Skip to content

features.md

Jump to bottom
Fame edited this page Aug 5, 2026 · 1 revision

Features

Everything you need for AI-assisted coding with structural quality guarantees.

Multi-Agent Orchestration

Nine canonical agents coordinate through ECHO with explicit separation of duties. Child agents receive only their authorized tool subset through strict allowlist filtering. Parallel agent work supports exploration, research, implementation, and independent review.

ECHO Perfection Loop

Every code change follows a formal Finite State Machine:

  1. RED — Identify ALL failures and issues with evidence
  2. GREEN — Implement minimal, surgical changes
  3. AUDIT — Independent verification by a separate agent
  4. SELF-CORRECT — Fix any blockers found during audit
  5. COMPLETE — Document results, archive tracking

No code is written without a plan. No plan is accepted without audit. No audit passes without evidence.

Tool Permission Boundaries

Each agent gets exactly the tools it needs via strict allowlist filtering. Detective gets read-only access. Forge gets write access. Thinker gets sequential thinking. No agent has more power than its role allows.

Context Compaction

Four-layer progressive auto-compaction keeps large repositories within model limits. Success is silent, failures are surfaced. Sessions can run through massive codebases without hitting context limits.

Checkpoint and Rewind

Each user turn can persist the pre-edit content of every first-touched file, including subagent writes. /rewind supports:

  • Code only — restore files while keeping the conversation
  • Conversation only — restore the transcript boundary without changing files
  • Both — restore code and conversation together
  • Fork — restore the selected turn into a fresh chat

Retention is bounded to the most recent 20 turns. No Git repository is required.

Fail-Closed Streaming

Incomplete or malformed tool calls are rejected, not coerced. Stale-fragment replacement for placeholder arguments. Tool errors, cancellation, retry, and child-agent failures are surfaced rather than silently treated as success.

Provider Flexibility

Works with multiple inference providers:

  • Ollama — Local-first, free, no API key required
  • OpenCode Go — Hosted gateway
  • OpenRouter — Multi-provider gateway
  • NVIDIA NIM — NVIDIA-hosted inference
  • CommandCode — OpenAI-compatible hosted inference
  • Custom endpoint — Any OpenAI-compatible API

Gateway model context lengths can be resolved from the live catalog.

Deep Research

The Researcher role ships a mechanical deep_research tool:

  • Multi-query web research with concurrency limits
  • URL dedup and domain scoring
  • Citations and graceful degradation
  • Pure search facade — no second LLM

GitHub Integration

A read-only GitHub helper connects to the official MCP server for:

  • PR/issue/CI review
  • Code search
  • Secret scanning
  • Audit trail

Requires SAVANT_CODE_GITHUB_TOKEN environment variable.

Database Helper

Four native tools with adapter-enforced safety:

  • list_tables — List all tables in the database
  • describe_table — Get schema information
  • execute_query — Run SQL queries (read-only by default)
  • analyze_query — Get query execution plans

Safety contract:

  • Read-only by default
  • LIMIT injection for queries without LIMIT
  • SQL redaction for telemetry
  • Destructive DDL blocking

Browser Automation

Supports:

  • Viewport presets (mobile/tablet/desktop)
  • Offline WCAG accessibility scan
  • Optional session persistence

HTML Export

/export writes a fully self-contained branded HTML report of the conversation:

  • Offline fonts (no network required)
  • Collapsible tool rows
  • Per-message and copy-all buttons
  • Branded with Savant Code identity

Slash Commands

Command Purpose
/help Show command help and tips
/new Start a fresh conversation
/history Browse and resume previous sessions
/copy Copy the complete conversation to the clipboard
/export Write a self-contained branded HTML report
/interview Create a structured specification
/plan Create an implementation plan
/review Review code changes
/goal Iterate toward a verifiable goal
/loop Schedule recurring checks
/verify Run typechecks
/permissions View or set the tool permission mode
/rewind Restore code and/or conversation from a prior turn
/health Check provider, Ollama, model, and permission status
/mode List the four modes and their contracts
/model Select or switch the active model
/provider Configure a hosted provider key
/bash Run a shell command
/image Attach an image for multimodal models
/init Create starter agent types and knowledge.md
/login / /logout Authenticate or end the current session

Learn More

Clone this wiki locally