Config (dependency injection)¶

Config (pycodeloop.core.config.Config) is a small dependency-injection container — build one, hand it a Provider and a tool set, pass it to CodeLoop:

config = Config(
    provider=GenericProvider.from_json("templates/anthropic.json"),
    tools=DEFAULT_TOOLS,
    system_prompt="You are a terse code reviewer.",
    max_turns=25,
)

Arg	Default	Purpose
`provider`	resolved from `PYCODELOOP_PROVIDER`/`PYCODELOOP_MODEL` env vars	LLM backend driving the agent
`tools`	`DEFAULT_TOOLS`	Tools exposed to the agent
`system_prompt`	`pycodeloop.core.agent.DEFAULT_SYSTEM_PROMPT`	Overrides the default instructions
`max_turns`	`25` (or `PYCODELOOP_MAX_TURNS`)	Hard cap on tool-use loop iterations
`max_history_turns`	`20`	Cap the session on the number of most recent user-initiated turns kept before each provider call; pass `None` for unbounded growth
`skills`	`False`	Discover Claude Code/Cursor/`AGENTS.md` skills on disk, expose a `read_skill` tool, and list them in the system prompt
`skill_sources`	all sources	Limit discovery to a subset (`"claude-skill"`, `"claude-memory"`, `"cursor-rule"`, `"agents-md"`)
`skills_refresh`	`False`	Skip the skills cache and force a full rescan
`storage`	`SqliteSessions()` (`~/.pycodeloop/pycodeloop.db`)	Persists sessions so `CodeLoop.run(prompt, session_key=...)` can resume across restarts; pass `False` for in-memory only

Config.__init__ validates that provider is an instance of the Provider ABC and raises NotProviderInstance immediately if not — a wrong object fails at construction time, not three tool calls into a run.

Default provider resolution¶

If provider is omitted, Config resolves one from pycodeloop.settings.Settings, which reads:

PYCODELOOP_PROVIDER — a path to a JSON config file (a bundled templates/anthropic.json-equivalent is the default), the bare string "generic" paired with an explicit url=, or a module.path:ClassName for a custom provider
PYCODELOOP_MODEL — overrides the model named in the JSON config (or passed to a custom provider)
The API key env var named by the JSON config's own api_key_env field (ANTHROPIC_API_KEY for the bundled default) — read automatically, no separate PYCODELOOP_* var needed

CodeLoop: Config + Agent + Session¶

CodeLoop (pycodeloop.core.codeloop.CodeLoop) is a thin wrapper: it takes a Config, builds the Agent and a Session, and exposes .run(prompt) that keeps conversation history across calls.

from pycodeloop import CodeLoop, Config
from pycodeloop.providers import GenericProvider

flow = CodeLoop(Config(provider=GenericProvider.from_json("templates/anthropic.json")))
flow.run("what does this repo do?")
flow.run("now add a test for it")  # remembers the first turn