Agent Sessions & Harnesses
Structured Agent Sessions
Section titled “Structured Agent Sessions”Apollo → Agent Session chrome and the conversation edge. An Agent Session shows the answering model and effort level, provides leased model and effort controls scoped to that session, derives context limits directly from the live stream rather than an estimate, and adds keyboard-reachable turn navigation colored by measured context consumption.
Unified sidebar row actions
Section titled “Unified sidebar row actions”Apollo sidebar → a chat or agent-session row → right-click or ···. Chat rows and agent-session rows share one keyboard-accessible actions menu for opening, reordering, pinning, renaming, grouping, archiving, and deleting. Archived sessions can return through the coding agent’s own saved memory without exposing which vendor or tool produced them.
Reliable live-agent messaging and transcript
Section titled “Reliable live-agent messaging and transcript”Project window → CLI → Message Agent / Interactive Terminal / Transcript. The CLI tab reports durable message-delivery states (so you know whether something you sent an agent actually reached it), supports managed attachments, exposes permission-specific actions, and gives you an independently selectable transcript with display recovery that doesn’t require restarting the agent.
MCP tool: orbit_amend_prompt.
Agent harness registry, routing, and permissions
Section titled “Agent harness registry, routing, and permissions”Settings → AI & Models; Pipeline diagnostics. Orbit detects which coding harnesses are registered on a machine and lets each model tier choose a primary and a fallback harness. Beyond Claude Code, Codex, and Gemini CLI, Orbit can run native CLIs for CodeWhale, Grok, and OpenCode with an explicit access posture and workspace/host boundaries, verified with a preflight check before a tier is allowed to route to them.
The included Luna model
Section titled “The included Luna model”Apollo footer and the mobile PWA’s Apollo header. Apollo identifies its Sivants-included model as Luna and does not offer a separate model selector for it — Apollo always answers as Luna unless you’ve configured a BYOK model in Preferences. Agent Sessions are unaffected by this: they continue to report whichever agent and model actually answered.