Skip to content

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

Fall back to the default model for memory panel calls

Fei2411fd0d7519e262
diff --git a/CHANGELOG.md b/CHANGELOG.md
index be0739b5..f7906d69 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,10 @@ All notable changes to DS Bot are listed here. The format follows [Keep a Change
## [Unreleased]
+### Fixed
+
+- The memory panel's summary, its "ask or update" and the memory review no longer stop when the Bot's model refuses the call or runs out of output budget: they try once more on the profile's default model. After you change the Bot's model, a failed summary is written again automatically. ([#116](https://github.com/Fei2411/DeepSeek-Bot/issues/116))
+
## [0.1.4] - 2026-10-13
### Added
diff --git a/ds-bot/src/host/memory.js b/ds-bot/src/host/memory.js
index aee6322d..4e4cf91f 100644
--- a/ds-bot/src/host/memory.js
+++ b/ds-bot/src/host/memory.js
@@ -936,6 +936,25 @@ export function install(rt) {
throw new Error(`The model's reasoning used up the output budget (${SUMMARY_MAX_TOKENS} and ${SUMMARY_RETRY_MAX_TOKENS} tokens); no complete answer was written`)
}
+ // A served model can still refuse the call (a key without access, say), and
+ // desiredModel only falls back for models that are not served; so a failed call
+ // is tried once more on the profile default.
+ async function botShot(bot, system, text, tag) {
+ const own = await modelFor(bot)
+ try {
+ return await oneShot(own, system, text, tag)
+ } catch (error) {
+ const fallback = splitRef(await rt.defaultModel())
+ if (fallback === undefined || (fallback.provider === own.provider && fallback.model === own.model)) throw error
+ try {
+ return await oneShot(fallback, system, text, tag)
+ } catch (again) {
+ const said = failure => failure?.message ?? String(failure)
+ throw new Error(`${own.provider}/${own.model}: ${said(error)}; ${fallback.provider}/${fallback.model}: ${said(again)}`)
+ }
+ }
+ }
+
/** @type {Map<string, {hash:string, text:string, at:number}>} */
const summaries = new Map()
const summarizing = new Map()
@@ -955,21 +974,23 @@ export function install(rt) {
const linesFor = files => allEntries(files).map(entry => `- ${entry.file.topic ? `[[${entry.file.topic}]] ` : ''}${entry.text}`)
function startSummary(bot, own, team, language, hash) {
+ const model = bot.model
const job = (async () => {
- const text = await oneShot(await modelFor(bot), SUMMARY_SYSTEM, summaryRequest({ name: bot.name, own: linesFor(own), team: linesFor(team), language }), { sessionId: rt.chatOf(bot.id), kind: 'memory' })
+ const text = await botShot(bot, SUMMARY_SYSTEM, summaryRequest({ name: bot.name, own: linesFor(own), team: linesFor(team), language }), { sessionId: rt.chatOf(bot.id), kind: 'memory' })
const value = { hash, text: text.slice(0, SUMMARY_TEXT_MAX), at: Date.now() }
summaries.set(bot.id, value)
summaryFailures.delete(bot.id)
if (rt.readOnly() === undefined) await writeText(summaryPath(bot.id), JSON.stringify(value, null, 2))
})().catch((error) => {
- summaryFailures.set(bot.id, { hash, message: error?.message ?? String(error) })
+ summaryFailures.set(bot.id, { hash, model, message: error?.message ?? String(error) })
rt.warn('ds-bot: memory summary for %s failed: %s', bot.name, error)
}).finally(() => summarizing.delete(bot.id))
summarizing.set(bot.id, job)
}
// The summary is written again when the entries or the interface language change;
- // a failed one waits for `regenerate` instead of retrying on every poll.
+ // a failed one waits for `regenerate`, or a new model for the Bot, instead of
+ // retrying on every poll.
