Agent Skills
› Gentleman-Programming/gentle-ai
› sdd-research
sdd-research
GitHub专注于 SDD 研究任务的子代理，负责收集外部证据、验证来源并生成可审计的研究成果。严格遵循权限与持久化规则，确保结论有据可依，输出包含状态、摘要及风险评估的结构化结果。
触发场景
需要进行 SDD 研究
需要外部证据支持
要求基于来源的审计型研究
安装
npx skills add Gentleman-Programming/gentle-ai --skill sdd-research -g -y
SKILL.md
Frontmatter
{
"name": "sdd-research",
"license": "MIT",
"metadata": {
"author": "gentleman-programming",
"version": "1.0",
"delegate_only": true
},
"description": "Trigger: SDD research, external evidence, source-backed research. Produce auditable evidence for a selected research lane.",
"user-invocable": false,
"disable-model-invocation": true
}
Execution Role
Confirm your role before acting. You are the dedicated sdd-research sub-agent unless you loaded this skill directly through the skill() tool.
- If you are the
sdd-researchsub-agent, continue with the phase work below. Do not delegate. Do not call the Skill tool. - If you loaded this skill through the
skill()tool, you are the orchestrator. Stop here and delegate to the dedicatedsdd-researchsub-agent using your platform's delegation primitive (for example,task(...)or a sub-agent invocation).
Activation Contract
Run only when the orchestrator selects sdd-research and supplies the change, questions, requested source classes, artifact store, and runtime capability declaration. Execute this phase directly; do not delegate.
Hard Rules
- Generated technical artifacts default to English. If technical artifacts are explicitly requested in another language, use a neutral/professional register. Public/contextual comments follow the target context language. Explicit user language or tone overrides win; otherwise use a neutral/professional register.
- Read
../_shared/research-lifecycle.mdand../_shared/sdd-phase-common.mdfirst. - Admit only
gentle-ai.sdd-research-capability/v1with exact declared grants fordocumentationoropen-web. - Never infer evidence capability from Bash, generic MCP, persistence access, filenames, or inherited unnamed tools.
- Denial, partial evidence, invalid sources, or persistence divergence emits no unvalidated claim and blocks proposal readiness.
- Keep evidence claims separate from non-authoritative product choices.
Decision Gates
| Condition | Outcome |
|---|---|
| Exact grants and complete mapped sources | done |
| Some questions remain unsupported | partial |
| Admission or persistence fails | blocked |
Execution Steps
- Retain the selected request and canonical desired content before source access or any write.
- Verify exact runtime grants for every requested class; stop on any denial.
- Collect sources and map each validated claim to source IDs, recording contradictions, uncertainty, and freshness.
- Persist
gentle-ai.sdd-research/v1and updategentle-ai.sdd-preproposal/v1using the active store contract. - In hybrid mode, write identical bytes to both stores. After a one-sided failure, use retained pre-write intent and canonical desired content—not either surviving store—to write a new positive revision to both stores, then read and compare both before readiness. If retained intent is unavailable, remain blocked and require explicit re-entry; never invent state.
Output Contract
Return status, executive_summary, artifacts, next_recommended, risks, and skill_resolution. Recommend orchestrator-owned product discovery only after done; otherwise recommend recovery.
References
../_shared/research-lifecycle.md../_shared/persistence-contract.md
版本历史
- cc4ed69 当前 2026-08-28 11:03
