What works today

kalku is pre-alpha. Everything this documentation describes runs today, except what this page marks as pending. The table comes from the README of the repository, which is where it is updated first.

Not yet

Piece	State
Parallel workers in an Elixir project (one build path, so one worker)	not yet
--watch and --ci	not yet
merge: combining sharded reports	not yet
macOS on Intel, Linux on arm64 (waiting on a kaikai toolchain for them)	not yet

Commands and core

Piece	State
kalku init: detects the language, writes .kalku.toml and the summoning script	done
kalku run <files>: measures those files and reports survivors	done
kalku cast <wekufe>: re-cast one wekufe, and kalku show to read it back	done
kalku propose-equivalent: an agent proposes, a person accepts	done
kalku info: spells, outcomes, formats and the agent loop, in the binary	done
kalku mcp: the four agent tools, over stdio	done
kalku serve: the client protocol on a 0600 Unix socket, streaming a run as it happens	done: one client at a time
Warm sessions: the pool a project was measured with survives the run, and is replaced when the project changes	done
Protocol: both directions, canonical NDJSON, fixtures for every message	done
Core: planner, test selection, outcomes, score, gate, equivalents, .kalku.toml, sharding, changed lines	done
Reports: human, JSON, GitHub annotations, and agent NDJSON with per-spell hints	done
Orchestrator: kalku pool, scheduler, timeout escalation (abort → reset → kill), restart budget, cancellation	done
--since <ref>: measure what a change touched, and block on holes it introduced	done
[score] ratchet: the score of the whole project may not drop below the baseline stored with --store-baseline	done
--all and --shard i/n: the whole project, in parts, for a nightly run	done: a shard says it is one and is not held to the threshold

The kalku of each language

Piece	State
Elixir kalku: sites, prepare, baseline with per-test coverage, cast, abort, reset, reload	done: every protocol message but delegate
kaikai kalku: the one kalku measures itself with	done
Rust kalku: sites for arm, compare, connect, negate, literal from syn, behind the kalku protocol	done: finds sites
Rust kalku: prepare, baseline and cast in the reni, abort, per-test coverage with LLVM	done
Rust kalku: kalku init for Cargo projects, shipped in the release	done
Python kalku: sites from ast, prepare/baseline/cast in forked children, abort, per-test coverage with sys.monitoring, kalku init, shipped in the release	done
Java kalku: sites for arm, compare, connect, negate, literal, call from javac, behind the kalku protocol, on Java 11, 17, 21 and 25	done: finds sites
Java kalku: prepare, baseline and cast in the reni for a Maven project	done
Java kalku: kalku init for Maven and Gradle projects, shipped in the release	done
Java kalku: abort, by ending the JVM the tests run in	done
Java kalku: per-test coverage, counted by JaCoCo and withheld wherever it cannot be trusted	done
Java kalku: Maven projects of several modules; a site is judged by the tests of every module that uses its own	done
Java kalku: Gradle builds, asked by running them with a script added from outside the project	done
Java kalku: projects that test with JUnit 4 or with TestNG, run through their JUnit Platform engines	done
Per-test coverage used by a run: a wekufe is cast against the tests that reach it	done: Elixir, Rust, Python, Java

Distribution

Piece	State
kalku_elixir on Hex: the Elixir kalku installs like any dependency	done
Release binaries: brew install, or a tarball per platform, built and checksummed by CI	done

kalku measured by kalku

kalku uses kalku to test its own code: it casts wekufe in its own sources through the kaikai kalku and reports what its suite did not notice. This is a report of that run:

kalku/core/shard.kai:30  literal  shard.wrap32/1
  - 4294967296
  + 0
  covered by 189 tests
  No test depends on the exact value `4294967296` in `shard.wrap32/1`.

score 40% · 2 killed · 3 survived · 1 compile error