Skip to content

Commit 15b3bef

Browse files
fix(codex): account for cache-write tokens (#1663)
* fix(codex): account for cache-write tokens Include Codex cache-write usage in costs, reports, and machine-readable output. Co-authored-by: KeiKawashima-HEROZ <[email protected]> * fix(codex): normalize cache usage deltas * test(codex): initialize cache-creation fixtures --------- Co-authored-by: KeiKawashima-HEROZ <[email protected]>
1 parent 809eeb6 commit 15b3bef

11 files changed

Lines changed: 753 additions & 100 deletions

File tree

‎rust/adapters/codex/src/aggregate.rs‎

Lines changed: 66 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -29,6 +29,7 @@ struct CodexEventKey {
2929
model_len: usize,
3030
input_tokens: u64,
3131
cached_input_tokens: u64,
32+
cache_creation_tokens: u64,
3233
output_tokens: u64,
3334
reasoning_output_tokens: u64,
3435
total_tokens: u64,
@@ -347,6 +348,7 @@ fn accumulate_codex_event_into_group(
347348
) {
348349
group.input_tokens += event.input_tokens;
349350
group.cached_input_tokens += event.cached_input_tokens;
351+
group.cache_creation_tokens += event.cache_creation_tokens;
350352
group.output_tokens += event.output_tokens;
351353
group.reasoning_output_tokens += event.reasoning_output_tokens;
352354
group.total_tokens += event.total_tokens;
@@ -361,6 +363,7 @@ fn accumulate_codex_event_into_group(
361363
let model_usage = group.models.entry(model.to_string()).or_default();
362364
model_usage.input_tokens += event.input_tokens;
363365
model_usage.cached_input_tokens += event.cached_input_tokens;
366+
model_usage.cache_creation_tokens += event.cache_creation_tokens;
364367
model_usage.output_tokens += event.output_tokens;
365368
model_usage.reasoning_output_tokens += event.reasoning_output_tokens;
366369
model_usage.total_tokens += event.total_tokens;
@@ -373,6 +376,7 @@ fn accumulate_codex_event_into_group(
373376
if is_long_context {
374377
model_usage.long_context_input_tokens += event.input_tokens;
375378
model_usage.long_context_cached_input_tokens += event.cached_input_tokens;
379+
model_usage.long_context_cache_creation_tokens += event.cache_creation_tokens;
376380
model_usage.long_context_output_tokens += event.output_tokens;
377381
}
378382
if record_service_tier {
@@ -395,20 +399,24 @@ fn accumulate_codex_event_into_usage_bucket(
395399
) {
396400
usage.input_tokens += event.input_tokens;
397401
usage.cached_input_tokens += event.cached_input_tokens;
402+
usage.cache_creation_tokens += event.cache_creation_tokens;
398403
usage.output_tokens += event.output_tokens;
399404
if is_long_context {
400405
usage.long_context_input_tokens += event.input_tokens;
401406
usage.long_context_cached_input_tokens += event.cached_input_tokens;
407+
usage.long_context_cache_creation_tokens += event.cache_creation_tokens;
402408
usage.long_context_output_tokens += event.output_tokens;
403409
}
404410
}
405411

406412
fn merge_codex_usage_bucket(target: &mut CodexUsageBucket, source: CodexUsageBucket) {
407413
target.input_tokens += source.input_tokens;
408414
target.cached_input_tokens += source.cached_input_tokens;
415+
target.cache_creation_tokens += source.cache_creation_tokens;
409416
target.output_tokens += source.output_tokens;
410417
target.long_context_input_tokens += source.long_context_input_tokens;
411418
target.long_context_cached_input_tokens += source.long_context_cached_input_tokens;
419+
target.long_context_cache_creation_tokens += source.long_context_cache_creation_tokens;
412420
target.long_context_output_tokens += source.long_context_output_tokens;
413421
}
414422

