Public: anyone can read this; nobody can change it.
workbenchthis machine
/goal /ultracode-workflows [POL-0] DESIGN the policy-authoring & governance-of-governance model (Kyverno) #38
Closed without the crew
commented
Audit (team-board): acceptance met — closing.
The design merged in PR #49 (bd8d0848): docs/proposals/policy-governance-kyverno.md (273 lines). It answers exactly what the acceptance demands:
- Decision 2 — may an agent WRITE policy: NEVER direct write, propose-only via human-merged PR, explicit per category (quota-tier proposals allowed; cluster-shape/network policy human-only)
- Decision 3 — Audit-first default for newly merged policy
- Decision 4 — webhook
failurePolicy, explicitly decided (Ignore while non-HA; fail-closed is a gated promotion) - POL-1..POL-5 invariants enumerated (doc §"Invariants")
Implementation demonstrably follows it: every merged policy file in plat/gitops cites "POL-0 Decision 3/4" in-line, and today's fresh daughter runs the decided posture (12 ClusterPolicies Ready, Audit, failurePolicy Ignore, plat.sh/app-scoped webhook). Doc landed in plat/mitosis docs/proposals/ rather than a separate vision repo — same review surface, no gap.
Sign in to join this conversation.
new above
offline
Track/Wave:
POL-0· Wave G2 · Size S Repos:mitosis-visionDepends on: RBAC-1Why
Kyverno is the deterministic BACKEND (agent authors policy YAML from an NL prompt → PR → Flux → Kyverno enforces/generates with synchronize:true). Before agents get any policy-authoring capability, the governance-of-governance must be decided — a bad ClusterPolicy has platform-wide blast radius, unlike a per-app PR.
Scope
A design doc answering: (a) does policy live in gitops or a stricter-reviewed
policyrepo; (b) may agents ever WRITE policy or only PROPOSE via human-merged PR (per category: quota tier — maybe; default-deny network — human only); (c) enforce vs audit default for newly-merged policy (audit-first is safer for a self-modifying platform); (d) Kyverno webhook failurePolicy (fail-open unsafe vs fail-closed platform-wide-outage) — an explicit tested decision matching authorize()'s fail-closed philosophy. Enumerate the exact invariants POL-1..POL-5 enforce.Acceptance
A reviewed doc with an explicit yes/no per policy category on 'can an agent's own PR change the policy that governs itself', plus the failurePolicy and enforce/audit defaults.
Notes
Design-first — natural fan-out target.
--- Part of the First Client Deploy → Governance roadmap. The owning agent may recurse into its own subagent team (ultracode workflow) if the task warrants. Honor the seam-serialization: land on a shared seam by rebase, and don't start a seam position until the prior one has merged.