Skip to content

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

fix: a switch is not failed by the model it is switching away from

srafyfd1a78d70c87
diff --git a/ds-bot/src/host/parts.js b/ds-bot/src/host/parts.js
index 574540a2..cb5af8df 100644
--- a/ds-bot/src/host/parts.js
+++ b/ds-bot/src/host/parts.js
@@ -182,6 +182,9 @@ export function install(rt) {
const NOTE_COOLDOWN_MS = [60_000, 300_000, 900_000]
const SWITCH_ATTEMPTS = 5
const OVERFLOW = 'CONTEXT_WINDOW_EXCEEDED'
+ // A checkpoint DSH asked for so a model switch could go ahead, written while no route
+ // could produce a summary. It is honest about that and points at the messages.
+ const NO_SUMMARY = '(No summary was written: no model was free to write it, so the messages the earlier part sent you are carried over word for word below. Read it from them, and read_own_chat finds the rest.)'
// Generic request and credential errors say nothing about the conversation.
const STUCK_CODES = ['IMAGE_OFFLOAD_REQUIRED']
const ORIGINAL_COMPACT = Symbol.for('ds-bot/compaction')
@@ -266,6 +269,9 @@ export function install(rt) {
const adoptedEngines = new Set()
// Sessions DS Bot itself is condensing right now.
const condensing = new Set()
+ // The range of a Bot Session DSH is condensing for a model switch, where the route the
+ // Session used last is the one being switched away from.
+ const switchRanges = new Set()
const refuseCompaction = () => Promise.reject(new Error(
'A Bot conversation makes room on its own, and read_own_chat searches all of it.'))
const isBotTarget = (target) => {
@@ -276,13 +282,15 @@ export function install(rt) {
// Condenses a range of a Bot Session the way DS Bot condenses its own parts: the
// session counts as condensing, so the summary hook writes DS Bot's checkpoint rather
// than the backend writing one of its own.
- async function condenseRange(target, run) {
+ async function condenseRange(target, run, { forSwitch = false } = {}) {
const sessionId = target.session.id
condensing.add(sessionId)
+ if (forSwitch) switchRanges.add(sessionId)
try {
return await run()
} finally {
condensing.delete(sessionId)
+ if (forSwitch) switchRanges.delete(sessionId)
}
}
// The engine that serves this agent, when DS Bot can condense through it.
@@ -316,7 +324,7 @@ export function install(rt) {
// Without the summary hook the backend would summarize a Bot conversation itself,
// so that case keeps the refusal.
if (original.summarize === undefined) return refuseCompaction()
- return condenseRange(target, () => original.compactRegion.call(this, start, end, target, signal))
+ return condenseRange(target, () => original.compactRegion.call(this, start, end, target, signal), { forSwitch: true })
}
if (original.summarize !== undefined) {
engine.summarize = function summarizeForBots(input, agent, signal) {
@@ -540,25 +548,7 @@ export function install(rt) {
const head = input.messages
const tail = checkpointTail(head, session.deriveMessages())
const prompt = tail.length === 0 ? CHECKPOINT_PROMPT : `${CHECKPOINT_PROMPT}\n\n${checkpointTailNote(tail)}`
- const assembler = new BlockAssembler()
- for await (const chunk of ctx.llm.stream({
- provider: route.provider,
- model: route.model,
- ...(route.reasoningEffort ? { reasoningEffort: route.reasoningEffort } : {}),
- messages: [...head, ...tail, { role: 'user', content: [{ type: 'text', text: prompt }] }],
- toolHistory: session.toolHistory(),
- ...(input.tools ? { tools: [...input.tools] } : {}),
- ...(maxTokens !== undefined ? { maxTokens } : {}),
- sessionId: session.id,
- purpose: 'compaction',
- signal,
- })) assembler.push(chunk)
- const finish = assembler.finish
- if (finish.kind === 'error' || finish.kind === 'aborted') throw new Error(finish.failure?.message ?? finish.kind)
- if (finish.kind === 'max-tokens') throw new Error('the checkpoint was cut off at the output limit')