function memoryView(body) {
const bot = rt.botOf(String(body.botId ?? ''))
if (bot === undefined) throw new Error('Unknown Bot')
@@ -979,8 +1000,9 @@ export function install(rt) {
const language = languageNow()
const hash = createHash('sha256').update(JSON.stringify([language, linesFor(own), linesFor(team)])).digest('hex')
const saved = savedSummary(bot.id)
- const failed = summaryFailures.get(bot.id)
- const wanted = body.regenerate === true || (body.summarize === true && saved?.hash !== hash && failed?.hash !== hash)
+ const prior = summaryFailures.get(bot.id)
+ const failed = prior?.hash === hash && prior.model === bot.model ? prior : undefined
+ const wanted = body.regenerate === true || (body.summarize === true && saved?.hash !== hash && failed === undefined)
if (wanted && entries.length > 0 && !summarizing.has(bot.id)) startSummary(bot, own, team, language, hash)
return {
botId: bot.id,
@@ -990,7 +1012,7 @@ export function install(rt) {
summary: entries.length > 0 && saved !== undefined ? { text: saved.text, at: saved.at } : null,
fresh: entries.length === 0 || saved?.hash === hash,
summarizing: summarizing.has(bot.id),
- error: failed?.hash === hash && !summarizing.has(bot.id) ? failed.message : null,
+ error: failed !== undefined && !summarizing.has(bot.id) ? failed.message : null,
}
}
@@ -1003,7 +1025,7 @@ export function install(rt) {
...allEntries(loadScope(ownScope(bot.id))).map(entry => ({ ...entry, scope: 'own' })),
...allEntries(loadScope(TEAM)).map(entry => ({ ...entry, scope: 'team' })),
]
- const answer = parseAsk(await oneShot(await modelFor(bot), ASK_SYSTEM, askRequest({ name: bot.name, numbered, language: languageNow(), text }), { sessionId: rt.chatOf(bot.id), kind: 'memory' }))
+ const answer = parseAsk(await botShot(bot, ASK_SYSTEM, askRequest({ name: bot.name, numbered, language: languageNow(), text }), { sessionId: rt.chatOf(bot.id), kind: 'memory' }))
const meta = { source: 'memory panel', by: 'user', added: today() }
const changes = []
for (const change of answer.changes.slice(0, ASK_CHANGES_MAX)) {
@@ -1236,7 +1258,7 @@ export function install(rt) {
const sizes = { own: ownFiles[0].content.length, team: teamFiles[0].content.length }
const earlier = batch.earlier.map(line => reviewLine(line.who, line.text, bot))
const language = writtenIn(read.filter(line => line.who === 'user').map(line => line.text.replace(/^User: /, '')), languageNow())
- const reply = await oneShot(await modelFor(bot), reviewSystem(writesTeam(bot.id)), reviewRequest({ bot, numbered, left, sizes, where: batch.where, earlier, fresh: fresh.texts, more, language }), { sessionId: rt.chatOf(bot.id), kind: 'memory' })
+ const reply = await botShot(bot, reviewSystem(writesTeam(bot.id)), reviewRequest({ bot, numbered, left, sizes, where: batch.where, earlier, fresh: fresh.texts, more, language }), { sessionId: rt.chatOf(bot.id), kind: 'memory' })
const changes = parseReview(reply)
if (changes === undefined) throw new Error(`the model's answer is not a list of changes: ${clip(reply.replace(/\s+/g, ' '), 120)}`)
const unknown = changes.find(change => change.op !== 'add' && numbered[Number(change.entry) - 1] === undefined)
diff --git a/ds-bot/test/memory.test.js b/ds-bot/test/memory.test.js
index d49c01b6..82273662 100644
--- a/ds-bot/test/memory.test.js
+++ b/ds-bot/test/memory.test.js
@@ -120,14 +120,15 @@ const ownChat = [
// The model answers each call with the next of `replies`. `history` holds the events
// of a Session, oldest first; each own chat has one user message, at seq 10, unless
// it says otherwise.
-function start({ home = mkdtempSync(join(tmpdir(), 'ds-bot-memory-')), replies = [], config = {}, history = {}, reasoning = { efforts: [{ id: 'off' }, { id: 'high' }] } } = {}) {
+function start({ home = mkdtempSync(join(tmpdir(), 'ds-bot-memory-')), replies = [], config = {}, history = {}, reasoning = { efforts: [{ id: 'off' }, { id: 'high' }] }, models = ['v4'], defaultModel } = {}) {
if (!existsSync(join(home, 'state.json'))) writeFileSync(join(home, 'state.json'), JSON.stringify(TEAM))
- const ctx = fakeContext()
+ const services = defaultModel === undefined ? {} : { agentDefaultModel: { currentSelection: () => ({ provider: 'deepseek', model: defaultModel }) } }
+ const ctx = fakeContext({ services })
const requests = []
Object.assign(ctx, {
llm: {
listProviders: () => [{ id: 'deepseek', name: 'DeepSeek' }],
- listModels: async () => [{ id: 'v4' }],
+ listModels: async () => models.map(id => ({ id })),
resolveModelInfo: async () => ({ reasoning }),
async* stream(request) {
requests.push(request)
@@ -135,7 +136,7 @@ function start({ home = mkdtempSync(join(tmpdir(), 'ds-bot-memory-')), replies =
if (reply instanceof Error) throw reply
const out = typeof reply === 'object' && reply !== null ? reply : { text: reply ?? '' }
yield { type: 'text-delta', index: 0, text: out.text ?? '' }
- yield { type: 'finish', reason: { kind: out.finish ?? 'stop' } }
+ yield { type: 'finish', reason: { kind: out.finish ?? 'stop', ...(out.failure ? { failure: { message: out.failure } } : {}) } }
},
},
sessionController: {
@@ -406,6 +407,65 @@ test('the panel call asks for the model\'s lowest thinking level when it has no
assert.equal(second.requests[0].reasoningEffort, undefined)
})
+const routes = requests => requests.map(request => `${request.provider}/${request.model}@${request.maxTokens}`)
+
+test('a summary the Bot\'s model refuses is written by the profile default model', async (t) => {
+ const replies = [{ text: '', finish: 'error', failure: '403 Forbidden' }, '## Overview\nYou like tea.']
