USER-INVOKED ONLY. Do NOT auto-trigger. Convene 3 fresh Codex subagents to give independent opinions on a question, idea, hypothesis, plan, or ambiguity. Use only when the user explicitly invokes this skill ("$council", "/council") or explicitly names it ("run council", "use the council skill", "convoque le council"). Never trigger on musings, opinions, or phrases like "je me demande si" — thinking out loud is not an invocation. If invoked with no extra prompt, infer the current subject from the active conversation and ask clean-context subagents to review it.
Drive and visually inspect a native GUI app (Iced, egui, GTK, winit…) on Hyprland through a headless output, leaving the user's desktop untouched (focus preserved, cursor restored). Use when asked to "test the GUI", "piloter l'app native", "screenshot the app", "voir le rendu", to verify a native app's rendering after a change, or to interact (keys, text, clicks) with a native window that exposes no AT-SPI tree. Do not use for web or Electron apps (use agent-browser), apps exposing a usable AT-SPI accessibility tree (prefer a semantic tool), or compositors other than Hyprland.
Use when building, operating, or resuming a deterministic Workflow() driver for a large autonomous multi-phase task — a full rewrite, a migration, a long audit. Triggers: design or implement a workflow driver script, translate a .agents/loops/*.md runner spec (looper output) into a Workflow() script, design phase gates or pause/resume mechanics, handle a paused driver run (retryStep bump, human arbitration, mutation suspected), harden a driver against non-deterministic agent behavior, or wrap an external CLI (e.g. codex exec) inside a workflow agent.
USER-INVOKED ONLY. Use only when Augustin explicitly invokes $skill-sync or asks to commit, push, and deploy agent skill changes live from agents-skills. Commits scoped source, pushes the default branch, publishes exact managed skill names through the shared post-push gate, and verifies live parity without touching project-pinned copies.
Portable discipline for designing, running, and interpreting skill evaluations with skill-creator. Use when defining eval success criteria, running or comparing trials, grading outputs or end states, reviewing traces, benchmarking a skill, iterating on evals, or maintaining eval artifacts. Separates experimental validity from artifact integrity and harness-specific formats.
Write, review, and harden non-trivial Bash and POSIX sh scripts for production. Use for shell files and CI, deployment, administration, or automation code involving untrusted input, destructive filesystem operations, privileges, pipelines, traps, signals, concurrency, portability, or failure handling. Skip trivial interactive commands and tasks better served by an existing project tool.
Generate a handoff prompt for another agent instance. Use when the user asks to write a prompt, meta-prompt, or briefing for a separate agent session — phrases like "write a prompt for another instance", "make a handoff prompt", "brief another agent", "prepare a prompt I can copy-paste", "write this up for another session", or when the user has been discussing/planning a task and wants to delegate execution to a fresh agent instance. Also triggers on "handoff", "meta-prompt", "delegate to another agent".
USER-INVOKED ONLY. Use only when the user explicitly invokes $premium-handoff or asks to turn a selected target, targetify seed, evidence summary, design/architecture problem, or premium-model target into a compact handoff prompt/artifact for a scarce frontier model. Do not target the whole repo, implement, or launch the premium model.