fix(session): break chunk runs on time gaps that cannot subtract exactly

Two safe-integer timestamps can differ by more than 2^53-1 (e.g.
MIN_SAFE_INTEGER to MAX_SAFE_INTEGER-1), so the dt subtraction rounds
and the packed row decodes to a timestamp one off the original --
violating the codec's lossless contract. Unreachable from a real clock
(the gap needs ~285k years) but reachable from hand-written fixtures,
and the decoder accepts hand-written rows.

continues() now refuses to extend a run across such a gap (the check is
exact both ways: an in-range true gap subtracts without rounding and
passes; an out-of-range one rounds to an out-of-range value and fails),
splitting the run instead -- whitelist philosophy, compression lost,
data never. The decoder tightens to match the encoder's image: time0
and dt must be safe integers, and reconstructed member seqs/times must
stay in safe range, so float reconstruction is exact wherever
validation passes.

The round-trip property now draws times from the full safe-integer
range (it previously generated only small gaps, which is how this
escaped); the bot's counterexample is pinned as an example test.

Found by ds-review-bot on #338.
This commit is contained in:
kingwl
2026-07-15 22:44:32 +08:00
parent f408d420da
commit 8f5c592b9f
2 changed files with 60 additions and 11 deletions

View File

@@ -135,6 +135,12 @@ function indexOf(event: DeltaEvent): number {
/** Whether `next` extends a run ending in `prev` (same kind already checked by the caller). */
function continues(prev: DeltaEvent, next: DeltaEvent, kind: DeltaKind): boolean {
if (next.seq !== prev.seq + 1) return false
// Two safe-integer times can sit further apart than a double subtracts
// exactly (2^53-1 and its negation differ by ~2^54); a rounded gap would
// decode to a different timestamp. The check is exact in both directions: a
// true gap within safe range subtracts without rounding and passes, while a
// true gap beyond it rounds to a value that is itself beyond and fails.
if (!Number.isSafeInteger(next.time - prev.time)) return false
if (next.data.turn !== prev.data.turn || next.data.step !== prev.data.step) return false
if (indexOf(next) !== indexOf(prev)) return false
if (kind !== 'tool-call-delta') return true
@@ -219,8 +225,8 @@ function malformed(tag: string, why: string): never {
throw new Error(`malformed ${tag} storage row: ${why}`)
}
/** Validate the shared run-data fields and the payload/dt arity. */
function validateRunData(tag: string, data: Record<string, unknown>, payloadKey: 'texts' | 'args'): void {
/** Validate the shared run-data fields and the payload/dt arity; returns the member payload. */
function validateRunData(tag: string, data: Record<string, unknown>, payloadKey: 'texts' | 'args'): string[] {
if (typeof data.turn !== 'number' || typeof data.step !== 'number' || typeof data.index !== 'number') {
malformed(tag, 'turn/step/index must be numbers')
}
@@ -229,12 +235,13 @@ function validateRunData(tag: string, data: Record<string, unknown>, payloadKey:
malformed(tag, `${payloadKey} must be a non-empty string array`)
}
const dt = data.dt
if (!Array.isArray(dt) || dt.some(gap => typeof gap !== 'number' || !Number.isFinite(gap))) {
malformed(tag, 'dt must be an array of finite numbers')
if (!Array.isArray(dt) || dt.some(gap => !Number.isSafeInteger(gap))) {
malformed(tag, 'dt must be an array of safe integers')
}
if (dt.length !== payload.length - 1) {
malformed(tag, `dt length ${dt.length} does not match ${payload.length} members`)
}
return payload as string[]
}
/** Validate a row-tagged parsed value's envelope and data, throwing on any malformation. */
@@ -245,11 +252,12 @@ function validateRow(value: Record<string, unknown>, tag: ChunkRow['type']): Chu
if (!Number.isSafeInteger(value.seq0) || (value.seq0 as number) < 0) {
malformed(tag, 'seq0 must be a non-negative safe integer')
}
if (typeof value.time0 !== 'number' || !Number.isFinite(value.time0)) {
malformed(tag, 'time0 must be a finite number')
if (!Number.isSafeInteger(value.time0)) {
malformed(tag, 'time0 must be a safe integer')
}
const data = value.data
if (!isRecord(data)) malformed(tag, 'data must be an object')
let payload: string[]
if (tag === 'tool-call-chunks') {
const withName = hasExactKeys(data, ['turn', 'step', 'index', 'id', 'name', 'dt', 'args'])
if (!withName && !hasExactKeys(data, ['turn', 'step', 'index', 'id', 'dt', 'args'])) {
@@ -258,12 +266,25 @@ function validateRow(value: Record<string, unknown>, tag: ChunkRow['type']): Chu
if (typeof data.id !== 'string' || (withName && typeof data.name !== 'string')) {
malformed(tag, 'id (and name when present) must be strings')
}
validateRunData(tag, data, 'args')
payload = validateRunData(tag, data, 'args')
} else {
if (!hasExactKeys(data, ['turn', 'step', 'index', 'dt', 'texts'])) {
malformed(tag, 'data must be exactly {turn, step, index, dt, texts}')
}
validateRunData(tag, data, 'texts')
payload = validateRunData(tag, data, 'texts')
}
// Reconstruction bounds. The encoder only packs runs whose member seqs and
// times are all safe integers, so a running value that leaves safe range is
// outside any encoder's image: float arithmetic would round it to a
// different number than exact arithmetic, a silent corruption. Within safe
// range every step is exact, so the first departure is always caught.
if (!Number.isSafeInteger((value.seq0 as number) + payload.length - 1)) {
malformed(tag, 'member seqs must stay safe integers')
}
let time = value.time0 as number
for (const gap of data.dt as number[]) {
time += gap
if (!Number.isSafeInteger(time)) malformed(tag, 'member times must stay safe integers')
}
return value as unknown as ChunkRow
}

