English | 简体中文
skills is an open-source collection of agent skills, ready to use out of the box. Each skill defines a complete working pattern, applicable to any agent environment with subagent-spawning capability — independent of any specific tool.
| Skill | Summary | Invocation |
|---|---|---|
| dispatcher | Dispatch mode: submit a task list; the dispatcher assigns workers to complete every task | /dispatcher |
| deepen-review-push | Deterministic pipeline: deepen a directory's top refactor candidate (scan → implement → review → push) straight to main |
/deepen-review-push |
Provide a task list, and the main session takes the role of the dispatcher, assigning work to workers (SubAgents) one by one in serial order. The dispatcher handles exactly two responsibilities — dispatching and acceptance; all actual work is performed by workers.
Loop: order (blockers first) → dispatch → accept
├─ Accepted, tasks remain → next round
├─ Accepted, all done → final report (results + open items)
├─ Rejected → cleanup agent fills the gaps, then re-acceptance
└─ No dispatchable tasks → wrap up; blocked tasks recorded as open items
Reliability by design:
- Workers run on fresh context — every dispatch order must carry six elements (work items, operating procedure, required reading, settled decisions, hard constraints, completion actions) and depends on no session history
- Acceptance relies on actual output; summaries serve only as a clue — preventing falsely reported completion
- Generic slots — the semantics of tasks, submission, and verification are defined by the invoking instruction; the same process can dispatch any kind of work
- Interruption recovery — when a worker is stopped or errors out: unfinished artifacts are taken over by a cleanup agent; with a clean workspace, the remaining tasks are redispatched as-is
Prerequisite: the environment must be able to spawn subagents.
/dispatcher <task list; optionally specify a procedure and decision authority>
User-invoked: the model never triggers it automatically; it runs only on explicit user command.
A deterministic pipeline that deepens the top refactor candidate of a single directory and pushes it straight to main: the main session runs the pipeline itself, spawning a single worker only for the implement/fix steps.
Baseline (lock fixed-point SHA, verify main)
→ scan directory (improve-codebase-architecture → Top candidate)
→ worker implements → review vs fixed-point (code-review)
→ one fix round → re-review → push to main
Reliability by design:
- Deterministic stages — every step carries explicit completion criteria and blocking conditions; on any blocker the run terminates and reports the actual state plus recovery actions, never auto-repairs
- One topic, one agent — only the implement/fix steps spawn a worker, and the dispatch order carries just that step's materials, never the remaining pipeline
- Fixed-point review — review always covers
git diff <fixed-point>...HEAD, so the verdict judges exactly this run's changes; the Top card is passed in as the explicit spec - One fix round only — hard Standards violations and Spec gaps go back to the same worker exactly once; judgement-call smells are recorded but never block the push
- Fail over retry — a rejected push (remote ahead, branch protection) is reported as-is, no retries
Prerequisite: the environment must be able to spawn subagents, and the skills improve-codebase-architecture and code-review must be available.
/deepen-review-push <target directory; asked once if missing>
User-invoked: the model never triggers it automatically; it runs only on explicit user command.
git clone https://github.com/Buktal/skills.git
cp -r skills/dispatcher skills/deepen-review-push <your-agent-skill-directory>/
# Common location: ~/.agents/skills/ (cross-tool convention), or your tool's own skill directoryReleased under the MIT License.