soturail
v1.4.0
Published
Local-first context rails for AI coding agents: reversible terminal compression, progressive repo reading, SDD workflows, hooks, benchmarks, memory and cache-friendly payloads.
Maintainers
Readme
SotuRail
SotuRail is a local-first Context OS for AI coding agents. It prepares compact project knowledge, evidence, workflows, reports and agent-safe exports without becoming an autonomous agent, cloud gateway or required server.
Install
npm install -g soturail
soturail --version
npx soturail --helpSotuRail requires Node.js 20 or newer. TypeScript is the portable default; native acceleration remains optional.
Quick Start
soturail status --agent
soturail knowledge compile project-guide README.md docs
soturail evidence collect
soturail report build
soturail dashboard build
soturail self readiness --v1 --strictAll generated artifacts stay local under .soturail/.
Stable Rails
| Rail | Purpose | | --- | --- | | Context and Project Brain | Progressive repo reading, memory, verified claims and agent briefs | | Knowledge and Evidence | Source-backed knowledge packs, provenance and honest verification status | | Workflow and Harness | Local workflow evidence, lifecycle state, handoffs and safe task coordination | | Evaluation and Skills | Deterministic datasets, golden checks and source-mapped Skill Rail 2.0 packs | | Reports and Dashboard | Local status, static dashboard, observability and redacted agent reports | | Release and Contracts | Schema checks, readiness gates, baselines and release evidence |
Main Commands
soturail knowledge estimate README.md docs
soturail knowledge compile project-guide README.md docs
soturail knowledge verify project-guide
soturail evidence collect
soturail evidence verify
soturail eval dataset init
soturail eval golden
soturail eval regression
soturail skills build project-guide README.md docs
soturail skills lint
soturail tasklet create review-docs
soturail tasklet run review-docs --dry-run
soturail report build
soturail release check --strictSee the documentation index and usage guide for the full surface.
Safety Model
- Local artifacts by default; no telemetry upload.
- No external LLM calls, paid embeddings or mandatory database.
- No arbitrary shell execution through MCP.
- No autonomous code rewriting.
- Evidence distinguishes verified, unverified, inferred and blocked states.
- Tasklets simulate and export work; they are not an agent runtime.
Read Security Boundaries, Security Model and Evidence Provenance Rail.
Agent Hosts
SotuRail exports provider-agnostic context for Codex, Claude, Cursor, OpenCode-compatible hosts, Gemini-compatible hosts, Antigravity-style hosts, DeepAgents-style targets and generic consumers.
soturail agents matrix
soturail agents export --host codex
soturail report agent --agent codexCompatibility is described conservatively in Agent Hosts.
Documentation
- Quickstart
- Documentation index
- v1 contract
- Stable command surface
- Knowledge Rail
- Evidence Provenance Rail
- Agent QA Rail
- Skill Rail 2.0
- Tasklet Rail
- Roadmap
Roadmap
v1.4.0 combines Knowledge, Evidence, Evaluation, Skills and Tasklets into one coherent local pipeline. The former v1.3.0 scope is absorbed into v1.4.0 so source-backed knowledge and provenance ship with their evaluation and reusable-work contracts.
Next directions:
- v1.5.0: Governance and Cost Rail
- v1.6.0: Optional Conductor experiments behind explicit approval gates
See Future Rails.
Development
npm install
npm run typecheck
npm run build
npm test
npm run docs:check
npm audit --omit=dev
npm run release:checkTests are local and deterministic. No network is required for normal commands or test suites.
License
Apache-2.0. See LICENSE.
