Terminology
Standardized terminology and translations for Aleph concepts
Terminology
This page establishes canonical names for Aleph concepts to ensure consistency across documentation, code, and UI.
Panel top-level modes (顶层模式): the desktop panel has seven, in nav order — Chat (
/chat), Dashboard (/dashboard), Memory (/memory), Agents (/agents), Teams (/teams), Extensions (/extensions), Settings (/settings). Settings groups MCP / Plugins / Skills under an Advanced group; the ClawHub settings tab was removed.
Core Concepts
| English | Chinese | Definition |
|---|---|---|
| Agent | 智能体 | An AI instance with a specific configuration (model, tools, prompts) |
| Harness | 执行驱动 | The Think→Act loop that drives agent execution |
| Thinker | 思考器 | The LLM interaction layer — prompt building, model routing, streaming |
| Orchestrator | 编排器 | Resolves AgentDef + FlowSpec, assembles dependencies, dispatches to Harness |
| Gateway | 网关 | WebSocket control plane — routing, sessions, events, authentication |
| Session | 会话 | A persistent conversation context identified by a session key |
| Turn | 轮次 | One complete Think→Act cycle within a session |
| Flow | 流程 | A single execution request from ingress to completion |
| FlowSpec | 流程规范 | Defines execution parameters for a single flow |
| AgentDef | 智能体定义 | Static configuration defining an agent's behavior |
The Three Twins (user-adjustable)
| English | Chinese | Definition |
|---|---|---|
| exec_tier | 执行等级 | Determines which tools may execute and how they're approved; the permission axis |
| think_level | 思考等级 | LLM reasoning budget and extended-thinking toggle; the inference axis |
| session_mode | 会话模式 | chat / work / code — statically partitions the tool presentation surface (progressive-disclosure core set + deferred-tool tier); does not affect permissions |
| session_set_mode | 模式切换工具 | Runtime tool called by the model to switch mode; the new value must be carried on the same send |
Loop-Graph Governance
| English | Chinese | Definition |
|---|---|---|
| Loop Graph | 治理拓扑 | The governance topology layer on top of the Harness (src/loop_graph/); a topological answer to the four single-loop failure modes (Goodhart, reference blindness, ring conflict, measurement decay) |
| Team Node | 团队节点 | Node kind fusing the multi-agent system into the graph; disband triggers victory-claim watchers |
| Audit Ring | 审计环 | Closed-set audit action log; objective ACL source for graph audits |
| Objective ACL | 客观 ACL | Access control list authorized objectively by graph node identity |
| Victory Claim | 胜利声明 | The action by which a team/loop declares completion; triggers watchers and audit |
| Watcher | 观察器 | Listens for governance events (victory claim, goal settled, team disbanded) and initiates audit |
| loop-auditor | 循环审计员 | Built-in protected agent; audit/supervision templates default to independent-context evidence |
| Goal | 目标 | Persistent goal managed by GoalStore; loop settlement can trigger it |
| LoopState | 循环状态 | State machine of the looping/ subsystem (Active / Paused / …) |
| Pause / Resume | 暂停 / 恢复 | Cross-session lifecycle control over loops and goals |
Memory System
| English | Chinese | Definition |
|---|---|---|
| Memory | 记忆系统 | Persistent knowledge storage (SQLite + sqlite-vec + markdown notes) |
| Memory Hub | 记忆中心 | Unified memory panel hosting Graph view and Table view in one toggle (/memory); node identity == note path |
| Graph View | 图谱视图 | Radial node-link canvas of memory notes (formerly the /memory canvas) |
| Table View | 表格视图 | Faceted list of memory notes (formerly the /dashboard/memory vault) |
| Card View | 卡片视图 | Card list with batch-action bar (the Vault panel's post-refactor list shape) |
| Deep Link | 深链 | In-page anchor that locates a specific note from a card or evidence chain |
| Evidence Chain | 证据链 | Sequence of notes that cite the current note; rendered in the Vault drawer |