+ const { tool, call, requests, stop } = start({ replies, models: ['v4', 'flash'], defaultModel: 'flash' })
+ t.after(() => stop())
+ await tool('remember', { text: 'Likes tea' }, 'b2')
+ await call('memory-view', { botId: 'b2', summarize: true })
+ await settle()
+ const done = (await call('memory-view', { botId: 'b2' })).value
+ assert.deepEqual([done.summary?.text, done.fresh, done.error], ['## Overview\nYou like tea.', true, null])
+ assert.deepEqual(routes(requests), ['deepseek/v4@2048', 'deepseek/flash@2048'])
+ assert.ok(requests.every(request => request.messages[0].content[0].text.includes('Likes tea')))
+})
+
+test('a summary cut off at both budgets on the Bot\'s model is written by the default model', async (t) => {
+ const replies = [{ text: '', finish: 'max-tokens' }, { text: '## half', finish: 'max-tokens' }, '## Overview\nYou like tea.']
+ const { tool, call, requests, stop } = start({ replies, models: ['v4', 'flash'], defaultModel: 'flash' })
+ t.after(() => stop())
+ await tool('remember', { text: 'Likes tea' }, 'b2')
+ await call('memory-view', { botId: 'b2', summarize: true })
+ await settle()
+ const done = (await call('memory-view', { botId: 'b2' })).value
+ assert.deepEqual([done.summary?.text, done.error], ['## Overview\nYou like tea.', null])
+ assert.deepEqual(routes(requests), ['deepseek/v4@2048', 'deepseek/v4@8192', 'deepseek/flash@2048'])
+})
+
+test('a summary that failed on both models is tried again once the Bot has another model', async (t) => {
+ const replies = [new Error('401: Unauthorized'), { text: '', finish: 'error', failure: '503 Service Unavailable' }]
+ const { tool, call, requests, stop } = start({ replies, models: ['v4', 'flash'], defaultModel: 'flash' })
+ t.after(() => stop())
+ await tool('remember', { text: 'Likes tea' }, 'b2')
+ await call('memory-view', { botId: 'b2', summarize: true })
+ await settle()
+ const failed = (await call('memory-view', { botId: 'b2', summarize: true })).value
+ assert.equal(failed.error, 'deepseek/v4: 401: Unauthorized; deepseek/flash: 503 Service Unavailable')
+ assert.equal(failed.summarizing, false)
+ assert.equal(requests.length, 2)
+
+ assert.equal((await call('update-bot', { id: 'b2', model: 'deepseek/flash' })).ok, true)
+ replies.push('## Overview\nYou like tea.')
+ const retried = (await call('memory-view', { botId: 'b2', summarize: true })).value
+ assert.deepEqual([retried.summarizing, retried.error], [true, null])
+ await settle()
+ const done = (await call('memory-view', { botId: 'b2' })).value
+ assert.deepEqual([done.summary?.text, done.error], ['## Overview\nYou like tea.', null])
+ assert.deepEqual(routes(requests).slice(2), ['deepseek/flash@2048'])
+})
+
+test('"ask or update" falls back to the default model when the Bot\'s model fails', async (t) => {
+ const replies = [new Error('403 Forbidden'), JSON.stringify({ reply: 'You like tea.', changes: [] })]
+ const { tool, call, requests, stop } = start({ replies, models: ['v4', 'flash'], defaultModel: 'flash' })
+ t.after(() => stop())
+ await tool('remember', { text: 'Likes tea' }, 'b2')
+ const answer = (await call('memory-ask', { botId: 'b2', text: 'What do I like?' })).value
+ assert.equal(answer.reply, 'You like tea.')
+ assert.deepEqual(routes(requests), ['deepseek/v4@2048', 'deepseek/flash@2048'])
+})
+
test('"ask or update" answers from the entries and changes them through the same rules', async (t) => {
const replies = []
const { tool, call, read, journal, requests, stop } = start({ replies })