kidwit

A complete coding-agent engine in one file of standard-library Python, around 500 lines of code. It works inside one directory and asks before it changes anything.

kidwit reads your code, edits files, runs commands and reports back, using native tool calling with Anthropic or any OpenAI-compatible API. It is a from-scratch build and a case study in how little a coding agent needs. It is not a replacement for mature agents.

Website: kidwit.sabbiramin.dev · PyPI: pypi.org/project/kidwit · Changelog · Issues

Demo

A real, unedited session: kidwit finds and fixes an off-by-one in a small project. The model is deepseek-chat through the OpenAI-compatible adapter. kidwit shows the diff and waits for y before it edits the file; running the tests needs no prompt, because plain test commands are on its safe list. Waits for the model longer than five seconds are cut to five, and the GIF plays at 0.8x. The recording is docs/demo.cast, which you can play with asciinema play docs/demo.cast, or watch with pause and copy at kidwit.sabbiramin.dev.

Install

You need Python 3.10 or newer. kidwit has no dependencies. The current release is a candidate (0.1.0rc2), so the version is pinned:

pipx install kidwit==0.1.0rc2       # or: uv tool install kidwit==0.1.0rc2, or: pip install kidwit==0.1.0rc2
export ANTHROPIC_API_KEY=sk-ant-...
kidwit "fix the failing test"

Or with one line, on macOS and Linux:

curl -fsSL https://kidwit.sabbiramin.dev/install | sh

The script is scripts/install.sh in this repository, so you can read it first. It never uses sudo and writes only inside your home directory. It installs with pipx or uv if you have one. Otherwise it downloads kidwit.py from the latest GitHub Release, checks its SHA-256 checksum, and puts a launcher at ~/.local/bin/kidwit. If there is no release file yet, it installs the PyPI package into its own virtual environment instead. Set KIDWIT_VERSION=0.1.0rc2 to pick a version, or KIDWIT_DRY_RUN=1 to see what it would do.

The whole engine is one file, so you can also run it from a clone with no install at all:

git clone https://github.com/sabbiramin113008/kidwit.git
cd kidwit
python3 src/kidwit.py "fix the failing test"

Without a task, kidwit starts an interactive session and, on a color terminal, shows a small banner. The banner never appears with --json, with a task on the command line, or when output is piped. /clear resets the conversation, /verbose toggles detailed output, and /exit (or /quit, or end of input) leaves.

Status

kidwit is early software. This is what has and has not been tested:

1. Tested live: the OpenAI-compatible adapter with DeepSeek (deepseek-chat) on macOS with Python 3.14. The 8 tasks in evals/ all pass end to end.
2. Not yet run against the real service: the Anthropic adapter. It is covered by the offline tests, but nobody has run it against the live Anthropic API yet.
3. Tests pass in CI on Linux, macOS and Windows, with Python 3.14. CI runs the offline tests and the eval self-test, not a live model. Nobody has run kidwit against a live model on Linux or Windows yet. On Windows the run tool uses cmd.exe, so commands the model writes in Unix shell syntax may fail there.
4. Python versions: CI runs 3.14 only. 3.10 to 3.13 passed once on Linux and macOS, and are not part of the regular checks yet. Python 3.9 is not supported from 0.1.0rc2 on.

If you try it somewhere on this list, please open an issue with what happened.

Configuration

Environment variables:

Variable	Meaning
ANTHROPIC_API_KEY	Key for the Anthropic Messages API. If set, Anthropic is the default provider.
OPENAI_API_KEY	Key for an OpenAI-compatible API. Optional for local servers.
AGENT_PROVIDER	anthropic or openai.
AGENT_MODEL	Model name. Defaults to claude-sonnet-5-5 for Anthropic. Required for openai.
AGENT_BASE_URL	API base URL, for OpenRouter, Ollama, LM Studio, vLLM or a proxy.

Provider choice follows this order: --provider, then AGENT_PROVIDER, then Anthropic if ANTHROPIC_API_KEY is set, then OpenAI-compatible.

Local model example:

AGENT_BASE_URL=http://localhost:11434/v1 AGENT_MODEL=qwen2.5-coder kidwit "add a docstring to main()"

Flags:

Flag	Meaning
--root DIR	Project directory. Default: the current directory.
--provider anthropic|openai	Choose the API.
--model M, --base-url URL	Override AGENT_MODEL and AGENT_BASE_URL.
--yes	Never ask for approval.
--read-only	Offer only the read tools. Writes and commands are not available.
--json	Print JSON-lines events instead of text.
--verbose	Show live command output and full tool arguments. By default kidwit prints one short dim line per tool call and then the answer.
--no-log	Do not write .agent/audit.jsonl.
--version	Print the version.

