Plugin operating rules — plugin/references/*.md
The plugin carries its own defaults for how a session works, so a Cyberine skill behaves the same on a machine whose Claude Code setup has no rules of its own (k6svf). Eight files under plugin/references/ hold them, each with a Ref code: line that changes whenever the file changes:
decision-policy.md: who decides what, never stopping the run, recording decisions withcyberine tasks decide, and the report-and-ack step.delegation.md: model and effort per spawn, at most two worktree agents at once, foreground versus background, stopping agents when done.session-scope.md: the session's scope, which project an artifact belongs to, and how to treat peer sessions and their messages.safety.md: git landing, remote systems, looking before deleting, and secrets.local-folder.md: what goes under.local/and the write-once naming that survives file sync between machines.development.md: no fake results, checks that count, and trusting the source over the docs.report-shape.md: bottom line first, the D-labelled layout for operator decisions, the handover brief after an absence, and relay text.audience-scaling.md: sizing a plan to who the project is for, and its walking skeleton.
Every skill has an ## Operating rules section naming the files it reads and when. In each file the order of precedence is: the operator's own message, then the host repository's rules, then the operator's own global instructions, then the file. The files fill in only where nothing else speaks. plan and board-run.md now point at decision-policy.md rather than carrying their own copy of it.
Run recipe: nothing to run. The skills read the files when they are invoked. To check that the plugin still validates after editing one:
bash scripts/plugin-check.sh # manifest, skills inventory, claude plugin validate --strict
npm run selftest:plugin-install # the installed symlink layout still exposes every skillKnown Limitations and Deferred Work. The files are prompt text: nothing enforces them, and a session that invokes no Cyberine skill never reads them. No hook loads them, because the plugin's hook list is fixed. Sessions driven by cyberine autopilot get their decision rules from src/autopilot/prompts.ts, not from these files. Workers get the code-owned safety floor and the worker defaults (k08sg) instead. The Ref code: lines are not checked by any script yet. The copies were taken from the author's own working rules on 2026-10-28 and are not synced with them, by design. The repo policy skills (test, CI, resources) are not ported yet.
--help on any subcommand is safe (kz4pg)
--help (or -h) anywhere before a -- prints the subcommand's usage and exits 0. Most handlers print their own richer help; agents, chat, tui, setup, reap, maintenance, status and version used to...
Plugin CLI cheatsheet — plugin/references/cli-cheatsheet.md
A Claude Code plugin skill cannot always shell out to cyberine guide before it needs the answer (a skill's own description budget is paid whether or not it ever runs a command), so the same...