Track B: agentic swarm (≥10 roles) + team scoring eval #82

Track B — Agentic swarm eval (≥10 specialists)

Status: unblocked for design; eng still after Track A review.

Track A landed

• Draft PR: Track A: Eval Harness + Scoring Protocol (Findings, Budget, Coverage) #83
• Branch: cursor/track-a-harness-eval-protocol-cdb2
• Provides: findings schema, validators, budget, coverage ledger, KBotEpisodeAdapter interface stub

Goal

Score the TEAM, not one chatty model:

• ≥10 role specialists (CPO locks: recon, web, auth/IDOR, API, priv-esc/Linux, network/pivot, binary/RE, cloud/misconfig, LLM/agent attack, supply chain)
• Thin coordinator + independent verifier (never the finder)
• Metrics: decomposition, handoffs, tool discipline, verified findings rate

Runtime preference

Drive/score kbot (Treelovah/kryptsec-kbot) via thin adapter — do not reinvent fleetd in OASIS TS.

Cost / weights

Research ≤$25/day formal runs. Swarm specialists default open ≤30B (SLM map v2).

Do not

Auto-merge. No CF branding. No Academy in this issue.