- const rawOutput = assembler.blocks()
- const summary = rawOutput.filter(block => block.type === 'text').map(block => block.text).join('\n').trim()
- if (summary === '') throw new Error('the checkpoint was empty')
+ const written = await summaryThrough(session, route, { input, head, tail, prompt, maxTokens, signal })
const place = placeOf(session.id)
const lines = await contextLines(session).catch(() => undefined)
const { kept, more } = await keptMessages(session.id, {
@@ -566,16 +556,69 @@ export function install(rt) {
group: place?.roomId !== undefined,
number: place === undefined ? 1 : partsOf(place.botId, place.roomId).length,
})
- const text = checkpointText({ number: checkpointCount(input.messages) + 1, summary, kept, more })
+ const text = checkpointText({
+ number: checkpointCount(input.messages) + 1,
+ summary: written?.summary ?? NO_SUMMARY,
+ kept,
+ more,
+ })
return {
summary: [{ type: 'text', text }],
- rawOutput,
- llmStreamCall: true,
- provider: route.provider,
- model: route.model,
+ rawOutput: written?.rawOutput ?? [{ type: 'text', text: NO_SUMMARY }],
+ llmStreamCall: written !== undefined,
+ provider: written?.provider ?? route.provider,
+ model: written?.model ?? route.model,
...(maxTokens !== undefined ? { maxTokens } : {}),
- ...(assembler.usage === undefined ? {} : { usage: assembler.usage }),
+ ...(written?.usage === undefined ? {} : { usage: written.usage }),
+ }
+ }
+
+ // The summary request itself. The Session's own route answers it, as it always has. A
+ // range DSH condenses for a model switch is different: the route the Session used last
+ // is the one being switched away from, and the switch can be because of that route (a
+ // model out of quota, for one). So the Bot's own model is tried next, and with no route
+ // left the checkpoint carries the messages without a summary rather than fail - a
+ // failure there is what DSH reports as MODEL_SWITCH_COMPACTION_FAILED, and it would
+ // leave the Bot on the model the user is trying to leave.
+ async function summaryThrough(session, route, { input, head, tail, prompt, maxTokens, signal }) {
+ const forSwitch = switchRanges.has(session.id)
+ const routes = [{ provider: route.provider, model: route.model, ...(route.reasoningEffort ? { reasoningEffort: route.reasoningEffort } : {}) }]
+ if (forSwitch) {
+ const bot = rt.selfOf(session.id) === undefined ? undefined : rt.botOf(rt.selfOf(session.id))
+ const ref = bot === undefined ? undefined : await Promise.resolve(rt.desiredModel?.(bot)).catch(() => undefined)
+ const wanted = splitRef(ref)
+ if (wanted !== undefined && routes.every(one => one.provider !== wanted.provider || one.model !== wanted.model)) routes.push(wanted)
}
+ let last
+ for (const one of routes) {
+ try {
+ const assembler = new BlockAssembler()
+ for await (const chunk of ctx.llm.stream({
+ ...one,
+ messages: [...head, ...tail, { role: 'user', content: [{ type: 'text', text: prompt }] }],
+ toolHistory: session.toolHistory(),
+ ...(input.tools ? { tools: [...input.tools] } : {}),
+ ...(maxTokens !== undefined ? { maxTokens } : {}),
+ sessionId: session.id,
+ purpose: 'compaction',
+ signal,
+ })) assembler.push(chunk)
+ const finish = assembler.finish
+ if (finish.kind === 'error' || finish.kind === 'aborted') throw new Error(finish.failure?.message ?? finish.kind)
+ if (finish.kind === 'max-tokens') throw new Error('the checkpoint was cut off at the output limit')
+ const rawOutput = assembler.blocks()
+ const summary = rawOutput.filter(block => block.type === 'text').map(block => block.text).join('\n').trim()
+ if (summary === '') throw new Error('the checkpoint was empty')
+ return { summary, rawOutput, provider: one.provider, model: one.model, usage: assembler.usage }
+ } catch (error) {
+ // A cancelled request is the caller's own decision, not a route that cannot answer.
