菜单
Operate and audit the xvision self-improvement flywheel across memory capture, autooptimizer, training run example pools, child training, lineage, velocity, and look-ahead-protection evidence.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | xvision-flywheel-ops |
| description | Operate and audit the xvision self-improvement flywheel across memory capture, autooptimizer, training run example pools, child training, lineage, velocity, and look-ahead-protection evidence. |
Use this skill when driving the closed loop: capture, observe, score, distill, optimize, train, gate, activate, recall, and retire.
Run this skill when auditing or driving the closed improvement loop: checking flywheel status and velocity, reading lineage, managing training-run example pools, or gathering weekly/release evidence across the full cycle.
xvision-autoresearch-opsxvision-memory-opsxvn flywheel status --agent <agent_id> --json
xvn flywheel velocity --agent <agent_id> --days 7 --json
xvn flywheel lineage --agent <agent_id> --limit 20 --json
Dashboard equivalents live in the memory/flywheel panels. MCP read equivalents:
xvn_flywheel_statusxvn_flywheel_velocityNote:
xvn optimize memory-demoshas been removed. Use the audit script directly and manage demo pools through the dashboard Optimizer panel.
bash scripts/audit-memory-demos.sh --target-agent <agent_id> --memory-agent <agent_id>
Use fresh-recorder only when non-reproducibility is intentional and recorded.
bash scripts/export-flywheel-velocity.sh --agent <agent_id> --days 7
bash scripts/leakage-regression.sh
cargo test -p xvision-mcp mcp_
Record which evidence is automated, which is manual, and which requires live-provider credentials.
fresh-recorder without rationale: Use fresh-recorder only when
non-reproducibility is intentional and recorded. Default to frozen-snapshot
for reproducible demo pools.
Wrong split for untouched period: The untouched split must be 70/15/15. Non-standard splits invalidate the look-ahead-protection evidence.
Skipping the MCP test suite: cargo test -p xvision-mcp mcp_ is part of
the release evidence checklist. Omitting it leaves MCP surface regressions
undetected.
Confusing velocity with lineage: Velocity is a rate metric (patterns/day, gate pass rate). Lineage is the audit trail of individual activations and retirements. Use the right command for the right question.
Running flywheel status during a live cycle: Flywheel reads are safe any time, but do not trigger optimize or training-run operations inside a live trading decision process.
flywheel-ops track (team/contracts/flywheel-ops.md)
Orient an agent USING xvision as a product — running `xvn` CLI verbs, the embedded dashboard, MCP tools, strategy/eval workflows, the live Tailscale nodes, and the remote CLI job surface. USAGE ONLY. If the task is editing crates/**, frontend/web/**, migrations, contracts, or CI/deploy scripts, use `xvision-dev` instead; do not load this skill just because the cwd is the xvision repo.
Use when QAing the xvision app through direct API or CLI calls, especially to verify Strategy, Scenario, and Eval create/edit/delete behavior, detect manifest drift, duplicate records, or invalid model resolution, and collect raw HTTP evidence without the browser UI.
Use when asked for a UX/UI audit, design review of the running product, "audit the site", or before a demo/launch — boots the real site, screenshots every page (desktop + mobile), audits like a senior design lead on day one, prioritizes P0-P3, fixes safe small issues on the spot, and writes an actionable report to docs/design-audit/.
Operate the xvision autooptimizer: distill Observations into candidate Patterns, run the gate against today and the untouched test period, record blind findings, activate or retire Patterns, and preserve the lineage evidence for audit.
This skill is for interface design — dashboards, admin panels, apps, tools, and interactive products. NOT for marketing design (landing pages, marketing sites, campaigns).
Operate xvision's persistent memory layer: inspect Observations and Patterns, preview recall safely, manage Pattern lifecycle, run leakage probes, and avoid F+L+T violations.