Improvement Orchestrator
v1.0.0当需要一键跑完「生成→评分→评估→执行→门禁」全流程、失败后自动重试、或批量改进多个 skill 时使用。不用于单独评估 skill 质量（用 improvement-learner）或手动打分（用 improvement-discriminator）。
⭐ 0· 58·0 current·0 all-time
by_silhouette@lanyasheng
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
The name/description (orchestrating a 5‑stage improvement pipeline) matches the actual behavior: the script dispatches Generator→Discriminator→Evaluator→Executor→Gate and writes state/artifacts and backups. All declared requirements (none) are appropriate for a local orchestrator that runs other local scripts.
Instruction Scope
SKILL.md and scripts explicitly instruct the agent to run local subprocesses, read feedback sources, and apply changes to the target skill (append markdown sections, create files) with backups/rollback. This is expected for an orchestrator, but it means the skill will read/write arbitrary files under the provided --target and --state-root and can forward failure traces into subsequent runs. The orchestrator itself does not call external endpoints, but it invokes other scripts (e.g., evaluator) which may call LLM CLIs or network services — review those scripts before use.
Install Mechanism
Instruction-only (no install spec). The bundle includes orchestration code only; nothing is downloaded or extracted from external URLs. Lowest install risk.
Credentials
The skill declares no required env vars or credentials, which is coherent. Caveat: the orchestrator spawns other local scripts (generator/discriminator/evaluator/executor/gate) that are expected to live in the repo; those sub-scripts may themselves require API keys or credentials (e.g., for LLMs) even though the orchestrator doesn't declare them. Confirm what the invoked scripts expect before running with real task suites.
Persistence & Privilege
always=false and no special platform privileges. The orchestrator writes persistent artifacts and backups to the user-supplied --state-root (normal for its purpose). It does not modify other skills' configuration beyond running the standard executor workflow for the provided --target, but it will apply changes to the target path (intended behavior).
Assessment
This skill is an on‑repo pipeline orchestrator: it will run local scripts (propose/score/evaluate/execute/gate), create state artifacts and backups, and may apply changes to files under the --target you provide. Before running: 1) Inspect the scripts it invokes in your repository (improvement-generator/discriminator/evaluator/executor/gate) to ensure they do only what you expect; 2) Run first with a disposable --state-root and use --eval-mock (avoid real LLM CLI calls) to observe behavior; 3) Backup your target skill or point --target at a copy if you are not ready for automatic modifications; 4) Be aware that the evaluator or other invoked scripts may require separate API keys or env vars — the orchestrator itself does not request credentials. If you want to be extra cautious, run the included tests and review the executor's logic to confirm it only performs allowed actions (append_markdown_section, create_file) for low‑risk categories.Like a lobster shell, security has layers — review code before you run it.
latestvk975camepbb5szv1ma8bv9ksas848ckq
License
MIT-0
Free to use, modify, and redistribute. No attribution required.