klasp

klasp.dev · crates.io · npm · PyPI · GitHub

Block AI coding agents (Claude Code, Codex, and Aider) on the same quality gates your humans hit at git commit.

What klasp does

You write one klasp.toml. You run klasp install. Every AI agent on the repo (Claude Code today, more coming) is now blocked on the same pre-commit, cargo clippy, pytest, or any-shell-command gate your humans see at git commit. The agent gets a structured "blocked, here's why" reply at its tool-call surface (Claude Code's PreToolUse hook) so it self-corrects rather than retrying with --no-verify. That retry path is the failure mode burning every team running agents at scale.

Why klasp

1. Stops the --no-verify escape hatch

Your agent runs the test suite, sees a red, decides the failure is "unrelated to my task", and commits with --no-verify (or amends past the hook). The bad path lands on main, CI catches it a few minutes later, and you're the one cleaning up.

klasp blocks at the Claude Code PreToolUse surface, before the agent's Bash call ever runs git commit. Failure detail is returned inline so the agent retries against the gate instead of around it.

2. CI parity at the agent surface

Agent ships a green-locally PR. CI runs the team's pre-commit, linter, and type-check. PR turns red 30 seconds later and you triage.

One klasp.toml declares the same checks. Pass-locally now means pass-CI.

# Typed recipe form (v0.2 W4) — preferred for pre-commit setups.
[[checks]]
name = "pre-commit"
triggers = [{ on = ["commit"] }]
[checks.source]
type = "pre_commit"           # optional: hook_stage, config_path

The typed type = "pre_commit" recipe handles the --hook-stage, --from-ref, --to-ref flags internally and parses pre-commit's per-hook output into structured findings the agent can act on. The v0.1 shell form still works:

[[checks]]
name = "pre-commit"
triggers = [{ on = ["commit"] }]
[checks.source]
type = "shell"
command = "pre-commit run --from-ref $KLASP_BASE_REF --to-ref HEAD"

3. Diff-scoped checks on big repos

Agent edits one file. The full test suite kicks off. 90 seconds later the agent has lost the thread and you're paying for tokens spent waiting.

Every shell check sees KLASP_BASE_REF, set to the merge-base of HEAD against the upstream tracking branch. Scope linters, formatters, and audits to the diff:

[[checks]]
name = "fallow-audit"
triggers = [{ on = ["commit", "push"] }]
[checks.source]
type = "shell"
command = "fallow audit --base $KLASP_BASE_REF"

4. Protected-path guards

Agent debugs a flaky test by reverting your schema migration. Or your i18n bundle. Or the generated API client. The PR looks clean and the regression surfaces in prod a week later.

Add a check that fails when a staged file lives in a path the agent shouldn't have touched without sign-off:

[[checks]]
name = "no-migration-edits-without-marker"
triggers = [{ on = ["commit"] }]
[checks.source]
type = "shell"
command = '''
  if git diff --cached --name-only | grep -q "^migrations/"; then
    test -f .agent-migration-allowed && exit 0
    echo "Migration files staged. If intentional, drop a .agent-migration-allowed marker and re-stage."
    exit 1
  fi
'''

5. Polyglot stacks

Agent onboards a repo with a TS frontend, Go API, and Python ML pipeline. Three test runners, three lint configs, three formatters. Agent guesses wrong and someone has to walk back its commits.

klasp.toml declares each surface once:

[[checks]]
name = "frontend"
triggers = [{ on = ["commit"] }]
[checks.source]
type = "shell"
command = "cd web && pnpm typecheck && pnpm lint"

[[checks]]
name = "api"
triggers = [{ on = ["commit"] }]
[checks.source]
type = "shell"
command = "cd api && go test ./... && go vet ./..."

[[checks]]
name = "ml"
triggers = [{ on = ["commit"] }]
[checks.source]
type = "shell"
command = "cd ml && uv run pytest && uv run ruff check"

Worked configs for pre-commit, fallow, pytest, cargo, ESLint/Biome, and ruff live in docs/recipes.md.

Install

Pick the package manager that matches your primary stack. All three ship the same binary:

cargo install klasp                        # Rust (also the right pick for x86 Macs — no prebuilt wheel)
npm i -g @klasp-dev/klasp                  # JS / TS (biome-style platform shim)
pip install klasp                          # Python (maturin wheel)

Prebuilt binaries cover darwin-arm64, linux-x64-gnu, linux-arm64-gnu, and win32-x64. On x86 Macs and other unsupported targets, cargo install klasp builds from source.

