Release Notes & Versioning
Versioning
Section titled “Versioning”Orbit’s product specification carries a version number that increments on each meaningful update, with the version and date recorded at the top of the spec. The current major line is the 3.0 architecture rebuild, which decoupled the Electron app from the server and made the server an optional sidecar.
The roadmap, at a glance:
- 2.0 — Foundation. Multi-project monitoring, server management, the Vault, the MCP pipeline, and the Electron shell.
- 3.0 — Architecture rebuild. Electron and server split into separate processes; the SQLite queue replaces the old flat file; mobile remote control becomes possible.
- 3.x — The platform. Model-agnostic chat, Image Studio, design integration, and the credits system.
- 4.0 — Mobile + multi-node. The mobile PWA and Capture SDK, plus running Orbit across several machines.
- 5.0 — Multi-node. Cross-node status aggregation, prompt routing, and aggregated logs and captures.
2026-09-04 — v6 reconciliation
Section titled “2026-09-04 — v6 reconciliation”The largest single documentation catch-up so far, folding in capabilities that had shipped but weren’t yet reflected on this site.
Boards — a durable, cross-project workspace for findings, plans, and running notes, with archive/restore, history, MCP and Apollo access, and a shared window that restores your layout between sessions. See Boards.
Sweep — the visual-review reviewer that powers the Convergence Gate is now documented as its own surface, including a full-canvas workspace mode. See Sweep.
Agent Sessions & Harnesses — structured Agent Session chrome (model/effort controls, measured-context turn navigation), unified sidebar actions across chats and agent sessions, durable live-agent messaging with a real transcript, and an agent harness registry that routes tiers to native CodeWhale, Grok, and OpenCode runners in addition to Claude Code, Codex, and Gemini CLI. See Agent Sessions & Harnesses.
Needs You and integration status — canonical checkout drift, dependency waits that couldn’t auto-resolve, and Needs Review dispositions now have one home, with truthful publication and integration-content status kept separate from that banner. See Needs You.
Git tab recovery surfaces — pull requests with guarded merge, dirty-pull recovery, recovery-ref discovery, unreachable-local-commit warnings, stash inventory, and publication repair doors all now live under one reference page. See Git Tab & Recovery.
Cortex decision memory — governance intake and decision lineage inside Memtrace: stable decision IDs, “why does this exist,” and governing decisions surfaced during approval. See Memtrace.
Multi-node releases — one annotated release tag drives durable per-node build intents, Build All joins Windows and macOS artifacts into a single GitHub draft, and each installation can carry its own header tint for quick visual identification. See Multi-Node, Relay & Bridge.
Smaller additions folded into existing pages: the Bridge relay’s two-step outbound approval and receiver-owned execution grants, Apollo’s included Luna model, prose benchmarks with user-authored rubrics, benchmark corpus contribution consent, the Destructive Command Guard, cache/build-info controls, the Developer API Usage Panel, and several onboarding and Git-tab details previously undocumented.
2026-07-06 — Reconciliation-era capabilities
Section titled “2026-07-06 — Reconciliation-era capabilities”This update consolidates features shipped across the v4 product line, covering quality-control loops, agentic benchmarking, spend controls, and an expanded MCP tool surface.
Convergence Gate — Sweep reviewer output now feeds into a bounded correction loop (Mode A: human-approved fix; Mode B: auto-approved and iterated). The loop exits on a clean review, oscillation, or a configurable max iteration count, then surfaces a pending human-merge card. Nothing reaches the main branch automatically.
Benchmark Mode — agentic runs — In addition to direct API calls, Benchmark Mode now supports real coding-agent spawns (Claude Code, Codex, or any configured runner), each in its own isolated worktree. Each candidate row carries independent tier, effort, context-mode (workspace vs. clean), and Memtrace settings, enabling direct effort-level comparisons within the same job.
Spend Gate and Tripwire Alert — A two-threshold daily cap (soft cap for subscription coverage, hard ceiling that blocks all Sivants-billed calls) gives predictable cost control. The Tripwire Alert adds a persistent UI banner that fires the moment any model outside the Sivants-covered allowlist generates a billing event.
Intelligence Toggles — Six per-category switches at Settings → Intelligence individually suppress background AI calls (impact pre-checks, completion-note enrichment, reference-doc updates, handoff updates, daily narrative, CLAUDE.md drift checks). Turning a toggle off fully skips the call — no API request, no usage row.
MCP tool family expansion — The built-in tool surface now includes: Firestore read tools (orbit_firestore_query, orbit_firestore_collections, orbit_firestore_schema); Godot Engine proxy tools (orbit_godot_*, ten tools covering project management, debug runs, and scene editing); Benchmark lifecycle tools (orbit_benchmark_*); and per-prompt effort override via the effort: routing tag.
Where changes are recorded
Section titled “Where changes are recorded”Per-project, every shipped prompt is recorded in that project’s COMPLETION_LOG.md with its number, description, commit hash, model, and notes — the canonical, prompt-by-prompt history. The About tab in Settings shows the current installed version and license summary.
Finding what shipped
Section titled “Finding what shipped”- In-app: the History/Activity sidebar and the Daily Summary modal summarize recent activity across projects.
- Per project: the Completions tab lists recent entries in reverse chronological order.