← 返回 Skills 市场
1705
总下载
3
收藏
6
当前安装
119
版本数
在 OpenClaw 中安装
/install turing-pyramid
功能描述
Prioritized action selection for AI agents. 10 needs with tension scoring, execution-gate tracking, optional continuity scripts, and an opt-in watchdog. Use...
安全使用建议
This skill appears to be what it claims: a local, stateful prioritization engine. Key precautions before installing or enabling persistence:
- Always set WORKSPACE to an isolated directory (never $HOME, /root, or any folder containing credentials).
- Review scripts (mindstate-watchdog.sh, mindstate-daemon.sh, run-cycle.sh, mark-satisfied.sh) if you plan to enable cron. The repo ships many scripts and tests — read them or run tests in a throwaway workspace.
- Do NOT enable allow_kill or allow_cleanup until you have audited the watchdog scripts and are confident about the exact PIDs/paths it will act on.
- external-model scanning is disabled by default; only enable it after documenting the API, the env var that will hold any key, and getting steward approval.
- Running the test-suite in an isolated WORKSPACE is recommended before using the skill in production.
If you follow those steps this skill is coherent and usable; treating the workspace boundary carefully mitigates the principal risk (accidental scanning of sensitive files).
功能分析
Type: OpenClaw Skill
Name: turing-pyramid
Version: 1.34.8
The Turing Pyramid skill is a sophisticated stateful motivation and action-prioritization engine for AI agents. While it possesses significant capabilities, including process monitoring (pgrep), process termination (kill), and persistence via cron, these are strictly scoped to its stated purpose of maintaining 'mindstate' continuity. The implementation demonstrates high security awareness: scripts like 'context-scan.sh' include path traversal validation using realpath, 'mindstate-watchdog.sh' anchors its kill scope to the skill's own directory to prevent collateral process termination, and 'mark-satisfied.sh' automatically scrubs sensitive patterns (IPs, emails, tokens) from audit logs. The documentation is exceptionally transparent regarding its security model and deployment risks.
能力标签
能力评估
Purpose & Capability
Name, description, SKILL.md, assets, and scripts all describe a local, stateful motivation/prioritization engine. Required binaries (bash, jq, bc, grep, find, flock, pgrep, df, kill, gzip) and the single required env var WORKSPACE are coherent with scanning, file-IO, and optional process-checking described in the docs. The actions labeled external in configs are presented as agent suggestions; core scripts claim no network I/O by default. No unrelated cloud keys or services are requested.
Instruction Scope
SKILL.md and included scripts clearly state the skill reads workspace files and writes state/audit files (needs-state.json, audit.log, followups.jsonl, MINDSTATE.md, watchdog logs). That behavior is expected for a stateful priority engine, but it is a material scope change compared to stateless skills: if WORKSPACE is pointed at a sensitive directory the skill will scan it. The runtime instructions explicitly warn about this and document opt-in features (allow_kill, allow_cleanup, external-model).
Install Mechanism
No install spec is present (instruction-only installer), and the bundle contains scripts and config files only. There are no external download URLs or extract steps. The package uses only standard POSIX utilities listed in metadata. This is low-risk from an install mechanism perspective.
Credentials
Only WORKSPACE is required; several benign optional env vars are documented (MINDSTATE_ASSETS_DIR, SKIP_SCANS, TURING_CALLER, etc.). No secrets, API keys, or unrelated service credentials are required by default. The assets include an 'external_model' config but it is disabled by default and requires explicit steward approval and an api_key_env to be populated.
Persistence & Privilege
The skill is not force-included (always:false). It can be invoked autonomously (default platform behavior) which is expected for skills. It writes persistent state inside WORKSPACE and documents cron-based continuity/watchdog modes. The potentially destructive features (killing processes limited to path-anchored mindstate-* processes, cleanup of orphan .tmp files) require explicit opt-in flags (allow_kill, allow_cleanup) and are documented as such — that makes the privilege request proportional, but these features should only be enabled after code review and testing in an isolated workspace.
