AdaL

Build me a multi-model coding agent CLI to replace AdaL, in an empty folder. Stack: Python 3.12, one file per module, no framework. - Providers: Anthropic, OpenAI and Google, behind a single `chat(model, messages, tools)` function. Keys from .env via python-dotenv, never hardcoded. A `--model` flag switches families mid-session without restarting. - Tools the agent can call: read_file, write_file, list_dir, run_shell, run_tests. run_shell prints the command and waits for y/n unless I pass --yolo. - An orchestrator loop: a planner call splits my request into numbered steps, then each step runs as a fresh sub-agent with its own context window and only the tools it needs. Sub-agent results append to a shared markdown scratchpad on disk. - Persist every session as JSONL under .agent/sessions/ so I can replay or resume. A `--resume <id>` flag reloads the scratchpad and continues. - After any step that edits files, automatically run the test command from pyproject.toml and feed failures back to the same sub-agent for one retry. - Print a running token and dollar total per provider at the end of every turn. - Out of scope, deliberately: browser automation, a GUI, team accounts, and any hosted service. Those are the parts the subscription is actually selling. - README: setup, the .env template, how to add a fourth provider, and one honest paragraph on where this loses to a tuned commercial harness on long tasks.