Glossary

Brisal uses each of these words to mean one specific thing, in the app, the docs, and this glossary. Definitions are tight on purpose; follow the link on each term for the full picture.

A

Add — The verb for pulling a system model into a workspace: it creates an owned, editable copy bound to a provider. (Providers are connected, agents are cloned — same mechanic, different verb.) See Models.

Agent — The assistant you actually chat with: a chosen provider and model plus a system prompt, a Thinking level, and generation settings. Where a model is what answers, an agent is who answers. See Agents.

Annotation — A line-by-line review you attach to an existing message in a session, tagging parts as Question / Edit / Note / Remove; when you send it, it becomes a new message to the agent. Reached from message-select mode; navigated by keyboard in annotate mode. See Keyboard shortcuts.

API key — The credential for reaching a provider. Stored in your OS secure credential store, never in Brisal’s config files. Not every provider needs one. See Providers.

Appearance — How Brisal looks for a workspace, across three independent axes: Mode (light/dark), Font size, and Palette (color). Set per-workspace, not globally. See Appearance & theming. Prefer appearance over “theme” for the look as a whole.

Archive — Set a workspace or project aside to declutter the list, without deleting it. Reversible for projects; a lightweight marker on workspaces.

B

Breadcrumb — The location trail at the top of a pane, starting at the Brisal logo. Distinct from the header’s back/forward navigation arrows. See Layout.

C

Capability chips — The small Tools / Vision / Streaming labels on a model; only the supported ones show.

Chord — One key plus any modifiers, written joined by + (ctrl+p, mod+m). The unit a sequence is built from. See Keyboard shortcuts.

Clone — The verb for copying a system agent into a workspace to get an owned, editable copy.

Clone-then-own (also copy-then-own) — The mechanic behind connect, add, and clone: the copy is a full, independent clone — not a live link and not an override. There is no auto-update from the original; the two evolve separately from the moment you copy.

Command palette — A searchable list of every keymap action, with each action’s current bindings shown. Opened with Ctrl+P; the fastest way to run or discover an action. See Keyboard shortcuts.

Connect — The verb for copying a system provider into a workspace to get an owned, editable copy. (The confirm dialog reads “Login to <Provider>”, but the action is connecting.)

Context window — The total tokens (prompt + reply) a model can hold, e.g. “128k context”.

Custom (provider / model / agent) — One defined from scratch rather than copied from a built-in, for something the catalog doesn’t cover. See Providers, Models, Agents.

D

Default agent — The agent a workspace reaches for first when starting a new session. Each workspace marks one of its own agents as default. Distinct from Le Chaton Fat (the primary built-in you clone).

Delete — Permanently erase a workspace or project and everything in it from disk. Irreversible. The selected workspace or project can’t be deleted (nor a locked project). Providers, models, and agents have no hard delete — they are disabled.

Disable / Enable — Set an item aside without removing it. A disabled provider, model, or agent drops out of the pickers; its file is preserved. For providers, disabling is the only write action allowed on a built-in.

disable-model-invocation — A skill frontmatter flag that keeps a skill off the automatic menu while leaving it reachable when invoked deliberately.

F

First-class entity — A description of models: each model is its own record with its own metadata and lifecycle, not a string buried inside a provider.

Font size — The appearance axis that scales the whole UI: an integer from 12 to 20 (default 14). Out-of-range values are clamped, never rejected.

Fork — Branch a session into a new, independent one. The fork keeps the parent’s worktree and tags but starts with an empty transcript — no messages are copied. Parent and child are independent afterward.

Frontmatter — The YAML block at the top of a skill file. description is required; name, disable-model-invocation, and allowed-tools are optional (allowed-tools is parsed but not enforced today).

G

Global — Applies across the whole app. Global controls live in the header — workspace switching and Settings. Contrast workspace-scoped.

GOTO — The keyboard jumper over sessions, workspaces, and projects. Opened with Ctrl+K (or a pre-scoped leader sequence). See Keyboard shortcuts.

H

Header — The full-width bar across the top of the window. Holds global controls: the workspace selector, Skills/Tools shortcuts, navigation arrows, and the Settings menu. See Layout.

Hidden skill — A skill that was discovered but isn’t offered to the agent, with a stated reason (shadowed by a higher scope, model invocation disabled, or a name mismatch). Surfaced, never dropped silently.

Hint bar (which-key menu) — The keymap’s live feedback surface: after you press the leader, it lists the keys you can press next and what each does. Its on-screen position is the hotkey menu position appearance setting. See Keyboard shortcuts.

