Glossary
Цей контент ще не доступний вашою мовою.
Agency
Section titled “Agency”Proactive mode where your agent watches file changes and surfaces relevant vault patterns without being asked. Enable with agency_enable, then configure watch paths and thresholds. See Capabilities, Agency.
Anti-pattern
Section titled “Anti-pattern”A known bad approach stored in the vault. Anti-patterns surface in searches to warn you away from repeating past mistakes. See types under Entry Types.
Archive
Section titled “Archive”The vault’s archival system. It stores snapshots of vault state, supports backup and restore, and handles vault optimization (compaction, cleanup). Managed via the archive facade. See Sync & Export and Capabilities, Archive.
Auth level
Section titled “Auth level”Permission required to call an operation: read (query data), write (add/modify data), or admin (delete data, reset state). Enforced at tool registration. See Security & Privacy.
Branching
Section titled “Branching”Create isolated copies of vault state to experiment without affecting the main vault. Branches can be merged back or discarded. See Vault Branching and Capabilities, Branching.
The intelligence layer on top of the vault. It tracks pattern strength, maintains a TF-IDF vocabulary, and provides recommendations for plans. See Under the Hood.
Chat facade
Section titled “Chat facade”The facade for chat transport integration: sessions, authentication, response chunking, voice transcription/synthesis, browser isolation, notifications, and message queue. It powers Telegram bots, web chat, and other conversational interfaces. See Capabilities, Chat.
Compound loop
Section titled “Compound loop”The self-reinforcing cycle: capture → vault → brain → plans → work → knowledge extraction → vault. Each cycle makes the next one better. See The Knowledge-Driven Approach.
Context facade
Section titled “Context facade”Handles entity extraction, knowledge retrieval, and context analysis. It analyzes prompts to extract files, functions, domains, and technologies, then retrieves relevant knowledge from the vault and brain. Used internally by the orchestrator. See Capabilities, Context.
Curator
Section titled “Curator”The automated maintenance system that keeps the vault clean: deduplication, contradiction detection, decay scanning, health audits, tag normalization, LLM enrichment. See Under the Hood.
The gradual reduction of an unused pattern’s strength score over time. This keeps stale knowledge from outranking actively useful patterns.
Automatic memory consolidation, vault cleanup, and maintenance. The dream facade runs background processes that deduplicate entries, archive stale knowledge, and resolve contradictions. See Capabilities, Dream.
Domain pack
Section titled “Domain pack”A standalone npm package that adds specialized operations and knowledge for a specific domain (e.g., @soleri/domain-design, @soleri/domain-code-review). Each pack has its own repository and release cycle. Install with soleri pack add <name> and register in agent.yaml. See Domain Packs and Customizing Your Agent.
Domain
Section titled “Domain”A knowledge area (e.g., frontend, security, infrastructure). Each domain gets its own search partition and facade with 5 operations. Added via npx @soleri/cli add-domain <name>. See Customizing Your Agent.
Domain facade
Section titled “Domain facade”An MCP tool entry point scoped to a single domain. Each domain facade exposes 5 operations: get_patterns, search, get_entry, capture, remove.
The gap between what a plan intended and what actually happened during execution. Measured during reconciliation. Low drift means the plan was accurate; high drift means reality diverged. See Planning.
Embedding
Section titled “Embedding”Manages the vault’s vector search layer. Provides status checks, batch rebuilds, and single-entry embedding operations. See Capabilities, Embedding.
Entry types
Section titled “Entry types”The kinds of knowledge stored in the vault: pattern, anti-pattern, rule, playbook, workflow, principle, reference.
Facade
Section titled “Facade”A single MCP tool entry point that dispatches to multiple operations via the op parameter. Every agent has 22 semantic facades plus one per domain. See API Reference.
SQLite’s Full-Text Search extension, version 5. It powers the vault’s text search with a porter tokenizer for stemming.
An approval checkpoint in a plan or loop. Plans have a two-gate system (approve plan, then approve tasks). Loops use gates to decide whether to continue iterating.
Governance
Section titled “Governance”Controls how knowledge enters the vault: capture quotas, proposal gates, duplicate detection, and decay policies. Presets: strict, moderate, permissive. See Customizing Your Agent.
A quality gate that runs automatically during development. Hooks check code against rules (e.g., no console.log, no any types) and block violations. Managed via npx @soleri/cli hooks. See Customizing Your Agent.
Intake
Section titled “Intake”The facade for ingesting external content (URLs, text, PDFs, books, batch imports) into the vault. Content is chunked, analyzed, and converted into structured knowledge entries. See Content Ingestion and Capabilities, Intake.
The vault’s linking system, inspired by Zettelkasten. Entries are linked with typed relationships (extends, supports, contradicts, sequences). Links enable graph traversal and orphan detection. See Entry Linking & Knowledge Graph and Capabilities, Links.
Knowledge extraction
Section titled “Knowledge extraction”After a plan completes, the brain examines the session (tools used, files modified, outcomes) and proposes reusable patterns back into the vault.