@@ -455,13 +463,19 @@ fn apply_recorded_usage_entries<'a>(
455463
let usage = CodexUsageBucket {
456464
input_tokens: key.input_tokens,
457465
cached_input_tokens: key.cached_input_tokens,
466+
cache_creation_tokens: key.cache_creation_tokens,
458467
output_tokens: key.output_tokens,
459468
long_context_input_tokens: if is_long_context { key.input_tokens } else { 0 },
460469
long_context_cached_input_tokens: if is_long_context {
461470
key.cached_input_tokens
462471
} else {
463472
0
464473
},
474+
long_context_cache_creation_tokens: if is_long_context {
475+
key.cache_creation_tokens
476+
} else {
477+
0
478+
},
465479
long_context_output_tokens: if is_long_context {
466480
key.output_tokens
467481
} else {
@@ -547,6 +561,7 @@ fn codex_event_key(
547561
model_len: model.len(),
548562
input_tokens: event.input_tokens,
549563
cached_input_tokens: event.cached_input_tokens,
564+
cache_creation_tokens: event.cache_creation_tokens,
550565
output_tokens: event.output_tokens,
551566
reasoning_output_tokens: event.reasoning_output_tokens,
552567
total_tokens: event.total_tokens,
@@ -564,6 +579,7 @@ fn merge_groups(target: &mut BTreeMap<String, CodexGroup>, source: BTreeMap<Stri
564579
let target_group = target.entry(period).or_default();
565580
target_group.input_tokens += group.input_tokens;
566581
target_group.cached_input_tokens += group.cached_input_tokens;
582+
target_group.cache_creation_tokens += group.cache_creation_tokens;
567583
target_group.output_tokens += group.output_tokens;
568584
target_group.reasoning_output_tokens += group.reasoning_output_tokens;
569585
target_group.total_tokens += group.total_tokens;
@@ -579,11 +595,14 @@ fn merge_groups(target: &mut BTreeMap<String, CodexGroup>, source: BTreeMap<Stri
579595
let target_usage = target_group.models.entry(model).or_default();
580596
target_usage.input_tokens += usage.input_tokens;
581597
target_usage.cached_input_tokens += usage.cached_input_tokens;
598+
target_usage.cache_creation_tokens += usage.cache_creation_tokens;
582599
target_usage.output_tokens += usage.output_tokens;
583600
target_usage.reasoning_output_tokens += usage.reasoning_output_tokens;
584601
target_usage.total_tokens += usage.total_tokens;
585602
target_usage.long_context_input_tokens += usage.long_context_input_tokens;
586603
target_usage.long_context_cached_input_tokens += usage.long_context_cached_input_tokens;
604+
target_usage.long_context_cache_creation_tokens +=
605+
usage.long_context_cache_creation_tokens;
587606
target_usage.long_context_output_tokens += usage.long_context_output_tokens;
588607
merge_codex_usage_bucket(
589608
&mut target_usage.recorded_standard_usage,
@@ -658,7 +677,7 @@ mod tests {
658677
use serde_json::json;
659678

660679
use crate::{
661-
PricingMap, cli::CodexSpeed, model_aliases::set_model_aliases_for_tests,
680+
CodexModelUsage, PricingMap, cli::CodexSpeed, model_aliases::set_model_aliases_for_tests,
662681
paths::CodexUsageSource,
663682
};
664683

@@ -965,6 +984,7 @@ mod tests {
965984
service_tier: &str,
966985
input_tokens: u64,
967986
cached_input_tokens: u64,
987+
cache_creation_tokens: u64,
968988
output_tokens: u64| {
969989
[
970990
json!({
@@ -986,6 +1006,7 @@ mod tests {
9861006
"last_token_usage": {
9871007
"input_tokens": input_tokens,
9881008
"cached_input_tokens": cached_input_tokens,
1009+
"cache_write_input_tokens": cache_creation_tokens,
9891010
"output_tokens": output_tokens,
9901011
"total_tokens": input_tokens + output_tokens,
9911012
},
@@ -1002,6 +1023,7 @@ mod tests {
10021023
"priority",
10031024
280_000,
10041025
20_000,
1026+
30_000,
10051027
500,
10061028
);
10071029
let standard_short = rollout(
@@ -1010,6 +1032,7 @@ mod tests {
10101032
"default",
10111033
100_000,
10121034
50_000,
1035+
10_000,
10131036
300,
10141037
);
10151038
let fixture = fs_fixture!({
@@ -1033,17 +1056,59 @@ mod tests {
10331056
let usage = &groups["2026-07-09"].models["gpt-5.6-sol"];
10341057

10351058
assert_eq!(usage.input_tokens, 380_000);
1059+
assert_eq!(usage.cache_creation_tokens, 40_000);
10361060
assert_eq!(usage.long_context_input_tokens, 280_000);
1061+
assert_eq!(usage.long_context_cache_creation_tokens, 30_000);
10371062
assert_eq!(usage.recorded_fast_usage.input_tokens, 280_000);
1063+
assert_eq!(usage.recorded_fast_usage.cache_creation_tokens, 30_000);
10381064
assert_eq!(usage.recorded_fast_usage.long_context_input_tokens, 280_000);
1065+
assert_eq!(
1066+
usage.recorded_fast_usage.long_context_cache_creation_tokens,
1067+
30_000
1068+
);
10391069
assert_eq!(usage.recorded_standard_usage.input_tokens, 100_000);
1070+
assert_eq!(usage.recorded_standard_usage.cache_creation_tokens, 10_000);
10401071
assert_eq!(usage.recorded_standard_usage.long_context_input_tokens, 0);
10411072
observed.push((usage.recorded_fast_usage, usage.recorded_standard_usage));
10421073
}
10431074

