kcmt 0.3.1

AI-powered atomic Git staging and committing tool with LLM-generated conventional commits

kcmt — AI-powered atomic Git staging and committing

kcmt is a small Python library and CLI that helps you:

• Parse and stage changes atomically (file-by-file).
• Generate clear, conventional-commit messages using an LLM.
• Commit safely with validation, retries, real-time progress, and helpful output.
• (Optional) auto-push successful commits to your remote.

It’s designed to be drop-in for your local repositories and integrates with multiple hosted LLM providers (OpenAI, Anthropic, xAI, GitHub Models).

Key features

• Atomic workflow: stage and commit per-file, with deletions handled first.
• LLM-assisted messages: conventional commit style with validation, retries, and auto-fixes.
• Strict failure on repeated invalid/empty LLM responses (no heuristic commit synthesis).
• Prepare phase aborts after 25 per-file failures/timeouts to avoid wasting additional requests.
• Built-in metrics summary (diff, queue, LLM, commit timings) to diagnose performance quickly.
• Connection pooling for provider APIs and batched diff collection for speed on large repos.
• Multi-provider support: OpenAI, Anthropic, xAI, and GitHub Models via a guided wizard.
• Parallel preparation: generate per-file commit messages concurrently with live stats.
• Automatic push to origin on success by default (use --no-auto-push to disable).
• Pricing-aware model selection and a cross-provider pricing board (--list-models).
• Built-in benchmarking across providers/models with a CLI leaderboard and optional JSON/CSV output.
• Small and composable core: use the CLI or import the library directly.

Supported Python versions

• Python 3.12+

Installation

Install from the repository with pip or uv:

• Latest from GitHub (subdirectory install):

  • pip:
    • pip install "git+https://github.com/djh00t/arby#subdirectory=kcmt"
  • uv:
    • uv pip install "git+https://github.com/djh00t/arby#subdirectory=kcmt"

• Local editable install (from the monorepo root):

  • pip:
    • pip install -e ./kcmt
  • uv:
    • uv pip install -e ./kcmt

Dependencies

• openai>=1.108.1 (shared client for OpenAI-compatible providers)
• httpx>=0.25.0 (Anthropic REST client)

Configuration

Run kcmt --configure inside a repository to launch a colourful wizard that:

• Detects available API keys (OPENAI_API_KEY, ANTHROPIC_API_KEY, XAI_API_KEY, GITHUB_TOKEN).
• Lets you choose the provider, tweak model/endpoint, and pick the env var to use.
• Persists the selection to ~/.config/kcmt/config.json (global, per-user).

Per-provider settings

kcmt maintains a per-provider settings map in ~/.config/kcmt/config.json. Each supported provider has:

• name: friendly display name
• endpoint: base URL for API calls
• api_key_env: environment variable that holds your API key/token (kcmt stores the variable name, not the secret)
• preferred_model: your saved default model for that provider

Example (~/.config/kcmt/config.json excerpt):

{
  "provider": "openai",
  "model": "gpt-5-mini-2025-08-07",
  "llm_endpoint": "https://api.openai.com/v1",
  "api_key_env": "OPENAI_API_KEY",
  "auto_push": true,
  "providers": {
    "openai": {
      "name": "OpenAI",
      "endpoint": "https://api.openai.com/v1",
      "api_key_env": "OPENAI_API_KEY",
      "preferred_model": "gpt-5-mini-2025-08-07"
    },
    "anthropic": {
      "name": "Anthropic",
      "endpoint": "https://api.anthropic.com",
      "api_key_env": "ANTHROPIC_API_KEY",
      "preferred_model": null
    },
    "xai": {
      "name": "X.AI",
      "endpoint": "https://api.x.ai/v1",
      "api_key_env": "XAI_API_KEY",
      "preferred_model": null
    },
    "github": {
      "name": "GitHub Models",
      "endpoint": "https://models.github.ai/inference",
      "api_key_env": "GITHUB_TOKEN",
      "preferred_model": null
    }
  }
}

First-use model selection

• When you first run kcmt against a provider that doesn't yet have a preferred_model, the CLI shows a pricing-aware model menu and saves your selection under providers[<provider>].preferred_model.
• Non-interactive runs skip the prompt and use the provider default (override with --model).

