AGENTS.md - burn-onnx

ONNX import for the Burn deep learning framework. Converts ONNX models to Rust source code and .bpk weight files.

For detailed architecture, pipeline phases, code examples, and step-by-step operator implementation walkthrough, read DEVELOPMENT-GUIDE.md.

Project Structure

crates/
├── burn-onnx/       # Converts ONNX IR to Burn Rust code (codegen)
├── onnx-ir/         # ONNX Intermediate Representation parser
├── onnx-ir-derive/  # Derive macros for onnx-ir
├── onnx-tests/      # End-to-end integration tests
├── burn-import/     # Legacy crate (deprecated, re-exports burn-onnx)
└── model-checks/    # Real-world model validation (excluded from workspace)

onnx-spec/ops/       # Per-operator markdown specs (reference material)

Architecture Rules

onnx-ir and burn-onnx have a strict separation of concerns:

Responsibility	Where
Parse ONNX protobuf into IR	onnx-ir
Extract ALL ONNX attributes faithfully	onnx-ir
Type inference, static shape inference	onnx-ir (NodeProcessor)
Structural validation (e.g. "got 3D, need 2D")	onnx-ir (ProcessError)
Translate ONNX semantics to Burn semantics	burn-onnx
Reject unsupported Burn features	burn-onnx (clear error, not panic)
Generate Rust code	burn-onnx (NodeCodegen)

Key principles:

• onnx-ir mirrors ONNX, not Burn. Config structs store original ONNX attributes. Do not pre-compute Burn-specific values (e.g., don't resolve auto_pad into padding values)
• No panics in codegen. Use ProcessError in onnx-ir for validation. Panics in burn-onnx crash the build with poor error messages
• Prefer Burn tensor APIs over manual loops. Check the Burn API before generating element-wise code. Native tensor ops are orders of magnitude faster. When in doubt about Burn APIs, search online rather than guessing
• Declarative node architecture. Framework code must not contain node-type-specific logic. All node-specific behavior lives in NodeProcessor implementations
• Don't reject unknown ONNX attributes. Extract what you need, ignore the rest. Future opsets may add new attributes

Coding Conventions

Rust

• Edition 2024
• No unsafe code. This project has no need for unsafe. Do not introduce unsafe blocks, unsafe fn, or unsafe impl. If a dependency requires unsafe, wrap it in a safe abstraction upstream
• #[derive(Debug, Clone)] on public types
• thiserror for errors, log for logging (not println!)
• /// doc comments on public APIs

onnx-ir

• Node processors are pub(crate); only node structs and configs are public
• Config structs: #[derive(Debug, Clone, Default)], include ALL ONNX attributes (Option<T> for optional ones)
• Set min_opset to the earliest opset that introduced the operator (check onnx-spec/ops/<OpName>.md)
• Every processor must be registered in registry.rs
• Use node.get_input(index) for optional inputs (returns None for absent/optional). Never check name.is_empty(); use is_optional() instead
• ProcessError has a Display impl. Format with {}, not {:?}
• See DEVELOPMENT-GUIDE.md for static_shape, constant lifting, and NodeProcessor trait details

burn-onnx

• Implement NodeCodegen directly on onnx-ir node types
• Use scope.arg() for inputs: handles clone tracking for on-device values (Tensor, ScalarTensor) and bare idents for host values (ScalarNative, Shape)
• Use arg_to_ident() only for outputs and host-side values. Never use it for ScalarTensor inputs (it skips clone tracking)
• Scope temporary variables in block expressions to avoid name collisions
• insta snapshot tests for ALL codegen branches (inline snapshots only: assert_snapshot!(code, @r"..."))
• Always specify explicit dtypes in generated code. Never rely on the device's default float/int dtypes (DeviceSettings::float_dtype / int_dtype) because they vary across CPU/GPU devices:
  • Use .cast(DType::XX) after .int() or .float() to preserve the ONNX-specified dtype
  • When creating tensors, pin the dtype with Tensor::from_data(data, (&device, dtype)) rather than the bare &device overload, which would resolve to the device's default
  • Never use bare .int() or .float() without a following .cast(target_dtype)
  • When multiple tensors interact in binary ops, ensure they share the same dtype (cast to a common dtype first)

