deepseek-harness/packages/llm/token-meter/src/index.ts

340 lines
13 KiB
TypeScript

/**
* Single replay-aware token-meter service for request and surface pressure.
*
* @module @deepseek-ai/dsh-token-meter
*/
import { Context, Service } from '@deepseek-ai/cordis'
import z from '@deepseek-ai/schemastery'
import { BlockAssembler } from '@deepseek-ai/dsh-llm'
import type { LlmImageRequestPricing, Message, TokenUsage } from '@deepseek-ai/dsh-llm'
import { deepFreeze } from '@deepseek-ai/dsh-util-values'
import type {
EpochHeader,
Session,
SessionEvent,
SessionLogOffset as SessionLogOffsetType,
SessionSeq as SessionSeqType,
} from '@deepseek-ai/dsh-session'
import { canonicalHeader, headerEquals, isSurfaceEvent, SessionLogOffset, SessionSeq } from '@deepseek-ai/dsh-session'
// Type-only: activates the `ctx.sessionProjections` Context declaration.
import type {} from '@deepseek-ai/dsh-session-projection'
import type {
TokenMeasurement,
TokenMeasurementBaseline,
TokenMeterConfig,
} from './types.ts'
import { contextBreakdownProjectionDefinition } from './breakdown-projection.ts'
import { contextPressureProjectionDefinition, tokenUsageProjectionDefinition } from './usage-projection.ts'
import { estimateContent, estimateHeader, estimateMessage, ROLE_OVERHEAD } from './estimate.ts'
import { commitSurfaceTokens, planSurfaceTokens } from './surface-fold.ts'
import type { MeterSurfaceNode } from './surface-fold.ts'
import { priceSurface } from './route-pricing.ts'
export type * from './types.ts'
// Module-edge re-export: forces the emitted index.d.ts to import the
// projection-unit modules, so their SessionProjectionStateMap augmentations load
// in aggregate programs that only import the package root.
export type * from './usage-projection.ts'
export type * from './breakdown-projection.ts'
/**
* Raw anchor facts captured at the latest successful call; the baseline is
* derived per measurement so the anchored surface reprices under the same
* route pricing as the current surface it is compared with.
*/
interface MeasurementAnchor {
readonly header: EpochHeader | undefined
/** Surface snapshot the anchored request was derived from. */
readonly nodes: readonly MeterSurfaceNode[]
/** Fixed-heuristic price of the call's provider output. */
readonly assistantTokens: number
/** Provider usage of the call, when it reported one under a known header. */
readonly usage: TokenUsage | undefined
}
interface ReplayState {
consumedEvents: SessionLogOffsetType
header: EpochHeader | undefined
surface: MeterSurfaceNode[]
stepStart: { turn: number; step: number; nodes: readonly MeterSurfaceNode[] } | undefined
anchor: MeasurementAnchor | undefined
}
/** Sum disjoint provider usage buckets without double-counting reasoning output. */
function usageTokens(usage: TokenUsage): number {
return usage.inputTokens
+ (usage.cacheReadTokens ?? 0)
+ (usage.cacheWriteTokens ?? 0)
+ usage.outputTokens
}
/** Compare optional envelopes so a headerless estimate can track later surface deltas. */
function optionalHeaderEquals(
left: EpochHeader | undefined,
right: EpochHeader | undefined,
): boolean {
if (left === undefined || right === undefined) return left === right
return headerEquals(left, right)
}
/** Reject stale or misspelled keys before defaults can hide them. */
function validateConfigKeys(config: TokenMeterConfig): void {
for (const key of Object.keys(config)) {
throw new Error(`TokenMeterConfig: unknown key "${key}" (no settings are supported)`)
}
}
declare module '@deepseek-ai/cordis' {
interface Context {
tokenMeter: TokenMeter
}
}
/** Replay owner for one service-wide estimator and isolated per-session folds. */
export class TokenMeter extends Service {
// Schemastery preserves untrusted loader keys on an empty object schema;
// the public type excludes settings while validateConfigKeys rejects them.
static Config: z<TokenMeterConfig> = z.object({}) as unknown as z<TokenMeterConfig>
static inject = ['sessionProjections']
private readonly states = new WeakMap<Session, ReplayState>()
constructor(ctx: Context, config: TokenMeterConfig = {}) {
super(ctx, 'tokenMeter')
validateConfigKeys(config)
ctx.sessionProjections.register(tokenUsageProjectionDefinition)
ctx.sessionProjections.register(contextPressureProjectionDefinition)
ctx.sessionProjections.register(contextBreakdownProjectionDefinition)
// Readers catch up independently, while eager observation bounds ordinary
// read latency without creating state for sessions no consumer has read.
ctx.on('session/event', (session) => {
if (this.states.has(session)) this._sync(session)
})
}
/**
* Measure current request pressure and surface through the durable tail.
