: number } | null;
+ reviewerStats?: { checks: number; allow: number; deny: number; unsure: number; tokens_in: number; tokens_out: number; cache_read?: number; cache_write?: number } | null;
}) {
const [open, setOpen] = useState(false);
// The Auto-Approve entry is gated on the server flag. Fetch once on first open; a session
@@ -962,7 +962,15 @@ function ModeMenu({
This session: {reviewerStats.checks} {reviewerStats.checks === 1 ? "check" : "checks"} ·{" "}
{reviewerStats.allow} cleared · {reviewerStats.deny} blocked · {reviewerStats.unsure} asked you
{reviewerStats.tokens_in + reviewerStats.tokens_out > 0 && (
- <> · ~{Math.round((reviewerStats.tokens_in + reviewerStats.tokens_out) / 100) / 10}k tokens</>
+ // Fresh and cached shown separately: cached input bills at ~10%, so
+ // one blended number would either overstate cost or hide most of the
+ // work. Two honest figures beat one misleading one.
+ <>
+ {" "}· ~{Math.round((reviewerStats.tokens_in + reviewerStats.tokens_out) / 100) / 10}k tokens
+ {(reviewerStats.cache_read ?? 0) > 0 && (
+ <> (+{Math.round((reviewerStats.cache_read ?? 0) / 100) / 10}k cached)</>
+ )}
+ </>
)}
</div>
</>
diff --git a/tests/test_auto_approve.py b/tests/test_auto_approve.py
index c838dca0..cb69bc29 100644
--- a/tests/test_auto_approve.py
+++ b/tests/test_auto_approve.py
@@ -9,7 +9,7 @@ from __future__ import annotations
import asyncio
import json
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
import pytest
@@ -942,3 +942,24 @@ def test_the_reviewer_is_never_asked_about_a_skill_save(tmp_path):
# The human decided, and not a single reviewer token was spent asking.
assert approvals == ["save_skill"]
assert engine.reviewer.asked == []
+
+
+def test_verdict_audit_rows_carry_the_verdicts_cache_figures(tmp_path):
+ # The Verdict's cache_read/cache_write must survive into the audit row —
+ # this is the exact seam where OPE-101's numbers were dropped one layer earlier.
+ engine, rows, _approvals = _engine(
+ tmp_path,
+ [_tool_turn(("run_shell", {"command": "pytest -q"})), AssistantTurn(text="ok", finish_reason="stop")],
+ )
+
+ class _CachedVerdictReviewer(_FakeReviewer):
+ async def review(self, **kw):
+ verdict = await super().review(**kw)
+ return replace(verdict, tokens_in=75, tokens_out=60, cache_read=1400, cache_write=25)
+
+ engine.reviewer = _CachedVerdictReviewer({"run_shell": "allow"})
+ _run(engine, "run the tests")
+
+ verdict_rows = [r for r in rows if r.get("stage") == "reviewer_verdict"]
+ assert verdict_rows and verdict_rows[0]["cache_read"] == 1400
+ assert verdict_rows[0]["cache_write"] == 25
diff --git a/tests/test_auto_approve_settings.py b/tests/test_auto_approve_settings.py
index f5dcd4cd..a7a37315 100644
--- a/tests/test_auto_approve_settings.py
+++ b/tests/test_auto_approve_settings.py
@@ -105,6 +105,7 @@ def test_reviewer_stats_aggregates_by_stage(tmp_path):
assert stats["live"] == {
"checks": 4, "allow": 2, "deny": 1, "unsure": 1,
"tokens_in": 380, "tokens_out": 90,
+ "cache_read": 0, "cache_write": 0,
}
assert stats["shadow"]["checks"] == 1 and stats["shadow"]["allow"] == 1
store.close()
@@ -146,3 +147,77 @@ def test_audit_migration_adds_columns_to_a_legacy_db(tmp_path):
def test_reviewer_stats_endpoint(client):
empty = client.get("/v1/sessions/nope/reviewer-stats").json()
assert empty["live"]["checks"] == 0 and empty["shadow"]["checks"] == 0
+
+
+def test_reviewer_stats_carry_the_cached_share(tmp_path):
+ # The badge could only ever see FRESH tokens (~75 of a ~1,500-token check once the
+ # provider caches the instruction prefix), so it under-reported cost by more the
+ # longer a session ran. The cached share now rides every verdict row into the sums.
+ from coworker.audit import AuditStore
+
+ store = AuditStore(tmp_path / "audit.db")
+ for _ in range(3):
+ store.append(
+ {
+ "session_id": "s1",
+ "tool": "run_shell",
+ "stage": "reviewer_verdict",
+ "status": "allow",
+ "tokens_in": 75,
+ "tokens_out": 60,
+ "cache_read": 1400,
+ "cache_write": 0,
+ }
+ )
+ live = store.reviewer_stats("s1")["live"]
+ assert live["checks"] == 3
+ assert (live["tokens_in"], live["tokens_out"]) == (225, 180)
+ assert (live["cache_read"], live["cache_write"]) == (4200, 0)
+
+
+def test_existing_databases_gain_the_cache_columns(tmp_path):
+ # A database created before 2026-08-22 has no cache columns. Opening it must migrate
+ # in place — old rows read as zero, new rows record the real figures.
+ import sqlite3
+
+ from coworker.audit import AuditStore
+
+ db = tmp_path / "audit.db"
+ con = sqlite3.connect(db)
+ con.execute(
+ """
+ CREATE TABLE audit_events (
+ id INTEGER PRIMARY KEY AUTOINCREMENT,
+ timestamp TEXT DEFAULT CURRENT_TIMESTAMP,
+ session_id TEXT, agent TEXT, workspace TEXT, connector TEXT,
+ tool TEXT, stage TEXT, status TEXT, approval TEXT, args TEXT,
+ result_preview TEXT, reason TEXT, resource TEXT, call_id TEXT,
+ tokens_in INTEGER DEFAULT 0, tokens_out INTEGER DEFAULT 0
+ )
+ """
+ )
+ con.execute(
+ "INSERT INTO audit_events (session_id, tool, stage, status, tokens_in, tokens_out)"
+ " VALUES ('s1', 'run_shell', 'reviewer_verdict', 'allow', 100, 50)"
+ )
+ con.commit()
+ con.close()
+
+ store = AuditStore(db) # migration happens on open
+ store.append(
+ {
+ "session_id": "s1",
+ "tool": "run_shell",
+ "stage": "reviewer_verdict",
+ "status": "allow",
+ "tokens_in": 75,
+ "tokens_out": 60,
+ "cache_read": 1400,
+ "cache_write": 25,
+ }
+ )
+ live = store.reviewer_stats("s1")["live"]
+ assert live["checks"] == 2
+ assert (live["tokens_in"], live["tokens_out"]) == (175, 110)
+ # The pre-migration row contributes zero cached, not garbage.
+ assert (live["cache_read"], live["cache_write"]) == (1400, 25)