orbit-setup

Orbit Setup

Help the user reach the workflow they want with the smallest necessary setup. Inspect what is already working before proposing changes. Choose a path below; loading this skill is not a request to install every optional component.

Establish the target

Use the request and local evidence to establish the machine/OS, repository, existing Orbit executable/version, and desired use: task tracking, agent-assisted work, unattended delivery, or remote access. Ask only for choices that cannot be inferred, such as the integration branch, provider, or which host owns a shared workspace. Preserve the user's installation method and existing config.

User need	Reference
First installation or another repository	First run; for an agent lane, continue with Provider setup
Add a Copilot, Antigravity or Muse crew; fix provider authentication	Configuration — Provider setup
Pass Claude or Codex credentials, choose a per-crew API key, or diagnose a stale service environment	Provider credentials
Every orbit command fails to load configuration because a crew name contains :; repair the defining file by hand	Configuration — Crew names containing a colon
Explain crew selection, provider aliases or a dispatch refusal	Configuration — Provider identity and resolution
Configure automatic per-draw effort (<crew>@auto bands, workflow.auto_effort.*), check which pools accept it, or verify a resolved effort without dispatching work	Configuration — Automatic effort selection
Configure or verify mid-run steering, including rollout timing and limits	Configuration — Mid-run steering; for the operator procedure, see orbit-orchestrate
Choose an MCP role, configure global MCP roles and toolsets, understand server flag precedence, or inspect the advertised surface	Configuration — MCP roles and toolsets; Remote access
Activity-scoped workspace knowledge: map exact activities in [workflow.agent_corpus] to a shared INDEX.md, verify activity.agent_corpus_applied in the run audit trail, or understand append proposals versus reviewed corpus changes	Configuration — Activity-scoped agent corpus
Experiments: [experiments] settings, approved config overlays, operator lifecycle, shadow/live modes and readouts	Configuration — Experiments
Discover, configure, inspect or stop experimental crew memory (crew_memory.enabled, orbit memory, shadow capture and live delivery)	Crew memory
Executor image input: declaring spec.image_input on an executor, vision = false on a crew, the canary and visual hold, or reading orbit agents vision-probe results	Configuration — Executor image input
Host capability labels: set machine.labels, understand the implicit OS label and requires_labels matching, or inspect placement waits	Host capability labels
Landed-delivery line classes (Analytics › Metrics › Lines changed): workflow.test_paths, workflow.non_code_paths, their defaults and verification; backfilling missing history with migrate --backfill-diff-stats	Configuration — Landed diff classification
Tune a repository's workspace policy from evidence: required validation and baseline commands, footprint locks and generated files, task types, type and complexity crew pools, purpose pools, crew selection, review layers	Workspace tuning
The fixed bounds of fleet notices, the attention feed, the resident evolution ledger and the failing-job warning, and how to verify each	Configuration — Fleet notices, attention, evolution ledger and failing jobs
Crews, workspace policy and settings, admission policy, host pressure, stalled providers, command timeouts, generated-file conflict recovery and prose-only validation	Configuration — Common settings; Crews; Recovery settings
Executor definitions: host OS support, auth probes, inspection tools and shared Cargo-cache grants	Configuration — Executor definitions
Per-tag instruction files and review-skip tag mappings	Configuration — Tag instructions; Review-skip tags
Review baseline commands and owed host evidence	Configuration — Review baseline commands and owed host evidence
Linux execution prerequisites or sandbox failure	Linux sandbox
Windows host (Orbit runs inside WSL2)	Windows through WSL2
macOS: Claude worker token and the clock env file	macOS
Headless browser kit (orbit config browser-kit) so workers can render and screenshot UI changes	Browser kit
Enable or disable a scheduled routine as an operator, verify it, or tell an enabled routine apart from a host-local pause or the host clock	Automation — Turning the scheduler on; Observe and control
Scheduled execution	Automation
Configure task-pilot eligibility or freshness, inspect readiness, or request a later assessment	Automation — Tuning task-pilot; Orchestration — task-pilot
Workspace scripts (target: script:<name> job steps): what a script is, granting or revoking one, grant staleness, network and root grants, or verifying a grant	Automation — Workspace scripts
Recurring QA and other task templates	Auto-tasks
Prompt-size cost estimates, token usage and unknown cost	Operational logs — Prompt-size cost accounting
Choose which shipped routines and auto-tasks an existing workspace carries, reconcile a policy change, or diagnose excluded defaults	Configuration — Seeding policy
MCP clients, SSH federation or remote dashboard	Remote access
Installing plugins, their orbit <ns> commands and dashboard panels, and the .orbit/plugins.yaml pin file	Plugins
Decision channels for plugin answer submission	Plugins — Decision channels
Register Grok's MCP server, understand its shared Claude config and compatibility cases, or verify which source Grok loaded	First run — Grok MCP registration
Set up a host-local application that acts on an operator's consent for exact workspace-bound tool calls, including credentials, callbacks, rotation and revocation	Scoped application authority
Manage a legacy external tool (orbit tool add, list, show, run, enable, disable, remove, doctor)	Plugins — Managing legacy external tools
Owner/replica roles, multiple machines, or coordinating a breaking workspace-layout upgrade across hosts	Multi-host
Single-owner distributed drain setup or claimed-attempt recovery	Distributed drain
Owner completion policy for distributed-drain handoffs (review versus automatic owner landing)	Configuration — Distributed-drain completion policy
Task snapshots, backup and restore	Publication
Upgrade, resource sync, doctor, worktree GC, workspace removal or teardown, or source-remote rebinding; inspect layout migrations or repair SQLite permissions offline	Maintenance
Inspect the installed skill catalog or user-level skill links: use orbit skill list and show, and orbit skill doctor to diagnose dangling links. If a client cannot find a skill, inspect the expected links or run orbit skill link; a healthy doctor result does not certify that every client link exists.	Maintenance — Skill management
Retired features and their leftover keys, commands or toolset names (explorer and agent invocation, crew ladder, regression-risk marks, explore brief, visual review, and the other removals): what to delete, what replaces it, the recognition deadline, and how to verify the cleanup with orbit doctor	Maintenance — Retired features
A removed command, MCP tool name or task input: orbit orchestrator, --role orchestrators, per-action MCP names such as orbit_friction_add, a bare string in a list field, or a model, agent or context task input. Find the replacement and what to delete	Maintenance — Retired features
Compare implementer models or crews now that the orchestrator scorecard is gone	orbit agents rank in model ranking; crew comparison in Configuration — Experiments
Inspect or compact admission receipt storage on an owner checkout	Maintenance — Admission receipt storage
Scheduler/service logs and host incidents	Operational logs
Inspect one CLI audit event by ID	Operational logs — Audit Event Log
A task held after a provider failure: which crews, accounts and hosts it excludes and when it clears	Operational logs — Provider failure holds