+ if (!forSwitch || signal?.aborted) throw error
+ last = error
+ rt.warn('ds-bot: no checkpoint from %s/%s for %s: %s', one.provider, one.model, session.id, error?.message ?? error)
+ }
+ }
+ rt.warn('ds-bot: the checkpoint for %s carries no summary; the switch goes ahead without one: %s', session.id, last?.message ?? last)
+ return undefined
}
// The steward's call: the latest messages of this part, and nothing of the Bot's
diff --git a/ds-bot/test/model-switch-compaction.test.js b/ds-bot/test/model-switch-compaction.test.js
index 8851ece7..d806ab63 100644
--- a/ds-bot/test/model-switch-compaction.test.js
+++ b/ds-bot/test/model-switch-compaction.test.js
@@ -27,14 +27,19 @@ const TEAM = {
questions: {},
}
-function start({ summarize = true } = {}) {
+function start({ summarize = true, outlet = false, stream } = {}) {
const home = mkdtempSync(join(tmpdir(), 'ds-bot-compaction-'))
writeFileSync(join(home, 'state.json'), JSON.stringify(TEAM))
- const bench = { regions: [], nows: [], summarised: 0 }
+ const bench = { regions: [], nows: [], summarised: 0, checkpoints: [] }
const compaction = {
async compactIfNeeded() { return null },
async compactNow(target, signal, commandId) { bench.nows.push([target?.session?.id, commandId]) },
- async compactRegion(start, end, target) { bench.regions.push([start, end, target?.session?.id]) },
+ // `outlet` is compaction-basic's own shape: the range is condensed through the
+ // engine's summarize, which is the hook DS Bot wraps.
+ async compactRegion(start, end, agent, signal) {
+ bench.regions.push([start, end, agent?.session?.id])
+ if (outlet) bench.checkpoints.push(await this.summarize({ messages: [] }, agent, signal))
+ },
}
if (summarize) compaction.summarize = () => { bench.summarised += 1; return { summary: [] } }
const ctx = fakeContext({
@@ -45,11 +50,17 @@ function start({ summarize = true } = {}) {
listProviders: () => [{ id: 'deepseek', name: 'DeepSeek' }],
listModels: async () => [{ id: 'v4' }],
resolveModelInfo: async () => ({ context: { contextWindow: 100_000 }, defaultMaxTokens: 8_000 }),
- stream() {
+ stream(options) {
+ if (stream !== undefined) return stream(options)
throw new Error('the checkpoint model is not called by these tests')
},
},
sessions: { messageProjections: [] },
+ // The checkpoint carries the latest messages of the part word for word.
+ sessionController: {
+ resolveAgent: async sessionId => ({ agent: { id: sessionId, session: { id: sessionId, seq: 2 } } }),
+ page: async () => ({ records: [], hasMore: false }),
+ },
workspaceRegistry: { archiveSession: async () => {} },
})
const rt = createRuntime(ctx, { home })
@@ -111,3 +122,42 @@ test('an explicit /compact on a Bot conversation is still refused', async (t) =>
)
assert.deepEqual(bench.nows, [])
})
+
+test('a checkpoint the route cannot write does not fail the switch', async (t) => {
+ // The reason for switching can be the route itself: a model out of quota answers every
+ // request with 429, the summary included. The range must still condense.
+ const asked = []
+ const { bench, compaction, adoptFor, agentOf, stop } = start({
+ outlet: true,
+ stream: options => {
+ asked.push(`${options.provider}/${options.model}`)
+ throw new Error('429: rate limited')
+ },
+ })
+ t.after(stop)
+ adoptFor('b2')
+ await compaction.compactRegion(3, 9, agentOf('b2'), new AbortController().signal)
+ assert.deepEqual(bench.regions, [[3, 9, 'b2']])
+ assert.deepEqual(asked, ['deepseek/v4'])
+ assert.equal(bench.checkpoints.length, 1)
+ assert.match(bench.checkpoints[0].summary[0].text, /^\[Checkpoint 1\]/)
+ assert.match(bench.checkpoints[0].summary[0].text, /No summary was written/)
+})
+
+test('a checkpoint the route cannot write still fails outside a switch', async (t) => {
+ // The steward's own condense path keeps its old behaviour: a model that cannot answer
+ // makes the part move on instead of leaving a checkpoint without a summary.
+ const { compaction, adoptFor, agentOf, stop } = start({
+ outlet: true,
+ stream: () => {
+ throw new Error('429: rate limited')
+ },
+ })
+ t.after(stop)
+ adoptFor('b2')
+ const session = agentOf('b2')
+ await assert.rejects(
+ () => compaction.compactNow(session, new AbortController().signal, 'command-1'),
+ undefined,
+ )
+})