orbit-pilot-operator

name	orbit-pilot-operator
description	Run Orbit Pilot launch workflows from Claude: CLI as source-of-truth contracts, optional browser (computer use / MCP) for manual sites, Playwright path via orbit work --playwright or orbit publish. Use when: user is launching a product, clearing the manual queue, or wants agent-driven orbit pipeline + mark-done. Packaged under apps/orbit-pilot/claude-skills/ for export with the app.
metadata	{"author":"Etc mono-repo","version":"1.0.2","companion_docs":["../../AGENTS.md","../../HUMAN_GUIDE.md","../../../../ideas/orbit-pilot/SPEC.md"]}
allowed-tools	["Read","Grep","Bash","Edit"]

Orbit Pilot operator (Claude skill)

Orbit Pilot is CLI-first: one launch.yaml, a platform registry YAML, generated packs under out/<campaign>/run-*/, JSON outputs, and audit. This skill is the high-level procedure; the executable contract is the CLI + bundled JSON Schemas.

Install in Claude Code: copy or symlink this folder to .claude/skills/orbit-pilot-operator/ in your project (see ../README.md).

Three layers (use together, catch drift)

Layer	What it is	Drift risk
This skill	Narrative workflow for Claude + optional browser tools	Doc lag vs CLI
CLI + schemas	orbit …, orbit validate-json, orbit schemas	Code is canonical for flags/output shape
Registry YAML	browser_form_selectors, submit_url, modes	Sites change DOM; selectors break

Drift checks (run periodically or when behavior feels wrong):

1. Read AGENTS.md and HUMAN_GUIDE.md in this package — if this skill disagrees, trust those docs, then update this skill.
2. orbit --help and orbit <cmd> --help — new flags win.
3. In package root: pytest -q — schemas and CLI payloads stay aligned in CI.
4. If Playwright autofill misbehaves but computer-use still works, suspect registry selectors or site DOM change — update seed_platforms.yaml / your registry copy, not only this skill.

Default agent workflow (machine-first)

Run from the directory that contains launch.yaml and the registry (or use absolute paths).

cd /path/to/orbit-pilot   # package root after pip install -e, or operator workspace
orbit pipeline --launch /path/to/launch.yaml --platforms /path/to/seed_platforms.yaml --out out/ --json

• Exit 0 only if ok_all is true (complete launch, doctor ready, check-run OK). Otherwise fix doctor / launch fields and rerun.
• Validate: orbit validate-json pipeline - < pipeline.json

Then either:

• API platforms: orbit publish --run <run_dir> --platform <slug> --json (dry-run) then --execute when ready.
• Manual / browser queue: loop orbit work --run <run_dir> --json (see below).

Human queue loop (orbit work)

JSON for agents:

orbit work --run /path/to/run-* --json

Parse the object:

• kind: "empty" — queue clear for manual/browser-assist items; switch to API publishes or stop.
• kind: "task" — use submit_url, prompt, payload, mark_done_command, optional playwright_assist_command, and operator_agent_guide (stable hints for local-agent vs Playwright paths; same contract for Codex/Cursor/local LLM shells).

Paths:

• --no-open — if you will open the URL via computer-use / MCP instead of the OS default browser.

With Claude computer-use / browser MCP (user’s Chrome):

1. Call orbit work --run … --json --no-open (or allow open — redundant if you also navigate).
2. Navigate to submit_url in the user’s browser; paste or type from prompt / pack text (public marketing copy only — never API tokens or private keys into third-party forms).
3. After the user confirms the listing is live, run the mark_done_command with the real URL (replace <URL>), or invoke orbit mark-done --run … --platform … --live-url … --json.

Playwright inside Orbit (Kernel / local Chrome):

• browser_assisted: from registry browser_fallback_opt_in + policy, or registry manual + risk.allow_browser_assist_manual + allow_browser_automation. Then orbit publish --run … --platform <slug> --execute --browser (secrets + optional ORBIT_BROWSER_CDP_URL). Shorthand: orbit work --run … --playwright — see AGENTS.md.
• Hosted / remote Chrome (CDP): set ORBIT_BROWSER_CDP_URL to the WebSocket or HTTP debugger URL (e.g. Kernel). Playwright uses connect_over_cdp; session cookies stay on that browser. Do not paste CDP URLs containing secrets into chats or commits.

Safety (non-negotiable)

• Do not publish --execute or drive browser posting without explicit user approval for that platform.
• Credentials: env + keyring for API publishers; never paste secrets into directory site forms.
• Respect site Terms of Service; autofill / automation can still be disallowed even with a human nearby.

Package layout (this repo)

• This app: apps/orbit-pilot/ (Python package, tests, AGENTS.md, HUMAN_GUIDE.md)
• Specs / matrix (mono): ideas/orbit-pilot/

Maintaining this skill

When you add CLI commands or change JSON shapes: update this file’s commands section and ensure ideas/orbit-pilot/SAMPLE_OUTPUTS.md (mono) / AGENTS.md match. Treat mismatches between this skill and orbit --help as a bug in the skill until fixed.