Flow Selector

> Trace. On load, print first: =============== [core/skills/flow-selector/SKILL.md] [core] ===============

Flow Selector

Classify a task into Quick / Standard / Extended by weighing boundary signals, then stop and wait for the human to confirm the flow before any downstream LSA ceremony fires. Per .lsa/VISION.md §4: "the orchestrator picks the flow by chain-of-thought, then states its reasoning and the human confirms or overrides."

Three flows, named by process shape:

• Quick — single pass, no LSA ceremony.
• Standard — light discovery + agent TDD + verify.
• Extended — full spec-driven flow: discover → specify → verify → delegate → reconcile.

Goal

Produce a flow label (Quick, Standard, or Extended) plus a 2–4-sentence rationale keyed to the boundary signals, present both to the human, and wait for an explicit confirmation before any LSA skill fires.

Input

• The user's most recent task description (one or more sentences in natural language).
• The current repo state (only as needed to answer "does a spec already exist for this module?").

Steps

1. List the boundary signals present in the task. Apply the five-item checklist at [../../../.lsa/VISION.md](../../../.lsa/VISION.md) §4 — new module · API/contract change · data-model change · ~5 files · no existing spec.

Observable result: a five-item bulleted checklist with yes / no next to each signal, derived only from the task description (and a minimal repo read if needed for the spec-exists question).

1. Apply the classification table at [../../../.lsa/VISION.md](../../../.lsa/VISION.md) §4 — four worked-example rows mapping pattern → flow.

Observable result: the matched row (or the closest analogue) named in the chain-of-thought.

1. State the chain-of-thought as a one-paragraph summary keyed to the signals from Step 1 and the matched row from Step 2. Observable result: a 2–4-sentence rationale printed back to the human.

1. Propose flow + rationale to the human and stop. Present:

• the proposed flow label (Quick / Standard / Extended)
• the 5-signal checklist from Step 1 (yes/no per signal)
• the rationale paragraph from Step 3
• decision. Prompt voice (per [../output/SKILL.md](../output/SKILL.md) Rule 5). Picker question: "Run `` as a Quick / Standard / Extended flow?". Option labels:

• [a] confirm ` flow → hand off to lsa:discover` (Standard / Extended) or direct response (Quick)
• [b] override to a different flow — I re-route accordingly
• [c] reconsider — I re-run the signal checklist

Format per [../output/SKILL.md](../output/SKILL.md); AskUserQuestion in Claude Code (per .lsa/VISION.md §2 principle 9, "Substrate-native first"). The checklist + rationale ride inside the picker (question text / option descriptions) per [../output/SKILL.md](../output/SKILL.md) Rule 5 "Self-contained gates" — same-turn prose before the picker may not render. Observable result: the proposal is on screen; control returns to the human; no downstream skill has fired.

1. On confirm, hand off per flow:

• Quick — return control to the agent for a direct single-pass response. ground-rules still applies.
• Standard — invoke lsa:discover for the light three-question probe.
• Extended — invoke lsa:discover to start the full loop. Vision's Extended loop is "discover → specify → verify → delegate → reconcile".

Observable result: the named downstream skill is invoked, or — for Quick — direct response begins.

Output

A flow label (Quick / Standard / Extended) and a 2–4-sentence rationale. Human-readable. Confirmed before any downstream LSA skill fires.

Constraints

• Do not start LSA ceremony before flow confirmation. No lsa:discover, lsa:specify, or any other LSA skill fires until the human responds.
• Do not invent boundary signals that are not actually present in the task description. If the task does not mention an API change, do not assume one.
• Do not silently choose a heavier flow than the human picks. If the human overrides downward (e.g., Extended → Standard), log the override in the rationale and proceed at the human's flow.
• Outputs follow [../output/SKILL.md](../output/SKILL.md) — citation by link, never restated.

---

On confirm, downstream LSA skills absorb the flow into their own gates. Every output still obeys core/ground-rules (content) and core/output (format).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: NVZver
• Source: NVZver/claude-marketplace
• License: MIT

Install and usage instructions live in the source repository linked above.