LLM and agent access

PyZNN publishes the same authoritative documentation in human-readable Docusaurus pages and clean Markdown artifacts for language models, coding agents, retrieval pipelines, and offline indexing.

Machine-readable endpoints

Endpoint	Purpose
https://pyznn.0x3639.com/llms.txt	Concise project summary and categorized Markdown link map
https://pyznn.0x3639.com/llms-full.txt	Complete documentation corpus in sidebar reading order
https://pyznn.0x3639.com/llms/{page}.md	Clean Markdown mirror of one documentation page
https://github.com/0x3639/znn-sdk-py	Canonical source repository

The root llms.txt follows the proposed llms.txt structure: one project H1, a summary blockquote, concise interpretation guidance, categorized link lists, and an Optional section for lower-priority context.

Which artifact to retrieve

Task	Best input
Orient to the project and choose relevant pages	llms.txt
Answer one focused implementation question	The linked page Markdown
Build a broad retrieval index	All files under /llms/
Supply one large context to a capable model	llms-full.txt
Verify an exact implementation detail	Repository source plus the generated reference

Prefer the smallest authoritative context that answers the question. The full corpus is intentionally comprehensive and may be unnecessary for a single API call.

Authority order

When two sources appear to disagree, use this order:

1. current Python source and generated runtime inventories;
2. the pinned stable zenon-sdk-spec revision named in CI;
3. generated API and model references;
4. conceptual guides and cookbook examples;
5. historical compatibility behavior.

The generated references and LLM artifacts are checked for staleness with:

python3.12 tools/generate_documentation.py --check

Facts code-generating agents must preserve

• Python 3.12 or newer is required.
• No public third-party node is configured by default.
• Use HttpClient for request/response RPC and WsClient for subscriptions.
• API facades send canonical positional parameters and parse typed results.
• Token amounts passed to transactions and builders are integer base units.
• Use znn.amount with strings or Decimal for display conversion; avoid binary floating point.
• Embedded builder methods return unsigned AccountBlock objects and perform no I/O.
• Transaction preparation and publication are distinct operations.
• Account-block signatures cover the raw canonical 32-byte hash.
• A custom PoW provider's nonce is verified locally before signing.
• Canonical publication success is JSON null.
• Subscriptions are async iterators over normalized individual updates.
• In-flight requests fail on connection loss; the client does not silently replay them.
• The removed embedded.plasma.getRequiredFusionAmount RPC must not be called.
• Tests and conformance verification do not require a live Zenon node.

Retrieval-friendly page structure

Each conceptual page provides:

• a unique H1 and descriptive front matter;
• explicit imports and complete Python blocks;
• input and environment requirements before online examples;
• result types and error behavior after examples;
• stable cross-links to detailed references;
• terms written out before abbreviations where practical.

Generated references provide exact method signatures, wire method names, declared result types, model wire keys, Python attributes, required flags, defaults, and enum values.

Suggested coding-agent prompt

Use https://pyznn.0x3639.com/llms.txt to select the minimum relevant PyZNN
documentation. Verify every method signature and model field against the linked
generated reference. Write Python 3.12+ async code, use an explicit trusted node
URL from configuration, preserve typed model parsing, use integer base units,
and close the client in finally. Do not invent RPC methods, model fields,
defaults, public nodes, or automatic retries for transaction publication.

Suggested review-agent prompt

Review this PyZNN code against https://pyznn.0x3639.com/llms-full.txt and the
current repository source. Check endpoint trust, request positional order,
nullable results, strict model fields, base-unit amounts, account-block hashing
and signing, PoW verification, publication null semantics, subscription cleanup,
and TransportError versus JsonRpcError handling. Report only actionable
discrepancies with exact source or documentation references.

Suggested retrieval chunks

For retrieval-augmented generation, split at H1/H2 headings and retain:

• the page title;
• source URL;
• nearest parent headings;
• fenced code language;
• tables as complete units rather than individual rows.

Do not split a code block from the paragraph that names its required environment variables, result type, or failure semantics.

Generated-file policy

Authoritative edits belong in documentation/website/docs. Do not hand-edit:

• docs/api-reference.md;
• docs/model-reference.md;
• static/llms.txt;
• static/llms-full.txt; or
• static/llms/*.md.

Regenerate them with tools/generate_documentation.py, build the site, and run the generator in --check mode before committing.

Search

The site includes local full-text search for documentation and regular pages. The index is generated at build time and loaded by the browser. Search queries do not require Algolia credentials and are not sent to an external search service.