Testing

• Unit tests in the same file as implementation
• Integration tests in crates/onnx-tests/tests/<op_name>/
• Bug fixes must include an integration test (write failing test first, then fix)
• Use torch.manual_seed(42) / np.random.seed(42) for reproducibility
• Cover at least one non-default configuration per operator
• Python test scripts use uv inline script format with onnx.reference.ReferenceEvaluator as ground truth

Consuming a generated model

• Do NOT use Model::new(&device) to run a generated model. new calls Param::uninitialized for every constant/weight, leaving them zeroed. Graphs whose forward uses self.<param>.val() (anything with ONNX Constant/Initializer nodes — including all attention _expanded variants and many normal models) then run with all-zero constants and produce wrong output, often as bizarre downstream shape errors (e.g. repeat([0, 0, 0]) collapsing a tensor to [0, 0, 0, 0])
• Use Model::from_file(bpk_path, &device) instead — it constructs via new then runs load_from(BurnpackStore), which is a no-op when there are no Param fields, so it is safe for graphs without constants too
• Model::default() works but pins the device to Device::default() and embeds the absolute bpk path captured at codegen time; prefer the explicit from_file form
• Test harnesses and demo binaries that construct generated models must follow this rule. The symptom of getting it wrong is "compare passes for ops that touch no constants, fails for anything reshape/tile/scatter-shaped"

Adding a New Operator

Read DEVELOPMENT-GUIDE.md for the full walkthrough with code examples. Checklist:

1. onnx-ir: crates/onnx-ir/src/node/<op>.rs

   • Read onnx-spec/ops/<OpName>.md for the full spec
   • Define config struct, implement NodeProcessor
   • Register in: node/mod.rs, ir/node.rs (macro), registry.rs

2. burn-onnx: crates/burn-onnx/src/burn/node/<op>.rs

   • Implement NodeCodegen, add insta snapshot tests
   • Register in: node/mod.rs, node_codegen.rs (dispatch macro)

3. onnx-tests: crates/onnx-tests/tests/<op>/

   • Python script (uv format) + Rust test
   • Register in: build.rs, test_mod.rs

4. Update SUPPORTED-ONNX-OPS.md

Code Review Checklist

• Config structs include ALL ONNX attributes (don't skip because burn-onnx doesn't use them yet)
• No unsafe code anywhere in the project
• No unwrap in library code (tests are fine)
• No panic! for structural validation (use ProcessError)
• Generated code compiles without warnings
• Snapshot tests cover each config variant and input type
• New operators have both unit and integration tests
• Processor is registered in registry.rs and dispatch macro

Common Commands

cargo test                          # Run all tests
cargo test -p onnx-ir               # Test specific crate
cargo test -p burn-onnx
cargo test -p onnx-tests
cargo xtask validate                # Format, lint, test
cargo run -p burn-onnx --bin onnx2burn -- model.onnx ./out  # Generate code from ONNX
cargo insta review                  # Review snapshot changes

Key Files

• crates/onnx-ir/src/processor.rs - NodeProcessor trait, ProcessError, DefaultProcessor
• crates/onnx-ir/src/registry.rs - Processor registration
• crates/onnx-ir/src/ir/node.rs - Node enum and define_node_enum! macro
• crates/burn-onnx/src/burn/node_codegen.rs - Codegen dispatch macro
• crates/burn-onnx/src/burn/graph.rs - Graph code generation
• crates/burn-onnx/src/burn/partition.rs - Submodule partitioning for large models
• DEVELOPMENT-GUIDE.md - Full implementation guide with code examples
• SUPPORTED-ONNX-OPS.md - Operator support table
• onnx-spec/ops/<OpName>.md - Official ONNX operator specs

Awesome Agent Conventions

A curated field guide to the convention files AI agents read, write, and act on.

22 conventions across 11 categories. From common project instruction files to newer agent-web discovery and trust formats.

Agent tools increasingly rely on plain files in a repository or website root: instructions, memory, rules, tool connections, prompt assets, discovery metadata, and protocol hints. The names are easy to mix up, and the adoption levels vary a lot.

This repo keeps the map practical:

• Know what a file is for. Each entry names the convention, usual filename, primary readers, and spec or source.
• Study real examples. Examples are fetched from public repositories by script, with provenance kept at the top of each file.
• Separate practice from proposal. Maturity labels show what is widely used, what is early, and what is still only proposed.

