MaxNow Last-30

Maintain rolling external AI signal context for the private MaxNow workstation.

This skill updates the Last-30 data files only. It does not update page code, product documents, or the main dashboard data unless explicitly instructed by the owner. Last-30 is for external AI signals, not MaxNow internal project logs.

Hard Boundary

Routine Last-30 runs may update only:

dash/data/last-30.json
dash/data/last-30.js

Routine Last-30 runs must not edit:

dash/index.html
dash/styles.css
dash/app.js
SPEC.md
README.md
DEPLOY.md
CONTEXT.md
IDEAS.md
UPDATE_LOG.md
dash/data/dashboard.json
dash/data/dashboard.js
dash/data/ai-news.json
dash/data/ai-news.js

If the data shape, page structure, or product boundary needs to change, stop and report it. Do not modify code or documentation during routine runs.

Goal

Keep a concise rolling Chinese frontier-news memory of:

• the latest 1-3 concrete AI releases
• this week's additional, non-duplicate frontier changes
• recent 30-day concrete milestones, not keyword buckets
• potential impact on MaxNow, Codex, OpenClaw, model choices, or token cost
• watch items and uncertain signals
• judgments that still need owner confirmation

Prefer free public sources first: official blogs/RSS, GitHub releases, Hacker News, Reddit/public community sources, arXiv, GDELT, or similar free indexes. X/Twitter is not a hard dependency; do not use paid X API unless the owner explicitly approves budget and account list.

Do not feed large full-text news dumps into the model. Use scripts or source snippets to prefilter candidates, then summarize only a small candidate set.

Data Files

dash/data/last-30.json is the source of truth. dash/data/last-30.js must contain the same object assigned to window.MAXNOW_LAST30_DATA.

Data Shape

Preserve this shape when possible:

{
  "updatedAt": "YYYY-MM-DD HH:mm",
  "sourceSummary": "Last-30 rolling context",
  "today": {
    "date": "YYYY-MM-DD",
    "title": "最新发布",
    "summary": "一句话摘要。",
    "items": [
      {
        "date": "YYYY-MM-DD",
        "title": "短标题",
        "summary": "事实或低风险摘要。",
        "source": "Owner / GitHub / OpenClaw / Token / Server",
        "confidence": "high",
        "needsOwnerConfirm": false,
        "url": ""
      }
    ]
  },
  "week": {
    "range": "YYYY-MM-DD/YYYY-MM-DD",
    "title": "本周前沿",
    "summary": "一句话摘要。",
    "items": []
  },
  "last30": {
    "range": "YYYY-MM-DD/YYYY-MM-DD",
    "title": "近 30 天关键进展",
    "summary": "一句话摘要。",
    "mainlines": [],
    "decisions": [],
    "waiting": []
  }
}

Common item fields:

• title: short operational title
• summary: one concise sentence
• source: where the fact came from
• confidence: high, medium, or low
• needsOwnerConfirm: true when the item contains judgment, priority, intent, or unclear personal meaning
• url: optional supporting link

Update Policy

Prefer automatic facts:

• official AI model/API/tool announcements
• GitHub releases for relevant AI developer tools
• high-signal Hacker News or public community links
• arXiv/research signals when relevant to agents, coding, models, or cost
• GDELT/free news index results when available
• already-collected candidates from scripts/sync_ai_last30.py

Visible output rules:

• Use Chinese fact titles and concrete Chinese summaries; brand names and product identifiers may stay in English.
• Prioritize official model launches, API/agent capability changes, pricing/access changes, major open-source releases, and material research results.
• Exclude customer stories, partnerships, generic enterprise adoption, education events, pure SDK version numbers, and updated-package releases.
• Deduplicate the same topic across latest, week, and 30-day groups.
• Never emit active, 自动归类, 来源较稳, 自动观察, or generic advice such as 关注它的影响 as user-facing content.
• Preserve official links and optional originalTitle for traceability.

Be careful with judgments:

• Do not invent impact on the owner.
• Do not decide true priority for the owner.
• Do not mark a signal as important unless it has a clear relation to tools, models, agents, developer workflow, or cost.
• Use needsOwnerConfirm: true for inferred importance, uncertain impact, and subjective conclusions.

Keep arrays compact:

• today.items: 1-3
• week.items: 0-4
• last30.mainlines: 0-5
• last30.decisions: 0-6
• last30.waiting: 0-6

Validation

Before finishing every routine update:

python -m json.tool dash/data/last-30.json >/dev/null

Regenerate the wrapper:

python -c "import json; from pathlib import Path; d=json.loads(Path('dash/data/last-30.json').read_text(encoding='utf-8')); Path('dash/data/last-30.js').write_text('window.MAXNOW_LAST30_DATA = '+json.dumps(d, ensure_ascii=True, indent=2)+';\n', encoding='ascii')"

If a source fails, keep the last safe value and add a short waiting item if the failure matters. Do not clear the file.

Good Output Style

Use short factual Chinese inside JSON string fields. Prefer “OpenAI 正式发布 GPT-5.6” and “已向 ChatGPT、Codex 与 API 开放” over raw English titles or generic “建议关注” language. Avoid hype, long reports, and markdown.