Back to the catalog

syntavra

Syntavra v0.0.1 pre-release runtime, sandbox, reversible compression, structural intelligence and unbounded external history

Open source Open in the app JSON README (API)

About

Syntavra v0.0.1 pre-release runtime, sandbox, reversible compression, structural intelligence and unbounded external history

Details

Kind
Plugins
Topic
Developer tools
Publisher
naveax
Origin
gemini
Category
ferramentas
Version
0.0.1
Stars
1
Open pull requests
3
Last push
2026-09-01T04:38:09Z
Repository state
ativo
Language
Python
License
MIT
Added
2026-08-30 14:13:39
Updated
2026-08-30 14:13:39
Origin id
naveax/syntavra

README

# Syntavra 0.0.1 — Pre-Release Token & Context Optimization Skill

Syntavra is a local-first **Agent Skill and runtime middleware** designed to reduce token and context overhead in existing AI coding agents. It does not replace Codex, Claude Code, Cursor, OpenCode or another agent/model.

Syntavra targets five cost surfaces while preserving exact recovery:

| Surface | What Syntavra does |
|---|---|
| Repository context | Retrieves exact definitions, impact paths, tests and verifiers before broad file reads |
| Tool output | Stores full stdout/stderr as exact evidence and returns bounded task-relevant views |
| MCP schemas | Exposes a small profile and deterministically compiles verbose discovery schemas |
| Session memory | Keeps active context bounded while exact external history remains searchable and recoverable |
| Measurement | Separates provider-observed usage from locally tokenized or estimated attribution |

> The only active product identity is **0.0.1 / pre-release**. The owner explicitly authorizes any version or release-channel change. External superiority, live certification, long-context quality, adoption and production maturity remain evidence-gated.

## Install

The registry package is prepared but not published yet:

```bash
npx @syntavra/install
```

Until registry publication, use the repository installer:

```bash
npx github:Naveax/Syntavra
```

The installer prefers checksum-verified portable binaries, falls back to Python 3.11+, configures detected hosts transactionally and runs a final health check.

## Engine selection

Syntavra is migrating to one dual-engine bundle containing the Python runtime, native Rust runtime and a native selector. Engine preference is explicit and persisted without hidden fallback:

```bash
syntavra engine status
syntavra engine use python --scope project
syntavra engine use rust --scope project
syntavra engine use auto --scope project

syntavra --engine python version
syntavra --engine rust version
```

Selection precedence is command override, `SYNTAVRA_ENGINE`, project configuration, user configuration and then the explicit `auto` policy. Selecting `python` never invokes Rust. Selecting `rust` never invokes Python; a command without an independent native Rust handler fails closed.

The canonical installed Python surface currently contains 245 executable public command paths. Full dual-engine parity inventory is tracked against those 245 paths in `contracts/engine/dual-engine-public-surface-v2.json`; bounded R23–R37 contract parity is not treated as a complete native product rewrite. The canonical route-coverage claim is:

```text
FULL_DUAL_ENGINE_PARITY_PROVEN
```

This route-level parity claim does not reactivate Rust feature/parity development and does not grant production-promotion credit. Rust development/resume authority and the 174/245 production-promotion boundary are governed separately.

## Product surface

```bash
syntavra setup                 # plan installation
syntavra setup --apply         # install and configure detected hosts
syntavra setup --repair --apply
syntavra status                # health, profile, evidence and observed savings
syntavra status --savings
syntavra status --doctor
syntavra run manifest
syntavra prove plan
```

Normal daily work still happens in the existing coding agent. Syntavra is intended to activate through its skill, MCP integration and host hooks rather than requiring every command to be prefixed manually.

For Codex, the managed MCP entry starts through the same Syntavra launcher using the internal `codex-mcp-bridge` route. User/global installs begin repository-unbound and require `syntavra.project.bind` before repository or process tools can run; trusted project-scope installs may auto-bind only to their exact project.

