Skip to content

v0.6.0 — build+verify harness + dedicated-server support

Choose a tag to compare

View all tags
@chapmanjw chapmanjw released this 25 May 06:44
· 29 commits to main since this release
v0.6.0
6d49ab8

Token-usage optimization + headless/dedicated-server support.

Build + verify harness (tools/builder/)

A stdlib harness that executes a plan.toon phase and mechanically verifies it against the live server outside the LLM context — plan steps are the code, acceptance + quality_contract are the assertions, the harness is the test runner, returning one compact digest instead of hundreds of in-context tool calls. Force-load-bracketed (auto-banded under the 256-chunk/dimension cap) and rate-limit resilient. Validated live against a dedicated server: all 9 plan ops and all 12 contract checks.

Model routing (off Opus where reasoning isn't needed)

  • blueprinter, philosopher, and natural-landmarks now fork to Sonnet (isolated context).
  • Orchestrator runs at effort: high.
  • philosopher returns drafted memory lessons for the orchestrator to persist; natural-landmarks returns a proposed composition for the orchestrator to confirm.

Dedicated / headless-server support

  • Step 0 detects dedicated-vs-single-player by sampling gameTime at 0 players.
  • Force-loading is a first-class concern (mandatory headless, where writes silently no-op in unloaded chunks); the registry records each build's envelope.
  • The "ticking freezes when unfocused" caveat is now single-player-only.

Also

  • researcher is WebFetch-first and pulls exact dimensions from Wikipedia/Wikidata REST (no new dependency).
  • planner pins the acceptance + envelopes schema; worker drives the harness; inspector consumes its verdict and focuses on perceptual/world-fit judgement.

See CHANGELOG.md for the full list.

Assets 2
Loading