Glossary

Glossary

Vocabulary used throughout the wiki. One sentence or paragraph per term.

---

Action log — A 500-entry ring buffer of every browser tool call (name, args, success/error, duration, timestamp). Lives in main process memory, streamed to the renderer for the activity ticker. AI can introspect via browser_get_action_log. See AI-Tools-Reference.

Agent — A pane's state record: role, status, current task, queue, shared context. Lives in clusterspace-agents.json. Not the same thing as an "AI agent" abstractly — it's specifically the per-pane state machine the orchestration tools manipulate. See Agent-Orchestration.

Allowed tools — Optional per-goal whitelist of tool names. When set, only listed tools are allowed; anything else prompts approval. More restrictive than risk. See Goal-Policy-and-Risk-Levels.

Approval gate — A modal prompt asking the user to allow / deny a risky AI tool call. Two systems: the legacy regex gate (for password fields, checkout URLs, file uploads on chat panel use) and the policy gate (for goal runs that exceed the declared risk). See Goal-Policy-and-Risk-Levels.

Broadcast mode — Toggle that fans your keystrokes out to every visible terminal pane in the workspace. Ctrl+B. See Broadcast-Mode.

Browser pane — A pane whose content is a Chromium webview with chrome controls, multi-tab, bookmarks, history, downloads. See Browser-Panes.

Bypass mode — Global setting that suppresses every approval gate. For automation rigs without a human. See Settings-and-Configuration.

Checkpoint — A persisted snapshot of a goal's state (id, goal text, criterion, policy, status, step counter, full step log, final report). Lives in clusterspace-goals.json. See Data-Storage-and-Migration.

Claim complete — A transient tool the AI calls to attempt finishing a goal. The runner verifies against the success criterion; if it fails, the loop resumes with the failure detail. See Success-Criteria.

Command palette — Ctrl+P searchable list of every action. See Command-Palette.

Context — In agent state: an array of snippets pushed by sibling agents via share_context. The destination sees them as injected context on its next turn. See Agent-Orchestration.

Conversation — An AIMessage[] history keyed by (provider × workspace × pane). Persisted in clusterspace-ai-memory.json. Per-pane isolation prevents agents in different panes from polluting each other. See AI-Overview.

Critic — A sibling model call fired every N steps during a goal run that judges progress (PROGRESSING/STUCK/ACHIEVED/MISLED). Steers the loop when needed. See Critic-and-Replan.

Denied tools — Optional per-goal blocklist. Always blocks regardless of risk. See Goal-Policy-and-Risk-Levels.

Discover — A button in AI provider settings that probes common ports on a given host to find local providers (Ollama, LM Studio, vLLM, text-generation-webui). See AI-Providers.

Fleet Dashboard — The multi-pane agent status view. Status counts, agent cards, active orchestration goal, event timeline. See Fleet-Dashboard.

Goal — In the autonomous runner: a (paneId, goal-text, success-criterion, policy) tuple the AI pursues until verified done. Different from an orchestration goal (multi-pane). See Goal-Runner-Overview.

Goal Dashboard — Live UI for autonomous goal runs. Ctrl+Shift+G. See Goal-Dashboard.

Goal Runner — The server-side loop that drives the AI through a goal until verification passes or the loop aborts. See Goal-Runner-Overview.

Grid — A workspace's layout: rows × cols + optional row/col fractional weights. 1×1 up to 4×4. See Workspaces-and-Layout.

IPC — Inter-process communication between renderer (Chromium) and main (Node.js). Goes through the preload bridge exposed as window.electronAPI. See IPC-Channels-Reference.

Legacy session name — A tmux session created before the fleet-term → clusterspace rename, named like fleet-term-pane-<short>. The picker offers a "Legacy suggestion" to reattach. See SSH-and-tmux.

Manual criterion — A success-criterion type that trusts the AI's claim_complete rationale verbatim. For exploratory goals where success isn't shell-verifiable. See Success-Criteria.

Model question — A success-criterion type that asks a judge model YES/NO about the AI's rationale. For subjective/visual outcomes. See Success-Criteria.