Contents

• Instruction & context · category page
  • 🟢 AGENTS.md
  • 🟢 CLAUDE.md
  • 🟢 Tool-specific instruction files
  • 🟠 OKF (Open Knowledge Format)
• Memory & state · category page
  • 🟢 MEMORY.md
  • 🟢 Memory Bank
• Spec-driven development · category page
  • 🟢 Spec Kit
  • 🟢 Kiro steering files
• Skills & prompt assets · category page
  • 🟢 SKILL.md
  • 🟢 Prompt asset files
  • 🟢 Claude Code commands
  • 🟢 Copilot prompt & instruction files
• Tooling & connections · category page
  • 🟢 MCP server config
• Rules & ignore files · category page
  • 🟢 Rules files
  • 🟢 AI ignore files
• Design · category page
  • 🟢 DESIGN.md
• Web & discoverability · category page
  • 🟢 llms.txt
  • 🟢 pricing.md
• Agent-web trust · category page
  • 🟠 auth.md
  • 🔵 ai.txt
• Identity & protocols · category page
  • 🟠 Agent Cards (A2A)
• Proposed namespace · category page
  • 🔵 The protocols.md namespace

What counts

This list is intentionally narrow. A file belongs here when it is a convention for agent behavior or agent-readable metadata: instructions, memory, skills, rules, tool config, prompt assets, or web-discovery hints.

Any file type can qualify - .md, .txt, .prompty, .json, dotfiles, or a directory pattern. Human-first project docs such as README.md, CONTRIBUTING.md, SECURITY.md, and CHANGELOG.md stay out unless the file has become an agent convention in its own right.

Maturity tiers

The badge is a claim about adoption, not quality. It keeps a proven convention from being presented the same way as a new idea.

Badge	Tier	Meaning
🟢	Adopted	Used in production by multiple tools, projects, or teams.
🟠	Emerging	Published by a real organization, but still early or limited in adoption.
🔵	Proposed	Publicly described, but without clear adoption beyond the proposal.

Instruction & context

Standalone page: categories/instruction-context.md

	Convention	Files	Read by	Spec
🟢	AGENTS.md	AGENTS.md	Most coding agents - OpenAI Codex, Cursor, Jules, Aider, Gemini CLI, Zed, and others	spec ↗
🟢	CLAUDE.md	CLAUDE.md	Claude Code, and tools that read the Claude memory convention	spec ↗
🟢	Tool-specific instruction files	GEMINI.md AGENT.md QWEN.md WARP.md CONVENTIONS.md copilot-instructions.md	Each file is read by its namesake tool - Gemini CLI, Amp, Qwen Code, Warp, Aider, GitHub Copilot - often alongside or as a bridge to AGENTS.md	spec ↗
🟠	OKF (Open Knowledge Format)	.md	Agents over MCP (okfy, openknowledge, superops okf CLIs); Google's knowledge-catalog ingests bundles	spec ↗

• AGENTS.md - A plain-Markdown "README for agents" - build/test commands, conventions, and gotchas an agent needs before touching the code. The most widely adopted cross-tool instruction file.
• CLAUDE.md - Anthropic's memory file for Claude Code - loaded automatically at session start to carry project commands, style rules, and standing instructions across turns.
• Tool-specific instruction files - Per-tool instruction files that predate or coexist with AGENTS.md. Some tools now default to AGENTS.md while keeping legacy filenames alive, so these variants still matter when auditing real repositories.
• OKF (Open Knowledge Format) - A machine-first organizational knowledge base: a version-controlled folder of typed Markdown files (one concept per file) that any agent reads as ground-truth context. Open-sourced by Google Cloud in 2026 as the content layer to MCP's transport.

Memory & state

Standalone page: categories/memory-state.md

	Convention	Files	Read by	Spec
🟢	MEMORY.md	MEMORY.md	Claude Code's auto-memory - the per-project MEMORY.md index it writes and re-reads each session	spec ↗
🟢	Memory Bank	projectbrief.md productContext.md activeContext.md systemPatterns.md techContext.md progress.md	Cline, Roo Code, and Cursor (via the Memory Bank custom-instructions pattern)	spec ↗

• MEMORY.md - A persistent, agent-maintained index of durable facts - written and re-read across sessions so an agent accumulates project memory instead of relearning each time.
• Memory Bank - Cline's structured memory system - a set of Markdown files an agent reads at the start of every task to reconstruct full project context after its session memory resets. The six files shown are Cline's set; tools like Roo Code use an overlapping but different variant.

