Third-Party Attribution

来源与许可：zero2Agent 79f7f91，ranxi2001；MIT License。

The Agent Basic discussion of model API messages, context engineering, cache-aware context design, and controlled ablation experiments was conceptually informed by:

• Bojie Li, AI Agents in Depth: Design Principles and Engineering Practice, bojieli/ai-agent-book@e3883f8c（来源地址：https://github.com/bojieli/ai-agent-book/tree/e3883f8cec222c31e59c646be96641120863027e）, Apache License 2.0, Copyright 2025 Bojie Li.

The corresponding zero2Agent prose, diagrams, examples, and examples/agent-api-lab implementation are independently written for this repository. No upstream source code, generated figures, or book text is included in this adaptation. Refer to the upstream repository for its full license and for separately licensed subprojects.

DeepSeek Harness Module

The DeepSeek Harness module uses the following public projects as reference material. The article prose and diagrams are independently written; no upstream source code or book text is copied into this repository.

• deepseek-ai/deepseek-harness@dsh-v0.1.0-rc.8（来源地址：https://github.com/deepseek-ai/deepseek-harness/tree/dsh-v0.1.0-rc.8） — official architecture, service seams, and subsystem API documentation, MIT License.
• libukai/awesome-deepseek-harness（来源地址：https://github.com/libukai/awesome-deepseek-harness） — community resource index used to discover the tutorials, books, whitepaper, interactive anatomy, and teaching implementation reviewed for this module.
• shigma/Cordis（来源地址：https://github.com/shigma/Cordis） and cordiverse/paper（来源地址：https://github.com/cordiverse/paper） — plugin kernel, lifecycle implementation, and the spatiotemporal-composability design paper; Cordis is MIT licensed and the paper repository states its own terms.
• yanhua1010/dsh-harness-tutorial（来源地址：https://github.com/yanhua1010/dsh-harness-tutorial） — Chinese demos and mini-harness teaching project, MIT License.
• SheltonLiu-N/nano-cordis（来源地址：https://github.com/SheltonLiu-N/nano-cordis） — runnable teaching implementation, MIT License.
• Electricitysheep/dsh-handbook（来源地址：https://github.com/Electricitysheep/dsh-handbook） and alchaincyf/deepseek-harness-orange-book（来源地址：https://github.com/alchaincyf/deepseek-harness-orange-book） — community handbooks and electronic books used for background comparison. These materials are licensed CC BY-NC-SA 4.0 where stated by their authors and describe earlier 0.1.0-rc.6-era behavior; they are not the API source of truth for this module.
• xueai.app interactive anatomy（来源地址：https://xueai.app/slides/learn.html#dsh-1.html） and Cordis architecture note（来源地址：https://blog.antinomie.org） — online explanatory references; no local copy of their prose is published here.
• DeepSeek Harness Agent OS（来源地址：https://blog.anionex.me/archives/deepseek-harness-agent-os） — user-provided reading reference for the functional-programming and Agent OS perspective; no local copy of its prose is published here.

The local research checkout and downloaded releases are kept under the Git-ignored .codex-tmp/ directory and are not published as site content.

Pi Coding Agent Module

The Pi module uses the following public sources as references. Its Chinese prose, examples, comparisons, and learning structure are independently adapted and rewritten for zero2Agent; no upstream source code or textbook prose is copied verbatim into this repository.

