Skip to content

A team of long-lived bots for DeepSeek Harness, with main bots, bot-to-bot messages and group chats.

Price replaced image history with the measured route

Fei2411a931b662bd42
diff --git a/ds-bot/src/host/parts.js b/ds-bot/src/host/parts.js
index 7a109ba3..fa6d15f3 100644
--- a/ds-bot/src/host/parts.js
+++ b/ds-bot/src/host/parts.js
@@ -192,9 +192,9 @@ export function install(rt) {
// A broken provider adapter must not crash every turn; estimate from the
// derived messages instead, with one warning.
let measureWarned = false
- const tokensOf = (session) => {
+ const measureOf = (session) => {
try {
- return ctx.get('tokenMeter')?.measure(session).totalTokens
+ return ctx.get('tokenMeter')?.measure(session)
} catch (error) {
if (!measureWarned) {
measureWarned = true
@@ -202,11 +202,12 @@ export function install(rt) {
}
}
try {
- return estimateTokens(session.deriveMessages().map(message => textOf(message.content)).join('\n'))
+ return { totalTokens: estimateTokens(session.deriveMessages().map(message => textOf(message.content)).join('\n')) }
} catch {
return undefined
}
}
+ const tokensOf = session => measureOf(session)?.totalTokens
const clock = createClock(config.timeZone)
// Replacing a handoff must shrink the request, not just create another Session.
async function canMakeRoom(agent, carry = []) {
@@ -229,10 +230,14 @@ export function install(rt) {
})
const meter = ctx.get('tokenMeter')
const size = message => meter?.estimateMessage(message) ?? estimateTokens(textOf(message.content))
- const history = messages.filter(message => message.role !== 'system').reduce((total, message) => total + size(message), 0)
+ const measured = measureOf(session)
+ // Images in the replaced history must use the same route pricing as the total.
+ const history = measured?.nodes === undefined
+ ? messages.filter(message => message.role !== 'system').reduce((total, message) => total + size(message), 0)
+ : measured.nodes.filter(node => session.eventAt(node.seq).type !== 'system/message').reduce((total, node) => total + node.tokens, 0)
const pending = carry.reduce((total, message) => total + size(message), 0)
const replacement = size({ role: 'user', content: [{ type: 'text', text }] }) + pending
- const tokens = tokensOf(session)
+ const tokens = measured?.totalTokens
// The system prompt and tool definitions stay; pending input is forwarded whole.
return replacement < history + pending
&& (lines === undefined || tokens === undefined || tokens - history + replacement <= lines.usable)
diff --git a/ds-bot/test/bot-messages.test.js b/ds-bot/test/bot-messages.test.js
index d57fc5d3..5bacc2d7 100644
--- a/ds-bot/test/bot-messages.test.js
+++ b/ds-bot/test/bot-messages.test.js
@@ -4,6 +4,9 @@ import { tmpdir } from 'node:os'
import { join } from 'node:path'
import { setTimeout as delay } from 'node:timers/promises'
import { test } from 'node:test'
+import { Context } from '@deepseek-ai/cordis'
+import { Session } from '@deepseek-ai/dsh-session'
+import TokenMeter from '@deepseek-ai/dsh-token-meter'
import { apply } from '../index.js'
import { fakeAttachments, fakeContext } from './fake-ctx.js'
import { install as installTurns } from '../src/host/turns.js'
@@ -954,6 +957,75 @@ test('a failed turn past the hard line keeps the real error instead of switching
})
for (const id of ['b', 'gb']) {
+ test(`model-priced images after a handoff can make room by switching (${id})`, async (t) => {
+ const seed = team({ group: true })
+ const segment = { sessionId: 'old', note: 'Keep working on the request.', endedAt: 1 }
+ if (id === 'gb') seed.rooms.r1.segments = { b: [segment] }
+ else seed.bots.b.segments = [segment]
+ const { ctx, agentOf, emit, fire, stop } = await start(seed)
+ t.after(stop)
+ const meterCtx = new Context()
