Skip to content

Commit adfc924

Browse files
authored
fix(server): preserve usage in oversized transcript records (pingdotgg#13650)
1 parent 38f3c62 commit adfc924

6 files changed

Lines changed: 636 additions & 37 deletions

File tree

Lines changed: 60 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,60 @@
1+
import { describe, expect, it } from "@effect/vitest";
2+
3+
import { createTranscriptJsonReader, TranscriptJsonLimitError } from "./AgentSessionJson.ts";
4+
5+
function read(
6+
json: string,
7+
size: number,
8+
select: (path: ReadonlyArray<string | number | null>) => boolean = () => true,
9+
) {
10+
const reader = createTranscriptJsonReader(() => {}, select);
11+
for (let offset = 0; offset < json.length; offset += size)
12+
reader.write(json.slice(offset, offset + size));
13+
return reader.finish();
14+
}
15+
16+
describe("transcript JSON projection", () => {
17+
it.each([1, 2, 7, 64, 1024])("matches JSON.parse across %i-character boundaries", (size) => {
18+
for (const text of [
19+
'{"a":1,"a":2,"b":"before","b":"after"}',
20+
'{"a":{"x":1},"a":{"y":2},"b":[],"c":{}}',
21+
'{"a":[null,true,false,1,-2.3e4,"😀\\u0061\\\\\\\"",{},[],[1,2]]}',
22+
'{"a":"s","a":null,"b":null,"b":"s","c":0,"c":false}',
23+
'{"__proto__":{"polluted":true},"constructor":1,"__proto__":2}',
24+
])
25+
expect(read(text, size)).toEqual(JSON.parse(text));
26+
});
27+
28+
it("projects siblings and array elements without merging repeated parent objects", () => {
29+
const text =
30+
'{"message":{"usage":{"input":100},"content":"large"},"message":{"usage":{"output":5},"content":[1,2]},"rows":[{"keep":1,"drop":2},{"keep":3}],"drop":{"keep":4}}';
31+
const projected = read(text, 1, (path) => {
32+
if (path[0] === "drop") return false;
33+
return !path.includes("content") && !path.includes("drop");
34+
});
35+
expect(projected).toEqual({
36+
message: { usage: { output: 5 } },
37+
rows: [{ keep: 1 }, { keep: 3 }],
38+
});
39+
});
40+
41+
it.each(['{"a":', '{"a":1} trailing', '{"a":1}{"a":2}', '{"a":"bad\\x"}', '{"a":[1,]}'])(
42+
"rejects malformed input %s",
43+
(text) => {
44+
expect(read(text, 1)).toBeUndefined();
45+
},
46+
);
47+
48+
it("retains the import allocation and depth limits", () => {
49+
const limited = createTranscriptJsonReader(
50+
() => {
51+
throw new TranscriptJsonLimitError("budget");
52+
},
53+
() => true,
54+
);
55+
expect(() => limited.write('{"a":1}')).toThrow(TranscriptJsonLimitError);
56+
expect(() => read("[".repeat(129) + "0" + "]".repeat(129), 10)).toThrow(
57+
TranscriptJsonLimitError,
58+
);
59+
});
60+
});

‎apps/server/src/project/AgentSessionJson.ts‎

Lines changed: 59 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,6 @@
11
import * as SchemaAST from "effect/SchemaAST";
2-
import { isMany, none, type Many } from "stream-chain/defs.js";
2+
import { none, type Many } from "stream-chain/defs.js";
33
import { Assembler } from "stream-json/core/assembler.js";
4-
import { filter } from "stream-json/core/filters/filter.js";
54
import * as StreamJson from "stream-json/core/parser.js";
65
import type { ParserOptions, Token } from "stream-json/core/parser.js";
76