Orchestration goal — A high-level multi-pane objective (separate from a GoalRunner goal). Tracks taskBreakdown across panes; surfaces in the Fleet-Dashboard. See Agent-Orchestration.

PagedTextResult — A standard tool result envelope for large text outputs. {content, hasMore, nextCursor?, totalBytes, truncated?}. Used by read_terminal_output, browser_get_content, browser_get_axtree. See AI-Tools-Reference.

Pane — A single cell in the workspace grid. Holds a terminal or a browser. May have multiple tabs (terminal panes only — browser tabs are different). See Workspaces-and-Layout.

Persona — A named bundle of system prompt + temperature + capabilities + skill references that shapes AI behavior. Six defaults, user-overridable. See Personas.

Plugin — A user-authored AI tool dropped as a .js file into <userData>/.../config/tools/. Hot-reloaded. See Plugin-Authoring.

Policy — Per-goal risk profile (risk ceiling, allowlist, denylist, sandbox dir). Gates every tool call during the goal's lifetime. See Goal-Policy-and-Risk-Levels.

PTY — Pseudo-terminal. The Node-side process behind every terminal pane. Spawned via node-pty (ConPTY on Windows). See Terminal-Panes.

Recipe — A stored sequence of browser tool calls that can be replayed deterministically. JSON-defined. Run via browser_run_recipe. See Browser-Recipe-System.

Renderer — The Chromium process running React. UI lives here. Talks to main via IPC. See Architecture-Overview.

Risk level — Per-tool tier in the policy ladder: read_only < write_local < network_get < network_write < spends_money. See Goal-Policy-and-Risk-Levels.

Sandbox dir — Optional per-goal directory prefix restricting file-touching tools. Tools wanting to write outside prompt the user. See Goal-Policy-and-Risk-Levels.

safeStorage — Electron's OS-keychain wrapper. Encrypts API keys, SSH passwords, browser saved logins. See Data-Storage-and-Migration.

Shell criterion — A success-criterion type that runs a shell command and checks the exit code. Most reliable. See Success-Criteria.

Skill — A reference document a persona's body can cite. Not auto-injected. Defaults: terminal-automation, claude-code-interaction. See Skills.

SSE — Server-sent events. The protocol used by OpenAI-compat providers to stream chat completions. See AI-Overview.

Step protocol — declare_step → action → verify_step cycle that personas mandate for non-trivial work. Forces the AI to articulate plan + observation. See AI-Tools-Reference.

Success criterion — How the Goal-Runner-Overview decides a goal is done. Four types: shell, model_question, json_predicate (deferred), manual. See Success-Criteria.

System prompt — The first message in every conversation. Per-provider, configurable. Default is a 700-line template covering environment, tools, step protocol, web patterns. See AI-Providers.

Tab (browser) — A page in a browser pane. Multi-tab, persisted with title/favicon/URL. See Browser-Panes.

Tab (terminal) — An independent PTY connection inside a terminal pane. Each tab = its own SSH+tmux session. See Per-Pane-Tabs.

Task template — A reusable multi-step workflow with per-step persona assignments. Markdown file with frontmatter. See Task-Templates.

Tmux — Terminal multiplexer used on remote hosts to keep sessions alive across pane close / network drops / app restarts. See SSH-and-tmux.

Tool — A function the AI can call. ~52 built in + any user plugins. Registered in toolRegistry. See Tool-Registry.

Tool context — The runtime services passed into every tool's run function (window, ptyManager, stores, vision helpers, state). See Tool-Registry.

Vision verifier — A model call that takes a screenshot and answers a strict YES/NO question. Used by browser_verify_visual_state. See Vision-Verification.

Wall clock cap — Per-goal hard time limit. Default 1 hour. Goal ends as failed if exceeded. See Goal-Runner-Overview.

Workspace — A named set of panes laid out on a grid. Independent grids per workspace. Switch with Ctrl+1-9 or the tab bar. See Workspaces-and-Layout.

---

See also

• Home — start here
• FAQ — short answers to common questions
• Troubleshooting — symptom-driven debugging