| Cross-Agent Selection | 跨代理选择 | Memory selection across multiple agent views; must go through atomic consistency |
| Provenance | 来源 | Source record for a note/evidence (writer, timestamp, source event) |
| Batch Export | 批量导出 | Export multiple cards at once as a deliverable |
| Dream Insights | 梦境洞察 | Read-only panel (Settings > Memory) backed by dreaming.list_insights — recent daily synthesis notes + dream-run history |
| Corrections | 纠正记录 | Read-only panel (Settings > Memory) backed by memory.list_corrections — surfaces the correction → distillation lifecycle |
| Raw Memory (L0) | 原始记忆 | Ephemeral session data before compression |
| Note (L1) | 笔记 | Persistent markdown knowledge files |
| Wikilink | 维基链接 | Obsidian-compatible [[note-name]] cross-references |
| Compression | 压缩 | Converting raw memories into structured notes |
| Dream Daemon | 梦境守护进程 | Background memory consolidation process |
| Scratchpad | 草稿板 | Per-session working memory (discarded on session end) |
| User Profile | 用户画像 | Dialectic user model (USER.md) |
| SkillOpt | 技能优化 | Self-evolution discipline — the memory side that bakes "good-enough stop" into the self-evolution pipeline |
Tool System
| English | Chinese | Definition |
|---|---|---|
| Tool | 工具 | A callable capability exposed to the LLM |
| ToolService | 工具服务 | Unified façade over all tool sources |
| Builtin Tool | 内置工具 | Native Rust tool implementation |
| MCP Tool | MCP 工具 | External tool via Model Context Protocol |
| Extension Tool | 扩展工具 | Plugin-provided tool (WASM / Node.js) |
| Skill | 技能 | A learned or installed capability that adds tools |
| Skill System | 技能系统 | Framework for skill discovery, registration, and lifecycle |
| Dispatcher | 调度器 | Tool registry, risk evaluation, and semantic retrieval |
| Hydration | hydrated tools | Semantic tool retrieval based on query context |
| Parallel Group | 并行组 | A bucket of tool calls in the Act phase, partitioned by resource claim and dispatched concurrently; groups are serial, intra-group concurrent |
| Resource Claim | 资源声明 | A tool's concurrency-conflict declaration over fs/network/process |
| Memo / Cross-batch Dedup | 备忘 / 跨批次去重 | Reuse of the first result for a duplicate (name, args) within or across batches |
| session_compact | 会话压缩 | session.compact tool — equivalent to /compact available on every surface |
| Slash Command | 斜杠命令 | Reference-driven single-source alias table supporting /help and friendly shortcuts |
Extensions
| English | Chinese | Definition |
|---|---|---|
| Extension | 扩展 | User-facing umbrella for Skill + Plugin + MCP server — exactly one kind per Extension. See Extensions Store |
| Extensions Store | 扩展商店 | Top-level panel mode (/extensions) for browsing/installing Extensions by functional category, with trust tiers and pre-install disclosure |
| ExtensionKind | 扩展类型 | Store-facing kind enum {Skill, Plugin, Mcp} — distinct from plugin-host runtime PluginKind {Wasm, Mcp, Static} |
| Aleph Hub | Aleph Hub | Single extension catalog source aleph-hub (hub.heyaleph.com/catalog.json), cold-start projected by hub::primer into ~/.aleph/hub_catalog.db |
| Hub tools | Hub 工具 | hub_catalog_sync / hub_fetch_docs / hub_resolve_spec / hub_install_run / hub_install_verify, registered in the main built-in tool registry and constrained by trust and user-consent gates |
| Trust Tier | 信任层级 | Source trust level: Official / Verified / Community / Unverified |
| Disclosure | 披露 | Mandatory pre-install summary (command+args, secrets, network/fs reach, pinned version + SHA256, tier) |
| Source | 源 | A catalog provider (plugin marketplaces, Official MCP Registry, Docker MCP Catalog) cached in SQLite for offline browse |
| Plugin.install (unified) | 插件统一入口 | plugin.install no longer guesses source; classifies the source (marketplace / zip / path) and routes uniformly |
| Plugin Marketplace | 插件市场 | Indexable directory exposed via plugin.marketplace.list / plugin.marketplace.add |
Security
| English | Chinese | Definition |