I

Inherit — In settings: a workspace setting left unset takes the system default’s value. The word for the absence of an override.

Interrupt — Stop a running turn — or a running ! / !! shell command — mid-flight. The transcript records that it was interrupted.

Invalid skill — A skill file that failed to parse or whose id is malformed. Set aside with a reason; it never crashes the skill list.

Isolation — See walled / isolated and organizational isolation.

K

Keybindings — A workspace’s overrides to the default keyboard shortcuts, edited on the Hotkeys tab of Settings and stored as a sparse keybindings.toml. Merge over the system defaults field by field, like appearance. See Config files.

L

Le Chaton Fat — The primary built-in agent and the app’s default: a sharp, complete general assistant, and the seed you clone your own agents from. Distinct from a workspace’s default agent. See also Le Lazy Chaton, the second built-in, tuned for terse output.

Leader (leader key) — The chord (Ctrl+B by default) that opens a multi-key sequence and the hint bar. Ctrl on every platform, tmux-style — not ⌘ on macOS. See Keyboard shortcuts.

Liveness — A session’s real-time state, shown as a dot in the sidebar: Running (a turn is executing), Blocked (waiting on a permission prompt), or Idle. Distinct from whether the session is merely open.

Lock / Unlock — Freeze a project against lifecycle changes: a locked project can’t be selected, archived, or deleted. Reversible, and a user-facing action (unlike the internal workspace lock).

M

Message-select — A modal keyboard mode for moving over the messages in a session transcript — to copy one, or to open annotate mode on it. Entered with ⌘/Ctrl+M. See Keyboard shortcuts.

Mid-session switch — Changing the agent — and so the model — partway through a session. The switch is recorded in the transcript, and both agents count as participants.

Mode — The light/dark axis of appearance: Light, Dark, or System (follow the OS — the default). Every palette defines both variants.

Model — One specific LLM (e.g. Mistral Medium 3.5): its context window, capabilities, pricing, and the provider-side name Brisal sends on a request. It belongs to exactly one provider. See Models.

N

New session funnel — The chip cascade at the top of the sidebar — project → worktree → agent — that sets the target for a new session. Changing a chip doesn’t navigate.

O

Organizational isolation — What a workspace’s wall provides: separation of configuration, not a security sandbox. Agent execution isn’t sandboxed, so never describe workspace isolation as a security boundary.

Orphaned session — A session whose worktree was deleted. It’s archived and set read-only, but its history is preserved. Distinct from a user-archived session, which keeps its worktree.

P

Palette — The color identity you pick by name (e.g. the built-in default), defined by a token-only theme file. The canonical noun for the color axis — not “skin” or “color scheme”. See Appearance & theming and the config-files reference.

Pane — A single content surface inside the PaneHost — the current view (workspace tabs, a project, a session chat). Today one pane fills the PaneHost. See Layout.

PaneHost — The region to the right of the sidebar that hosts panes. Holds one pane today; a later release can split it into several.

Precedence — For skills with the same id across tiers: project > workspace > external > builtin. The higher tier wins; the loser becomes a hidden skill (never dropped).

Private bash (!!) — A shell command you run in a session’s worktree whose output is saved to the transcript but withheld from the model on later turns. “Private” means hidden from the agent, not unsaved.

Project — A unit of work inside a workspace: it holds its own sessions, project-scoped agents and skills, and a source. It can wrap a codebase but need not. See Projects. Not “repo” (it may have no code) and not “folder”.

Provider — An LLM service Brisal talks to (e.g. Mistral), configured with the endpoint, limits, and — for most — an API key. Every model, agent, and session ultimately runs against a provider. See Providers.

Public bash (!) — A shell command you run in a session’s worktree; its output is saved and included in the agent’s context on the next turn.

R

read_skill — The built-in tool an agent calls to pull in a skill’s full body by id. The agent sees a menu of skill descriptions; it reads a body only when it calls this tool.

Reasoning level — The thinking effort chosen per message in the session composer (Off / Low / Medium / High). Distinct from an agent’s own default Thinking setting — this one applies to the single message you’re about to send.

S

Select — Make a workspace or project the current one, deselecting the previous. Selecting a workspace swaps the whole context (sessions, agents, theme).

Selected project — The one project you’re working in within the selected workspace; its sessions are the ones you see and create. Exactly one at a time. Badged Selected.

Selected workspace — The one workspace you’re currently working in; everything below the header is scoped to it. Exactly one is selected at a time. Badged Selected.

Sequence — An ordered run of chords separated by spaces (ctrl+b w n), the multi-key shortcuts the leader opens. See Keyboard shortcuts.

