# finolabs-vibecode

The finolabs-vibecode plugin is installed.
Your first step on any coding request (a bug fix, feature, refactor, test or review), before reading code or writing a plan, plan mode included: invoke the `coding-router` skill. It picks the workflow and depth and routes to the plugin's other skills.
A delegated run with a complete brief (checkout, task, rules, tests to run) follows the brief instead, and says so in its report.
Invoke `delegate` before handing work to another agent.
A report is a claim, yours or another agent's: before saying work is done, fixed or passing, run the targeted tests yourself and read the output (counts, failures, skips); a missing setup, service or credential is BLOCKED, not PASS (advisory, agent judgment).
Invoke `checks` before accepting a delegated run, and when choosing which tests a change needs (advisory, agent judgment).
Name this session when its first task starts, when the task or phase changes and after every compaction, not for every step: `<emoji> <project> <scope>: <activity>`, near 50 characters, scope optional, activity verb-led, one emoji for the kind of work (🐛 fix, ✨ feature, 🔍 review, 🧭 plan; the full set is in `tidy` references/titles.md). Where the host supports titles, set it with the Claude Code desktop app's session-title tool (load it through tool search when it is deferred); in a terminal ask the user to type `/rename <title>`; where the host has none, say so in one line. A delegated or headless run is named at launch instead (advisory: no hook renames).
Ask the user's decisions through the host's native question tool (`AskUserQuestion` on Claude Code, `request_user_input` on Codex, `ask_question` on Antigravity), in the language the user is using: a question, description or choice label a skill quotes is the meaning to convey, so translate it, keeping command names, flags, paths and skill names unchanged. Put the recommended option first (a plain yes/no confirmation is neutral, with nothing recommended), even when another skill's text only says "ask", and use plain text only when there is no such tool, it fails, nobody can answer, or the answer is free text (advisory, agent judgment).
