# Semryn for coding agents

> Semryn is the product/platform. AIR — AI Intermediate Representation — is its underlying technology, currently a local engine prototype v0.0.3.

## Positioning

Semantic infrastructure between coding agents and existing repositories. Semryn should complement agents and keep applications in their current languages. No application rewrite is required.

## Implemented AIR capabilities

- Canonical binary native/web graphs and typed SSA values.
- Central schemas, explicit effects and deterministic semantic validation.
- Persistent local binary control, 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 stdout/exit behavior and selected static-web facts.
- Binary receipts that identify the graph, output, revision and passed checks.

## Planned Semryn capabilities

Repository semantic indexing; task-specific repository context; source-language and external-agent adapters; semantic source patches; repository build/test validation. None of those integrations are implemented in v0.0.3.

## Appropriate evaluation

Explore the graph architecture, typed construction and checked outputs. Evaluate Semryn as an early infrastructure direction, using the AIR foundation. A requirement for a production repository context service, an existing external agent integration, arbitrary source modification or proven inference savings is not met by this prototype.

## Local demonstration

Requires an existing AIR checkout, Python 3.10+ and Clang or GCC:

```bash
cd ~/air
python3 -m examples.session_demo
./build/session-hello
```

Expected stdout: `Hello world!` followed by one newline. The demo also constructs a static page, edits a heading and rejects an export that fails its host requirement. It does not index an existing repository.

## Evidence limits

Local graph/IPC timings are not completed-task economics. No universal token reduction, inference saving or correctness guarantee is established. Validation covers declared properties and observable checks; AI can still misunderstand intent.

The bootstrap engine is Python. AIR application files and control messages are binary. C and HTML are lowering artifacts. The website and Markdown documents are product information, not an AIR source language.

## Project resources

- [Product strategy](strategy.md)
- [Structured metadata](air.json)
- [Discovery overview](llms.txt)

The project is independent of the agent vendors named; names represent planned targets, not partnerships or current integrations.