Set up a repo

cd your-project
klasp init                                 # writes klasp.toml at repo root
$EDITOR klasp.toml                         # add your checks (see below)

# For Claude Code
klasp install --agent claude_code          # writes .claude/hooks/klasp-gate.sh + merges .claude/settings.json

# For Codex
klasp install --agent codex               # writes .codex/git-hooks/<gate>.sh

# For Aider
klasp install --agent aider               # edits .aider.conf.yml commit-cmd-pre

# Or all three at once
klasp install --agent all                 # installs Claude Code + Codex + Aider in one step

klasp doctor                               # verify the install is healthy

Uninstall

klasp uninstall --agent claude_code        # removes the hook + settings entry, preserves siblings
klasp uninstall --agent aider              # removes commit-cmd-pre entry from .aider.conf.yml

Migration from v0.2.x

After upgrading the binary to v0.2.5 (cargo install klasp, npm i -g @klasp-dev/klasp, or pip install -U klasp), re-run klasp install in each enrolled repo so the hook shim picks up KLASP_GATE_SCHEMA=2. Old shims fail open — they print a schema-mismatch notice and let the commit proceed — so there is no silent breakage, but the gate is inactive until the hook is regenerated. Configs written for v0.2 with no parallel or policy field continue working unchanged.

klasp install --agent all                  # regenerates the shim(s) with KLASP_GATE_SCHEMA=2

What works today

v0.3 ships the third agent surface (Aider), the plugin extensibility protocol, and a public conformance matrix so "klasp supports agent X" means the same thing for every X. See docs/roadmap.md for the full milestone plan.

Example klasp.toml

This is klasp's own dogfood config (also at /klasp.toml):

version = 1

[gate]
agents = ["claude_code"]
policy = "any_fail"

# Fast type/borrow check on every commit attempt.
[[checks]]
name = "cargo-check"
triggers = [{ on = ["commit", "push"] }]
timeout_secs = 60
[checks.source]
type = "shell"
command = "cargo check --all-targets --workspace"

# Lint with -D warnings on commit and push.
[[checks]]
name = "cargo-clippy"
triggers = [{ on = ["commit", "push"] }]
timeout_secs = 180
[checks.source]
type = "shell"
command = "cargo clippy --all-targets --workspace -- -D warnings"

# Full workspace tests on push only (slower than clippy).
[[checks]]
name = "cargo-test"
triggers = [{ on = ["push"] }]
timeout_secs = 300
[checks.source]
type = "shell"
command = "cargo test --workspace"

Every shell check sees KLASP_BASE_REF in its environment, set to the merge-base of HEAD against the upstream tracking branch (falling back to origin/main, origin/master, then HEAD~1). Diff-aware tools (pre-commit run --from-ref, fallow audit --base) can scope themselves to the diff without an agent-side wrapper. See docs/recipes.md for worked examples in pre-commit, fallow, pytest, ESLint/Biome, ruff.

Parallel check execution (v0.2.5+)

By default, klasp runs checks sequentially. When individual checks are long-running and independent (test runners, type-checkers), you can enable rayon-based parallel execution by adding parallel = true to the [gate] section:

[gate]
agents = ["claude_code"]
parallel = true

With parallel = true, all triggered checks for a gate run concurrently via rayon's work-stealing thread pool. A 5-check workload that takes 25 seconds sequentially completes in ~5 seconds in parallel mode.

Important contract: checks must be stateless when parallel mode is enabled. Reading shared input (source files, config) is fine. Writing to shared output — the same temp file, the same database row, process-global state — will race and produce non-deterministic results. klasp does not detect or prevent this. The default remains false so existing v0.2 configs continue working unchanged.

Verdict policies

The [gate].policy field controls how multiple check results are combined into a single gate outcome. Three values are available:

Warn verdicts are never counted in the decisive majority or unanimity tests — they pass through as informational noise regardless of the policy.

See docs/recipes.md for selection guidance.

Documentation

• docs/design.md — v0.1 architecture, abstractions, and rationale
• docs/recipes.md — worked klasp.toml examples for pre-commit, fallow, pytest, cargo, ESLint/Biome, ruff; verdict policy guidance
• docs/plugins.md — plugin authoring guide; fork examples/klasp-plugin-pre-commit/
• docs/plugin-protocol.md — plugin wire-format specification
• docs/conformance-matrix.md — per-surface support matrix (Claude, Codex, Aider all-green; Cursor NO-GO)
• docs/roadmap.md — milestones from v0.1 → v1.0
• CHANGELOG.md — release notes

Repository layout

License

Apache-2.0. See LICENSE.