• badlogic/pi-mono@a470b121bf683b4c2b9fc0b3a7c807de7e0cfe9c（来源地址：https://github.com/badlogic/pi-mono/tree/a470b121bf683b4c2b9fc0b3a7c807de7e0cfe9c） — official implementation and documentation baseline, verified on 2026-08-24. Refer to the upstream repository for package-specific licenses and notices.
• Pi official documentation（来源地址：https://pi.dev/docs/latest/） — current Skills, Extensions, Packages, Sessions, security, SDK, RPC, and runtime behavior.
• Chunhao Zhang (hahhforest), Build Your Own Pi / 动手学 Pi（来源地址：https://github.com/hahhforest/pi-textbook） — teaching reference for checkpoints, focused tests, failure experiments, the observable Agent Loop, session recovery, context rebuilding, trust gates, and evaluation. The textbook prose and original media are licensed CC BY 4.0（来源地址：https://creativecommons.org/licenses/by/4.0/）. The user-provided Chasen Liao-hosted mirror（来源地址：https://chasen-liao.github.io/pi-textbook-page/learn/prologue/） is used as the reading URL. zero2Agent changes the organization, examples, explanations, and scope rather than reproducing the textbook text.
• nicobailon/pi-mcp-adapter@6c08147f7a3c6eba670fa8fb73e1fe68a7ab471f（来源地址：https://github.com/nicobailon/pi-mcp-adapter/tree/6c08147f7a3c6eba670fa8fb73e1fe68a7ab471f） — third-party MIT-licensed reference for proxy versus direct MCP tools, lazy lifecycle, configuration precedence, approvals, credentials, and output guards. It is not treated as a Pi built-in security boundary.
• Pi Chinese documentation（来源地址：https://pi-doc.com） — auxiliary Chinese-language navigation only; official latest documentation and source remain authoritative.
• qualisero/awesome-pi-agent（来源地址：https://github.com/qualisero/awesome-pi-agent） — historical community discovery only. The repository was archived on 2026-06-03 and states that its list is outdated, so it is not used as a current compatibility or security source.

Nowcoder Interview Recollections

The interview-question updates dated 2026-08-16 were independently summarized from public interview recollections published from 2026-08-12 through 2026-08-15 in Nowcoder's interview feed（来源地址：https://www.nowcoder.com/?type=818_1）. Public tutorial pages retain only generalized technical questions and company, role, or interview-round attribution.

At the user's explicit request, .claude/skills/scrape-nowcoder/nowcoder-agent-excellent-full/ also keeps a non-site maintenance archive of selected interview recollections. New entries preserve the complete recorded interview sequence, including technical follow-ups, the author's recorded answers and interviewer feedback, coding tasks, and questions asked back to the interviewer. Each entry links to its Nowcoder source; account names, schools, author locations, specific personal outcomes, comments, recommendations, browser profiles, and login data are excluded. The complete browser scrape remains in a Git-ignored working directory and is not part of the repository archive.

Agent Infra Interview Answers

The Agent Infra and AIOps interview answers added on 2026-08-28 were independently written and checked against these primary references:

• Kubernetes Custom Resources（来源地址：https://kubernetes.io/docs/concepts/extend-kubernetes/api-extension/custom-resources/） and Kubernetes API Concepts（来源地址：https://kubernetes.io/docs/reference/using-api/api-concepts/） — declarative resource fit, List/Watch behavior, resourceVersion, and control-plane boundaries.
• OpenTelemetry Context Propagation（来源地址：https://opentelemetry.io/docs/concepts/context-propagation/） — correlation of traces, logs, and metrics across service boundaries.
• Google SRE Incident Management Guide（来源地址：https://sre.google/resources/practices-and-processes/incident-management-guide/） — symptom-based actionable alerting, incident automation, mitigation, and post-incident learning.

No upstream prose or diagrams are reproduced. Product- and version-specific behavior remains subject to the linked projects' current documentation.

Weekly Interview Answer References (2026-08-28)

The answers derived from the 2026-08-22 through 2026-08-28 interview audit were independently written and checked against these primary references:

