Skip to content

v6.2.0 — Loop Discipline & Read-Only Verification

Latest

Choose a tag to compare

@TheWayWithin TheWayWithin released this 20 Jun 19:48
· 110 commits to main since this release

The thing that judges the work must be read-only to the thing doing it. With a trustworthy judge in place, agents can iterate in safe, measured, reverted-unless-better loops. Shipped as a four-sub-sprint umbrella (6a–6d).

Headline changes

  • Read-only quality gates (6a): .quality-gates.json and gates/ are unwritable by every agent via permissions.deny at the tool layer, not a prompt convention. Live-demoed. Paired with a default-fail verification contract (a criterion flips to pass only on captured command output).
  • Ratchet mission-optimize (6b): optimisation as a measured loop in an isolated worktree, median-of-3 baseline, keep-or-revert, JSONL .loops/ log, hard caps, never auto-merged. Plus a new code-review-loop skill and a loop-discipline field-manual guide.
  • Phase-gated meta-loop (6c): /coord continue converges on two clean rounds, spends a per-phase error budget then escalates, takes condensed subagent returns, and restarts from the last evidence-passed gate. A Bash gate-guard hook closes the route the deny rules missed.
  • Bulk-ops toolkit: tier-aware, registry-driven fleet operations.

Honest framing

Loops are watched-first and human-merged: diligent, safe iteration with a trustworthy judge, not unattended autonomy.

Full notes: docs/RELEASE-HISTORY.md