Configure API keys for multiple providers

• kcmt --configure-all lets you pick which providers to configure and select which environment variable holds each API key. This updates both the legacy provider_env_overrides and the new providers[prov].api_key_env entries.
• kcmt --verify-keys prints a concise table of providers, the env var in use, whether it’s set, and any detected alternatives in your environment.

OpenAI batch mode

• Enable with --batch (or via the configure menu) to route commit message generation through the OpenAI Batch API.
• Pick the batch model interactively; defaults to your OpenAI preferred model. Override on the CLI with --batch-model.
• kcmt shows a spinner while the batch job runs and polls for up to 5 minutes by default (tunable via --batch-timeout or KCMT_BATCH_TIMEOUT).

Provider defaults

Provider	Default model	Default endpoint
OpenAI	gpt-5-mini-2025-08-07	https://api.openai.com/v1
Anthropic	claude-3-5-haiku-latest	https://api.anthropic.com
xAI	grok-code-fast	https://api.x.ai/v1
GitHub	openai/gpt-4.1-mini	https://models.github.ai/inference

You can still override values at runtime:

kcmt --provider openai --model gpt-5-mini-2025-08-07 --endpoint https://api.openai.com/v1 \
     --api-key-env OPENAI_API_KEY --repo-path .

Additional environment tweaks remain available:

• KLINGON_CMT_LLM_MODEL
• KLINGON_CMT_LLM_ENDPOINT
• KLINGON_CMT_GIT_REPO_PATH
• KLINGON_CMT_MAX_COMMIT_LENGTH (applies to subject line validation; body is no longer truncated)
• KCMT_PROVIDER – force provider selection for one-off runs without editing ~/.config/kcmt/config.json (useful when CI supplies secrets that differ from the persisted repo defaults) Deprecated: KLINGON_CMT_ALLOW_FALLBACK previously enabled a heuristic fallback subject after repeated LLM failures. This path has been removed; kcmt now fails fast with an explicit LLMError so you never get an invented message. The variable is ignored if set.

Additional LLM behaviour environment variables:

• KCMT_LLM_REQUEST_TIMEOUT – per-request HTTP timeout (seconds, default 5)
• KCMT_PREPARE_PER_FILE_TIMEOUT – per-file generation timeout in atomic workflow
• KCMT_PREPARE_WORKERS – override number of concurrent LLM preparations (or use --workers)
• KCMT_OPENAI_DISABLE_REASONING – disable reasoning / chain-of-thought (default on)
• KCMT_OPENAI_MINIMAL_PROMPT – force minimal prompt style (adaptive toggle)
• KCMT_OPENAI_MAX_TOKENS – max completion tokens for OpenAI-like providers
• KCMT_FAST_LOCAL_FOR_SMALL_DIFFS – opt-in local conventional subject for tiny diffs (<=3 changed lines)
• KLINGON_CMT_AUTO_PUSH=0|1 (disable or enable automatic git push; default is enabled)

List models and pricing

kcmt --list-models prints a simple cross-provider pricing board ordered by output (completion) price per 1M tokens, including context window and max output when available. Use --debug to see the raw structured listings.

Example:

kcmt --list-models

Benchmarking

kcmt --benchmark

Run a local benchmark across providers/models using a fixed set of example diffs. kcmt measures latency, estimates cost, and scores conventional-commit quality with lightweight heuristics.

Basic usage:

kcmt --benchmark --benchmark-limit 5

Options:

• --benchmark-limit INT – max models per provider (default 0 = all models)
• --benchmark-timeout SECS – per-call timeout
• --benchmark-json – also emit machine-readable JSON
• --benchmark-csv – also emit CSV rows

Snapshots are saved under ~/.config/kcmt/repos/<repo-id>/benchmarks/benchmark-<timestamp>.json for later comparison.

