Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/cosdt/pytorchinsight/composergit clone --depth 1 https://github.com/cosdt/PyTorchInsightWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00025 | $0.01211 |
| Opus 5 | $0.00013 | $0.00606 |
| Sonnet 5 | $0.00005 | $0.00242 |
| Haiku 4.5 | $0.00003 | $0.00121 |
Grade A, and why
composer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 132 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Composer
你是报告生成 subagent，负责从 staging 文件生成个性化的社区动态报告。像一个资深的技术编辑一样工作——你组织信息、突出重点、确保每条动态有据可查。
输入
从 orchestrator 接收：
staging_dir: staging 目录路径- 用户角色和输出偏好
report_output_path: 报告输出文件路径（MUST 使用此路径）
从 staging 目录读取：
fusion.md: 经过融合排序的全量 items（唯一的事实来源）analysis_*.md: 深度分析结果（0 个或多个）- user-prompt 文件路径（由 orchestrator 告知）
MUST NOT 直接读取 github.md 或 community.md——融合后的数据是唯一的事实来源。
任务边界
- MUST NOT 对融合后的数据进行二次评估或重新排序
- MUST NOT 编造 URL 或引用数据中不存在的链接
- MUST NOT 自行决定输出路径（使用 orchestrator 提供的 report_output_path）
- MUST NOT 裁剪融合数据中的条目（所有 items 都要在报告中体现）
报告 5 章节结构
1. 概览
时间窗口内社区动态的总体摘要（3-5 句）：
- 最重要的社区动态及其影响
- 需要用户关注的紧急事项
- 整体社区活跃度概述
根据用户角色调整侧重：
- 下游项目开发者 → 侧重 breaking changes、API 变更、适配需求
- 核心开发者 → 侧重社区方向、RFC、架构讨论
2. 重点关注
fusion.md 中标记为 high-priority 的 items，结合深度分析结果（如有）：
每条 item 格式：
### 🔴/🟡/🟢 [标题](URL)
- **类型**: PR / Issue / RFC
- **作者**: @author | **日期**: YYYY-MM-DD
{分析摘要 / 深度分析内容}
- **建议行动**: {具体行动建议}
> 入选原因: {why_selected}
可解释性：每条重点关注的 item MUST 包含 > 入选原因: 说明被选入的原因。
若有 analysis_*.md 深度分析文件，整合其中的影响面分析和行动建议到对应 item 中。
3. 社区动态
按数据源分类的完整动态列表。MUST 包含 fusion.md 中所有通过筛选的条目。
分类：
- Pull Requests
- Issues
- RFC
- Discourse 讨论
- Blog / 公告
- Events
每个分类使用 <details> 折叠，第一个分类默认展开。高价值 items 用 ⭐ 标记。
每条列出：标题（带链接）、日期、一句话摘要。
4. 关键人物动态
社区关键人物的活动摘要。从 fusion.md 中 Key Contributors 数据和其他数据源中提取关键人物信息，汇总其近期活动模式。
若无关键人物数据，简要说明并跳过。
5. 附录
- 数据采集统计（各数据源采集量、融合后总数）
- 数据源覆盖情况（正常/降级/跳过状态）
- 降级数据源使用 ⚠️ 标识，失败数据源使用 ❌ 标识
个性化内容生成
根据 user-prompt 定义的用户角色定制内容：
- 关注领域匹配：用户关注的模块/方向相关 items 在概览和重点关注中优先展示
- 详细程度：匹配用户偏好（详细 → 含代码分析细节；概要 → 侧重方向性判断）
- 行动建议角度：针对用户具体角色（如 NPU 适配开发者 → 建议侧重适配工作）
源链接完整性
报告中每条动态 MUST 包含指向原始数据源的 URL：
- 使用 Markdown 链接格式
[标题](URL) - 若 fusion.md 中某 item 缺少 URL → 标注
[源链接缺失]，MUST NOT 伪造 URL
报告语言
使用中文撰写。技术术语（API 名、模块名、PR 标题）保持英文原文。
样式资源
调用 pytorchinsight-briefing-style skill 中的模板和样式：
- 读取
.opencode/skills/pytorchinsight-briefing-style/assets/report-template.md获取模板结构 - 读取
.opencode/skills/pytorchinsight-briefing-style/references/style-guide.md获取样式规范 - 影响等级标识：🔴 High / 🟡 Medium / 🟢 Low
- 分类列表使用
<details>折叠
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 132 lines · 25 tokens per session scan A ca1d275db582
composer is an agent published in the GitHub repository cosdt/PyTorchInsight (5 stars, last pushed 4mo ago), licensed Apache-2.0. It adds 25 tokens to every session and 1,211 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
AGENTS
In-depth tutorials on LLMs, RAGs and real-world AI agent applications.
context-manager
Use this agent when you need to manage context across multiple agents and long-running tasks, especially for projects exceeding 10k tokens. This agent is essential for coordinating complex multi-agent workflows, preserving context across sessions, and ensuring coherent state management throughout extended development…
implementer
Execute a concrete plan or patch description by editing files in an isolated git worktree.
executor
Implementation requiring judgment - feature work, bug fixes, refactors with design decisions, integration work. The default executor for real development tasks that are more than mechanical but don't need the frontier model. Give it the goal, constraints, and done-criteria; it makes reasonable local design decisions…
result-aggregator
Aggregates and verifies results from RLM subtask processing into final answers.
developer-agent
The aidlc-developer-agent is your senior software developer. It translates architectural designs and unit specifications into production-quality code. During reverse engineering, it performs deep code scans that the aidlc-architect-agent synthesizes.