View File

@@ -106,6 +106,22 @@ describe('packChunkRuns', () => {
expect(packChunkRuns(events)).toStrictEqual(events)
})
it('breaks a run on a time gap beyond safe-integer range (subtraction would round)', () => {
// Both endpoints are safe integers, but their true difference (~2^54)
// exceeds exact double range: b - a rounds, so a + (b - a) !== b and a
// packed row would decode to a different timestamp.
const a = Number.MIN_SAFE_INTEGER
const b = Number.MAX_SAFE_INTEGER - 1
expect(a + (b - a)).not.toBe(b) // the rounding this guard exists for
const events = [
chunkEvent(0, a, { type: 'text-delta', index: 0, text: 'x' }),
chunkEvent(1, b, { type: 'text-delta', index: 0, text: 'y' }),
chunkEvent(2, b + 1, { type: 'text-delta', index: 0, text: 'z' }),
]
expect(packChunkRuns(events)).toStrictEqual(events) // split at the gap; halves too short
expect(decodeAll(packChunkRuns(events))).toStrictEqual(events)
})
it('stores a delta with an off-whitelist data envelope verbatim (parsed-fixture shapes)', () => {
const mk = (seq: number, data: unknown): SessionEvent =>
({ type: 'assistant/chunk', seq, time: 1000, data } as SessionEvent)
@@ -144,11 +160,15 @@ describe('decodeStorageRecord', () => {
['an envelope with extra keys', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [], texts: ['a'] }, extra: 1 }],
['a negative seq0', { type: 'text-chunks', seq0: -1, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [], texts: ['a'] } }],
['a non-finite time0', { type: 'text-chunks', seq0: 0, time0: Infinity, data: { turn: 1, step: 1, index: 0, dt: [], texts: ['a'] } }],
['a fractional time0', { type: 'text-chunks', seq0: 0, time0: 1.5, data: { turn: 1, step: 1, index: 0, dt: [], texts: ['a'] } }],
['a data shape mismatch', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [], args: ['a'] } }],
['a non-string member', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [], texts: [7] } }],
['an empty member list', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [], texts: [] } }],
['a dt arity mismatch', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [1, 2], texts: ['a', 'b'] } }],
['a non-finite dt gap', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [NaN], texts: ['a', 'b'] } }],
['a fractional dt gap', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [0.5], texts: ['a', 'b'] } }],
['a member seq leaving safe range', { type: 'text-chunks', seq0: Number.MAX_SAFE_INTEGER, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [0, 0], texts: ['a', 'b', 'c'] } }],
['a member time leaving safe range', { type: 'text-chunks', seq0: 0, time0: Number.MAX_SAFE_INTEGER, data: { turn: 1, step: 1, index: 0, dt: [1], texts: ['a', 'b'] } }],
['a non-numeric turn', { type: 'text-chunks', seq0: 0, time0: 1, data: { turn: 'x', step: 1, index: 0, dt: [], texts: ['a'] } }],
['a tool-call row without id', { type: 'tool-call-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, dt: [], args: ['a'] } }],
['a tool-call row with non-string id', { type: 'tool-call-chunks', seq0: 0, time0: 1, data: { turn: 1, step: 1, index: 0, id: 7, dt: [], args: ['a'] } }],
@@ -183,11 +203,19 @@ const boundaryChunkArb: fc.Arbitrary<StreamChunk> = fc.oneof(
fc.record({ type: fc.constant<'finish'>('finish'), reason: fc.constant({ kind: 'stop' as const }) }),
)
/** Batches with contiguous seqs, arbitrary gaps in time, mixed chunk kinds and turn/step placement. */
/**
* Batches with contiguous seqs, arbitrary timestamps, mixed chunk kinds and
* turn/step placement. Times draw from the FULL safe-integer range (not just
* realistic clocks) so the property exercises the gap-overflow guard: two safe
* endpoints can differ by more than a double subtracts exactly.
*/
const batchArb: fc.Arbitrary<SessionEvent[]> = fc.array(
fc.record({
chunk: fc.oneof({ weight: 4, arbitrary: deltaChunkArb }, { weight: 1, arbitrary: boundaryChunkArb }),
gap: fc.integer({ min: -5, max: 200 }),
time: fc.oneof(
{ weight: 4, arbitrary: fc.integer({ min: 995, max: 9000 }) },
{ weight: 1, arbitrary: fc.integer({ min: Number.MIN_SAFE_INTEGER, max: Number.MAX_SAFE_INTEGER }) },
),
turn: fc.nat(1),
step: fc.nat(1),
}),
@@ -196,7 +224,7 @@ const batchArb: fc.Arbitrary<SessionEvent[]> = fc.array(
// plain objects real log events are (the log is JSON), so equality compares
// values, not prototypes.
).map(entries => JSON.parse(JSON.stringify(
entries.map((entry, k) => chunkEvent(k, 1000 + entry.gap * k, entry.chunk, entry.turn, entry.step)),
entries.map((entry, k) => chunkEvent(k, entry.time, entry.chunk, entry.turn, entry.step)),
)) as SessionEvent[])
describe('chunk-row codec properties', () => {