|---|---|---|
| Sandbox | 沙箱 | Per-session execution isolation with capability enforcement |
| Capability | 能力 | Permission declaration (filesystem, network, process) |
| Capability Ledger | 能力账本 | Per-install visible capability list rebuilt from disk (used in pre-install disclosure) |
| Approval Gate | 审批门 | Human-in-the-loop approval for elevated permissions |
| fail-closed | 失败关闭 | Refuses authorization when a policy file is missing or permissive_default disagrees with Default::default() |
| IdentityContext | 身份上下文 | Immutable identity snapshot flowing through execution |
| Per-Agent Signing Key | 每代理签名密钥 | Each agent owns an independent signing key that writes the operation ledger |
| Operation Ledger | 操作账本 | Signed operation sequence owned by the signer; replayable against the agent's identity chain |
| Delegated Role Chain | 委托角色链 | Delegation chain; lifecycle is closed within the chain |
| Signer Ownership | 签名者所有权 | The private key is held by the corresponding agent; the ledger is replayable against the chain |
| PolicyEngine | 策略引擎 | Stateless permission checker |
| GuestScope | 访客范围 | Tool-level restrictions for guest sessions |
| Pairing | 配对 | Device trust establishment via PIN/code |
| TOFU | 首次使用信任 | Trust-on-First-Use: first encounter with a self-signed cert prompts the user to approve and pin fingerprint + SAN |
| SAN Drift | SAN 漂移 | Change in the host's non-loopback interface IP set; triggers atomic cert regen (never bricks boot) |
| Self-Signed TLS | 自签名 TLS | The gateway terminates TLS in-process; remote connections must use wss:// |
| Trusted Proxy | 可信代理 | Reverse-proxy trust list — decides whether the gateway reads the real client IP |
| Trusted Origin | 可信来源 | Origin allowlist — used by browsers and WebViews for cert trust and microphone permission |
| Guardian Judge | 守护者判官 | The LLM judge used by heartbeat probes; its payload must mask secrets |
Communication
| English | Chinese | Definition |
|---|---|---|
| Interface | 接口 | A client connection type (CLI, Telegram, Discord, iMessage, etc.) |
| Channel | 频道 | A specific conversation endpoint within an interface |
| Session Key | 会话密钥 | Unique identifier for a session (e.g., agent:main:main) |
| Event Bus | 事件总线 | Pub/sub system for real-time event distribution; broadcast::RecvError::Lagged must recover rather than die permanently |
| JSON-RPC | JSON-RPC | Request/response protocol over WebSocket |
| Trace | 追踪 | Execution event stream for observability |
| Error Receipt | 错误回执 | Single-sourced user-readable error returned by the gateway; the raw error chain never leaks out |
| Dead Letter | 死信 | Messages that failed to deliver via channel.*; re-drivable from the Panel |
| Webhook Backpressure | Webhook 反压 | Webhook receiver returns 503 under congestion rather than silently dropping |
AI/LLM
| English | Chinese | Definition |
|---|---|---|
| Provider | 提供商 | LLM API vendor (OpenAI, Anthropic, Gemini, etc.) |
| Protocol | 协议 | Adapter for a specific LLM API format |
| Model | 模型 | A specific LLM instance (e.g., gpt-4o, claude-sonnet) |
| Prompt | 提示词 | Input text sent to the LLM |
| PromptLayer | 提示层 | One section of the assembled system prompt |
| Streaming | 流式传输 | Real-time token-by-token response delivery |
| Thinking | 思考 | Extended reasoning mode (Claude's extended thinking) |
| In-core Pricing | 核心定价 | session.usage is computed in-core from a single pricing table; the shell no longer carries a duplicate price table |
Advanced Features
| English | Chinese | Definition |
|---|---|---|
| A2A | A2A | Agent-to-Agent protocol for inter-agent communication |
| ACP | ACP | Agent Client Protocol for agent discovery and delegation |
| ClawHub | ClawHub | Legacy registry — no longer a first-class Extensions source; settings tab removed, subsumed into the Extensions Store (long-tail only, read via the Store Agent) |
| Team | 团队 | A leader + members group; on first message the leader plans first ("Team Strata") and the plan is welded into every member's prompt |