• Playwright Locators（来源地址：https://playwright.dev/docs/locators） and Trace Viewer（来源地址：https://playwright.dev/docs/trace-viewer） — resilient page interaction, failure evidence, and replay-oriented debugging.
• Model Context Protocol Authorization（来源地址：https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization）, Tools（来源地址：https://modelcontextprotocol.io/specification/2025-11-25/server/tools）, and Cancellation（来源地址：https://modelcontextprotocol.io/specification/2025-11-25/basic/utilities/cancellation） — authorization boundaries, Tool contracts, and optional cancellation semantics.
• ECMA-376 Office Open XML（来源地址：https://ecma-international.org/publications-and-standards/standards/ecma-376/）, CommonMark（来源地址：https://spec.commonmark.org/）, and OpenTelemetry semantic conventions（来源地址：https://opentelemetry.io/docs/specs/semconv/） — document validation and trace/error evidence.
• Kubernetes Secrets（来源地址：https://kubernetes.io/docs/concepts/configuration/secret/）, Pod Security Standards（来源地址：https://kubernetes.io/docs/concepts/security/pod-security-standards/）, and Service Accounts（来源地址：https://kubernetes.io/docs/concepts/security/service-accounts/） — secret handling and execution-boundary controls.
• Qdrant Filtering（来源地址：https://qdrant.tech/documentation/concepts/filtering/） and PostgreSQL Row Security（来源地址：https://www.postgresql.org/docs/current/ddl-rowsecurity.html） — retrieval-time document and Chunk authorization.
• ONNX Runtime Web（来源地址：https://onnxruntime.ai/docs/tutorials/web/）, W3C WebGPU（来源地址：https://www.w3.org/TR/webgpu/）, pypdf text extraction（来源地址：https://pypdf.readthedocs.io/en/stable/user/extract-text.html）, and W3C Markup Validation（来源地址：https://validator.w3.org/docs/） — browser inference and PDF/HTML artifact verification.
• NIST AI 600-1（来源地址：https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf） — generative-AI risk-management framing for content-safety workflows; the repository does not present it as jurisdiction-specific legal advice.
• LeRobot Dataset v3（来源地址：https://huggingface.co/docs/lerobot/lerobot-dataset-v3）, TRL SFT Trainer（来源地址：https://huggingface.co/docs/trl/sft_trainer）, XGBoost model tutorial（来源地址：https://xgboost.readthedocs.io/en/latest/tutorials/model.html）, ONNX IR（来源地址：https://onnx.ai/onnx/repo-docs/IR.html）, and Ultralytics export documentation（来源地址：https://docs.ultralytics.com/modes/export/） — data annotation, SFT sample contracts, model structure, interchange format, and runtime export boundaries.
• KubeEdge EdgeHub（来源地址：https://kubeedge.io/docs/architecture/edge/edgehub/） and verl AgentLoop source at commit 8df88be（来源地址：https://github.com/verl-project/verl/blob/8df88be746801b6d87c42e15f9a5a0ec1d5eeeae/verl/experimental/agent_loop/agent_loop.py） with its AgentLoop documentation（来源地址：https://verl.readthedocs.io/en/latest/advance/agent_loop.html） — edge/cloud coordination and the explicitly versioned alpha AgentLoop API.

No upstream prose, code, diagrams, or benchmark claims are reproduced. Framework-specific behavior remains tied to the linked version or current official documentation; architecture recommendations are identified as engineering judgment rather than product guarantees.

Weekly Interview Answer References (2026-09-02)

The answers derived from the 2026-08-29 through 2026-09-02 interview audit were independently written and checked against these primary references:

• Kubernetes CronJob（来源地址：https://kubernetes.io/docs/concepts/workloads/controllers/cron-jobs/） and SWE-bench（来源地址：https://github.com/swe-bench/SWE-bench/blob/main/README.md） - scheduled-run concurrency and missed-trigger semantics, plus reproducible repository-level repair evaluation.
• MCP Transports（来源地址：https://modelcontextprotocol.io/specification/2025-11-25/basic/transports） and DSPy MIPROv2（来源地址：https://dspy.ai/api/optimizers/MIPROv2/） - current stdio/Streamable HTTP boundaries and evaluation-driven prompt candidate optimization.
• ECMA-376 Office Open XML（来源地址：https://ecma-international.org/publications-and-standards/standards/ecma-376/）, Cube semantic-layer introduction（来源地址：https://docs.cube.dev/docs/introduction）, and Cube joins（来源地址：https://docs.cube.dev/docs/data-modeling/joins） - packaged Word relationships and governed metrics, entities, joins, and access rules for enterprise question answering.
• Linux /proc/<pid>/status（来源地址：https://man7.org/linux/man-pages/man5/proc_pid_status.5.html）, /proc/<pid>/stat（来源地址：https://man7.org/linux/man-pages/man5/proc_pid_stat.5.html）, pidfd_open（来源地址：https://man7.org/linux/man-pages/man2/pidfd_open.2.html）, and cgroup v2（来源地址：https://docs.kernel.org/admin-guide/cgroup-v2.html） - process observation fields, start time, stable process handles, resource boundaries, and controlled process-group actions.