Task tracking alone does not require a provider, PR credentials, a scheduler or remote hosts. Add those only for the requested workflow. Agent execution needs an authenticated supported provider and working sandbox; PR delivery also needs its Git/PR credentials. Task search uses local SQLite FTS5 and requires no download.

Apply and verify

1. Inspect installed command help, effective config and relevant prerequisites. Use the owning host's registered tools for durable state; see the shared tool surface for authority and routing.
2. Make the authorized configuration changes. Existing approval carries forward; do not ask again for routine steps already covered. New scheduling, downloads, remote roles or credentials are separate choices when the request omits them. Keep secrets out of task records and examples.
3. Verify the requested workflow through its actual client and destination. A version string is not proof of a working MCP connection; discover the intended workspace and exercise an audited operation. Use disposable fixtures for execution checks unless running real work is authorized.
4. Report what changed, what works, and any specific remaining operator action. Distinguish installed, configured and enabled. Return to orbit for task work or orbit-orchestrate for authorized dispatch.

Preserve existing work

An upgrade must preserve running-client compatibility or refuse before changing live state. Check long-lived MCP and service processes, not only active jobs. Never delete admission/identity files, bypass schema guards, or start a shadow store to recover access. Follow maintenance.

Seeded schedules are disabled until an operator deliberately enables them, except the worktree-gc, task-pilot and crew-memory-dream routines and the host-health auto-task, which seed enabled when the seeding policy selects them. Seeded enabled does not turn the experimental crew-memory feature on: crew-memory-dream mints nothing while crew_memory.enabled is false or no running memory-arm experiment applies. Configure, inspect and stop crew memory with Crew memory. Preserve existing schedule and enablement choices on an established host. Shared repositories need an explicit owner; copying a checkout does not migrate its control plane.