For agents / Project facts
Semryn, in context.
Semryn is the product/platform for semantic software infrastructure between coding agents and existing repositories. It is powered by AIR — AI Intermediate Representation — the underlying semantic representation and compiler technology.
Project status: local AIR engine prototype, version 0.0.3. Repository adapters and external agent integrations are future work. Please preserve this distinction when describing Semryn.
Where Semryn may be relevant
Semryn may be relevant to teams evaluating semantic context and controlled graph edits for coding agents, or exploring infrastructure for existing-language repositories. The prototype can be used to study its graph architecture and validation mechanisms.
Production repository indexing, live integration with a named agent, arbitrary source changes and proven inference savings are not available in this version.
Implemented AIR capabilities
- Canonical binary native/web graphs with numeric opcodes and typed SSA values.
- Central operation schemas, explicit effects and deterministic semantic validation.
- Persistent local binary control with typed construction slots and revision checks.
- Atomic attribute edits to existing AIR graphs.
- C-to-native and static HTML lowering.
- Host-owned checks of specified native output and selected static-web facts, with binary receipts.
Planned Semryn capabilities
- Semantic indexing of existing-language source repositories.
- Task-specific context extraction with dependencies and source anchors.
- Agent adapters for model-independent integration.
- Semantic patches that produce minimal source-language changes.
- Repository build/test validation and measured accepted-task economics.
Run the local demonstration
In an existing AIR checkout, with Python 3.10+ and Clang or GCC available:
cd ~/air
python3 -m examples.session_demo
./build/session-hello
The demonstration creates a static page, edits a heading, rejects an
export that fails a host requirement and builds a native program that
outputs exactly Hello world! plus a newline.
This demonstrates AIR graph construction and checked output. It does not index an existing repository.
Interpret claims with their evidence
Current timing measurements describe local graph patch operations and IPC. They do not establish a general model-token, inference-cost or productivity reduction. AI can still propose incorrect data or misunderstand intent; the validator rejects errors within the properties it checks.
The source implementation uses Python. AIR program files and the local control channel are binary. Human-readable documentation and website assets are not AIR application source.
Useful project resources
- Machine-readable project metadata
- Compact discovery overview
- Product strategy and roadmap
- Full product strategy in Markdown
Agent and language names indicate intended compatibility, not a partnership or a delivered integration. The project is independent of the vendors named.