| Team Strata | 团队战略协调 | Leader-first strategic planning that coordinates a team's members |
| Coord Task | 协调任务 | A member's assigned task with status Pending → InProgress → WaitingReview → Completed, reviewed by the leader via task_review |
| Group Chat | 群聊 | The three-panel team chat window (roster · attribution-bubble message flow · workspace), powered by the durable per-team thread (teams.chat.*) |
| WorldModel | 世界模型 | Environmental state tracking for proactive behavior |
| Proactive Dispatcher | 主动调度器 | Daemon that dispatches proactive tasks based on world state |
| Desktop Bridge | 桌面桥 | UDS JSON-RPC protocol for desktop automation |
| Deliverable | 交付物 | Work product produced by the artifact_publish tool — landed as ArtifactOrigin::Deliverable; Panel "Deliverables" pane pins and opens it |
| Artifact Store | 工件存储 | src/artifacts/ — unified backend for deliverables (MAX_ARTIFACT_BYTES = 50 MB + MAX_ARTIFACTS_PER_SESSION = 200) |
| ArtifactOrigin | 工件来源 | Inbound / Outbound / Export / Deliverable four origins (wire-vocabulary single source: aleph_protocol::artifact) |
| Export Subsystem | 导出子系统 | src/export/ — renders deliverables as standalone HTML / Markdown documents (self-contained CSS) |
| Voice Subsystem | 语音子系统 | Streaming ASR (Deepgram / WhisperLiveKit, local aleph-voice sherpa-onnx + mock engine) + TTS (multi-provider + fallback cap 2) |
TranscriptDelta | 转写增量 | Streaming transcript contract (committed / interim / utterance_end / error); Panel VAD authoritative |
| Vocabulary Biasing | 域词偏置 | [voice] vocabulary config — translated per ASR backend to hotwords / initial_prompt / prompt / keywords |
StreamingTranscriber | 流式转写器 trait | Provider-neutral contract in src/gateway/voice/streaming/mod.rs |
| Voice-as-Context | 语音即上下文 | ASR-transcribed signal into context aggregation (per-turn transient recall), not persisted as user message |
| Loop | 循环 | src/looping/ — current-session sustained repetition (9 actions); LoopRegistry is in-process only |
| Goal | 目标 | src/goal/ — autonomous objective pursuit (persisted GoalStore + objective gate + TreeBudget) |
| Strategy | 策略 | src/strategy/ — team or naked-loop plan (composite-keyed StrategyStore) |
LoopState | 循环状态 | Active / Paused / Stopped three-state (in-process) |
GoalStatus | 目标状态 | Active / Paused / Completed / Aborted |
GateOutcome | 门判定 | Goal subsystem objective gate type-state (maker / checker dual confirmation) |
| TreeBudget | 预算树 | Goal and loop shared budget tree; any sub-goal exceeding budget forces entire tree to stop |
| Composite Key | 复合键 | goal_key / loop_key / session_key / team_key four namespaces |
| Plan Team / Naked Loop | 团队/裸循环策略 | [strategy] plan_team / plan_naked_loop config switches |
| Atomic Put-if-Absent | 原子首发 | StrategyStore fire-once mechanism (race protection) |
| Run Mode Pin | 运行模式钉 | Leader stamps usage_mode; subagents inherit; rejects session_set_mode mid-run |
| Loop Auditor | 循环审计员 | Built-in protected agent; audit / supervision templates default to independent-context evidence |
loadable | 可加载容器 | data.rs + loader.rs — all fetches route through it; failed fetch unrepresentable as empty |
| Vault Card View | Vault 卡片视图 | cards.rs — three-state shell (Loading / Empty / Content) |
| Evidence Chain | 证据链 | notes_citing — render reverse links into drawer |
| Provenance | 来源三元组 | writer / timestamp / source event (provenance.rs) |
notes_sources / notes_provenance | 笔记溯源表 | 26.7.22+ new; basis of graph cache and insights |
| Louvain | Louvain 社区检测 | Hand-rolled implementation (no external crate, R3 compliant) |
| 4-signal Recall | 四信号召回 | Node relevance + Leiden refinement + wikilink skeleton + co-recall |
| Phone / Tablet Platform | 手机 / 平板平台 | interfaces/webchat/src/platform/{phone,tablet}/ — phone has 8 submodules; tablet currently mainly skeleton |
