Skip to content
Back to skills

Roundnext

ASecurity

Run one full "what should happen next" round over the live board — ripeness-check planned items and start what is ready, re-examine blockers and unblock or derive unblocking work, scan the project for genuinely new next work across multiple dimensions, and route anything needing a human call to the decisions panel as a Hyperbrief briefing. Invoke when a work track closes, when idling feels wrong, or on a cadence.

  • 8 stars
  • 0 votes
  • 0 copies
  • 4 views
  • Added September 29, 2026
ai-agentsrustgosecurity

Works with

  • mcp

Security analysis

A100/100

Scanned September 29, 2026

npx -y skills add SoliEstre/EstreGenesis --skill roundnext --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Roundnext?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Roundnext
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/soliestre-roundnext/badge)](https://www.skillsdirectory.com/skills/soliestre-roundnext)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: roundnext
description: Run one full "what should happen next" round over the live board — ripeness-check planned items and start what is ready, re-examine blockers and unblock or derive unblocking work, scan the project for genuinely new next work across multiple dimensions, and route anything needing a human call to the decisions panel as a Hyperbrief briefing. Invoke when a work track closes, when idling feels wrong, or on a cadence.
---

# /roundnext — one planned/blocked/new-work round

Mechanizes the composition of Constellation §13.16.6 element 5 (planned-queue scan), §13.20 (blocker tracking + nudge), and §13.18 (autonomous execution) into a single invocable round. The output of a round is: work started, queues re-sorted, blockers acted on, new work registered, and human-gated items routed to the decisions panel — never a silent "nothing to do".

## Procedure (5 steps, in order)

**0. Load board truth** — read the live board state (`board_state_get` MCP tool, or the board's `state.json` directly). You need: `planned[]` (with `blocked`/`blockReason`/`when`), `current`, `decisions[]` (open ones), and recent `done[]` for context.

**1. Ripeness check on unblocked planned items** — for each planned item that is not blocked, ask: has its start condition arrived? (`when` reached, predecessor in `done[]`, ordering satisfied, no fresher priority above it.) If ripe → **start it now** (§13.18: a defined-next-step proceeds without asking). When the started item is a multi-stage track (3+ phases, or spanning sessions/external waits), register it on `current[]` **with a `stages` timeline** (§13.31.4 — `{label, status: 'pending'|'active'|'done'}`); this start step is the schema's natural entry hook. If not ripe, record why in one line (this becomes the priority rationale in step 2's sort).

**2. Blocker re-examination on blocked planned items** — for each `blocked: true` item, re-verify the blocker *now* (don't trust the recorded `blockReason` — blockers go stale silently: the external dependency may have shipped, the gate may have been decided, the precondition may have been met by unrelated work). If the blocker is gone → flip to unblocked; if it is also ripe, start it; otherwise merge into the priority sort. If still blocked, keep it and pass to step 3.

**3. Derive unblocking work for still-blocked items** — per §13.20: for each surviving blocker, ask *what would it take to clear this, and which part of that can be done from here, now?* (e.g. evidence the other side is waiting for, a prerequisite refactor, a reproduction, a nudge message per the §13.20 BlockerNudge cadence). Anything actionable becomes a new planned item (priority-sorted) — or starts immediately if it is small and ripe. If nothing is actionable from this side, ensure the blocker is visible on the board (BlockerManifest) and nudge cadence is armed.

**4. New-work scan (multi-dimensional)** — look beyond the queue: does the *current project state* imply next work not yet registered? Sweep several dimensions rather than one: pending release/version sync surfaces · doc/spec drift vs shipped code · verification gaps (things claimed but never machine-checked) · security/hygiene follow-ups · dogfood evidence worth capturing · cross-module interactions newly possible · user-stated intents not yet decomposed. Register findings as planned items with one-line rationale. Do NOT inflate: only work that is real now, not speculative backlog padding.

**5. Route human-gated items to the decisions panel** — anything from the queue or step 4 that needs a user call (loss/external-publish, a genuine design fork, strategy value judgments) must NOT be silently deferred or started: register it as a decisions-panel entry with a **Hyperbrief briefing** — default tone **L1.1.1** (plain-language audience, abbreviations expanded, jargon unpacked) unless the project specifies another level. Structure: §1 situation in plain words → §2 why it needs a call → §3 the question → §4 options with costs/benefits and a recommendation → §5 urgency note. Per §13.17, the decisions panel is the escalation route — no inline structured-choice prompts in main chat.

## Closing the round

- Re-sort `planned[]` by the priority rationale gathered above and write the board update (directly or via the board-worker delegation pattern, per workspace configuration).
- Report the round's delta in a few lines: started N / unblocked N / new-planned N / decisions-registered N / still-blocked N (with nudge state). A round that changes nothing should say *why* nothing changed — that is itself signal.
- This skill does not override §13.18 gates: it never pushes/deploys/sends/deletes on its own authority, and a genuinely new major fork goes to step 5, not to execution.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.