skill

agent

Event-driven task orchestration: Kafka events → Kubernetes Jobs running AI agents.

bborbe/agentcommunitylistedunreviewed

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/bborbe/agent itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 22, 2026

agent

SDK + Claude Code plugin for the bborbe agent platform.

Go SDK — shared types, schemas, and runtime helpers consumed by every agent + task-system service in the ecosystem
Claude Code plugin — /launch-agent slash command for interview-driven scaffolding of new agents

Quick start

Install the plugin (in Claude Code):

claude plugin marketplace add bborbe/agent
claude plugin install agent

Scaffold a new agent:

/launch-agent <name>

Walks you through the [[Agent Design Guide]] interview, recommends a reference shape (claude/code/gemini/pi), clones the matching template repo via gh repo create --template, customizes the clone, writes vault artifacts (knowledge page, goal, scenario, NEXT-DIRECTIONS), and prints a deploy checklist. See commands/launch-agent.md for the workflow.

What's in here

Single Go module at github.com/bborbe/agent with these subpackages:

Package	Role
`agent` (root)	`Agent`, `Phase`, `Step`, `Status`, `Task`, `TaskFrontmatter`, parser/markdown helpers — the runtime contract every agent honors; also the backend-independent `Session` / `SessionFactory` seam for interactive conversations
`claude/`	Claude Code runner helpers (used by agent-claude template); also the `Session` implementation that holds one long-lived `claude` process per session over the CLI's stream-json protocol, with mid-turn tool-permission requests routed out to a `PermissionDecider`
`pi/`	MiniMax `pi` runner helpers (used by agent-pi template)
`command/`	CQRS command shapes (`task.CreateCommand`, `task.UpdateFrontmatterCommand`, `task.IncrementFrontmatterCommand`) + `ErrTaskAlreadyExists` sentinel
`delivery/`	`ResultDeliverer` interface (Kafka + file deliverers)
`envparse/`	Env-var parsing helpers for agent main.go bootstraps
`healthcheck/`	Generic agent liveness handler
`interactive/`	Shared interactive HTTP surface — readiness, metrics, prompt intake, the permission endpoint and an A2A surface (an Agent Card at `/.well-known/agent-card.json` and the `/a2a` JSON-RPC endpoint) over a per-session cache of long-lived conversations that drops a session once it has gone the `sessionIdleTimeout` passed to the constructor without serving a turn, and drops the least recently used session when the cache is at the `maxSessions` passed to the constructor (defaulting to `interactive.DefaultMaxSessions`, 8, when non-positive), rebuilding either on that session's next turn; the prompt, permission and `/a2a` routes require an `Authorization: Bearer <token>` header while readiness, metrics and the Agent Card stay open; both constructors also take an optional trailing `interactive.WithPushNotifications(store, sender)` that wires the caller's own push config store and sender through the SDK and makes the Agent Card advertise `capabilities.pushNotifications` — derived from the option, never hardcoded — with the detailed contract in `docs/interactive-service.md`
`metrics/`	Prometheus metrics for agent + executor runtime
`mocks/`	counterfeiter-generated test doubles

Consumers

Each agent + task service lives in its own repo and imports this SDK:

Repo	Role
agent-claude	AI-heavy reference template (Claude Code) — `is_template: true`
agent-code	Pure-Go reference template (deterministic phases)
agent-gemini	Boundary-translator reference (Gemini at planning edge)
agent-pi	Tier-D LLM reference (MiniMax Pi)
agent-task-controller	Single git writer for the vault (Kafka → git via git-rest)
agent-task-executor	Kafka event consumer + per-task K8s Job spawner

Producers (emit task commands)

recurring-task-creator — Schedule CR → cron tick → CreateCommand
maintainer — watcher/github-pr / github-build / github-release → CreateCommand
manual — /vault-cli:create-task via vault-cli

Use as a Go module

go get github.com/bborbe/agent

import (
    "github.com/bborbe/agent"
    "github.com/bborbe/agent/delivery"
    "github.com/bborbe/agent/command/task"
)

History

This repo was a monorepo through 2026-06-24, hosting task/controller/, task/executor/, and 4 reference agents under agent/{claude,code,gemini,pi}/ as Go sub-modules + the SDK under lib/. On 2026-06-25 each service and reference agent was extracted to its own repo (see Consumers table above), the SDK was promoted from lib/ to the repo root, and the module identity collapsed from github.com/bborbe/agent/lib to github.com/bborbe/agent. See CHANGELOG.md ## v0.70.0 for the migration guide.

Older import path github.com/bborbe/agent/lib/... continues to resolve via historical tags (latest v0.69.0) for any consumer not yet migrated; new development happens at the flat path.

Plugin layout

.claude-plugin/, commands/, agents/, skills/, scenarios/ at repo root — same convention as bborbe/coding and bborbe/dark-factory.

Path	Role
`.claude-plugin/plugin.json` + `marketplace.json`	Plugin metadata
`commands/launch-agent.md`	The `/launch-agent` slash command (thin dispatcher to the skill)
`agents/agent-shape-picker.md`	Sonnet subagent: use case → shape recommendation with reasoning
`skills/launch-agent/SKILL.md`	8-phase orchestrator (interview → shape → clone → customize → render templates → commit → checklist)
`skills/launch-agent/references/`	shapes.md + interview.md + 4 output templates (config-crd, vault-page, goal, scenario) + next-directions
`scenarios/001-launch-agent-happy-path.md`	End-to-end smoke test of the scaffolding flow

Architecture references

docs/kafka-schema-design.md — Kafka topic + command schema design
docs/task-flow-and-failure-semantics.md — task lifecycle + failure modes
docs/agent-job-interface.md — what every agent main.go must implement
docs/agent-job-lifecycle.md — phase + step lifecycle
docs/dod.md — definition of done
docs/deployment.md — platform deploy notes
Agent Hub — full architecture catalog (in personal vault)
Quick-Launch New Agents — the goal this split landed under

License

BSD-2-Clause.