variant_b // terminal_editorial // animations: none
One CLI conducts three vendors. Gates make it safe.
agentflows fans tasks out to Kimi K3, MiniMax-M3, and claude workers under a single schema-validated contract. Everything below was measured in this repo on 2026-07-17 — this page is a spec sheet, not a pitch.
| metric | method | value |
|---|---|---|
| repo tests | npm test at audit time | 100/100 pass |
| swarm audit | 4 read-only K3 subagents, disjoint scopes | 3 P0 · 8 P1 found |
| citation check | spot-verified against source | 6/6 verified |
| caveman condensing | K3, verbose doc → 6 bullets | −46% words |
| caveman generation | K3, dense handoff prose | +11% (worse — generation stays normal) |
Capabilities
4 entries · each runnableK3 first pass, evidence-backed findings; MiniMax-M3 cross-check verifies and hunts misses.
Parallel read-only subagents on disjoint scopes; lead synthesizes one verified answer.
402 / entitlement classifies as quota; chain degrades Kimi → MiniMax-M3 unassisted.
Static page concepts as safe, reviewable artifacts inside the lab boundary.
The loop
structural gates, not vibesDownstream agents never spawn until the upstream gate clears. Rejections loop back with findings attached, then halt for a human. The same loop reviewed this page before you read it.
[dispatch] ──▶ [review: kimi-k3] ──▶ [crosscheck: minimax-m3] ──▶ [gate: pass]
│ │ │
│ one task contract │ verify + hunt misses │ severity-ranked
│ schema-validated out │ corroborated > single │ human approves
▼ ▼ ▼
findings.md verified set ship or halt
Run it
expected: halt at gate on fixture bugnode bin/aflow.js workflow run top-review-lab --variant strict
# or the two-model corroboration loop:
aflow flow run crosscheck --diff HEAD~1 "review this change"
Boundaries
- No credentials, transcripts, or session state in any demo page.
- No mutation outside
experiments/skills-workflows-hooks-lab/. - Static HTML/CSS only — no backend, no trackers, no animation budget.