10441075
assert_eq!(observed[0], observed[1]);
10451076
}
10461077

1078+
#[test]
1079+
fn merges_cache_creation_usage_into_groups_and_recorded_buckets() {
1080+
let usage = CodexModelUsage {
1081+
cache_creation_tokens: 30,
1082+
long_context_cache_creation_tokens: 20,
1083+
recorded_standard_usage: CodexUsageBucket {
1084+
cache_creation_tokens: 30,
1085+
long_context_cache_creation_tokens: 20,
1086+
..CodexUsageBucket::default()
1087+
},
1088+
..CodexModelUsage::default()
1089+
};
1090+
let mut source_group = CodexGroup {
1091+
cache_creation_tokens: 30,
1092+
..CodexGroup::default()
1093+
};
1094+
source_group.models.insert("gpt-test".to_string(), usage);
1095+
let source = BTreeMap::from([("2026-07-09".to_string(), source_group)]);
1096+
let mut target = BTreeMap::new();
1097+
1098+
merge_groups(&mut target, source);
1099+
1100+
let merged = &target["2026-07-09"];
1101+
let merged_usage = &merged.models["gpt-test"];
1102+
assert_eq!(merged.cache_creation_tokens, 30);
1103+
assert_eq!(merged_usage.cache_creation_tokens, 30);
1104+
assert_eq!(
1105+
merged_usage
1106+
.recorded_standard_usage
1107+
.long_context_cache_creation_tokens,
1108+
20
1109+
);
1110+
}
1111+
10471112
#[test]
10481113
fn dedupes_copied_token_usage_after_model_alias_resolution() {
10491114
let _aliases = set_model_aliases_for_tests([("private-alpha", "gpt-5.2")]);

‎rust/adapters/codex/src/lib.rs‎

Lines changed: 81 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -69,6 +69,7 @@ mod tests {
6969
use crate::cli::SharedArgs;
7070
use crate::{CodexModelUsage, CodexServiceTier, CodexTokenUsageEvent, CodexUsageBucket};
7171
use ccusage_test_support::fs_fixture;
72+
use serde_json::json;
7273

7374
#[test]
7475
fn loads_directory_groups_with_date_filter_without_global_event_vector() {
@@ -132,6 +133,7 @@ mod tests {
132133
model: Some("gpt-5".to_string()),
133134
input_tokens: 100,
134135
cached_input_tokens: 90,
136+
cache_creation_tokens: 0,
135137
output_tokens: 5,
136138
reasoning_output_tokens: 0,
137139
total_tokens: 105,
@@ -161,6 +163,77 @@ mod tests {
161163
assert_eq!(report["daily"][0]["models"]["gpt-5"]["cacheReadTokens"], 90);
162164
}
163165