Spec-driven development

Standalone page: categories/spec-driven-development.md

	Convention	Files	Read by	Spec
🟢	Spec Kit	constitution.md spec.md plan.md tasks.md	GitHub Spec Kit's slash-command agents (Copilot, Claude, Gemini, Cursor, and more)	spec ↗
🟢	Kiro steering files	product.md structure.md tech.md	AWS Kiro (steering files are largely Kiro-specific)	spec ↗

• Spec Kit - GitHub's spec-driven workflow - a constitution plus per-feature spec → plan → tasks files that drive an agent through structured, reviewable implementation.
• Kiro steering files - Kiro's always-on steering docs - product, structure, and tech files that give the agent persistent project context outside of any single spec.

Skills & prompt assets

Standalone page: categories/skills-prompt-assets.md

	Convention	Files	Read by	Spec
🟢	SKILL.md	SKILL.md	Claude Agent Skills, Claude Code, Amp, Agent Skills-compatible tools	spec ↗
🟢	Prompt asset files	.prompty .prompt system_prompt.txt	Prompty tooling, Azure AI / Semantic Kernel, and apps that load externalized prompts	spec ↗
🟢	Claude Code commands	.md	Claude Code - project .claude/commands/ and user ~/.claude/commands/	spec ↗
🟢	Copilot prompt & instruction files	.prompt.md .instructions.md	GitHub Copilot in VS Code / Copilot CLI	spec ↗

• SKILL.md - A self-contained, model-invoked capability file that tells an agent when to load a reusable procedure and how to execute it.
• Prompt asset files - Externalized prompt files - Prompty's YAML-front-mattered .prompty, plain .prompt templates, and system_prompt.txt - that pull the prompt out of source code so it can be versioned and edited on its own. Only .prompty has a formal spec (prompty.ai); .prompt and system_prompt.txt are ad-hoc externalized-prompt filenames.
• Claude Code commands - A Markdown file Claude Code exposes as a /slash-command - a reusable, version-controlled prompt workflow, with optional frontmatter (allowed-tools, model, argument-hint) and $ARGUMENTS and shell placeholders (@file references are a general Claude Code prompt feature, not command-specific). Now converging with Agent Skills, but still widely committed in its own right.
• Copilot prompt & instruction files - Modular, path-scoped Copilot context: *.instructions.md auto-attach to matching files via an applyTo glob, while *.prompt.md are reusable prompts you invoke by name - the granular cousins of a single .github/copilot-instructions.md.

Tooling & connections

Standalone page: categories/tooling-connections.md

	Convention	Files	Read by	Spec
🟢	MCP server config	.mcp.json	Claude Code, Cursor, VS Code / Copilot, and Claude Desktop - every MCP host reads the same mcpServers schema, though the filename and path differ per tool	spec ↗

• MCP server config - A JSON file that tells an agent which Model Context Protocol servers to launch and how (command, args, env) - making a project's tool and data integrations portable, shareable, and version-controlled across every MCP-capable client.

Rules & ignore files

Standalone page: categories/rules-ignore-files.md

	Convention	Files	Read by	Spec
🟢	Rules files	.cursorrules .mdc .clinerules .clinerules/ (pattern) .windsurfrules	Cursor (.cursorrules / .mdc), Cline (.clinerules/ and legacy .clinerules), Windsurf (.windsurfrules)	spec ↗
🟢	AI ignore files	.aiignore .cursorignore .codeiumignore .aiexclude	JetBrains Junie (.aiignore), Cursor (.cursorignore), Codeium/Windsurf (.codeiumignore)	spec ↗

• Rules files - Per-tool rule files that scope agent behavior - older single-file forms (.cursorrules, .clinerules, .windsurfrules) and newer directory-based, glob-scoped forms (.cursor/rules/.mdc, .clinerules/, .windsurf/rules/.md).
• AI ignore files - gitignore-syntax files that fence an AI agent out of paths - secrets, vendored code, generated output - so they're never sent to the model as context.

Design

Standalone page: categories/design.md

	Convention	Files	Read by	Spec
🟢	DESIGN.md	DESIGN.md	Google Stitch natively; and coding agents (e.g. Claude Code) when pointed at it as design context	spec ↗

• DESIGN.md - A structured, machine-readable design specification - tokens, components, and layout intent - that an agent reads to generate or keep UI consistent with an established system. Open-sourced by Google Labs in 2026 as a cross-tool draft spec.

