| name | nw-mode-select |
| description | Choose human-on-the-loop vs auto mode for a piece of work, classify it S/M/L, and pick the matching path before starting. Load at the START of any nWave-adjacent task, before dispatch, when the mode/size has not already been declared by the user in this conversation. |
| user-invocable | true |
nw-mode-select — activation and mode UX (K3-A)
This is a decision skill, not a runtime. It does not sequence, gate, or ledger anything — it tells YOU, the LLM, which of three conversational postures to take before you start real work, using only instructions you execute yourself.
Step 1 — Is a mode already explicit?
If the user already said "human", "auto", "direct", "just do it", "walk me through it", or otherwise pinned a mode in this conversation, do NOT re-ask. A generic authorization to act autonomously counts as auto: phrases like "work autonomously", "make reasonable choices/decisions", or "use your best judgment" pin the mode exactly as explicitly as the literal word "auto" — proceed as auto without asking again. Direct mode is legal only for S: a single small, unambiguous, already-scoped action is named "direct mode" in your reply and proceeds. If an explicit direct request classifies M or L, say that direct cannot bypass the M/L floor, preserve its act-now intent as explicit auto, and emit auto M or auto L — never an illegal direct M/direct L marker and never a second question.
An explicit mode still invokes this skill once, every size included: a mode pinned in conversation removes only the re-ask question in Step 3 — it never removes the one required nw-mode-select invocation before dispatch, S included. Never re-ask a pinned mode.
Step 2 — Classify size (S/M/L)
Ask yourself, from the request text and repo evidence — not from vibes:
- S (small): a bounded edit that preserves product behavior and needs no delivery authority or architecture decision; reversible and testable in one pass.
- M (medium): one independently deliverable vertical, including any new or changed product behavior, even if it touches one file.
- L (large): spans an epic/mission, multiple independently deliverable outcomes, or work whose shape is not yet known and needs DISCUSS/decomposition; DESIGN remains gap-triggered, never automatic.
State the classification and the ONE observable reason for it (behavior change,
authority gap, outcome count, or unknown shape) before proceeding.
Then emit exactly one standalone line, once:
NW-MODE-SELECTED: <direct|human|auto> <S|M|L>
Valid combinations are direct S, human S|M|L, and auto M|L. This is an
ephemeral handoff observation, never a file, setting, receipt, or workflow state.
A root tool action with a missing, malformed, or conflicting line is blocked
because the selected route would be ambiguous.
Step 3 — Choose the path
| Size | Default mode | What that means operationally |
|---|
| S | direct | Invoke this skill once, classify S, then exit directly and act, reporting the diff — S never delegates to nw-auto. |
| M | ask: human or auto | State the classification and ask ONE question: "human-on-the-loop (I project each stage to HTML and wait for your GO) or auto (I ask once, then run with minimal interaction)?" |
| L | human-on-the-loop by default | Large/uncertain shape defaults to staged human review; auto only if the user explicitly overrides. |
Never infer "auto" for an L-classified request from silence — silence on an L request means ask, not assume. Silence or the absence of a reply channel never manufactures authorization, on M or L: an unattended or headless session with no reply forthcoming still asks and waits. Only an explicit phrase already given by the user (Step 1) pins the mode without asking.
Human-on-the-loop: what "project to HTML" means here
Human mode is a sequence of LOCAL HTML documents you produce and hand to the Artifact tool (or write to disk if there is no artifact channel), one per stage the work actually needs — stopping at each stage for the user's GO. Reuse an existing generator instead of inventing a new templating layer: scripts/gen_status_dashboard.py (backlog SSOT → Artifact HTML) covers this shape; if a stage doesn't fit it, write a small self-contained HTML file directly (inline CSS, no build step) — do not build a template engine or a second dashboard generator. If no reusable mechanism applies, fall back to a plain markdown turn instead of manufacturing new infrastructure.
Auto mode: what "minimal context" means here
Auto mode asks for authorization exactly once — naming what you are about to do and its blast radius — then executes with minimal further interaction: no per-stage HTML, just the working diff and a short end-of-run summary. After classification, delegate explicit Auto M/L to nw-auto; that skill is the sole route authority. Do not restate or execute its M/L algorithm here, and do not route Auto through nw-deliver, nw-distill, or a generic wave command. Direct S and Human routes are unchanged.
After NW-MODE-SELECTED: auto M or NW-MODE-SELECTED: auto L, the next tool
call is exactly Skill(nw-auto). Do not inspect the repository, dispatch a
role, or mutate anything between the marker and that Skill call.
What this skill is explicitly NOT
- Not a sequencer, not a state machine, not a new CLI verb.
- Not a ledger or receipt system — direct-mode and auto-mode choices are conversational, never persisted as a workflow record.
- Not a fork of the wave spine (
/nw-*). Human-on-the-loop may continue through its existing wave route; direct S remains direct. Auto M/L delegates only to nw-auto as stated above.
- Not a new gate — the existing root activation path (the reused
PreToolUse/Agent hook) that surfaces this skill adds no hook code of its own; it may block the first mutation until this one selection is observed, then stays silent.