| UnifiedSaveBar | 统一保存条 | interfaces/webchat/src/platform/wide/views/settings/ — unified save mode for all settings pages |
| Phase 3 / Phase 4 Settings | 第三 / 四阶段设置 | SearchRegistry UI + PII migration + Provider Presets Swift UI |
behavior.input_mode | 输入行为模式 | cut / copy / halo three options (CGEventTap hotkey + clipboard context) |
Naming Conventions
Code → Docs Mapping
| Code Symbol | Document Name |
|---|---|
AgentHarness | Harness / 执行驱动 |
HarnessDeps | Harness Dependencies |
HarnessRunner | Harness Runner |
FlowRequest | Flow Request |
FlowOutcome | Flow Outcome |
SessionService | Session Service |
ToolService | Tool Service |
AiProvider | AI Provider |
StopHookHandler | Stop Hook |
ContextBudget | Context Budget |
SkillPrefetcher | Skill Prefetcher |
TraceSink | Trace Sink |
InProcessActorSessionService | In-Process Session Service |
WorkspaceSandbox | Workspace Sandbox |
OsSandboxDriver | OS Sandbox Driver |
ApprovalGate | Approval Gate |
PolicyEngine | Policy Engine |
IdentityContext | Identity Context |
GuestScope | Guest Scope |
MultiProviderRegistry | Multi Provider Registry |
ProtocolRegistry | Protocol Registry |
PromptPipeline | Prompt Pipeline |
MemoryEnvelope | Memory Envelope |
WorkingMemoryAssembler | Working Memory Assembler |
HybridAssembler | Hybrid Assembler |
NoteFactRetrieval | Note Fact Retrieval |
CompressionService | Compression Service |
DreamDaemon | Dream Daemon |
DesktopBridgeClient | Desktop Bridge Client |
LoopGraph | Loop-Graph Governance |
GoalStore | Goal Store |
LoopState | Loop State |
ArtifactStore | Artifact Store |
ProjectStore | Project Store |
Hub | Hub (Extensions / Secrets / Trust catalog) |
SignerKey | Per-Agent Signing Key |
OperationLedger | Operation Ledger |
Verification | Verifier / Turn-verify context |
Guardrail | Guardrail (Block / Sanitize / Pass) |
Deprecated Terms (Do Not Use)
| Old Term | Replacement | Reason |
|---|---|---|
| OTAF | Think→Act | Architecture changed from OTAF to Harness |
| Agent Loop | Harness | Unified under Harness terminology |
src/agent_loop/ | src/harness/ | Harness migration; the old directory has been removed |
| LanceDB | SQLite + sqlite-vec | Storage backend changed |
| JSON5 config | TOML config | Config format changed |
aleph binary | aleph-server | Binary renamed |
| Skill System v1 | Skill System v2 | Complete rewrite |
| EvolutionTracker | Registry | Part of old skill system |
| SolidificationDetector | EligibilityService | Part of old skill system |
| SubagentTool | (retained as the sub-agent spawn tool) | Coexists with FlowRunTool (src/orchestrator/flow_run_tool.rs); the former handles sub-agent spawn, the latter handles flow orchestration |
| LoopToolRegistry | ToolService | Unified tool façade |
| ClawHub settings tab | Extensions Store | ClawHub demoted to long-tail; settings tab removed |
teams.run | teams.chat.send / Coord Tasks | Method never existed — teams use the chat thread + tasks |
coordination: sequential|parallel|hierarchical | Team Strata (leader-first planning) | Coordination modes were fictional |
| DiminishingReturnsDetector | (removed) | Harness ratchet tightened to the current CEILING (src/harness/tests/budget.rs:335); the hard stop inside the loop violates R10 |
| Telegram local pairing store | Router-singleton pairing store | Channel-local store never took effect |
| Discord parallel nested config layers | Router-singleton config | Parallel layers removed |
| Provider failover deep-clone whole session | Atomic regen marker + single clone | Performance and correctness fix |
aleph.svg hardcoded static path | core_path("/aleph.svg") | Follows the core migration |
Hard stops outside max_iterations | max_iterations + tool-loop verifier only | R10 forbids deterministic completion judgments inside the loop |
EphemeralSession / dual-layer session | SessionEventRecord + TurnId | SessionEvent system |
See Also
- Architecture Overview — Where these components fit in the system
- Concepts — High-level explanation of core ideas