Skip to content

v5.3.0

Choose a tag to compare

@github-actions github-actions released this 19 Jul 22:34
· 1117 commits to main since this release

[5.3.0] - 2026-07-20

Single-version harness lore makes the installed package, runtime authority,
validation evidence, correction review, and audit promotion path explicit:

  • activation-check inventories Flow activation across OpenCode config,
    plugin directories, and package cache. activation-apply plans by default,
    then uses an exact canonical pin, backups, quarantine, and a recovery journal
    with --apply; ambiguous or externally managed sources require manual
    remediation. Post-mutation failure attempts exact safe rollback and records
    whether recovery converged or needs journal-backed manual repair.
  • Process-global runtime leadership permits one operational Flow instance.
    Duplicate versions fail closed; a deterministic highest-version identity is
    diagnostic only and cannot silently take control.
  • Three bounded harness profiles (control, standard, and assurance) and
    three rollout modes (control, observe, and enforce) now drive optional
    worker admission through a trusted runtime-policy footer. Worker model and
    current OpenCode steps routing are configurable by role.
  • Validation is runtime-attested: flow_validation_start binds the current run
    and source to the exact next Bash command, which emits an immutable receipt
    reference consumed through flow_review_start.request.validationRefs.
    Failed, incomplete, stale, altered, or duplicate receipts cannot become
    review evidence.
  • Correction review binds the latest recorded failure to authoritative source
    manifests and a deterministic delta. Narrow correction context is used only
    when complete and safe; broad, security-sensitive, persistence-sensitive,
    missing, unavailable, or oversized context falls back to full review. A
    bounded correction-only public-contract/cross-layer hint can elevate semantic
    scope to full without overriding more specific runtime reasons. The existing
    two-failure run-scoped cap remains authoritative.
  • AuditLedgerV1 now provides bounded typed findings, conservative severity
    rules, explicit refutations and falsifiers, derived summaries, and
    deterministic reconciled Markdown through flow_audit_render.
  • Privacy-safe bounded host observation and the sanitized full-repository audit
    oracle separate observed zero from unavailable data and require same-source,
    same-model quality parity plus lower observed work before a candidate profile
    can be promoted. The checked-in standard and assurance observations remain
    unavailable, so enforcement is not yet a release claim.