*
* The effective envelope's routed provider/model selects the request-image
* pricing every node is priced under: a route whose adapter declares image
* pricing charges each retained image its visual tokens plus its
* model-visible text, while other routes keep the fixed heuristic. Provider
* usage is reused only when the latest successful call's canonical request
* envelope matches `requestHeader` and its total is no lower than that
* call's full route-priced anchor; otherwise the complete envelope and
* surface are repriced.
*
* `requestHeader` replaces the latest logged envelope for pressure and node
* pricing; the node set always describes the current session surface. Every
* call clones those positional nodes, so measurement is O(surface).
*
* @param session - session to replay through its current durable tail.
* @param requestHeader - optional effective request envelope replacing the latest logged header.
* @returns a detached deeply immutable pressure and surface measurement.
*/
measure(session: Session, requestHeader?: EpochHeader): TokenMeasurement {
const state = this._sync(session)
const header = requestHeader === undefined
? state.header
: canonicalHeader(requestHeader)
const pricing = this._routeImagePricing(header)
const surface = priceSurface(state.surface, pricing)
const anchor = state.anchor
let baseline: TokenMeasurementBaseline
let surfaceDeltaTokens: number
if (anchor !== undefined && optionalHeaderEquals(anchor.header, header)) {
// Matching headers share one route, so the anchored snapshot reprices
// under the same pricing as the current surface and the signed delta
// compares like with like.
const anchorSurfaceTokens = priceSurface(anchor.nodes, pricing).surfaceTokens
+ anchor.assistantTokens
const estimatedAnchorTokens = estimateHeader(header) + anchorSurfaceTokens
const usage = anchor.usage
// Signed heuristic deltas remain conservative only from an anchor
// that is at least as large as the matching full heuristic price.
baseline = usage !== undefined && usageTokens(usage) >= estimatedAnchorTokens
? { kind: 'usage', tokens: usageTokens(usage), usage }
: { kind: 'estimated', tokens: estimatedAnchorTokens }
surfaceDeltaTokens = surface.surfaceTokens - anchorSurfaceTokens
} else if (header === undefined && surface.surfaceTokens === 0) {
baseline = { kind: 'none', tokens: 0 }
surfaceDeltaTokens = 0
} else {
baseline = {
kind: 'estimated',
tokens: estimateHeader(header) + surface.surfaceTokens,
}
surfaceDeltaTokens = 0
}
return deepFreeze(structuredClone({
logRevision: state.consumedEvents,
baseline,
surfaceDeltaTokens,
totalTokens: Math.max(0, baseline.tokens + surfaceDeltaTokens),
surfaceTokens: surface.surfaceTokens,
nodes: surface.nodes,
}))
}
/** Resolve the routed model's image pricing, when the llm service and route declare one. */
private _routeImagePricing(header: EpochHeader | undefined): LlmImageRequestPricing | undefined {
const config = header?.config
if (config === undefined) return undefined
return this.ctx.get('llm')?.imageRequestPricing(config.provider, config.model)
}
/**
* Heuristically price one model-visible message (instance face of the pure
* `estimateMessage` export from `estimate.ts`).
* @param message - message to price without mutation.
* @returns content and role-framing tokens under the fixed service heuristic.
