dev/mock-llm.mjs: 长任务 N with K omitted emits no text block per step — .map(Number) turns the missing K capture into NaN, defeating the size = 4 default #55

What happened?

A 长任务 N turn in the dev mock model (ds-bot/dev/mock-llm.mjs) with the optional K omitted — documented default 4 — does not grow the context after the opener step. Every step after the opener returns no text block at all: 0 characters of 步骤N filler instead of the documented ~4000. There is no error to copy; the per-step filler is silently missing, so the multi-step context-growth scenario this command exists for does nothing.

• What I did: started the mock model with node dev/mock-llm.mjs and drove its /v1/messages endpoint through a 长任务 3 turn (K omitted), answering each list_bots tool call with an empty result.
• What I expected: steps 1 and 2 each add about 4000 characters of filler. dev/mock-llm.mjs:19 documents 长任务 N [K] (N steps of K×1000, default 4), and the branch comment at dev/mock-llm.mjs:159 repeats "(default 4)".
• What happened instead: the opener (步骤0) correctly emits 4000 characters, but every later step emits no text block; only the closing 长任务做完了，一共 3 步。 line arrives. With K given explicitly (长任务 3 4) the same steps work and each emits 4000 characters.

Actual output, verbatim from two independent runs of the reproduction script below:

opener step0 text length: 4000 (expected ~4000, 步骤0 filler)
step1 text: null
step1 text length: 0 (documented: ~4000 chars of 步骤1 filler; default K = 4)
control (K given as 4): step1 text length: 4000

=== K omitted ("长任务 3", documented default K=4) ===
[长任务 3] step 1: textBlocks=0 chars=0 toolCalls=list_bots
[长任务 3] step 2: textBlocks=0 chars=0 toolCalls=list_bots
[长任务 3] step 3: textBlocks=1 chars=14 firstText="长任务做完了，一共 3 步。" toolCalls=none
---
K omitted total text chars: 4014 (documented: 3 steps x ~4000 + finisher)
K=4 explicit total text chars: 12014
VERDICT: K-omitted turn is missing its per-step filler

(The 4014 total is just the 4000-character opener plus the 14-character finisher; the two missing per-step fillers would have added about 8000 more.)

Steps to reproduce

From a checkout of this repository at commit bdc5fda64e2ea1e83a4d146440072a2413025df4, save and run this script with Node (node repo-longtask-e2e.mjs). It starts the mock model itself on port 8791 and talks to its /v1/messages endpoint; no DSH install and no API key are involved.

// repo-longtask-e2e.mjs — "长任务 3" with K omitted must grow the context ~4k
// characters per step; instead every step after the opener returns no text block.
import { spawn } from 'node:child_process'

const PORT = 8791
const child = spawn('node', ['dev/mock-llm.mjs'], {
  cwd: 'ds-bot', // the plugin folder inside this repository
  env: { ...process.env, MOCK_PORT: String(PORT), MOCK_DELAY_MS: '10' },
  stdio: ['ignore', 'pipe', 'pipe'],
})
await new Promise(resolve => child.stdout.on('data', resolve))

async function call(messages) {
  const response = await fetch(`http://127.0.0.1:${PORT}/v1/messages`, {
    method: 'POST',
    headers: { 'content-type': 'application/json' },
    body: JSON.stringify({ model: 'mock-1', stream: true, max_tokens: 8192, messages }),
  })
  const text = await response.text()
  const events = [...text.matchAll(/data: (.+)\n/g)].map(m => JSON.parse(m[1]))
  const blocks = []
  for (const event of events) {
    if (event.type === 'content_block_start') blocks.push({ ...event.content_block, index: event.index })
    if (event.type === 'content_block_delta' && event.delta.type === 'text_delta') blocks[event.index].text = (blocks[event.index].text ?? '') + event.delta.text
    if (event.type === 'content_block_delta' && event.delta.type === 'input_json_delta') blocks[event.index].input = JSON.parse(event.delta.partial_json)
    if (event.type === 'message_delta') blocks.stop_reason = event.delta.stop_reason
  }
  return blocks
}

// 1. opener turn
const first = await call([{ role: 'user', content: '长任务 3' }])
const step0 = first.find(b => b.type === 'text')
const toolUse = first.find(b => b.type === 'tool_use')
console.log('opener step0 text length:', step0.text.length, '(expected ~4000, 步骤0 filler)')

// 2. first tool result -> should be 步骤1 filler (~4000 chars), actually is empty
const second = await call([
  { role: 'user', content: '长任务 3' },
  { role: 'assistant', content: [step0, { type: 'tool_use', id: toolUse.id ?? 'call-1-1', name: 'list_bots', input: toolUse.input ?? {} }] },
  { role: 'user', content: [{ type: 'tool_result', tool_use_id: toolUse.id ?? 'call-1-1', content: '[]' }] },
])
const step1 = second.find(b => b.type === 'text')
console.log('step1 text:', JSON.stringify(step1?.text ?? null))
console.log('step1 text length:', step1?.text?.length ?? 0, '(documented: ~4000 chars of 步骤1 filler; default K = 4)')

// 3. control: with K given explicitly, the same step works
const third = await call([
  { role: 'user', content: '长任务 3 4' },
  { role: 'assistant', content: [{ type: 'text', text: 'x' }, { type: 'tool_use', id: 'call-2-1', name: 'list_bots', input: {} }] },
  { role: 'user', content: [{ type: 'tool_result', tool_use_id: 'call-2-1', content: '[]' }] },
])
const step1k = third.find(b => b.type === 'text')
console.log('control (K given as 4): step1 text length:', step1k.text.length)

child.kill()
process.exit(0)

It prints the first output block above: opener 4000, step 1 missing (null, 0 chars), control 4000.

Diagnostics

• Repository commit: bdc5fda64e2ea1e83a4d146440072a2413025df4 ("Let every Bot update its own settings with update_bot"), main branch.
• Node.js v22.23.2, macOS 26.6.2 (arm64).
• Settings → Feedback diagnostics do not apply here: the defect is in the dev mock model itself (ds-bot/dev/mock-llm.mjs), reached by running it directly; no DSH install, no real model, no API key, and no error text — the failure is silent.

How did you install DS Bot?

Git address (a development checkout of this repository itself, for working on the plugin).

Privacy

I checked that this report has no API keys, tokens or personal data.

Root cause

ds-bot/dev/mock-llm.mjs:162, in the tool-result branch of decide():

const [, steps, size = 4] = /^长任务\s*(\d+)(?:\s+(\d+))?/.exec(textOf(messages[openerAt].content)).map(Number)

For 长任务 3, .exec() returns ['长任务 3', '3', undefined], and .map(Number) turns the omitted K capture into NaN. A destructuring default only replaces undefined, not NaN, so size = 4 never applies. Then filler(NaN * 1000, ...) (dev/mock-llm.mjs:78) builds Array.from({ length: Math.ceil(NaN / 20) }) — an empty array — and returns '', and the SSE writer skips empty text blocks entirely (if (block.text === '') continue, dev/mock-llm.mjs:541), so the step emits only the tool call and no text.

Expected behaviour, per the file's own documentation (dev/mock-llm.mjs:19 and the comment at dev/mock-llm.mjs:159): the omitted K falls back to 4, and the sibling opener branch at dev/mock-llm.mjs:341 already parses the same regex correctly with Number(order[2] ?? 4) — which is why the opener step works with K omitted.

I have a small fix ready that mirrors that same ?? 4 pattern (keep the match object, convert each capture separately), and am happy to be assigned to it.

Related issues

I searched the open and closed issues and pull requests for 长任务, mock-llm and NaN, and found no existing report of this.