Web & discoverability

Standalone page: categories/web-discoverability.md

	Convention	Files	Read by	Spec
🟢	llms.txt	llms.txt llms-full.txt (pattern)	Docs sites publish it for LLM tools and crawlers - though no major provider has confirmed reading it	spec ↗
🟢	pricing.md	pricing.md	Agents and LLM browsers fetching a clean, parse-able pricing page	spec ↗

• llms.txt - A proposed-turned-widely-published standard: a root-level Markdown file giving LLMs a curated, link-rich map of a site's docs. Published across hundreds of developer-docs sites - though whether the major LLM providers actually read it remains unproven.
• pricing.md - The Markdown twin of a pricing page - same URL with a .md suffix - so an agent gets structured plans and numbers instead of scraping marketing HTML. A concrete, shipping instance of the page.md pattern.

Agent-web trust

Standalone page: categories/agent-web-trust.md

	Convention	Files	Read by	Spec
🟠	auth.md	auth.md	Agents discovering how to authenticate to a service (early adopters)	spec ↗
🔵	ai.txt	ai.txt	AI training/data-mining crawlers that voluntarily honor AI usage preferences; crawler support is not yet reliable	spec ↗

• auth.md - A Markdown file that tells an agent how to authenticate with a service - discovery of auth endpoints and flows. Shipped by WorkOS as a real, working convention, but adoption beyond it is still early.
• ai.txt - A text file declaring machine-readable consent, licensing, or policy preferences for AI training and data-mining. Spawning popularized the deployed root-file pattern, and a 2026 Internet-Draft now proposes a well-known URI; adoption and crawler obedience are still thin, so it stays 🔵.

Identity & protocols

Standalone page: categories/identity-protocols.md

	Convention	Files	Read by	Spec
🟠	Agent Cards (A2A)	agent-card.json agent.json (pattern)	A2A-compatible agents discovering another agent's capabilities	spec ↗

• Agent Cards (A2A) - The Agent2Agent (A2A) capability card - a JSON document at a well-known path advertising an agent's skills, endpoints, and auth so other agents can discover and call it. Now a Linux Foundation project at v1.0; adoption is growing but early.

Proposed namespace

Standalone page: categories/proposed-namespace.md

	Convention	Files	Read by	Spec
🔵	The protocols.md namespace	proof.md	- (no demonstrated readers; aspirational)	spec ↗

• The protocols.md namespace - A single maintainer's pre-registered namespace of ~74 aspirational .md "protocols" (proof.md, signature.md, reputation.md, …) staked as Schelling points for a future agent web. Published concept, no demonstrated adoption - see the page for the audited, honest caveats.

Maintaining examples

Example files are fetched, not invented. The extractor pulls them from public sources, stores them under conventions/<slug>/examples/<source>/<filename>, and adds a line-1 provenance comment. The examples remain under their upstream owners' licenses and terms; see THIRD_PARTY_EXAMPLES.md before reusing them.

To refresh everything:

pip install -r scripts/requirements.txt
python scripts/extract.py          # fetch real files + rebuild each convention's README
python scripts/build_readme.py     # rebuild this README from scripts/targets.json

Re-running is idempotent. A missing target prints a miss and is skipped. Examples are representative samples: any file over 256 KB (for example, a multi-MB llms-full.txt) is truncated with a marker pointing back to the full source. scripts/targets.json remains the source of truth for conventions that have not been migrated yet; the skill-md pilot uses local convention metadata instead. Edit the relevant source and re-run both scripts.

Shortcut targets are available in the Makefile:

make verify          # schema + generated files + example provenance + links
make extract         # refetch public examples and rebuild generated docs
make license-report  # summarize upstream licenses for vendored examples

CI keeps the generated files and links honest. The verify workflow checks that generated docs match catalog metadata and migrated local metadata, and that every spec, example, and instance URL still resolves on each pull request and weekly. Run the same link check locally with python scripts/check_links.py.

Contributing

Read CONTRIBUTING.md. In short: an entry must pass the filter above and carry evidence for its maturity tier. Add sources to scripts/targets.json for non-migrated conventions, run the scripts, and open a PR. The skill-md pilot uses local convention metadata instead. Do not hand-write example files.

Before proposing adjacent standards, check WATCHLIST.md. Project direction lives in ROADMAP.md.

License

The curation, scripts, and original prose in this repository are MIT. Vendored example files remain under their upstream owners' licenses and terms; see THIRD_PARTY_EXAMPLES.md.