Knowledge pack
Section titled “Knowledge pack”A bundle of pre-built knowledge entries (patterns, anti-patterns, principles) that can be installed into an agent. See Customizing Your Agent.
An iterative validation cycle (do → validate → fix → repeat) that converges on a quality target. Built-in modes: token-migration, contrast-fix, component-build, plan-iteration, custom. See Validation Loops.
Operator
Section titled “Operator”The engine module that tracks the human operator’s profile: expertise levels, corrections, interests, and work patterns. It learns silently from interactions and adapts agent behavior over time. See Operator Learning and Capabilities.
Pack authoring
Section titled “Pack authoring”The process of creating installable capability bundles (knowledge packs, domain packs, hook packs, skill packs) for distribution. Packs are npm packages or local folders that follow the Soleri pack schema. See Creating Packs.
MCP (Model Context Protocol)
Section titled “MCP (Model Context Protocol)”The protocol that connects your agent to your AI editor. Your agent registers tools over MCP; your AI editor calls them based on your conversation.
MCP Bridge
Section titled “MCP Bridge”A chat facade subsystem that allows local tool execution from chat platforms (Telegram, web). Registered tools can be called via the bridge with output compression for display. See Capabilities, Chat.
Pattern
Section titled “Pattern”A proven approach stored in the vault. Patterns surface in searches and brain recommendations to guide future work.
Pattern strength
Section titled “Pattern strength”A score (0.0 to 1.0) tracked by the brain. It increases when a pattern is used successfully, and decreases with disuse (decay) or drift. Higher strength means more confident recommendations.
Playbook
Section titled “Playbook”A multi-step procedure with validation criteria at each step. Created from repeated workflows and run on demand. Supports start, step, complete, and match operations. See Planning.
RateLimiter
Section titled “RateLimiter”Transport-level request throttling. It tracks per-client request counts within configurable time windows and blocks requests that exceed the limit. See Transports.
Review
Section titled “Review”A governance workflow where vault entries can be submitted for review, approved, or rejected before becoming active. Used in strict governance mode to gate knowledge quality. See Knowledge Review Workflow and Customizing Your Agent.
Routing
Section titled “Routing”The control module’s intent detection system. It analyzes user messages to classify intent (BUILD, FIX, REVIEW, PLAN, IMPROVE, DELIVER) and route to the appropriate workflow. See Capabilities, Control.
Reconciliation
Section titled “Reconciliation”The step after plan execution where the agent compares what was planned vs. what happened. It produces a drift report and accuracy score, then feeds the knowledge loop.
SessionManager
Section titled “SessionManager”Transport-level session tracking. It manages client sessions with TTL-based expiry, unique ID generation, and automatic reaping of expired sessions. See Transports.
Skills
Section titled “Skills”SKILL.md files that define reusable capabilities for an agent. Each skill has YAML frontmatter (name, trigger phrases, description) and a markdown body with instructions. Skills are installed in the agent’s skills/ folder. See Skills Catalog.
The facade for synchronizing vault state with external systems: git push/pull for vault backup, Obsidian sync for knowledge management interop. See Sync & Export and Capabilities, Sync.
Severity
Section titled “Severity”Priority level of a vault entry: critical (must follow), warning (should follow), suggestion (nice to have). Severity affects search ranking, with critical entries always surfacing first.
Multi-tier vault connections. You can connect external knowledge sources (other vaults, databases) and search across all tiers simultaneously. See Capabilities, Tier.
TF-IDF
Section titled “TF-IDF”Term Frequency-Inverse Document Frequency. A text ranking algorithm that weights terms by rarity. “Authentication” scores higher than “the” because it carries more meaning. The brain maintains a TF-IDF index across all vault entries.
Transport
Section titled “Transport”The communication layer between your agent and its clients. Soleri supports four transports: stdio (MCP for your AI editor), HTTP/SSE (REST APIs with Server-Sent Events), WebSocket (bidirectional streaming), and LSP (Language Server Protocol for editors). See Transports.
The agent’s long-term knowledge store. It’s a SQLite database with FTS5 full-text search that stores structured entries (patterns, anti-patterns, rules) organized by domain and severity. Supports branching, scoping, sharing, and multi-tier connections. See Under the Hood.
Workspace
Section titled “Workspace”The project directory where an agent is scaffolded and runs. Contains agent.yaml, instructions/, workflows/, knowledge/, skills/, and auto-generated files (CLAUDE.md, AGENTS.md).
YOLO mode
Section titled “YOLO mode”Autonomous execution mode where the agent skips plan approval gates but preserves all safety invariants (tests, verification). Launch with soleri yolo or activate in conversation by saying “go yolo”. See YOLO Mode.
Vault branch
Section titled “Vault branch”An isolated copy of vault state for experimentation. Create a branch, add entries, and merge back, or delete it without affecting the main vault. See Vault Branching and Capabilities, Branching.
Back to Your Agent, Quick Reference or explore Capabilities for the full operation list.