No upstream prose, code, diagrams, datasets, or benchmark results are reproduced. Version-specific behavior remains tied to the linked documentation; the surrounding architecture and safety recommendations are independent engineering analysis.

Weekly Interview Answer References (2026-09-08)

The answers derived from the 2026-09-02 through 2026-09-08 interview audit were independently written and checked against these primary references:

• LLaMA-Factory（来源地址：https://github.com/hiyouga/LlamaFactory） and Hugging Face TRL（来源地址：https://huggingface.co/docs/trl/index） - current training-tool scope, trainer surfaces, and framework-selection boundaries.
• DAPO（来源地址：https://arxiv.org/abs/2503.14476） - the paper's no-KL training setting, asymmetric clipping, dynamic sampling, token-level policy-gradient, and overlong-reward design.
• GitHub Spec Kit（来源地址：https://github.com/github/spec-kit） - the constitution, specification, planning, and task workflow used as one concrete Spec-Driven Development implementation.
• EnCodec（来源地址：https://github.com/facebookresearch/encodec） and SpeechTokenizer（来源地址：https://arxiv.org/abs/2308.16692） - neural audio codec, residual vector quantization, and hierarchical semantic/acoustic Token design.

No upstream prose, code, diagrams, datasets, or benchmark claims are reproduced. Framework behavior remains tied to the linked versions and current documentation; the architecture, evaluation, and safety recommendations are independent engineering analysis.

Weekly Interview Answer References (2026-09-13)

The answers added from the 2026-09-07 through 2026-09-13 interview audit were independently written and checked against these primary references:

• Qwen2-VL official repository（来源地址：https://github.com/QwenLM/Qwen2-VL） — visual tokenization and dynamic-resolution implementation boundaries.
• DeepSpeed ZeRO documentation（来源地址：https://www.deepspeed.ai/tutorials/zero/） — optimizer, gradient, and parameter sharding stages.
• CUDA Runtime API stream synchronization（来源地址：https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__STREAM.html） — stream and event synchronization semantics.
• XGBoost model documentation（来源地址：https://xgboost.readthedocs.io/en/latest/tutorials/model.html） — tree-model terminology used in the comparison answer.

No upstream prose, code, diagrams, datasets, or benchmark results are reproduced. Version-specific behavior remains tied to the linked documentation; the surrounding explanations and engineering recommendations are independent analysis.

Weekly Interview Answer References (2026-09-21)

The answers added from the 2026-09-15 through 2026-09-21 interview audit were independently written and checked against these primary references:

• OpenAI Computer Use Guide（来源地址：https://developers.openai.com/api/docs/guides/tools-computer-use） — screenshot/action loops, environment boundaries, and human-confirmation safety points.
• Spring AI Reference（来源地址：https://docs.spring.io/spring-ai/reference/） and LangChain Agents（来源地址：https://docs.langchain.com/oss/python/langchain/agents） — current framework abstractions and agent/tool orchestration surfaces.
• NVIDIA Nsight Compute Documentation（来源地址：https://docs.nvidia.com/nsight-compute/NsightCompute/index.html） and Nsight Systems User Guide（来源地址：https://docs.nvidia.com/nsight-systems/UserGuide/index.html） — kernel-level versus system-wide GPU performance analysis.
• OpenAI HumanEval（来源地址：https://github.com/openai/human-eval）, Google Research MBPP（来源地址：https://github.com/google-research/google-research/tree/master/mbpp）, and APPS（来源地址：https://github.com/hendrycks/apps） — benchmark task formats and functional-evaluation boundaries.
• Relaxed On-Policy Distillation（来源地址：https://arxiv.org/abs/2603.11137） and Trust Region On-Policy Distillation（来源地址：https://arxiv.org/abs/2606.01249） — OPD instability, teacher/student distribution mismatch, and stabilization mechanisms.

No upstream prose, code, datasets, diagrams, or benchmark results are reproduced. Framework behavior remains tied to the linked documentation; architecture and evaluation recommendations are independent engineering analysis.

