Skip to content

1.0.2

Latest

Choose a tag to compare

@akseolabs-seo akseolabs-seo released this 03 Jul 16:28

Cuts the always-on token cost by ~40% with no loss of function, verified by evals.

  • New references/soul-compact.md: every rule compressed to its core imperative; the rationalization table, red flags, and GREEN-tested sentences kept verbatim. Measured ~4,000 → ~2,400 tokens per session.
  • The sync script now installs the compact rendering into global instruction files; skill copies keep the full version. A structural parity gate (rule count, table rows, red-flag count) refuses to sync on mismatch, with a unit test behind it.
  • Equivalence verified on a Haiku-class model: the manufactured-findings and stale-green scenarios both hold with the compact body (recorded in references/evals.md).
  • Documented a measured eval methodology caveat: subagents inherit the parent session's startup snapshot of global files, so new rules must be tested in-prompt or in a fresh session.

Full details in CHANGELOG.md.