← 返回 Skills 市场
Miroprism
作者
Jeremy Knows
· GitHub ↗
· v1.4.0
· MIT-0
261
总下载
0
收藏
1
当前安装
4
版本数
在 OpenClaw 中安装
/install miroprism
功能描述
MiroPRISM — Adversarial two-round review protocol. Extends PRISM with a mandatory second round where every reviewer must respond to all R1 findings with evid...
安全使用建议
Before installing, confirm where this skill will read and write: it will create analysis/miroprism runs, write logs and archives, and expects to read files you point it at (Evidence Rules require quoting file/line). Verify the repository source (README suggests github.com/jeremyknows/MiroPRISM) and only clone from a trusted location. Ensure your agent/platform provides the LLM access this protocol assumes (Sonnet/Haiku) and that any API keys are managed by the platform — the skill declares no credentials itself. Consider running initial tests on non-sensitive artifacts, enable the --review-digest manual pause to inspect sanitized digests before R2, and review the sanitization/transparency log to ensure sensitive snippets (API keys, secrets) are not being retained or broadcast. If you have low tolerance for local archival of potentially sensitive content, run MiroPRISM in an isolated project workspace.
功能分析
Type: OpenClaw Skill
Name: miroprism
Version: 1.4.0
MiroPRISM is a sophisticated orchestration protocol for multi-round AI reviews, designed to reduce consensus bias. The skill demonstrates high security awareness by implementing a mandatory sanitization phase (Phase 2 in SKILL.md) that strips code blocks, URLs, and structured data from reviewer outputs to prevent prompt injection between rounds. It uses standard shell-based concurrency management (PID-based .lock files) and local logging (metrics.tsv) within its designated workspace, with no evidence of data exfiltration, unauthorized network access, or malicious intent.
能力评估
Purpose & Capability
The name/description match the SKILL.md: it implements a two‑round adversarial review flow and explains R1/R2 behavior. However, the README references cloning a GitHub repo for installation while the registry lists Source: unknown — a minor mismatch. The docs also assume access to specific LLM models (Sonnet/Haiku) and spawning subagents, but the skill declares no environment variables or credentials for model APIs. That implicit dependency is unexplained.
Instruction Scope
SKILL.md is prescriptive: it tells the agent to create run directories under analysis/miroprism, manage .lock files, check PIDs (kill -0), read specific files to produce evidence, and store R1/R2 outputs and transparency logs. Those actions align with an orchestrator but do grant the skill the ability to read project files and write local archives; the Evidence Rules require quoting specific file lines which could cause broad file access if the artifact boundaries aren't enforced. The collision handling (PID checks) touches system process state — reasonable for concurrency, but worth reviewing for platform compatibility.
Install Mechanism
This is instruction-only (no install spec, no code files executed by the platform). README shows manual git clone instructions pointing to github.com/jeremyknows/MiroPRISM — that's a normal install route, but it is an external source the user must trust. No downloads from arbitrary URLs or archive extraction are embedded in the skill itself.
Credentials
The skill expects to run multiple reviewers on specific LLM models and reports token/cost estimates, yet requires no environment variables or API credentials in the metadata. If your agent/platform needs API keys (Anthropic/other) to spawn subagents, those credentials are implicit but not declared. The Evidence Rules also ask reviewers to 'read at least 3 files' and cite file/line numbers — that can justify filesystem access, but the skill does not limit which files, so privilege to read arbitrary repo files is effectively requested without explicit declaration.
Persistence & Privilege
The skill writes outputs under analysis/miroprism/<runs|archive>, creates .lock files, and keeps a transparency log (SHA256 + sanitization counts). It does not set always:true and does not request system-wide config changes. Storing sanitized and raw reviewer outputs locally is expected, but verify that the chosen storage path is acceptable (sensitive content may be archived).
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install miroprism - 安装完成后，直接呼叫该 Skill 的名称或使用
/miroprism触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v1.4.0
5 P0 fixes: UNCERTAIN evidence gate (>=50 chars or INCOMPLETE), PID-only slug collision detection (mtime removed), per-reviewer counts removed from transparency log, CHANGELOG section, mode selection decision tree
v1.3.0
v1.3.0: P0 anti-sycophancy fixes — evidence rejection <100 chars, silence=[INCOMPLETE], R2 re-sanitization (Step 9.5), 250-char descriptions, embedded Evidence Rules
v1.2.0
v1.2.0: Self-review R2 fixes — spawn mechanism docs, PID lock, sanitization example, corrected cost table, verdict drift threshold, sycophancy claim updated, Evidence Rules dedup, references/ split
v1.1.0
v1.1.0: embedded R1 prompts (self-contained, no PRISM required), fixed clone URL, added Quick Start, metrics tracking (metrics.tsv), large artifact cost warning
元数据
常见问题
Miroprism 是什么？
MiroPRISM — Adversarial two-round review protocol. Extends PRISM with a mandatory second round where every reviewer must respond to all R1 findings with evid... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 261 次。
如何安装 Miroprism？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install miroprism」即可一键安装，无需额外配置。
Miroprism 是免费的吗？
是的，Miroprism 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Miroprism 支持哪些平台？
Miroprism 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Miroprism？
由 Jeremy Knows（@jeremyknows）开发并维护，当前版本 v1.4.0。
推荐 Skills