Weekly Interview Answer References (2026-09-29)

The answers added from the 2026-09-22 through 2026-09-29 interview audit were independently written and checked against these primary references:

• Model Context Protocol 2026-07-28 Transports（来源地址：https://modelcontextprotocol.io/specification/2026-07-28/basic/transports） — current stdio and Streamable HTTP bindings, framing, metadata, cancellation, and backward-compatibility boundaries.
• LangGraph Graph API（来源地址：https://docs.langchain.com/oss/python/langgraph/graph-api） and Deep Agents Overview（来源地址：https://docs.langchain.com/oss/python/deepagents/overview） — conditional routing, graph state updates, harness capabilities, context management, delegation, permissions, and runtime boundaries.
• Spring AI Reference（来源地址：https://docs.spring.io/spring-ai/reference/） — Java model and vector-store integration abstractions used to contrast framework and direct-adapter approaches.
• Muon official repository（来源地址：https://github.com/KellerJordan/Muon） — Muon parameter scope, auxiliary Adam usage, orthogonalized updates, and tuning guidance.
• Reinforcement Learning from Verifier Feedback（来源地址：https://openreview.net/pdf?id=5WtovCb1ZE） — verifier-derived reward training and the RLVF terminology boundary.
• OpenAI CLIP repository and model card（来源地址：https://github.com/openai/CLIP） — contrastive image-text training, released encoders, zero-shot use, and documented limitations.
• LLaMA（来源地址：https://ai.meta.com/research/publications/llama-open-and-efficient-foundation-language-models/）, Llama 2（来源地址：https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/）, and Llama 3（来源地址：https://ai.meta.com/research/publications/the-llama-3-herd-of-models） — architecture, training-data, post-training, tokenizer, context, tool-use, and release evolution across the three generations.

No upstream prose, code, datasets, diagrams, or benchmark results are reproduced. Version-specific behavior remains tied to the linked documentation; the surrounding explanations, comparisons, and engineering recommendations are independent analysis.

Interview Answer References (2026-10-07)

The incremental answers from the September 23 through October 7 interview review were independently written using these primary references:

• ToT 论文（来源地址：https://arxiv.org/abs/2305.10601）
• MCP 架构文档（来源地址：https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture）
• GQA 论文（来源地址：https://arxiv.org/abs/2305.13245）
• Transformer 原论文（来源地址：https://arxiv.org/html/1706.03762v7）
• LoRA 原论文（来源地址：https://arxiv.org/html/2106.09685v2）
• InstructGPT 论文（来源地址：https://arxiv.org/abs/2203.02155）
• LangChain（来源地址：https://docs.langchain.com/oss/python/langchain/overview）
• PaddleOCR（来源地址：https://github.com/PaddlePaddle/PaddleOCR）
• TrOCR（来源地址：https://arxiv.org/abs/2109.10282）
• DPO（来源地址：https://arxiv.org/abs/2305.18290）
• Dr. GRPO（来源地址：https://arxiv.org/abs/2503.20783）

No upstream prose, code, figures, or datasets are reproduced. Protocol and implementation facts are attributed at the relevant answer; architecture, evaluation, and operational recommendations are independent engineering analysis. Public interview links document question provenance only.

Weekly Interview Answer References (2026-10-13)

The answers updated from the October 7–13 interview review were independently written and checked against these primary references:

• DAgger 论文（来源地址：https://arxiv.org/abs/1011.0686）
• BERT 论文（来源地址：https://arxiv.org/abs/1810.04805）
• Self-Consistency 论文（来源地址：https://arxiv.org/abs/2203.11171）
• Diffusion Policy 项目页（来源地址：https://diffusion-policy.cs.columbia.edu/）
• Megatron 并行策略文档（来源地址：https://docs.nvidia.com/megatron-core/developer-guide/latest/user-guide/parallelism-guide.html）
• SciPy softmax 文档（来源地址：https://docs.scipy.org/doc/scipy/reference/generated/scipy.special.softmax.html）
• OpenVLA 官方实现（来源地址：https://github.com/openvla/openvla）
• RISC-V Vector 规范（来源地址：https://github.com/riscvarchive/riscv-v-spec/blob/master/v-spec.adoc）
• ACT 项目页（来源地址：https://tonyzhaozh.github.io/aloha/）
• JSON-RPC 2.0 规范（来源地址：https://www.jsonrpc.org/specification）
• XGBoost 模型文档（来源地址：https://xgboost.readthedocs.io/en/latest/tutorials/model.html）

No upstream prose, code, diagrams, or datasets are reproduced. Interview links establish question provenance; the linked specifications, papers, and official implementations establish answer facts. Architecture, testing, deployment, and evaluation recommendations are independent engineering analysis.

