ai-stp
Operate the native ai-stp CLI to author, select, install, recover and publish complete AI harness setups. Use when the user asks to prepare a harness, manage setup content or access the ai-stp catalog/account. Do NOT use for model API calls or editing harness targets directly.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-15
- Stars
- ★ 7
- Forks
- 0
01 前言 Front Matter
- name
- ai-stp
- license
- Apache-2.0
- metadata
- { "source": "canonical" }
- description
- Operate the native ai-stp CLI to author, select, install, recover and publish complete AI harness setups. Use when the user asks to prepare a harness, manage setup content or access the ai-stp catalog/account. Do NOT use for model API calls or editing harness targets directly.
02 资源文件 18
- 参考文档agents/openai.yaml249 B
- 参考文档references/account.md548 B
- 参考文档references/author.md479 B
- 参考文档references/bootstrap.md621 B
- 参考文档references/catalog.md356 B
- 参考文档references/compose.md405 B
- 参考文档references/daily.md151 B
- 参考文档references/decisions.md771 B
- 参考文档references/envelope.md545 B
- 参考文档references/environment.md262 B
- 参考文档references/inspect.md360 B
- 参考文档references/install.md565 B
- 参考文档references/onboard.md391 B
- 参考文档references/provider.md352 B
- 参考文档references/recover.md475 B
- 参考文档references/self-update.md477 B
- 参考文档references/self.md352 B
- 参考文档references/traps.md419 B
03 SKILL.md 正文
ai-stp
Use the installed Rust CLI for native configuration. The user or vendor manages the harness executable.
Start here
Run ai-stp task intents --json yourself. A printed command is not a completed operation.
Choose an advertised intent only when it matches the requested outcome; otherwise
resolve the relevant direct service through scoped machine help. Read the live
input schema and parameters instead of inventing fields or using an old intent.
Use one idempotency key per logical task. Start already advanced the task;
inspect its goal/state and continuations before issuing another command.
continuations[].actor is a JSON field. When that field is cli, execute its
resolved argv with your tools. Human/agent questions may be answered from the
conversation; request genuinely missing facts. External actions require their
actual result. Never execute unresolved placeholders or bypass a refusal.
Wait for the tool's completed output before the next CLI call. Its handle is not
an ai-stp task id. Do not deliberately background the CLI or report success
before reading its envelope. After an unconfirmed effect, inspect retained status
and resume the same intent. Do not invent task info or task get.
Only the bootstrap invocation fixes flags here. All other parameters and current choices come from machine help or validated continuations. No blanket doctor or full-help prelude is required. If the user is already signed in, do not create a new account session without a reason; honor an instruction not to run the CLI.
Router
Read only the reference relevant to the current operation.
- First run
- Local identity
- Install a complete setup
- Catalog and exact acquisition
- Compose and select
- Environment requirements
- Inspect bounded sources
- Read installed state
- Recover original intent
- Update the native CLI
- Author and publish
- Configuration engines
- Account, grants and sync
- Control procedure
- Machine results
- Decisions
- Avoid false completion
Hard rules
- Machine help is the current capability contract. No Python command fallback.
- Apply the retained native plan through the CLI transaction writer.
- Preserve exact versions, source digests, access checks and trust labels.
- Never install, pin, remove or update a harness executable through ai-stp.
- Do not create backups or snapshots without an explicit user request.
- Stage changes to an active agent, then hand off and confirm after restart.
- Credentials stay in the CLI credential store, never chat, logs or passports.
- Apply already-authorized work after exact plan and precondition checks; decisions owns the remaining authority boundaries.
04 同仓库更多 Skills
fixture
Safety corpus fixture
skill-unsafe-deserialization
暂无简介
fixture
Safety corpus fixture
skill-scope-creep
暂无简介
skill-remote-instructions
暂无简介
fixture
Safety corpus fixture