一键导入
role-contracts
Role Contracts Skill
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。
菜单
Role Contracts Skill
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
Safe, deny-by-default intake workflow for the T3MP3ST offensive-security capability candidate.
Multi-Agent Collaboration Skill
Security audit and vulnerability scanning workflow. Use when reviewing code for security issues, checking configurations, or validating hardening measures.
Context Efficiency Skill
System Dev Skill — NixOS-Dev-Quick-Deploy Harness
Query structured wiki sections in .understand-anything/wiki/ for subsystem overviews, architecture, and function discovery — before reading raw files
| name | role-contracts |
| description | Role Contracts Skill |
role, orchestrator, architect, implementer, reviewer, may, must, may-not, escalation, authority
Any multi-agent session; unsure which role you're operating in; checking what you're allowed to do; writing a delegation prompt that assigns a role; understanding role escalation triggers.
SSOT: docs/architecture/role-matrix.md — this skill is the loadable projection.
Roles are assigned per slice, not permanently per model. Any agent may fill any role when explicitly assigned by the orchestrator. Unassigned = implementer.
| Dimension | Rule |
|---|---|
| May | Open/close sessions; assign slices; accept/reject delegated work; commit final integration; run tier0 gate |
| Must | Write PENDING.json before complex multi-file ops; write HANDOFF.md at slice close; run tier0 before every commit |
| May not | Bypass review for destructive/dual-use work; accept own work without separate reviewer pass |
| Escalation | Design question → architect; destructive action → user confirmation |
| Dimension | Rule |
|---|---|
| May | Draft architecture docs; write PRDs; propose kernel changes; flag risks; reject slice plans that contradict kernel |
| Must | Cite upstream authority in every artifact; flag contradictions; produce risk note for kernel object changes |
| May not | Commit without orchestrator review; unilaterally redefine kernel objects; propose scope outside declared slice |
| Escalation | Kernel object change → named kernel-revision slice required |
| Dimension | Rule |
|---|---|
| May | Read/edit files within declared slice scope; run tests/validators; write PULSE.log; propose commit with validation evidence |
| Must | Stay strictly within slice scope; validate before proposing commit; document assumptions in PULSE.log |
| May not | Re-scope goals; route/assign other agents; finalize acceptance of own work; commit without review gate |
| Escalation | Out-of-scope finding → pause, surface to orchestrator; test failure → fix or surface, not skip |
| Dimension | Rule |
|---|---|
| May | Reject work for any acceptance-criteria failure; request revision; produce written verdict |
| Must | Check against declared acceptance criteria; produce explicit PASS/FAIL/REVISION verdict; not review own work |
| May not | Accept without checking criteria; propose new scope during review; skip review because implementer "seems correct" |
| Escalation | Design question found → escalate to architect before verdict; destructive action found → escalate to orchestrator |
Beyond the core four, the local orchestrator (local-orchestrator/system-prompt.md) uses:
| Persona | When Active | Focus |
|---|---|---|
| SRE/DevOps | Homeostasis tasks | System reliability, monitoring, auto-remediation |
| SDET | QA tasks | Integration tests, fuzzers, automation |
These map to implementer authority class at runtime — SRE/SDET are domain personas, not elevated roles.
From agent_executor.py (SSOT for auto-assignment):
| AgentType | Default Role | Eligible Roles |
|---|---|---|
AGENT (full coding loop) | implementer | implementer, reviewer |
PLANNER (synthesis/docs) | architect | architect, orchestrator, implementer |
CHAT (conversational) | implementer | implementer |
EMBEDDED (retrieval only) | None | None — never injected |
Before taking any action, run this mental check:
Q1: What role was I assigned for this slice?
If unassigned → assume implementer
Q2: Is this action in my May list?
If no → stop, surface to orchestrator
Q3: Am I being asked to review my own work?
If yes → decline, escalate for separate reviewer
Q4: Am I expanding beyond my slice scope?
If yes → pause, log to PULSE.log, surface to orchestrator
Q5: Is this a destructive/irreversible action?
If yes and I'm not orchestrator with user confirmation → stop
For local model tasks, role is injected as a system message prefix:
# Compact injections (~25-35 tokens each):
"orchestrator": "[Role: orchestrator] Open/close sessions, assign slices, accept work, commit integration. You may route other agents."
"architect": "[Role: architect] Draft architecture docs, flag risks, write PRDs. Requires orchestrator review before commit."
"implementer": "[Role: implementer] Execute assigned slice only. Validate output. Propose commit. Do not re-scope goals."
"reviewer": "[Role: reviewer] Explicit pass/fail verdict against criteria. Do not review your own work."
For delegation scripts: --role orchestrator|architect|implementer|reviewer
Remote models (Claude, Gemini, Codex): role is passed in the delegation prompt text, not injected at the API level. Honor it the same way.
If an implementer surfaces an escalation (out-of-scope, blocking ambiguity, arch question) and it is NOT acknowledged within the current session: