delegate-kit
Powerjackie/delegate-kit
Cross-platform skill suite for delegating work to subagents in Claude Code / OpenCode / Codex. Classifier + Research/Executor/Reviewer brief templates with structured JSON outputs.
Instalação
npx convez-ai Powerjackie/delegate-kit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
delegate-kit
A cross-platform skill suite for delegating work to subagents in Claude Code and OpenCode. Classifier + Research / Executor / Reviewer brief templates with structured JSON outputs.
Status: v0.1 — scaffolding complete, two self-bootstrap evals run
(see examples/); not yet validated against unrelated
production workflows. Use at your own discretion; expect rough edges.
Feedback and example briefs welcome (see CONTRIBUTING.md).
Runtime (Claude Code): delegate-flow runs these same briefs and schemas on the Workflow engine — file-ownership disjointness becomes a hard gate, output shape is enforced at the tool layer, and research findings are adversarially verified. delegate-kit defines what to say to a subagent; delegate-flow defines how to run a fleet of them, deterministically.
Why this exists
When you delegate a task to a subagent, three things go wrong by default:
- The subagent has no parent-chat context — phrases like "as discussed above" silently fail.
- There's no enforced output shape — the main agent has to re-read long transcripts to merge results.
- There's no agreed file-ownership contract — parallel subagents conflict on shared files.
delegate-kit is a four-skill suite that gives a personal-development
agent loop (Claude Code, OpenCode, similar tools) a small, explicit
delegation protocol to address all three.
What's inside
| Skill | Triggers on | Use it to write a brief for… |
|---|---|---|
delegate-orchestrate | "delegate", "parallelize", "spawn agents", "audit", "派 subagent", "并行", "调研" | The main agent's classifier — decides whether to delegate, picks role, picks model tier, points at the role-specific skill |
delegate-research | "investigate", "explore", "audit", "trace", "调研" | A read-only subagent that returns findings + evidence as JSON |
delegate-execute | "implement", "fix in ", "refactor", "add tests for" | An Executor subagent with explicit file_ownership and validation commands |
delegate-review | "review", "audit", "score", "validate", "审查", "评审" | A Reviewer subagent that scores artifacts against a rubric and gives a merge recommendation |
Each role skill ships a brief template and a required output JSON schema. Brief templates are Markdown (human-writable, copy-pasteable); outputs are fenced JSON (machine-mergeable by the main agent).
Install
Claude Code (user-level, default)
git clone https://github.com/<your-fork>/delegate-kit ~/delegate-kit
~/delegate-kit/install/claude-code.sh
Copies the four skills into ~/.claude/skills/. Reload Claude Code
to pick them up.
Claude Code (project-level)
~/delegate-kit/install/claude-code.sh --project /path/to/your/project
Installs into <project>/.claude/skills/ — useful when the skills should
travel with a repo.
OpenCode
OpenCode reads ~/.claude/skills/ by default, so the Claude Code install
usually covers it. If you've set OPENCODE_DISABLE_CLAUDE_CODE_SKILLS=1,
install explicitly:
~/delegate-kit/install/opencode.sh
OpenAI Codex CLI
Codex uses the same SKILL.md format as Claude Code. Install into
~/.codex/skills/:
~/delegate-kit/install/codex.sh
Codex auto-discovery from ~/.codex/skills/ is unverified at v0.1 — if
the skills do not load after install, the alternative is to register
delegate-kit as a Codex plugin marketplace (codex plugin marketplace).
Tracked as a v0.2 candidate.
Dev mode (hacking on this repo)
Use --symlink so edits to skills/*/SKILL.md go live without re-running
the installer:
~/delegate-kit/install/claude-code.sh --symlink
Windows
The install scripts are Bash. On native Windows, use WSL / Git Bash, or
manually copy skills/delegate-*/SKILL.md into
%USERPROFILE%\.claude\skills\. A PowerShell installer is tracked for v0.2.
How it triggers
When you ask the main agent to do something delegation-shaped — e.g., "audit how token refresh works across these modules" or "并行检查这三个 分支的认证逻辑" — Claude reads the matching skill's description and loads its body. The skill then guides the main agent through:
- Classify (orchestrate): delegate or not? Which role?
- Write brief (role-specific skill): copy template, fill
task_id,parent_task,file_ownership,acceptance_criteria, etc. - Spawn subagent via the agent platform's Agent / Task tool with the appropriate tool permissions.
- Merge by parsing the subagent's required JSON output.
Skills are soft policy — they guide the agent's behavior when matched,
but they are not hard sandboxes. Real permission boundaries live in the
platform's tool / permission settings (Claude Code allowed-tools,
OpenCode permission.task, etc.).
Design choices
- Markdown briefs, JSON outputs. Asymmetry is deliberate: humans write briefs (Markdown reads well), machines merge outputs (JSON parses cleanly).
- File ownership is the parallel-safety primitive. Two subagents may
run in parallel iff their
exclusive_writesets are disjoint. No separate "parallel constraint" bookkeeping. - Three roles, not ten. Research / Executor / Reviewer cover most delegation needs. Add more only after observing repeat patterns.
- Cross-platform from the start, lightly tested. Skill bodies avoid OS-specific commands and require subagents to use repo-relative paths. Honest status: developed on macOS, designed portable, Linux/Windows untested. PRs documenting platform gotchas welcome.
Limitations
- v0.1 has only been validated against two self-bootstrap evals (the
kit was used to scaffold and document itself, see
examples/). Brief-template fields are likely to shift after the first delegation on an unrelated production codebase. - Native Windows install requires WSL or manual copy.
- No automated SKILL.md linter yet (planned for v0.2).
License
Apache License 2.0 — see LICENSE.