Quick start (CLI)

kcmt --configure              # guided setup -> ~/.config/kcmt/config.json
kcmt                          # per-file atomic commits with live stats
kcmt --workers 8              # explicitly set parallel LLM preparations
kcmt --oneshot --verbose      # single best-effort commit
kcmt --file README.md         # commit a specific file
kcmt --provider xai --model grok-code-fast --api-key-env XAI_API_KEY

Exit codes

• 0 on success
• 1 on workflow error (input/validation/LLM/Git failures)
• 2 on configuration error (no usable API key)

CLI reference

kcmt accepts the following common options:

• --configure – launch the interactive setup wizard.
• --configure-all – select provider(s) and set which env var holds each API key.
• --verify-keys – show which env var is used per provider and whether it’s set.
• --provider, --model, --endpoint, --api-key-env – override saved provider details.
• --repo-path PATH – target repository (defaults to current working directory).
• --max-commit-length INT – validate (not hard truncate) the subject line length (default 72 for legacy compatibility; body is preserved).
• --auto-push / --no-auto-push – enable/disable automatic push (default: enabled; can also set KLINGON_CMT_AUTO_PUSH).
• --max-retries INT – retries when Git rejects (default 3).
• --oneshot – stage all changes, pick one file, and commit it once.
• --file PATH – stage & commit an explicit file.
• --no-progress – disable the live stats bar.
• --verbose, -v – emit detailed logs and per-file results.
• --list-models – show a pricing comparison board of models across providers.
• --benchmark – run the model benchmark and show leaderboards. Combine with --provider / --model to focus a subset; see docs/benchmark.md for the scoring rubric and pass/fail guidance. Missing credentials are listed in the "Excluded Models" table at the end of each run.
• --benchmark-json / --benchmark-csv – print results in machine-readable formats.

Conventional commit automation

kcmt now ships with a Commitizen configuration that mirrors the LLM-generated commit format. After installing the project you can:

• run cz check to validate a message before committing manually;
• run cz commit to invoke Commitizen's prompt flow while still benefitting from kcmt's validation rules and version tracking (it watches kcmt/__init__.py).

The configuration lives in pyproject.toml under [tool.commitizen], so any repository that adopts kcmt inherits the same conventional commit guardrails automatically.

Library usage examples

Basic: generate a message from staged changes

• from kcmt.config import load_config
• cfg = load_config()
• from kcmt.commit import CommitGenerator
• gen = CommitGenerator(repo_path=cfg.git_repo_path, config=cfg)
• msg = gen.generate_from_staged(context="Refactor widgets", style="conventional")
• print(msg)

Generate from working tree changes

• from kcmt.config import load_config
• cfg = load_config()
• from kcmt.commit import CommitGenerator
• gen = CommitGenerator(repo_path=cfg.git_repo_path, config=cfg)
• msg = gen.generate_from_working(context="Work in progress", style="conventional")
• print(msg)

Run the full atomic workflow

• from kcmt.config import load_config
• cfg = load_config()
• from kcmt.core import KlingonCMTWorkflow
• wf = KlingonCMTWorkflow(repo_path=cfg.git_repo_path, max_retries=3, config=cfg)
• results = wf.execute_workflow()
• print(results["summary"])
• for r in results.get("file_commits", []):
  • print(r.success, r.commit_hash, r.message)

Using GitRepo directly

• from kcmt.config import load_config
• cfg = load_config()
• from kcmt.git import GitRepo
• repo = GitRepo(cfg.git_repo_path, cfg)
• print(repo.get_working_diff())
• if repo.has_working_changes():
  • repo.stage_file("README.md")
  • repo.commit("docs: update readme")

API documentation (high-level)

Exceptions

• kcmt.exceptions.KlingonCMTError: Base error
• kcmt.exceptions.GitError: Git command errors
• kcmt.exceptions.LLMError: LLM call errors
• kcmt.exceptions.ConfigError: Config/ENV errors
• kcmt.exceptions.ValidationError: Validation failures

Configuration

• kcmt.config.Config
  • Fields: provider, model, llm_endpoint, api_key_env, git_repo_path, max_commit_length, auto_push, providers (per-provider map), provider_env_overrides (legacy mapping)
• kcmt.config.load_config(overrides=None)
  • Merge global ~/.config/kcmt/config.json, environment, and optional overrides.
• kcmt.config.save_config(config, repo_root=None)
• kcmt.config.get_active_config() / set_active_config()

LLM

• kcmt.llm.LLMClient
  • generate_commit_message(diff: str, context: str = "", style: str = "conventional") -> str

