Appendix: Glossary
| Term | Definition |
|---|---|
| Workspace | The smallest unit of execution; binds a local directory and a config set; called a "project workspace" in the UI |
| Session | One continuous conversation context, persisted locally |
| Work mode | One of five: auto / plan / prepare / execute / assess |
| Assistant / Persona | The prompt set that decides AI style and boundaries |
| Tool | A capability the AI can invoke (read file, run command, inspect sample, etc.) |
| Skill | A prompt-template pack giving the assistant domain-expert capability |
| Subtask | A subordinate job the assistant dispatches to an isolated sub-assistant |
| Plan / Todo | Structured task list; preserved across compaction |
| History compaction | Compresses older messages into a summary; originals are kept |
| Approval | The deny / allow-once / always-allow decision before tool execution |
| Companion messaging | Driving the AI assistant from an IM channel |
| Plan quota | Model-call quota that comes with your account's plan, metered per period (5-hour / 7-day) |
| Error drawer | The panel that aggregates recent tool and service errors |
| Binding code | 6-digit / 10-min code for IM binding |
| SBOM | Software bill of materials — lists the project's third-party components (CycloneDX / SPDX / DSDX / SWID formats) |
| VEX | Vulnerability Exploitability eXchange (OpenVEX) — states whether a given vulnerability actually affects this project |
| OSV / KEV | OSV: open-source vulnerability database; KEV: CISA's Known Exploited Vulnerabilities catalog |
| Self-repair loop | The mechanism that auto-picks a repair strategy on errors (compact / switch provider / backoff / terminal) — §8.5 |
| Parallel fan-out | Using AgentParallel to split a task into multiple legs running at once, then aggregate — §15.1.3 |
| Auto-reflect | Automatic retrospective (GRAI+KISS) at task wrap-up, sedimenting lessons to memory for review — §17.6.4 |
| Restate intent | Before starting on the session's first message, the AI restates the goal / plan for your confirmation — §5.4.1 |
| Whole-repo search / RAG | code.search / code.ask: non-vector (zero-embedding) symbol-level whole-repo search and Q&A — §10.10.5 |
| Trace | Step-level replay of one run: timeline + aggregate metrics + per-tool metrics, splittable by run and drillable into subagents — §9.4.2 |
| Local mode | A session-level switch that disables the built-in internet tools in one click (shell and external MCP network access are not blocked) — §13.2.2 |
| Image input | Sending images with a message to a model that supports vision input; images go out only with their own turn — §5.3.3 |
| Sandbox attachment | Letting a session temporarily reach directories / files outside the workspace; the scope is either "this session only" or "project-workspace-wide" — §5.6.1 |
| Self-reported pause | The internal marker the AI emits when it holds off on todos as you asked, which switches the automatic check off — §16.4.1 |
| Live guidance | Messages you add while the Agent is running; they go into the "Pending" queue and can be sped up with "Prioritize" or "Handle now" — §5.4.2 |
| Marked messages | Messages you mark via right-click, located / copied / exported through the "Marked" strip — §5.4.3 |
| Auto-repair | The strategy of switching to another provider with the same-named model when the selected model is unavailable or the upstream fails — §8.2.4 |
| Effort | Landi N2.7's thinking level (Off / Low / High / Extra high / Max) — §8.2.5 |
| Repetition spin | The model repeatedly outputs similar content without calling tools; cut off within a turn, halted across turns — §9.4.1.3 |
| Session routine | A timed prompt the AI sets in the current session with RoutineCreate — §15.6.1 |
License: AVL Code Proprietary Software License — Copyright © 2024–2026 Antiy. All Rights Reserved. Document version: 2026-10-15, matching app version v0.10.15-alpha.1442.