Command map for an orchestrating session — cyberine guide

cyberine guide prints an intent-to-command cheatsheet for an ORCHESTRATING agent session (any harness with a shell, not only Claude Code) — the recipe to run first in a fresh session before guessing at a subcommand. It renders from the same command registry (src/core/commandRegistry.ts) that cyberine --help renders from, so the guide can never name a command the CLI does not actually have.

cyberine guide              # every command, grouped by intent (get ready, run one job, research, ...)
cyberine guide research     # filter to rows matching a keyword (name, usage, description, intent, cost class)
cyberine guide --json       # the raw registry rows + the fixed notes, for a script to parse

Each row carries a cost class read from what the command's implementation actually does: free-read (no side effects), writes-local (mutates local files/config/git tree, never dispatches to a paid provider), or spends-quota (can invoke the Orchestrator/adapters — costs money or subscription quota). The guide always prints two fixed notes regardless of the filter: bare cyberine tasks is the board view (cyberine tasks status), and the legacy .local/dogfood/backlog-*.md listing is cyberine tasks dogfood; and anything marked [spends-quota] needs the human operator's explicit yes first, every time.

Known Limitations and Deferred Work. The guide is for the orchestrating session only — it is never injected into fleet workers, and the MCP server's own instructions string only points at it in three lines rather than inlining the cheatsheet (see "Expose Cyberine as an MCP server" below). Cost class is assigned per top-level command from its default/primary behavior; a command whose subverbs mix classes (e.g. cyberine next reads by default but --dispatch <id> spends quota) carries a one-line note instead of a second row — read the note, not just the tag. Keyword filtering is a plain case-insensitive substring match over name/usage/description/intent label/cost class/note, not a fuzzy search.