Glossary

Agent (coding agent) — the tool that actually runs a prompt and writes code: Claude Code, Codex CLI, Gemini CLI, or a local/custom model. At least one must be registered before any prompt runs.

Apollo — Orbit’s built-in operational co-pilot. Knows your projects live and executes real Orbit operations on confirmation. See Apollo.

Benchmark Mode — a job runner that fans a single prompt across multiple model or agent configurations in isolated worktrees, then compares pass rates, cost, and quality to surface the best run. Supports API mode (direct provider calls) and agentic mode (real coding-agent spawns with tier and effort settings). Nothing applies to your project until you approve the winning run. See Benchmark Mode.

Bridge — a human-gated message channel between two AI instances; every reply waits in an Outbox until approved.

Complexity tier — the abstract routing label on a prompt (Low, Medium, High, Coordinator) that maps to a concrete agent and model in settings.

Completion log — the append-only COMPLETION_LOG.md audit trail of shipped prompts (number, description, commit, model, notes).

Convergence Gate — a bounded correction loop that drives Sweep reviewer output through iterative fix cycles until the reviewer clears, findings oscillate, or a maximum iteration count is reached. When the loop exits it surfaces a pending human-merge card rather than silently committing to your main branch. See Convergence Gate.

Coordinator Mode — a mode that decomposes a large task into parallel subtasks across git worktrees, with two approval gates. See Coordinator Mode.

Firestore tools — the orbit_firestore_query, orbit_firestore_collections, and orbit_firestore_schema MCP tools. All three are read-only: query a collection with optional filters and ordering, list top-level collections, or sample documents to infer field types. Requires a Firestore service account configured in the Vault.

Ghost ship — a prompt that reports success without committing a real change. See Ghost Ships & Validation.

Godot tools — the orbit_godot_* MCP tools, backed by a lazy-initialized godot-mcp proxy process. Cover project listing and metadata, launching and running projects in debug mode, reading debug output, and scene file creation and editing. Requires godotMcp configuration in ~/.orbit/config.json.

Handoff — a structured session summary stored under context/handoffs/ that carries project state between sessions.

Intelligence Toggles — per-category switches at Settings → Intelligence that fully suppress specific background AI calls (code-impact pre-checks, completion-note enrichment, reference-doc updates, handoff updates, daily narrative, CLAUDE.md drift checks). When a toggle is off, no API request is made and no usage is recorded. See Intelligence Toggles.

MCP (Model Context Protocol) — the protocol Orbit uses to expose its operations to AI clients and to consume external tools. See Connections & MCP.

Memtrace — Orbit’s semantic code search and Code Graph. See Memtrace.

Node — a machine running Orbit as a peer on your account. See Multi-Node Relay.

Orbit Chat — the multi-provider planning conversation. See Orbit Chat.

Per-prompt effort override — an effort: tag in a prompt’s routing header (after complexity:) that sets the model’s reasoning depth for that specific prompt, overriding the tier default in Settings. Valid values are low, medium, high, xhigh, and max. Can also be set per candidate row in a Benchmark Mode job.

Pipeline — the live cross-project feed of staged, pending, running, completed, and failed prompts.

Project Database — per-project relational or document database access, configured via connection string in project settings. The matching MCP tool families (orbit_neon_* for Neon Postgres, orbit_supabase_* for Supabase Postgres, orbit_firestore_* for Firestore) give agents read-only query and schema-inspection access to a project’s database from Claude Chat without exposing credentials directly.

Prompt — a single numbered unit of work for an agent, carrying a routing tag and waiting in the queue until approved.

Sandbox — an isolated git worktree where a prompt runs without touching the main branch, then is promoted or discarded. See Sandbox.

Sivants Credits — the prepaid, never-expiring balance for AI work that runs through Sivants’ proxy.

Spend Gate — a two-threshold daily cap on Sivants-billed AI calls. Spend below the soft cap is subscription-covered; spend above it draws Sivants Credits; once the hard ceiling is reached all Sivants-billed calls are blocked for the rest of the day. BYOK calls are never gated. See Spend Gate.

Tripwire Alert — a persistent banner that fires the moment any model outside the Sivants-covered allowlist generates a billing event. Appears at the top of the Orbit UI without requiring a log check, and persists until dismissed. See Tripwire Alert.

Vault — local, encrypted storage for API keys and secrets, never sent to Sivants. See The Vault.

Wake Receiver — the small service that lets your phone bring the Orbit server up on demand.