~/daily/work 2026-11-09 · published 2026-11-21

hakoniwa, day one: from ADRs to a world

What the git history of the test engine shows for 2026-11-09: 282 commits in one day, the order things landed in, and the benchmark that did not hit its target.

The numbers

hakoniwa is the test engine. Its repository history starts on 2026-11-09, and 282 commits carry that date: 237 without merges, plus the merges of pull requests #5 to #13. Three non-merge commits are dated the 10th and 38 the 13th.

That is a count of commit dates, not a measure of effort. Many of those commits come from parallel branches merged the same day and named by version: v2, v3, v4, v5.

The order things landed

  1. 1Decisions firstADR-0001 to ADR-0005 are among the first commits: the neutral core, a run lifecycle owned by the engine with leased workers, admission gates against false tests, the worker and its evidence, and the engineering principles.
  2. 2Lifecycle and gatesA test state machine where only a person can admit, a run state machine with leases and step-boundary pause, and the admission gates G1 to G5: compile, clean replays, a seeded fault that must turn the test red, assertion provenance, and quiet implicit oracles.
  3. 3Storage and surfacesA Postgres schema with forced row-level security, an in-memory store with the same conformance suite, then one operation table served over HTTP, gRPC and MCP, plus a CLI derived from it.
  4. 4A speed target and a benchmarkADR-0011 set the goal: at least 50% less wall time and CPU than plain Playwright with the same verdicts. A 17-spec benchmark app and harness followed the same day.
  5. 5Versions v2 to v5A commit titled Decide v2 to v5 and plan their build opens the second half of the day: v2 soft steps and check runs, v3 recording runs and locator healing, v4 the world host with a clock seam, fork and quiescence probe, v5 shortcuts, a result cache, suites and a signed inbox.
  6. 6Spikes on the worldSpikes S8, S10 and S11 measured the clock on Postgres, the quiescence probe and control flake. They are the subject of Testing with a frozen clock.

The benchmark, the same day

The benchmark result landed the same day it was built, and it missed the target. Medians of 3 healthy repetitions, 4 CPUs and 6 GB per container: the engine used 37 to 39% less CPU than plain Playwright, and wall time was within noise at -3 to -5%. The 50% target was not met. The full reading is in a separate post, Where the wall time went.

Verification in Docker

The same day, v2 and v3 were verified in a container with a script replaying fixed actions over the HTTP API and no model. The v3 run recorded five sessions, compiled them, passed gates G1 to G5 and replayed them, with 0 tokens spent. A renamed test id turned a replay into healed with one locator proposal, which a person accepted.

It also found five bugs in the v2 and v3 code. The one I like best: the runner kept the metadata of its first configuration, so the second and later runs of a worker got no evidence directory. The fix reads the current run file at test time.