yuri-shura

6-perspective adversarial review for high-stakes turns (architecture decisions, refactor planning, deployment review) — fans out architect, adversary (7-vector attack), maintainer, ops, product, and s

SKILL.md

GitHub

Yuri Shura -- 6-Perspective Review

Inspired by Istishraf (MIT-licensed Claude plugin). Original implementation adapted to Yuri OS's pulse-cortex.

When to invoke

Model-invocable for high-stakes strategic review (architecture, refactor planning, deployment); or /shura <topic>. Advisory only — never holds implementation authority.

6 Perspectives (native Workflow fan-out)

Run via the Workflow tool: 6 parallel agents, each adopting one perspective lens (Opus/Sonnet, self-selected per depth). Each returns a bounded structured finding; the main thread consolidates. No external model lanes.

Perspective Lens
Architect soundness of design
Adversary 7-vector attack protocol (see below)
Maintainer long-term cost
Ops production readiness
Product user impact
Security risk surface

Output format

Six bounded sections (50 lines each) consolidated by main thread. Each perspective answers: assessment, risks, recommendation.

Adversary Protocol — 7-Vector Attack Checklist

The @ds-pro adversary lane runs this checklist against every artifact. Derived from Anderson's security framework + Ibn Khaldun's resilience/decay model + SOUL.md reversibility principle. Source: strategic-thinker MIT brief (2026-05-16).

# Vector Question
1 Policy Does this violate an explicit rule, boundary, or operating constraint?
2 Mechanism How could the implementation fail, hang, corrupt state, or be exploited?
3 Assurance What cannot be monitored, audited, or verified after the fact?
4 Incentives What system or actor benefits from this failing silently?
5 Fragility Where are the single points of failure, tight coupling, or cascade paths?
6 Scope creep What silent expansion or unapproved scope increase does this permit?
7 Irreversibility How hard is rollback if this is wrong? Is there a tested recovery path?

Adversary prompt injection: for each vector, deliver one finding or "clean" — no filler. Total output ≤ 35 lines.

Quarantine rules

Advisory output only — no implementation authority. Findings are consolidated for the owner; any resulting change goes through the normal owner-gated commit path.

Trigger phrases (auto-classify)

  • "architecture review"
  • "should we refactor"
  • "deployment plan"
  • "high stakes"
  • "before we ship"

Session Notes

2026-05-17

  • session: 113m | peak ctx: 0% | compacts: 0
  • tools: Bash×168, Edit×41, Read×33, mcp×33, Write×11, ToolSearch×2, Skill×1, AskUserQuestion×1, ExitPlanMode×1
  • corrections: none
  • errors: none

2026-05-16

  • session: 61m | peak ctx: 0% | compacts: 0
  • tools: Bash×50, Edit×28, Write×13, Read×12
  • corrections: none
  • errors: none