Git operations

• kcmt.git.GitRepo
  • has_staged_changes() -> bool
  • get_staged_diff() -> str
  • has_working_changes() -> bool
  • get_working_diff() -> str
  • get_commit_diff(commit_hash: str) -> str
  • get_recent_commits(count: int = 5) -> list[str]
  • stage_file(file_path: str) -> None
  • unstage(file_path: str) -> None
  • commit(message: str) -> None
  • process_deletions_first() -> list[str]
  • push(remote: str = "origin", branch: str | None = None) -> str

Commit generation

• kcmt.commit.CommitGenerator
  • generate_from_staged(context: str = "", style: str = "conventional") -> str
  • generate_from_working(context: str = "", style: str = "conventional") -> str
  • generate_from_commit(commit_hash: str, context: str = "", style: str = "conventional") -> str
  • suggest_commit_message(diff: str, context: str = "", style: str = "conventional") -> str
  • validate_conventional_commit(message: str) -> bool
  • validate_and_fix_commit_message(message: str) -> str

Core workflow

• kcmt.core.FileChange
  • file_path: str
  • change_type: str # "A" | "M" | "D"
  • diff_content: str
• kcmt.core.CommitResult
  • success: bool
  • commit_hash: str | None
  • message: str | None
  • error: str | None
• kcmt.core.KlingonCMTWorkflow(repo_path: str | None = None, max_retries: int = 3)
  • execute_workflow() -> dict
    • deletions_committed: list[CommitResult]
    • file_commits: list[CommitResult]
    • errors: list[str]
    • summary: str

Notes on behavior

• Deletions are grouped and committed first with a generated message.
• Then remaining file changes are parsed from git diff and committed per-file.
• Commit messages are validated and may be LLM-fixed on failure; retries are applied.
• Enrichment pass: for substantial diffs (>=10 changed lines) a second LLM call may add a concise multi-line body explaining what and why while preserving the original header.
• Fast fail: after retry exhaustion (currently 3 attempts) kcmt raises an LLMError; no heuristic commit message is generated.
• If --auto-push is enabled and at least one commit succeeds, kcmt attempts git push origin <current-branch> and records the result (results['pushed']=True).

Development

Prereqs

• Python 3.12+
• A virtual environment
• Git

Set up

• uv venv && source .venv/bin/activate # or python -m venv .venv
• uv pip install -e ./kcmt # or pip install -e ./kcmt
• uv pip install pytest # or pip install pytest

Run tests (from project root)

• make test # quiet pytest run via Makefile
• uv run -m pytest tests # or: pytest

Lint/format (recommendations)

• Use Black and isort (not enforced by this package directly).
• Keep lines <=80 chars where practical, file size <500 lines for readability.

Publishing

• Version is defined in kcmt/init.py (managed by hatch).
• Build and publish via hatch (example):
  • uv pip install hatch
  • hatch build
  • hatch publish # configure as needed

Security notes

• Your GITHUB_TOKEN is used as the API key for the configured endpoint.
• Treat it like any other secret; avoid checking it into logs or code.
• Prefer passing the token via the CLI or securely provisioned environment.

Changelog

• 0.1.2 — Update default OpenAI model to gpt-5-mini-2025-08-07 (automatic migration from legacy gpt-5-mini), test helper env var KCMT_TEST_DISABLE_OPENAI to bypass real API calls in isolated tests.
• 0.1.1 — Auto-push option, improved retries, subject-only length enforcement.
• 0.1.0 — Initial release: CLI + atomic workflow + LLM commit generation.

License

MIT (see repository license)

Strict testing and CI quick reference

Environment

• uv venv && source .venv/bin/activate
• uv sync --group dev
• uv pip install -e .

Lint/format/type-check

• uv run ruff check kcmt tests
• uv run black --check kcmt tests
• uv run isort --check-only kcmt tests
• uv run mypy kcmt

Run tests

• Basic: uv run pytest -ra -vv tests
• Strict CI-like run (parallel, warnings as errors, coverage): uv run pytest -n auto -ra -vv -W default -W error::DeprecationWarning -W error::ResourceWarning --strict-config --strict-markers --cov=kcmt --cov-branch --cov-report=term-missing:skip-covered --cov-fail-under=85 tests
  • On Windows/PowerShell use a single line (no trailing \) or replace the line continuation with a backtick (`); PowerShell treats \ as a literal character and will otherwise break up the arguments.

Make targets

• make test
• make test-verbose
• make test-strict
• make check # lint + typecheck + strict tests