## MCP profiles

| Profile | Purpose | Maximum public surface |
|---|---|---:|
| `minimal` | Default hot-loop token saver | 10 tools |
| `balanced` | Repository context, output, memory and provider controls | 36 tools |
| `audit` | Full inspection and administration | Entire installed catalog |

`tiny`, `optimized` and `full` remain compatibility aliases only. The installed profile, listed tools, callable tools and benchmark profile are derived from the same canonical registry.

## Implemented surfaces and evidence boundary

The repository contains implementations for command rewriting and output compaction, structural repository indexing, bounded session memory, host/MCP integration, provider-usage receipts, and optional local UI or editor integrations.

Implementation presence is not evidence of real-world savings, integration certification, adoption, reliability, or superiority. Registry counts are treated as internal inventory only and are intentionally omitted from the public overview because they can drift and do not measure task success.

```bash
syntavra run rewrite -- git status
syntavra run code-intel report
syntavra run memory-search "cache decision"
syntavra prove plan
```

Evidence rules are defined in `docs/CLAIMS_POLICY.md`; the detailed implementation inventory is in `docs/IMPLEMENTATION_STATUS_001.md`.

## Exact-recovery rule

Syntavra does not treat deletion as compression:

```text
full repository/tool/session evidence
  -> content-addressed local artifact
  -> bounded task-relevant view
  -> recovery handle
  -> exact reveal/verification when required
```

Correctness, verifier success and security boundaries outrank nominal token reduction.

## Measurement rule

The primary optimization metric is:

> **provider-observed cost per verified successful task**

Raw byte reduction is reported separately. Source-level attribution identifies schema, repository, tool-output, memory, conversation and output tokens with an explicit confidence level:

- `PROVIDER_OBSERVED`
- `LOCALLY_TOKENIZED`
- `ESTIMATED`
- `UNKNOWN`

Synthetic fixtures and internal component tests cannot open public superiority claims.

## Primary certification targets

The first live-certification targets are:

1. Codex
2. Claude Code
3. Cursor

Other adapters remain contract-tested or declared bridges until external execution receipts prove live behavior.

## Benchmarking

SignalBench compares independent external arms under the same frozen repository, prompt, model, reasoning mode, context window, verifier, permissions, timeout, cache policy and hardware class. Templates include plain host, Caveman, RTK, Token Savior, repository-context tools, a combined competitor pack and Syntavra minimal/balanced arms.

Missing competitors, provider usage or verifier output fail closed; they are never replaced with synthetic results.

## Canonical documentation

- `docs/001_PRE_RELEASE.md`
- `docs/ARCHITECTURE.md`
- `docs/TOKEN_SAVER_PLAN_001.md`
- `docs/COMPETITIVE_GAP_CLOSURE_001.md`
- `docs/SECURITY_MODEL.md`
- `docs/ADAPTER_PLATFORM.md`
- `docs/SIGNALBENCH.md`
- `docs/OPERATIONS.md`
- `docs/CLAIMS_POLICY.md`
- `docs/IMPLEMENTATION_STATUS_001.md`

## Current claim boundary

```text
FULL_DUAL_ENGINE_PARITY_PROVEN
EXTERNAL_SUPERIORITY_NOT_PROVEN
LONG_CONTEXT_QUALITY_NOT_PROVEN
MEASURED_AGENT_BENCHMARK_NOT_PROVEN
LIVE_INTEGRATION_CERTIFICATION_NOT_PROVEN
DAILY_CODING_AGENT_READINESS_NOT_PROVEN
PUBLIC_PRODUCT_MATURITY_NOT_PROVEN
```

<!--
Internal and simulated Roblox Studio evidence registry markers. These markers identify
claim records only; they do not upgrade any claim to live, public, or independent proof.
[claim:roblox.activation]
[claim:roblox.task_state]
[claim:roblox.capabilities]
[claim:roblox.simulated]
[claim:roblox.transcript]
[claim:roblox.live]
[claim:roblox.datastore]
[claim:roblox.external_engines]
[claim:roblox.tests]
-->

More