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 turing-pyramid - 安装完成后，直接呼叫该 Skill 的名称或使用
/turing-pyramid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
v1.34.8
Integrity actions audit: 9→7 concrete verifiable actions. New INTEGRITY_CHECKPOINTS.md with structured entries + INTEGRITY_CHECKPOINT_TEMPLATE.md with gate thresholds. scan_integrity.sh now reads checkpoints file directly. 28/28 tests.
v1.34.7
Deliberation discipline tracking: new deliberation.log, tiered observables (absent/minimal/substantive/deep), classify_deliberation_level() helper, deliberate.sh now logs invocations, gate-resolve records level, run-cycle shows discipline warnings, mindstate-freeze adds discipline mirror. 28/28 tests.
v1.34.6
Connection actions expanded: 6 new actions (Moltbook post, AgentGram post, steward ping variants, emoji react, voluntary check-in). Weight recalibration: light actions raised, heavy actions lowered, steward issue ping gets priority weight 65.
v1.34.5
Deferred backlog pressure: scan_closure now sees deferred count (pressure bands), gate-resolve --close abandons deferred items, new closure action for deliberate backlog review. 28/28 tests.
v1.34.4
Integrity scanner rebuilt: was a pure time-decay stub, now reads memory files for alignment/drift signals + checkpoint age + SOUL engagement checks. Ritual loop broken. Config declaration fixed. 27/27 tests green.
v1.34.3
Self-throttle race fix: daemon --force flag, boot always gets fresh reality. Test isolation: state reset per test, SKIP_DAEMON_FORCE. 27/27 + 2/2 tests green.
v1.34.2
Execution gate continuity fix: daemon tracks pending gate status, freeze prioritizes pending gate actions in open_threads, boot warns on blocked execution gate.
v1.34.1
Runtime housekeeping, decisions.log, git drift scanner, PII cleanup
v1.33.9
ClawHub safety clarification: confirmed safe publish defaults (allow_kill=false, allow_cleanup=false), clarified stateful writes, isolated WORKSPACE requirement, external-model and destructive watchdog paths explicitly opt-in
v1.33.8
Deliberation Phase 2A: [DELIBERATIVE] tag in run-cycle, --conclusion in mark-satisfied+gate-resolve, action_mode stored at propose time, scrub_sensitive moved to shared utils. All 27 deliberation tests green.
v1.33.7
Critical init/install bugfix: repaired needs-state template/schema drift, hardened init.sh with fallback rebuild from needs-config, and added regression test for fresh init + stale-template recovery.
v1.33.6
Scanner false-positive audit: autonomy infra-vs-permission blocked distinction; closure blocked keyword hardening; added regression tests for scanner false positives.
v1.33.5
Compaction continuity (current-task.md recovery in mindstate-boot), freeze grep -oP crash fix, 27/27 tests green, gate isolation in test suite
v1.33.3
Docs: fd allocation table, onboarding updated for deliberation+association scan, description updated, version consistency, test gaps documented
v1.33.1
PII scrub + 3 bugfixes: freeze SESSION_START, exclude-source, test validation
v1.33.0
Continuation Phase 2: boot contextual recall, freeze deliberation residuals, 19 association tests
v1.32.0
Continuation & Association Scan: contextual recall, concluded+action warning, followup 2w/1m, 17 new tests
v1.31.3
ClawHub security review fixes: network claims, WORKSPACE isolation warning
v1.31.2
Bugfixes: defer path scrub, template --save writes file, CHANGELOG updated, settings.version synced, scrub_sensitive consistent across gate-resolve and mark-satisfied. 28 tests green.
v1.31.1
Deliberation Protocol Phase 1+2: 22 deliberative actions, deliberate.sh (template/validate/validate-inline), --conclusion in mark-satisfied + gate-resolve, outcome types taxonomy, route reframe, 27 tests. Watchdog safe defaults.
元数据
常见问题
Turing Pyramid 是什么？
Prioritized action selection for AI agents. 10 needs with tension scoring, execution-gate tracking, optional continuity scripts, and an opt-in watchdog. Use... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 1705 次。
如何安装 Turing Pyramid？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install turing-pyramid」即可一键安装，无需额外配置。
Turing Pyramid 是免费的吗？
是的，Turing Pyramid 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Turing Pyramid 支持哪些平台？
Turing Pyramid 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 Turing Pyramid？
由 TensusDS（@tensusds）开发并维护，当前版本 v1.34.8。
推荐 Skills