goal-flight · git:20260902.6897489 · 2026-09-02 · sha256 60b4a8b9fd893333
goal-flight git:20260902.6897489A
Immutable. This exact content is served forever at /api/v1/blob/60b4a8b9fd893333.
--- name: goal-flight description: "Use when the user invokes /goal-flight or asks for Goal Flight to plan, dispatch, review, recover, or resume a long-running orchestrated repository run from file-backed state." --- # goal-flight Use this skill when a repository task needs durable planning, resumable work, worker dispatch, review flights, or handoff notes that survive context loss. ## Load Order 1. Read the repository `AGENTS.md` first when present. 2. Read `<skill-root>/SKILL.md` as the canonical Goal Flight workflow, where `<skill-root>` is resolved per `## Skill root` below. 3. Read `protocols/guidance-extended.md` by default — non-frontier controllers benefit from its worked examples and expanded rationale for the core's rules; skip only when context is tight (the core alone is complete). 4. Read the newest relevant `docs-private/RESUME-NOTES-*.md` only when the task asks to resume prior Goal Flight work. ## Operating Rules - Treat repository Markdown plus Git as canonical state. - Keep queue, ledger, status, review, and handoff artifacts file-backed. - Treat host-native memory, chat history, and plugin config as advisory. - For significant coding, decompose the task before implementation. - For review flights, write prompts and outputs to files, then summarize only decisions and findings into the live conversation. - Do not rewrite the root `SKILL.md` during setup. Setup registers checked-in wrappers and config only. ## Skill root Live controllers load the installed skill pin at `$GOALFLIGHT_ROOT` or `~/.goal-flight/skill/`, not a live source checkout. Run every `python3 <skill-root>/scripts/...` command from that pin (`$GOALFLIGHT_ROOT/scripts` or `~/.goal-flight/skill/scripts`). When working inside the goal-flight repository itself, `<skill-root>` is the repository root. The project tree is `$GOALFLIGHT_PROJECT_ROOT`, a named checkout, or the live controller project — pass `--project-root` when the cwd is not that tree. The pin is a detached `git worktree` of a goal-flight checkout at a release tag; nothing in install or setup creates it. Create it once with `git -C <goal-flight-checkout> worktree add ~/.goal-flight/skill <tag>` (or point `$GOALFLIGHT_ROOT` at any checkout), then `python3 <skill-root>/scripts/goalflight_skill_link.py --pin`. Doctor warns when this installed wrapper has drifted from the repo copy; resync it with the setup command in `## Setup` below. ## Setup Use setup for host registration and machine bootstrap. Use init for project-local state and execution readiness. From the cloned Goal Flight repository, run: ```shell ./setup.sh --agent codex ./setup.sh --apply --yes --agent codex ./setup.sh --cursor ./setup.sh --apply --yes --cursor ``` Dry-run output must show every planned copy, merge, link, or registration before mutation. Apply requires explicit approval and writes a machine-local backup manifest for rollback. For Codex, setup registers the desktop-facing plugin and checks the CLI worker surface. Codex Desktop is the likely orchestrator; `codex exec` remains the worker path. Codex setup registers codedb by default when available; context-mode registration is opt-in with `GOALFLIGHT_CODEX_CONTEXT_MODE=1` because dispatched Codex workers disable context-mode by default. For Cursor, setup installs global agent instructions, a personal skill, rules, and Cursor MCP config for context-mode. Use `--cursor-project <project>` for per-project `.cursor/` wrappers, `--cursor-agents-standard` for `~/.agents/skills/`, or `--cursor-link-claude --addons ''` to symlink Cursor to an existing Claude skill checkout. Codex plugin autocomplete lists skills, not nested subcommands. Use `goal-flight-doctor` and `goal-flight-init` when you want command-shaped autocomplete entries, or invoke `goal-flight` with the command text. After setup and host restart, run `/goal-flight init <topic>` in the target project. Init runs doctor, capacity checks, worker readiness checks, and writes compact project-local caveats.