+ meterCtx.provide('sessionProjections', { register() {} })
+ meterCtx.provide('llm', {
+ imageRequestPricing: () => ({ priceImages: images => images.map(() => ({ visualTokens: 2_000, text: '' })) }),
+ })
+ await meterCtx.plugin(TokenMeter)
+ t.after(() => meterCtx.fiber.dispose())
+ const meter = meterCtx.tokenMeter
+ const get = ctx.get
+ ctx.get = name => name === 'tokenMeter' ? meter : get(name)
+ ctx.llm.resolveModelInfo = async () => ({ context: { contextWindow: 50_000 } })
+ let made = 0
+ ctx.sessionController.create = async () => ({ sessionId: `s${++made}` })
+ const agent = agentOf(id)
+ const session = Session.create(id)
+ agent.session = session
+ const events = []
+ const record = (type, data, surface) => {
+ events.push(session.append(type, data, surface ? { surfaceOp: 'append' } : undefined))
+ emit(id, type, data)
+ }
+ ctx.sessionController.page = async ({ address, beforeSeq }) => ({
+ records: address.sessionId === id ? events.filter(event => beforeSeq === undefined || event.seq < beforeSeq).map(event => ({ event })) : [],
+ hasMore: false,
+ })
+ const header = { config: { provider: 'custom', model: 'image-model' } }
+ const system = { role: 'system', content: textBlock('System instructions. '.padEnd(36_000, 's')) }
+ record('request/header', { header })
+ record('system/message', { turn: 1, step: 1, message: system }, true)
+ record('user/message', { source: { kind: SOURCE_KIND, handoff: true }, content: textBlock('Earlier context. '.padEnd(12_000, 'x')) }, true)
+ record('step/start', { turn: 1, step: 1 })
+ record('assistant/message', { turn: 1, step: 1, ...assistantText('Ready to work with the images.'), stream: [] }, true)
+ record('step/end', { turn: 1, step: 1 })
+ record('user/message', {
+ source: { kind: 'user' },
+ content: [
+ ...textBlock('Inspect these images.'),
+ ...Array.from({ length: 30 }, () => ({ type: 'image', attachment: { attachmentId: `sha256:${'1'.repeat(64)}`, mediaType: 'image/png', bytes: 100, width: 1600, height: 1600 } })),
+ ],
+ }, true)
+ const before = meter.measure(session)
+ const history = before.nodes.filter(node => session.eventAt(node.seq).type !== 'system/message').reduce((total, node) => total + node.tokens, 0)
+ const heuristic = session.deriveMessages().filter(message => message.role !== 'system').reduce((total, message) => total + meter.estimateMessage(message), 0)
+ assert.ok(before.totalTokens > 50_000)
+ assert.ok(history > 60_000)
+ assert.ok(before.totalTokens - heuristic > 50_000, 'structural image references cannot stand in for the removed visual tokens')
+
+ await fire('agent/request-error', { agent, failure: { code: 'CONTEXT_WINDOW_EXCEEDED' }, signal: AbortSignal.timeout(60_000) })
+ for (let count = 0; count < 200 && !agentOf('s1').calls.some(([, message]) => message.source?.handoff); count += 1) await delay(5)
+ const handoff = agentOf('s1').calls.find(([, message]) => message.source?.handoff)?.[1]
+ assert.ok(handoff, 'the shorter part can replace the model-priced images')
+ assert.equal(made, 1)
+ const replacement = Session.create('replacement')
+ replacement.append('request/header', { header })
+ replacement.append('system/message', { turn: 1, step: 1, message: system }, { surfaceOp: 'append' })
+ replacement.append('user/message', handoff, { surfaceOp: 'append' })
+ const after = meter.measure(replacement).totalTokens
+ assert.ok(after < 50_000, 'the same system prompt and the real handoff fit')
+ assert.ok(after < before.totalTokens)
+ t.diagnostic(`pressure ${before.totalTokens} -> ${after}; replaced history ${history} (heuristic ${heuristic}); new parts ${made}`)
+ })
+
test(`a fresh part with oversized tools stops until the replacement context can fit (${id})`, async (t) => {
const { ctx, agentOf, emit, fire, stop } = await start(team({ group: true }))
t.after(stop)