A task given on the command line runs once. The exit code is 0 on success and 1 on error, so it works in scripts and benchmarks.

Reading the terminal

On a color terminal each kind of text looks different. Your prompt is a bold green >>. The model's replies have a magenta ▌ gutter. Tool calls are one dim line with the tool name in cyan. Permission requests start with a yellow ┌ permission needed line, followed by the diff or command and a yellow └ allow? prompt. Errors are red. With NO_COLOR, piped output or --json, none of this styling is printed.

Tools

The model gets six tools: list_files, search, read_file, write_file, edit_file and run. Inside a git repository, list_files and search honour .gitignore, so build output and similar files stay out of the way. Edits are exact-string replacement: a match that differs only in trailing whitespace is accepted, and when nothing matches the error names the closest line so the model can correct itself. old_text must match exactly once unless replace_all is set, and files with Windows line endings keep them.

If the project root has an AGENTS.md (or else a CLAUDE.md), its first 20,000 characters are added to the system prompt.

Approval modes

1. ask (default). Before every write_file, edit_file and run, kidwit shows the diff or command and waits for y, a, n, or any other text. Other text is sent to the model as the reason for the denial. a (always) approves that kind of action for the rest of the session: file changes, or commands, but not both.
2. auto (--yes). No prompts, except for dangerous commands.
3. read-only (--read-only). The write tools are not offered to the model, and are refused if called anyway.

Some things are handled for you in every mode:

1. In ask mode, harmless commands run without a prompt: ls, pwd, git status, git diff, git log, git show, git branch, and test runners such as python3 -m unittest, pytest, npm test, go test and cargo test. They must use plain relative arguments with no shell operators. cat, grep and head are not on the list on purpose: the path jail does not cover run, so they could read files outside the project.
2. Dangerous commands always ask, even with --yes or after a: rm -rf, sudo, curl ... | sh, git reset --hard, git clean, git push, chmod -R and similar. With no one to answer (for example --json with closed stdin) they are denied.
3. /undo in an interactive session reverts the last file write or edit. It does not undo commands.
4. If the model's reply is cut off at the token limit, the half-written tool call is not run. The model is told to do less per call.

Every write, edit and command is appended to <root>/.agent/audit.jsonl. kidwit creates .agent/.gitignore so the log stays out of git.

Limits you should know about

The path jail covers the file tools only. read_file, write_file and edit_file refuse paths outside the project root, including ../ and symlinks that lead out, and list_files and search skip such symlinks. run executes real shell commands, and those can reach anything your user account can reach. The protections in v1 are the approval prompt and --read-only. There is no OS-level sandbox, so review each command before you approve it, and avoid --yes on a machine or directory you care about.

Long sessions are kept within the model's context in two steps. First, old tool output is dropped. If the history is still too large (over about 200,000 characters), kidwit asks the model to summarize the older half of the conversation and continues from that summary.

Other deliberate omissions: no streaming of model responses (command output does stream), no sub-agents, no MCP, no saved sessions, and no UI beyond plain terminal text.

Event protocol

With --json, kidwit writes one JSON object per line on stdout. Tasks and approval answers are read from stdin, one line each. End of input during an approval counts as "no". A UI can drive the engine through this protocol.

type	Fields
info	message
text	text
tool	id, name, input
output	line (live command output)
approval	action (write, edit or run), target, detail (diff or command)
result	id, output, is_error
done	turns, input_tokens, output_tokens
error	message

Tests

The tests use only unittest and need no API key. A fake local HTTP server stands in for both APIs.

scripts/check.sh                              # everything below in one go
python3 -m unittest discover -s tests -t .    # just the tests

scripts/check.sh checks the Python version, compiles src/kidwit.py, confirms it imports only the standard library, runs the tests, and checks the size (grep -vcE '^\s*(#.*)?$' src/kidwit.py). Over 500 lines prints a warning, and over 700 fails.

Evals

evals/ holds 8 small tasks, each a tiny repo with a failing check: an off-by-one, a name error, implementing a function, a root cause in a second file, a rename across files, a CRLF file, a missing config default, and a new CLI flag. They show whether a change to kidwit makes it better or worse.

python3 evals/run_evals.py --selftest     # checks the tasks themselves, no model needed
python3 evals/run_evals.py                # runs every task against your configured model
python3 evals/run_evals.py --task crlf-file -- --provider openai --model deepseek-chat

The runner prints pass or fail, turns, tokens and time per task, and saves the results to evals/results/. Be aware that it runs the agent with --yes, so the model's commands run as you with no sandbox. Dangerous commands still need approval, and because the runner closes stdin they are denied. Run it only where that is acceptable.

License

MIT.