overnightworks

Agent Providers — Development skill for Claude Code

Development community

Agent provider layer (Claude, Grok, Codex over subscription-login CLIs / APIs): routing, live catalogs, turn stream, tool loop, bounded process runner, status probes, sandbox building blocks.

How to install Agent Providers

This entry records only its repository, not the path inside it, so there is no exact command to give. Open overnightworks/agent-providers and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Agent Providers does

Agent provider layer (Claude, Grok, Codex over subscription-login CLIs / APIs): routing, live catalogs, turn stream, tool loop, bounded process runner, status probes, sandbox building blocks. Extracted from songmaker (#825).

Alternatives in Development

Connect — Fancy stream processing made operationally mundane 8608 114 1 8.6k ★
14 Dev Runner — Prompt 14: Create Development Runner 2.2k ★
Ctf 24h Fleet — 24h unattended multi-target Web CTF fleet loop runner 1.1k ★

README

agent-providers

The reusable agent provider layer: it routes a turn to Claude, Grok or Codex — over each provider's CLI or its HTTP API — and owns the machinery a host application should not rebuild. Live model catalogs and readiness probes, the streaming turn transport, the tool loop, a bounded subprocess runner, and the sandbox directory-name constants all live here; everything that touches a host's own truth reaches the library through the ports the host supplies.

Distribution `overnightworks-agent-providers`, import package `agent_providers`.

What it owns, and what it leaves to the host

It owns the machinery that is the same in every deployment:

Module	Owns
`agent_providers.config`	The deployment facts every module reads, installed once per process via `configure()`
`agent_providers.catalog`	Live provider/route readiness probes and model listings (`probe_provider_route`, `list_provider_models`)
`agent_providers.dispatch`	Routing one turn to a provider and route, and the cover-image capability
`agent_providers.events`	The `StreamEvent` union a turn emits (assistant text, tool call, tool result, final)
`agent_providers.tool_loop`	The provider-agnostic tool loop driving a `ToolExecutor` over a `ToolTransport`
`agent_providers.tools`	The `ToolCatalog` and its Anthropic/OpenAI schema projections
`agent_providers.process`	`run_cli_bounded`: the one bounded subprocess layer (start, stdin, first line, deadline, process-group cleanup)
`agent_providers.images`	The `ImagePolicy` an image turn is held to
`agent_providers.sandbox`	The directory-name prefix constants a host's kernel sandbox profile must permit

It leaves the application everything the library cannot know: the MCP tool set and its execution, the prompts, where settings are stored, and domain ownership. The library depends on `pydantic`, `httpx` and `anyio` at runtime and on nothing else — no host application, no web framework, no d