菜单
Pause execution and synthesize the live session state. Reads the Mentat insight bus, summarizes recent state transitions, lists drift events, surfaces Q-table route hints, and returns a structured reflection block.
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
Derive, harden, audit, port, and maintain dense single-file agent constitutions, system prompts, prompt frameworks, and agent skill instructions. Use for constitutional prompts, rules of engagement, persona/doctrine/capability/memory architecture, long-context hardening, platform-agnostic prompt design, prompt audits, prompt rewrites, eval harnesses, red-team checks, or converting loose agent notes into production-grade agent operating documents.
Package skills, agent prompts, plugin wrappers, MCP/server reference cards, and local registry metadata into a private multi-client marketplace for Codex, Claude Code, and OpenCode. Use when building or auditing plugin manifests, marketplace cards, registry indexes, client adapters, package boundaries, install docs, dry-run install plans, or promotion gates for local/private agent artifacts.
Sovereign Intelligence Workspace for strategic reverse engineering, OSINT/SIGINT forensics, supply chain attack analysis, malware attribution, and full-spectrum threat intelligence. Generates exhaustive SSDS Sovereign Intelligence Casefiles (SIC). Powered by the Varys strategic cognition framework. Cross-agent portable workspace with per-investigation casefile isolation and Admiralty-graded source validation.
Package skills, agent prompts, plugin wrappers, MCP/server reference cards, and local registry metadata into a private multi-client marketplace for Codex, Claude Code, and OpenCode. Use when building or auditing plugin manifests, marketplace cards, registry indexes, client adapters, package boundaries, install docs, dry-run install plans, or promotion gates for local/private agent artifacts.
Derive, harden, audit, port, and maintain dense single-file agent constitutions, system prompts, prompt frameworks, and agent skill instructions. Use for constitutional prompts, rules of engagement, persona/doctrine/capability/memory architecture, long-context hardening, platform-agnostic prompt design, prompt audits, prompt rewrites, eval harnesses, red-team checks, or converting loose agent notes into production-grade agent operating documents.
CTMv3 workspace activation cognitive layer. Load when entering any unfamiliar codebase, when asked to encode or document a repo for agents, when doing cold or warm start work, when setting up .github/.claude/.codex/.sovereign directories, or when building ARCHITECTURE_MAP.md, AGENTS.md, TOPOLOGY.md, or PROVENANCE.md. Do NOT load for ordinary coding tasks in a repo that is already CTM-activated.
| name | mentat-reflect |
| description | Pause execution and synthesize the live session state. Reads the Mentat insight bus, summarizes recent state transitions, lists drift events, surfaces Q-table route hints, and returns a structured reflection block. |
| when_to_use | Trigger phrases: "where are we", "status", "what's the state", "reflect", "let's pause", "step back", "are we drifting", "context check", "show me the insight log", "session status". Also fire automatically when the session has had 8+ consecutive EXECUTING tool uses without a verify step. |
| allowed-tools | Bash(python3 ${CLAUDE_PLUGIN_ROOT}/bin/mentat*) |
You are running inside a Claude Code session that the Mentat plugin is observing. Mentat tracks an explicit state machine over your work (planning, exploring, executing, verifying, reflecting, blocked, drifting, compacting), a TD-learning Q-table over (state, tool) pairs, and a structured insight bus.
When this skill fires, do the following in order:
Read the live session state by running:
python3 ${CLAUDE_PLUGIN_ROOT}/bin/mentat status --json
Read the last 30 insights:
python3 ${CLAUDE_PLUGIN_ROOT}/bin/mentat tail --n 30 --json
Synthesize a reflection in this exact shape:
## Mentat Reflection
- Current state: <STATE>
- Transitions so far: <N>
- Last tool: <TOOL> (success=<bool>)
- Drift events: <N>
- Q-route hint for this state: <TOOL> (Q=<value>)
### Recent insights worth surfacing
- <type>: <one-line summary>
- ...
### Recommendation
<one paragraph: stay the course / verify / pivot / re-plan / ask user>
If state is DRIFTING, do not propose any work that would write or shell-out.
Reaffirm scope first: print the contents of .mentat/scope.md if it exists,
and ask the user whether to update the scope or revert the in-flight change.
If state is BLOCKED, propose a sub-agent dispatch (mentat-cartographer for read-only mapping or mentat-crucible for adversarial pushback) before continuing.
Emit your reflection as a normal assistant message — no need to write a file. The act of reflecting itself is logged to the insight bus by the hooks.