*/
estimateMessage(message: Message): number {
return estimateMessage(message)
}
/** Catch one session's fold up to the current durable tail. */
private _sync(session: Session): ReplayState {
let state = this.states.get(session)
if (state === undefined) {
state = {
consumedEvents: SessionLogOffset(0),
header: undefined,
surface: [],
stepStart: undefined,
anchor: undefined,
}
this.states.set(session, state)
}
while (state.consumedEvents < session.seq) {
// oxlint-disable-next-line typescript/no-non-null-assertion -- contiguous session seqs index the durable log
const event = session.eventAt(SessionSeq(state.consumedEvents))!
this._foldEvent(session, state, event)
state.consumedEvents = SessionLogOffset(state.consumedEvents + 1)
}
return state
}
/**
* Run every fallible step — surface plan and anchor validation — before
* mutating replay state, so a malformed event remains unread on every
* retry instead of half-applying.
*/
private _foldEvent(session: Session, state: ReplayState, event: SessionEvent): void {
let nextHeader = state.header
let nextStepStart = state.stepStart
let nextAnchor = state.anchor
switch (event.type) {
case 'request/header':
nextHeader = canonicalHeader(event.data.header)
break
case 'step/start':
if (state.stepStart !== undefined) {
throw new Error(
`token meter: step/start at seq ${event.seq} arrived before turn ${state.stepStart.turn}/step ${state.stepStart.step} ended`,
)
}
nextStepStart = { ...event.data, nodes: [...state.surface] }
break
case 'step/end':
if (state.stepStart === undefined
|| state.stepStart.turn !== event.data.turn
|| state.stepStart.step !== event.data.step) {
throw new Error(`token meter: step/end at seq ${event.seq} has no matching step/start event`)
}
nextStepStart = undefined
break
default:
break
}
const plan = isSurfaceEvent(event)
? planSurfaceTokens(state.surface, event)
: undefined
if (event.type === 'assistant/message') {
const stepStart = state.stepStart
if (stepStart === undefined
|| stepStart.turn !== event.data.turn
|| stepStart.step !== event.data.step) {
throw new Error(`token meter: assistant/message at seq ${event.seq} has no matching step/start event`)
}
// assistant/message is surface-mandatory at every append/seed boundary.
// oxlint-disable-next-line typescript/no-non-null-assertion
const eventTokens = plan!.tokens
if (event.data.usage !== undefined && nextHeader !== undefined) {
nextAnchor = {
header: nextHeader,
nodes: stepStart.nodes,
assistantTokens: this._estimateProviderAssistant(session, event, eventTokens),
usage: event.data.usage,
}
} else {
nextAnchor = {
header: nextHeader,
nodes: stepStart.nodes,
assistantTokens: eventTokens,
usage: undefined,
}
}
}
state.header = nextHeader
state.stepStart = nextStepStart
if (plan !== undefined) {
commitSurfaceTokens(state.surface, plan)
}
state.anchor = nextAnchor
}
/**
* Reassemble provider output from the exact cited chunk seqs for a usage anchor.
* Missing legacy source seqs conservatively treat the durable output as the
* provider output; an explicit empty list prices a known empty stream.
*/
private _estimateProviderAssistant(
session: Session,
event: SessionEvent<'assistant/message'>,
durableEventTokens: number,
): number {
const sourceSeqs = event.sourceEventSeqs
if (sourceSeqs === undefined) return durableEventTokens
const assembler = new BlockAssembler()
const seen = new Set<SessionSeqType>()
for (const seq of sourceSeqs) {
if (seq >= event.seq) {
throw new Error(`token meter: assistant/message at seq ${event.seq} source seq ${seq} is not earlier`)
}
if (seen.has(seq)) {
throw new Error(`token meter: assistant/message at seq ${event.seq} repeats source seq ${seq}`)
}
seen.add(seq)
// Session construction validates contiguous seqs, and the explicit
// earlier-than-assistant check above therefore guarantees existence.
const source = session.eventAt(seq)
// oxlint-disable-next-line typescript/no-non-null-assertion
const sourceEvent = source!
if (sourceEvent.type !== 'assistant/chunk') {
throw new Error(`token meter: assistant/message at seq ${event.seq} source seq ${seq} is not assistant/chunk`)
}
if (sourceEvent.data.turn !== event.data.turn || sourceEvent.data.step !== event.data.step) {
throw new Error(`token meter: assistant/message at seq ${event.seq} source seq ${seq} belongs to another step`)
}
assembler.push(sourceEvent.data.chunk)
}
const providerContent = assembler.blocks()
return providerContent.length === 0 ? 0 : estimateContent(providerContent) + ROLE_OVERHEAD
}
}
export default TokenMeter