| name | onboard-swarm |
| description | The parallel per-agent swarm that beats cold-start. After install, each bounded agent reads its canonical source (git/claude-code → Analyst, granola → SDR, cross-stream → Auditor) in parallel and posts a candidate — the one fact, commitment, or risk worth acting on — before any deliberation. Read-only, LLM-optional. Run this to show the vault won't start empty. |
| disable-model-invocation | true |
| allowed-tools | Bash(node *) |
/onboard-swarm — beat cold-start
The first thing Liminal does after install is not ask you for a brief. It reads. Each bounded agent reads its canonical source in parallel and posts a candidate — the one fact, commitment, or risk worth acting on — so the first deliberation starts warm instead of from a blank table.
This is the full deliberation bootstrap. It is read-only (it reads sources and posts candidates in memory; it does not write the vault or mutate any source — the daemon does the real ingest) and LLM-optional (each agent deliberates live on Opus when a credential is present; otherwise it falls back to a labeled fixture so the beat always renders).
Flow
1. Run the swarm
node ${CLAUDE_PLUGIN_ROOT}/skills/onboard-swarm/run.js
The runner returns JSON:
mode — live if the agents read on Opus, fixture if no Anthropic credential was available (say so honestly), cold if no source has signal.summary — { streams_total, streams_live, candidates_total, cold_start }.candidates — each { source, agent_owner, status, interpretation, refused, mode }. The owning agent reads its own source, so candidates are in-lane work, not refusals.
For the lightweight "what's here?" probe (no agent reads, no LLM), pass --scan — it returns { summary, streams } where each stream is { source, agent_owner, status, detail, count }.
2. Render the swarm
Group candidates by agent_owner, in bounded order (Analyst, SDR, Auditor). Lead each block with the agent and the source it read, then its candidate.
ANALYST (facts: what was built, what changed)
from git [the candidate fact]
from claude-code [the candidate fact]
SDR (commitments: who, when, the next move)
from granola [the candidate commitment]
AUDITOR (risks: gaps and drift across the streams)
from cross-stream [the candidate risk — reads the other agents' candidates]
If mode is fixture, add one plain line: "(candidates shown from a bundled sample — set ANTHROPIC_API_KEY or run claude setup-token to run this live on Opus.)"
3. Print the close line
If candidates were posted, end with:
The swarm read your streams and posted candidates — the first deliberation won't start cold. Install the desktop app to let the daemon ingest these on a loop and deliberate over them.
If mode is cold (no source has signal — common in a fresh CI checkout), say so honestly:
No live streams here yet. On a real machine the swarm reads git, Claude Code history, and meeting notes in parallel; here there's nothing to read. Run /try-liminal to see the loop on a sample brief.
Voice rules
- Lead with the agent that owns each stream — the bounded geometry is the point.
- Candidates are in-lane work; the refusal beat lives in /try-liminal, not here.
- No emoji, no exclamation marks.
What this skill is not
- Not ingest — it posts candidates in memory; the daemon (
bin/liminal-substrated.js) persists ingested signal.
- Not a vault write — read-only by design, so a first run never touches the real vault.
- Calendar has no adapter yet (the remaining "next" in
docs/specs/2026-06-13-swarm-onboarding.md), so it never posts a candidate.
