Glossary
The glossary has two columns on purpose. The reader-facing shelves may use only the screen word;
docs/build/ uses the implementation word. Keeping the mapping in
one place is what lets a support conversation and a stack trace be about the same
thing.
| Screen | Implementation | Means |
|---|---|---|
| Workspace | container / task | One person’s private environment: their repositories, work in progress and sessions. Runs as a dedicated container, or as sandboxed host processes on the Docker-less target. The Japanese UI always writes it in katakana (ワークスペース), never Latin “Workspace” |
| Workspace action bar | — | The strip that starts and stops the workspace and holds Start, Preview and pane splitting |
| Session | session | One task’s conversation, working location and execution state. It does not imply a terminal |
| Execution method | driver | How Agent Fleet runs an agent and delivers instructions to it |
| Managed | paneless / managed driver | Runs on a shared runtime, operated from the chat view. No terminal |
| Terminal (CLI) | tmux pane / PTY | You drive the agent’s own interactive screen in a terminal |
| Agent | kind | A CLI coding AI: claude, codex, opencode and the rest. See agents.md |
| Agent memory | AF memory (memoryx, af/ in the memory repo) |
Agent Fleet’s own memory, shared by every kind through af’s memory_* tools. Running text says “Agent Fleet memory”; “Agent memory” is the settings tab |
| Agent’s own memory | native memory (claude/codex roots, ?native=1) |
The memory an agent kind keeps by itself: Claude Code’s auto-memory, Codex’s memories, Muse Code’s memory. Not Agent Fleet memory; Memory snapshots version it. Headings may shorten it to “Claude Code’s own memory”. Never “native memory” on screen; “auto-memory” only where the sentence names Claude’s feature |
| Harness | kind | The program that runs the agent loop: Claude Code, opencode, codex, lcpp. This is the session kind, unchanged; “Agent” above is the screen word for it today. Not on screen as “harness” yet (ADR 0128) |
| Route | route (usagex.Record.Route) |
Where a model call went and with which credential: OpenCode Zen, the OpenCode Go subscription, the deployment’s engine (af-engine-llm), a member’s llama.cpp (member-llamacpp). Recorded per call in the usage ledger from the call’s own evidence, never from the session’s current settings; empty means not recorded. Not on screen yet (ADR 0128) |
| Maker | maker / “provider” in model_provider.go |
Who built the model: Anthropic, OpenAI, Zhipu. Not the route: one maker’s model can be reached over several routes. The code’s word “provider” means the maker here. Not on screen yet (ADR 0128) |
| Billing | billing (usagex.Record.Billing) |
How one call was charged: subscription-included, subscription-extra (usage credits past a window), metered, self-hosted, free, or unknown. A call whose billing cannot be told from evidence is unknown (also what an empty value reads as), never “included”. Not on screen yet (ADR 0128), and not the “opencode.ai billing” choice on the opencode card, which picks a route |
| Assistant | assistant chat | A purpose-specific chat that uses no repository. Not a session |
| Mirror | transcript / mirror | The rendered view of a running or stopped conversation |
| Working copy | working copy / dir | The folder of a repository inside the workspace that you actually edit |
| worktree | git worktree | An independent working copy of the same repository, so parallel work does not collide |
| Parent | parent clone | The working copy a worktree was made from, and what status displays compare against |
| Pane | pane | One subdivision of the main area |
| Working set | working set | A named grouping of repositories, conversations, sessions and schedules that narrows the left pane. It moves and copies nothing |
| Shared session | session share | Showing a conversation read-only to another member of the same tenant |
| Handoff | handoff | Passing a conversation to a new session, or to another member |
| Fork | fork at message | Starting a new session from a past point in an existing conversation |
| Child session | spawn / origin session |
A session another session started and handed a task to. The parent may list, watch, stop, resume and rename the children it started, and nothing else; deleting one is yours. See Start child sessions from a session |
| Work item | work item | An issue, ticket or pull request pulled in from a provider. See repos.md |
| Memo queue | memo | Instructions parked now and sent to a session later, in a batch |
| Cleanup / trash | cleanup / shelf | The sweep of stopped sessions, stale worktrees and merged branches. What it removes is stashed and can be restored |
| Browser pane | browser pane | The workspace’s own Chromium rendering 127.0.0.1:{port} into a pane. See browser-pane.md |
| Lightweight preview | preview proxy | The same service opened in a new tab under a /preview/{port}/ sub-path. WebSocket and SSE pass through, but an app emitting absolute-path assets breaks |
| Preview subdomain | preview subdomain | A https://<random>-<port>.<domain>/ URL issued every time the workspace starts. The app is served at the root and several ports are open at once. Not issued on every deployment |
| Shared preview | shared preview | A preview a member of your tenant turned “Show it to your tenant” on for. It opens after you sign in, and not while their workspace is stopped |
