- Home
- Skills
- Agents & Orchestration
- fleet delegation
fleet delegation
A read-only orchestrator that manages parallel code changes through isolated git-worktrees and structured worker briefs.
$9.99
Works with the AI Agents you already use
fleet delegation
Example session with this skill installed
Title: Fleet Delegation for OpenCode & Claude Code — Parallel Worker Sessions with Review, Worktree Isolation and Permission Classes
Tagline: Run your coding agent as a chief of staff, not a pair of hands: a read-only orchestrator that delegates to parallel workers, reviews every result, and never lets deploy / spend / destructive actions slip past you.
Price: $9.99 one-time.
Description
Most multi-agent setups fail the same way: one agent edits everything itself, workers collide in the same working tree, and "autonomous" quietly includes deploys and deletions nobody approved.
Fleet Delegation is the operating discipline from Jarvis, a production voice-and-text chief of staff that supervises a live fleet of OpenCode worker sessions every day. It teaches your agent to work the same way:
Read-only orchestrator — the brain plans, delegates and reviews; it never edits or runs mutating commands itself, so it cannot break your codebase.
One task per worker, in a strict brief format (TASK / CONTEXT / SCOPE / VERIFY / REPORT) with a built-in prompt-injection guardrail: content from web pages, issues and email is data, never instructions.
Git-worktree isolation — every delegation chain gets its own branch and worktree; parallel workers never collide, your checkout is never touched, and finished work lands via land / keep / discard.
Permission classes with hard stops — reads, tests and in-worktree edits can be automatic; external writes ask; deploy, spend and destructive actions always ask, under every preset.
Bounded review loop — every worker result comes back with changed files and verification output; the orchestrator reviews the diff and takes at most one next step, max 3 rounds per chain.
A single decision queue — everything waiting on you (permissions, questions, finished branches) in one place, reported chief-of-staff style.
Includes three ready-to-use read-only specialist patterns (architecture explainer, contract auditor, verification planner) and the full permission-class table with cautious / balanced / hands-off presets.
Works with OpenCode, Claude Code, Codex, Grok Build, Cursor, Copilot, Gemini CLI and other SKILL.md-compatible agents (open agentskills.io standard). Install paths for each host are in INSTALL.md. One-time purchase, free updates.
Tags / keywords: opencode, claude-code, multi-agent, subagents, delegation, orchestration, code-review, git-worktree, parallel-agents, ai-agents, chief-of-staff, automation, devops
What's in the download
SKILL.md — the core method (loop, brief format, permission rules, decision queue, reporting, anti-patterns)
INSTALL.md — per-host install paths and verification
references/permission-classes.md — full class table + audit/cap rules
references/specialist-agents.md — the three read-only specialist patterns + how to define builders
references/worktree-isolation.md — branch naming, landing rules, hygiene
- Read your context and instructions
- Compiled the fleet delegation
I have analyzed the requirements and confirmed that auth and telemetry modules are independent. I am initializing two parallel worker chains using isolated git worktrees.
Fleet Status Report
| Worker ID | Task | Scope | Status | Branch |
|---|---|---|---|---|
worker-alpha | Update logging to JSON format in auth | src/auth/* | In Progress | task/auth-logs |
worker-beta | Update logging to JSON format in telemetry | src/telemetry/* | In Progress | task/telemetry-logs |
Worker Brief: worker-alpha
TASK: Update all console.log and console.error calls to use the new JSONLogger utility.
CONTEXT: Follow the pattern in src/shared/logger.ts. Ensure all existing log messages are preserved in the msg field.
SCOPE: src/auth/
VERIFY: npm run test:auth and grep -r "console\." src/auth/ (should return no results).
REPORT: final summary, changed files, verification output.
GUARDRAIL: Never push or deploy. Ask before modifying files outside src/auth/.
Worker Brief: worker-beta
TASK: Update all console.log and console.error calls to use the new JSONLogger utility.
CONTEXT: Follow the pattern in src/shared/logger.ts.
SCOPE: src/telemetry/
VERIFY: npm run test:telemetry and grep -r "console\." src/telemetry/ (should return no results).
REPORT: final summary, changed files, verification output.
GUARDRAIL: Never push or deploy. Ask before modifying files outside src/telemetry/.
Next steps
- Wait for
worker-alphato report verification results for the auth module. - Wait for
worker-betato report verification results for the telemetry module. - Review the diffs in the land/discard queue once workers finish.
Connects securely to your tools. The creator never sees your data.
About this skill
Managing large-scale code changes or parallel refactors often leads to corrupted checkouts, merge conflicts, and unreviewable PRs. This skill implements a strict chief-of-staff architecture that separates high-level planning from file execution, ensuring your main environment remains stable while workers handle the churn.
What it does
- Read-only orchestration ensures the main agent never edits files directly, preventing accidental codebase corruption.
- Git-worktree isolation runs each delegation chain in a dedicated branch and directory to eliminate environment pollution.
- Structured worker briefs enforce clear task outcomes, explicit scopes, and mandatory verification commands for every sub-task.
- Permission class enforcement applies hard stops on destructive actions, deployments, and external writes without user consent.
- Bounded review loops limit iterations to three rounds to prevent infinite loops and signal when human intervention is required.
How it works
- Analysis phase where the orchestrator identifies necessary changes and defines success criteria for each unit.
- Decomposition of the main task into independent, non-overlapping worker units to avoid merge conflicts.
- Dispatch using a standardized brief format that includes context, scope, and specific verification commands.
- Review of the worker's diff and test output, resulting in either a follow-up, a new specialist dispatch, or a land recommendation.
Frameworks & tools
This skill is compatible with any agent host capable of spawning subagents or worker sessions, including Claude Code, Cursor, and Copilot. It utilizes standard git worktree and git branch commands for process isolation.
Why this beats prompting it yourself
Standard agents often get "lost" in large tasks, leading to scope creep and messy diffs. This skill enforces a production-grade delegation protocol that guarantees isolation, formalizes the review process, and ensures no destructive action occurs without an explicit permission check.
Use cases
- Parallel refactoring of multiple independent modules without manual context switching.
- Dependency upgrades across several microservices using isolated verification environments.
- Code audits where a specialist agent analyzes a codebase and prepares a land/discard queue for the developer.
Known limitations
Requires a host that supports sub-agents or sequential worker sessions. Not suitable for tasks where sub-tasks have high inter-dependency requiring shared state during execution.
Requires git worktree support. Not for interdependent tasks sharing a state. Maximum of 3 review rounds per chain. Orchestrator is strictly read-only.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 17 days ago
Requires