166+
#[test]
167+
fn reports_codex_cache_write_tokens_and_cost_for_gpt_5_6_terra() {
168+
let fixture = fs_fixture!({
169+
"session.jsonl": [
170+
json!({
171+
"timestamp": "2026-08-20T05:49:00.000Z",
172+
"type": "turn_context",
173+
"payload": { "model": "gpt-5.6-terra" },
174+
})
175+
.to_string(),
176+
json!({
177+
"timestamp": "2026-08-20T05:49:12.034Z",
178+
"type": "event_msg",
179+
"payload": {
180+
"type": "token_count",
181+
"info": {
182+
"last_token_usage": {
183+
"input_tokens": 935_040,
184+
"cached_input_tokens": 875_306,
185+
"cache_write_input_tokens": 57_610,
186+
"output_tokens": 11_150,
187+
"reasoning_output_tokens": 1_141,
188+
"total_tokens": 946_190,
189+
},
190+
"total_token_usage": {
191+
"input_tokens": 935_040,
192+
"cached_input_tokens": 875_306,
193+
"cache_write_input_tokens": 57_610,
194+
"output_tokens": 11_150,
195+
"reasoning_output_tokens": 1_141,
196+
"total_tokens": 946_190,
197+
},
198+
},
199+
},
200+
})
201+
.to_string(),
202+
]
203+
.join("\n"),
204+
});
205+
let shared = SharedArgs {
206+
single_thread: true,
207+
timezone: Some("UTC".to_string()),
208+
..SharedArgs::default()
209+
};
210+
211+
let groups =
212+
load_groups_from_directory(fixture.root(), &shared, AgentReportKind::Daily).unwrap();
213+
let report = report_from_groups(
214+
&groups,
215+
AgentReportKind::Daily,
216+
&PricingMap::load_embedded(),
217+
CodexSpeedPolicy::Forced(CodexServiceTier::Standard),
218+
);
219+
let daily = &report["daily"][0];
220+
let model = &daily["models"]["gpt-5.6-terra"];
221+
222+
assert_eq!(daily["inputTokens"], 2_124);
223+
assert_eq!(daily["cacheCreationTokens"], 57_610);
224+
assert_eq!(daily["cacheReadTokens"], 875_306);
225+
assert_eq!(daily["totalTokens"], 946_190);
226+
assert_eq!(model["inputTokens"], 2_124);
227+
assert_eq!(model["cacheCreationTokens"], 57_610);
228+
assert_eq!(model["cacheReadTokens"], 875_306);
229+
230+
let expected_cost =
231+
2_124.0 * 4e-6 + 875_306.0 * 0.4e-6 + 57_610.0 * 5e-6 + 11_150.0 * 18e-6;
232+
let actual_cost = daily["costUSD"].as_f64().unwrap();
233+
assert!((actual_cost - expected_cost).abs() < 1e-12);
234+
assert!((report["totals"]["costUSD"].as_f64().unwrap() - expected_cost).abs() < 1e-12);
235+
}
236+
164237
#[test]
165238
fn reports_codex_model_aliases_without_raw_model_names() {
166239
let _aliases = crate::model_aliases::set_model_aliases_for_tests([
@@ -176,6 +249,7 @@ mod tests {
176249
model: Some("private-codex-alpha".to_string()),
177250
input_tokens: 100,
178251
cached_input_tokens: 10,
252+
cache_creation_tokens: 0,
179253
output_tokens: 5,
180254
reasoning_output_tokens: 0,
181255
total_tokens: 105,
@@ -188,6 +262,7 @@ mod tests {
188262
model: Some("private-codex-beta".to_string()),
189263
input_tokens: 50,
190264
cached_input_tokens: 5,
265+
cache_creation_tokens: 0,
191266
output_tokens: 3,
192267
reasoning_output_tokens: 0,
193268
total_tokens: 53,
@@ -308,9 +383,11 @@ mod tests {
308383
recorded_fast_usage: CodexUsageBucket {
309384
input_tokens: 300_000,
310385
cached_input_tokens: 40_000,
386+
cache_creation_tokens: 0,
311387
output_tokens: 800,
312388
long_context_input_tokens: 300_000,
313389
long_context_cached_input_tokens: 40_000,
390+
long_context_cache_creation_tokens: 0,
314391
long_context_output_tokens: 800,
315392
},
316393
..CodexModelUsage::default()
@@ -596,6 +673,7 @@ mod tests {
596673
model: Some("gpt-5.3-codex".to_string()),
597674
input_tokens: 140,
598675
cached_input_tokens: 40,
676+
cache_creation_tokens: 0,
599677
output_tokens: 5,
600678
reasoning_output_tokens: 2,
601679
total_tokens: 147,
@@ -608,6 +686,7 @@ mod tests {
608686
model: Some("gpt-5.3-codex".to_string()),
609687
input_tokens: 70,
610688
cached_input_tokens: 70,
689+
cache_creation_tokens: 0,
611690
output_tokens: 10,
612691
reasoning_output_tokens: 0,
613692
total_tokens: 80,
@@ -620,6 +699,7 @@ mod tests {
620699
model: Some("gpt-5-mini".to_string()),
621700
input_tokens: 10,
622701
cached_input_tokens: 0,
702+
cache_creation_tokens: 0,
623703
output_tokens: 2,
624704
reasoning_output_tokens: 0,
625705
total_tokens: 12,
@@ -632,6 +712,7 @@ mod tests {
632712
model: None,
633713
input_tokens: 999,
634714
cached_input_tokens: 0,
715+
cache_creation_tokens: 0,
635716
output_tokens: 999,
636717
reasoning_output_tokens: 0,
637718
total_tokens: 1_998,

0 commit comments

Comments
 (0)