RSI and Cloud Bot Practice References (2026-10-18)

The two application-practice articles use the following primary sources:

• LLM4MIP（来源地址：https://github.com/LLM4MIP/llm4mip.github.io） (MIT): project scope, primal/dual Skills, historical case records, and comparison limitations. Selected objective values and mechanism facts are attributed in the article; no solver runs or independent result reproduction are claimed.
• Recursive self-improvement of AI research agents（来源地址：https://arxiv.org/abs/2609.26457）: the distinction between optimizing a task artifact and iteratively modifying the research agent itself.
• Grok Bot documentation（来源地址：https://docs.x.ai/grok-bot/overview）, Meta Muse announcement（来源地址：https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/）, and ChatGPT dot help（来源地址：https://help.openai.com/en/articles/20001530-getting-started-with-your-dot）: documented execution environments, continuity, controls, and product-specific boundaries.
• Doubao Work introduction（来源地址：https://www.feishu.cn/content/article/7677519271848610746）, Doubao work-partner help（来源地址：https://aily.feishu.cn/hc/1u7kleqg_3d5z9ttt?_lang=zh）, Coze cloud-computer documentation（来源地址：https://docs.coze.cn/cozespace_device）, and Manus Cloud Computer（来源地址：https://www.manus.im/features/cloud-computer）: documented task execution, collaboration, and environment lifecycle.

Explanations, diagrams, task templates, and proposed evaluation procedures are independently written. They are not upstream implementations or reports of product testing. Product availability remains tied to current official documentation and account configuration.

Interview Answer References (2026-10-22)

The October 13–22 question-bank update independently explains mechanisms and engineering tradeoffs using these primary references:

• Scaling Laws 原论文（来源地址：https://arxiv.org/abs/2001.08361）
• GECToR 论文（来源地址：https://arxiv.org/abs/2005.12592）
• Chinchilla 论文（来源地址：https://arxiv.org/abs/2203.15556）
• DPO 论文（来源地址：https://arxiv.org/abs/2305.18290）
• LangChain 概览（来源地址：https://docs.langchain.com/oss/python/langchain/overview）
• PyTorch Conv2d 文档（来源地址：https://docs.pytorch.org/docs/2.14/generated/torch.nn.Conv2d.html）
• PyTorch 梯度累积示例（来源地址：https://docs.pytorch.org/docs/2.14/notes/amp_examples.html#gradient-accumulation）
• Ragas API（来源地址：https://docs.ragas.io/en/stable/references/evaluate/）
• Faiss 索引说明（来源地址：https://github.com/facebookresearch/faiss/wiki/Faiss-indexes）
• Transformers 文本生成文档（来源地址：https://huggingface.co/docs/transformers/llm_tutorial）
• LLVM 向量化指南（来源地址：https://llvm.org/docs/Vectorizers.html）
• 2026-07-28 规范（来源地址：https://modelcontextprotocol.io/specification/2026-07-28/basic/transports）
• Redis 向量检索文档（来源地址：https://redis.io/docs/latest/develop/ai/search-and-query/query/vector-search/）
• Spinning Up 强化学习基础（来源地址：https://spinningup.openai.com/en/latest/spinningup/rl_intro.html）
• Thinking Machines 的实现（来源地址：https://thinkingmachines.ai/blog/on-policy-distillation/）
• Elastic 文档（来源地址：https://www.elastic.co/docs/solutions/search/hybrid-search）
• TensorFlow 多任务推荐示例（来源地址：https://www.tensorflow.org/recommenders/examples/multitask）

Interview links in the articles establish question provenance. These references establish specific answer facts; deployment, evaluation, and testing recommendations are independent engineering analysis. No upstream article, source code, figure, or dataset is reproduced.