| Connection | connection / secrets | A credential you attached from the Console: an agent, a git provider, a tracker |
| Tenant | tenant | One team or department. Members of different tenants are invisible to each other |
| Slot | slot | On the EC2 target, one pooled instance a workspace can be placed on |
| Deployment | deployment | One installation of Agent Fleet. One company runs one |
| Start / spawn | start / spawn | A person starting a session or workspace is “start” (起動). Only a session starting a child session is “spawn” (起こす) |
| Sign in | sign-in (IdP) | Authenticating to Agent Fleet itself. Authenticating to an external CLI or service (claude, Civitai, gcloud) is “log in” (ログイン) in the Japanese UI |
| Log in | login | Authenticating to an external CLI or service (claude, Civitai, gcloud). Authenticating to Agent Fleet itself is “sign in” (サインイン) |
| Deploy | deploy | Used only as a verb (デプロイする). The noun is “deployment” (配備): “deployment default”, “deployment-wide” |
| This browser | browser-local | The scope of settings stored in the browser you are using (Japanese: このブラウザ, never この端末). “端末” is reserved for the terminal emulator |
| Usage limit | usage limit / quota | The cap on how much you may use (利用枠). The cap on the number of child sessions is “child cap” (子の上限); “slot” is only the EC2 sense |
| Child cap | child session cap | The most child sessions one session may start (子の上限) |
| Collapse / fold | collapse / fold | Folding something away on screen only (畳む). Ending a session is written “stop” or “end” (停止 / 終了) |
| On / Off | on / off | Switches in prose are “オン／オフ”, never Latin ON / OFF. When the control’s own name is “有効” / “無効” (enabled / disabled), follow it |
| Default | default | A setting’s initial value is 既定 (never デフォルト) |
| Inference engine | engine / role (llm, image) |
A model server the deployment runs, points at, or borrows (one for chat, one for images). What the top-bar pills report on |
| Engine row | engine row / lifecycle |
One declared engine: the deployment’s own GPU, a ComfyUI on your network (external), one borrowed from another deployment (remote), or any OpenAI-compatible image server. Which row draws a picture is decided by the row, not by the server’s type |
| Borrowed engine | remote engine | An engine another deployment runs and lends through its gateway. Its catalogue is a read-only copy here; starting, stopping and editing happen over there |
| Model catalogue | catalog | The checkpoints, LoRAs and VAEs the engines can load, one row per model, one catalogue per deployment |
| Family | family / base_model |
The checkpoint’s lineage: SD 1.5, SDXL, SD 3.5, FLUX.1, FLUX.2 klein, Z-Image, Anima, Krea 2. It decides the workflow, the default size and which settings are read. The image-generation pane’s card for it |
| Plan card | ingest plan / plan_token |
The quote shown before a model is taken in: every file it needs, what each costs (a download in MiB, or nothing when the bytes are already held), the licence and the warnings. Not the chat plan card a session asks you to approve |
| S3 Bucket | ledger / object store | The tab (“Bucket”) that lists what the deployment’s storage actually holds (including objects no catalogue row declares) and lets you register, move or delete them |
| Image generation (pane) | imagegen pane | The pane that makes pictures on the deployment’s ComfyUI without an agent: trial run, batches, seeds, LoRAs |
| Studio | image studio / ImageStudio / studios/<id>.json |
One image-generation draft and the zero or one session bound to it. It has an id of its own, so switching agents keeps the draft, the edit history and the versions. A person presses generate; the agent refines the draft and may trial one picture |
| Draft | draft | What is in a studio’s form (prompt, negative, knobs, size, LoRAs, operation, reference images, model). Both the agent and you write it; a per-field lock stops the agent’s writes |
| Edit history | draft log / studios/<id>.log.jsonl |
One entry per change to the draft (time, writer, each changed field before and after). “Restore this point” deletes nothing; the restore is itself an entry |
| Version | version / press line |
A copy of the draft at the moment trial, enqueue or the agent’s trial was pressed. A picture remembers the version it came from, and “Restore these settings” returns to that version’s draft |
| Input set (not on screen) | input set / ~/.cache/agent-fleet/generated/console/inputs/<set>/ |
The fixed copy of the reference and mask images taken at enqueue. The renderer reads the copy, never the original, so swapping the original afterwards changes nothing. It goes when its group finishes |
| Image gallery | gallery pane | A folder’s pictures as cards, with folders, covers, counts and an enlarged view |
| Fleet graph | fleetgraph pane | One lane per session, time running left to right: which session started which, what passed between them, and when each was working, waiting or idle. The elapsed counterpart of the sessions overview |
| LoRA | LoRA adapter | A small add-on trained against one family that steers a checkpoint’s style or subject. Listed only for the family it matches |
| Trigger words | trigger words | The words a LoRA needs in the prompt to do anything. Shown on its row before you pick it, and added as chips when you do |