@@ -55,6 +54,7 @@ export class TranscriptJsonLimitError extends Error {}
5554
export function createTranscriptJsonReader(
5655
reserve: (bytes: number) => void,
5756
selectPath: (path: JsonPath) => boolean,
57+
options?: { readonly maxDepth?: number },
5858
) {
5959
// The synchronous tokenizer is exported at runtime in 3.6.0, but omitted
6060
// from its bundled types. Unlike parser(), it does not wrap tokens in an
@@ -65,9 +65,6 @@ export function createTranscriptJsonReader(
6565
) => (input: string | typeof none) => Many<Token> | typeof none;
6666
};
6767
const tokenize = jsonParser({ packValues: false });
68-
const select = filter({ filter: selectPath, streamKeys: false }) as (
69-
input: Token | typeof none,
70-
) => Token | Many<Token> | typeof none;
7168
const assembler = new Assembler();
7269
let key: string | null = null;
7370
let value = "";
@@ -100,13 +97,62 @@ export function createTranscriptJsonReader(
10097
assembler.consume(token);
10198
}
10299
};
100+
// Forward actual selected keys instead of reconstructing them from path
101+
// changes: adjacent duplicate keys have the same path but JSON.parse keeps
102+
// the last value. Reconstructing paths can silently retain the first value.
103+
const stack: Array<{ path: JsonPath; key: string | number | null; selected: boolean }> = [];
104+
let selectedValue = false;
105+
const startValue = () => {
106+
const parent = stack.at(-1);
107+
const path = parent?.selected ? [...parent.path, parent.key] : [];
108+
const selected = (parent?.selected ?? true) && selectPath(path);
109+
if (selected && typeof parent?.key === "string") {
110+
assemble({ name: "keyValue", value: parent.key });
111+
}
112+
return { path, selected };
113+
};
114+
const endValue = () => {
115+
const parent = stack.at(-1);
116+
if (parent && typeof parent.key === "number") parent.key++;
117+
};
103118
const selectToken = (token: Token | typeof none) => {
104-
const selected = select(token);
105-
if (selected === none) return;
106-
if (isMany(selected)) {
107-
for (const item of selected.values) assemble(item);
108-
} else {
109-
assemble(selected);
119+
if (token === none) return;
120+
switch (token.name) {
121+
case "keyValue": {
122+
const parent = stack.at(-1);
123+
if (parent) parent.key = token.value;
124+
return;
125+
}
126+
case "startObject":
127+
case "startArray": {
128+
const frame = startValue();
129+
stack.push({ ...frame, key: token.name === "startArray" ? 0 : null });
130+
if (frame.selected) assemble(token);
131+
return;
132+
}
133+
case "endObject":
134+
case "endArray":
135+
if (stack.pop()?.selected) assemble(token);
136+
endValue();
137+
return;
138+
case "startString":
139+
case "startNumber":
140+
selectedValue = startValue().selected;
141+
if (selectedValue) assemble(token);
142+
return;
143+
case "endString":
144+
case "endNumber":
145+
if (selectedValue) assemble(token);
146+
endValue();
147+
return;
148+
case "nullValue":
149+
case "trueValue":
150+
case "falseValue":
151+
if (startValue().selected) assemble(token);
152+
endValue();
153+
return;
154+
default:
155+
if (selectedValue) assemble(token);
110156
}
111157
};
112158
const consume = (input: string | typeof none) => {
@@ -116,8 +162,8 @@ export function createTranscriptJsonReader(
116162
if (tokens === none) return;
117163
for (const token of tokens.values) {
118164
if (token.name === "startObject" || token.name === "startArray") {
119-
if (++depth > 128)
120-
throw new TranscriptJsonLimitError("Transcript JSON nesting exceeds 128 levels");
165+
if (++depth > (options?.maxDepth ?? 128))
166+
throw new TranscriptJsonLimitError("Transcript JSON nesting exceeds the depth limit");
121167
} else if (token.name === "endObject" || token.name === "endArray") {
122168
if (--depth === 0) complete = true;
123169
}

‎apps/server/src/usage/UsageService.test.ts‎

Lines changed: 53 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -645,6 +645,59 @@ describe("UsageService", () => {
645645
}).pipe(Effect.scoped),
646646
);
647647

648+
it.live(
649+
"keeps large-record totals and costs exact through append, dedupe, restart and cleanup",
650+
() =>
651+
Effect.gen(function* () {
652+
const { transcript, settings, home } = yield* setup;
653+
const large = claudeLine(1, 9900).replace(
654+
'"message":',
655+
'"padding":' + encodeUnknownJsonString("x".repeat(9 * 1024 * 1024)) + ',"message":',
656+
);
657+
yield* Effect.promise(() => NodeFSP.writeFile(transcript, large));
658+
yield* Effect.gen(function* () {
659+
const service = yield* UsageService.make;
660+
const first = yield* service.readSummary(WINDOW);
661+
assert.strictEqual(totalOutputTokens(first), 9900);
662+
assert.closeTo(
663+
first.buckets.reduce((sum, bucket) => sum + bucket.costUsd, 0),
664+
0.4951,
665+
1e-12,
666+
);
667+
const warm = yield* service.readSummary(WINDOW);
668+
assert.deepStrictEqual(warm.buckets, first.buckets);
669+
// The repeated content block has the same message/request identity.
670+
yield* Effect.promise(() => NodeFSP.appendFile(transcript, large + claudeLine(2, 100)));
671+
const appended = yield* service.readSummary(WINDOW);
672+
assert.strictEqual(totalOutputTokens(appended), 10000);
673+
assert.strictEqual(
674+
appended.buckets.reduce((sum, bucket) => sum + bucket.totals.uncachedInputTokens, 0),
675+
20,
676+
);
677+
const restarted = yield* UsageService.make;
678+
const restored = yield* restarted.readSummary(WINDOW);
679+
assert.deepStrictEqual(restored.buckets, appended.buckets);
680+
yield* Effect.promise(() => NodeFSP.rm(transcript));
681+
const afterCleanup = yield* UsageService.make;
682+
assert.deepStrictEqual(
683+
(yield* afterCleanup.readSummary(WINDOW)).buckets,
684+
appended.buckets,
685+
);
686+
}).pipe(
687+
Effect.provide(
688+
serviceLayers({
689+
prefix: "usage-service-large-record-test",
690+
home,
691+
settings,
692+
ratesDocument: {
693+
"claude-fable-5": { input_cost_per_token: 1e-5, output_cost_per_token: 5e-5 },
694+
},
695+
}),
696+
),
697+
);
698+
}).pipe(Effect.scoped),
699+
);
700+
648701
it.live("preserves saved tokens, costs and sessions after transcript cleanup and restart", () =>
649702
Effect.gen(function* () {
650703
const { transcript, settings, home } = yield* setup;

0 commit comments

Comments
 (0)