Session — One conversation with an agent inside a project: messages, tool calls, and shell runs accrue as a durable transcript. Created on your first message; survives restarts. See Sessions. Not “chat” or “thread” as the entity.

Sessions jumper — The workspace-wide session list in the sidebar, below the new-session funnel: search, group-by, and per-row actions.

Setting — A single piece of user configuration: mode, font size, palette, or a shortcut. See Settings.

Sidebar — The left column: the jumper for starting and switching between sessions across every project in the current workspace. See Layout.

Skill — A self-contained, discoverable instruction set an agent can pull in for a task — a Markdown file with a description that’s always visible and a body loaded on demand. Where a system prompt is always on, a skill is pulled in only when relevant. See Skills.

SKILL.md — The canonical filename for a package-style skill (skills/<id>/SKILL.md); a flat skills/<id>.md file works too.

Source — Where a project’s files come from, fixed at creation: Scratch (empty, no code), Local (a link to a directory on disk), or Remote (a git clone). Only Scratch is offered in the app today.

System (tier / layer) — The app-wide set of built-in, read-only providers, models, and agents every workspace can draw on, like templates. A workspace sits above this layer. (Skills use their own four-tier model instead — see Tier (skill).)

System (provider / model / agent) — The built-in, read-only, shipped-with-the-app version, acting as a template. You can enable/disable it but never edit or delete it. See Providers, Models, Agents.

System default — The machine-wide baseline settings a workspace starts from and overrides per field; unset fields inherit it.

System prompt — An agent’s standing instructions — its role, tone, and rules — carried into every conversation. Optional, Markdown.

T

Tab — A switcher inside a pane that selects one section of a view (a workspace’s Details / Projects / Providers / Agents / Sessions / Skills / Tools). Tabs stay within one pane; they aren’t top-level navigation. See Layout.

Theme file — The on-disk CSS file that defines a palette (themes/<name>.css), holding only :root (light) and .dark (dark) color tokens. “Palettes are data”: the palette is its theme file. See the config-files reference.

Thinking — An agent-wide reasoning-effort setting: Off (default), Low, Medium, High, or XHigh. A standing knob on the agent — distinct from the per-message reasoning level in the composer.

Tier (skill) — Where a skill lives, in increasing precedence: builtin (ships inside Brisal, always on), external (your ~/.agents/skills folder, opt-in), workspace (shared in one workspace), or project (shipped in a project’s tree, loaded only once the project is trusted). Not an ownership tier — a skill is authored in place, not copied between tiers. See Skills.

Tool — A concrete action an agent can take (read_skill, running a command, …), as opposed to a skill, which is guidance it reads. Reading a skill is itself a tool call.

Tool policy — What an agent is allowed to do beyond plain chat. It — not a skill’s allowed-tools — governs whether the agent may use tools at all.

Transcript — The full, ordered history of a session — every message, tool call, and result — persisted as the record of what happened.

Trust — A per-project safety gate deciding whether Brisal loads that project’s own skills and hooks (which run as you). Deny-by-default. It gates loading, not execution — it is not a sandbox.

Trust decision — Your recorded verdict for a project — trusted or denied — remembered against its location. Re-pointing the source to a different path makes the old decision stale (reverts to deny).

Turn — One round of a session loop: your message plus everything the agent does in response (thinking, tool calls, results, and the final reply).

U

Unread — A session whose latest activity is newer than the last time you opened it — the “a background session finished” signal in the sidebar.

Usage tier — A model’s cost label: Free, Paid, Enterprise, or Experimental, shown alongside its per-token pricing.

W

Walled / isolated — Scoped strictly to one workspace and invisible to others. Projects, provider keys, agents, skills, and theme are all walled; a leak between workspaces is a bug.

Wire name — The provider-side model name Brisal sends on a request (e.g. mistral-medium-3.5), distinct from a model’s internal id.

Workspace — Brisal’s top-level container. Everything you work with — projects, providers, models, agents, skills, sessions, and the theme — lives inside a workspace and is walled off from every other. Its purpose is compartmentalization. See Workspaces.

Workspace (provider / model / agent) — One owned by a single workspace, fully editable and isolated from other workspaces. Obtained by connecting, adding, or cloning the built-in — or creating a custom one. See Providers, Models, Agents.

Worktree — The isolated git checkout and branch a session’s changes live in. A session is bound to one worktree when it’s created.

Workspace-scoped — Applies only to the selected workspace. These controls live inside the pane (the workspace tabs). Contrast global.