kerpopule/hermes-jev-skills v0.23.0

0.23.0

on GitHub

Released by the daily Jev steward: 9 commit(s) on main since v0.22.2.

• Add an offline-by-default bounded design-choice reference example (jevkit.design_choice), with complete preset IDs, escaped local copy, contrast/bounds checks, opt-in validated Jev selection and exact baseline fallbacks. Add offline transport/CLI tests and an optional 12-case real-browser QA script. No live selection benchmark, new skill, routing hook or default change; see docs/design-choice.md.

• Fix #36: make the Hermes manifest provider-neutral instead of prompting every installation for a TypeSafe-only key. Document runtime provider/key resolution and preserve no-key fail-open behavior. Add OpenRouter-only disposable-install and offline no-key regressions; no key flow, endpoint or provider default changes.

• Reconcile the previously installed opt-in jevkit.local_goal pilot with the public source: whole prevalidated action tables, exact non-sensitive plaintext bindings, protected-field/stale-result refusal and offline regressions. Preserve historical synthetic-engine qualification receipts without asserting new native/end-to-end performance. No default driver, confidence, routing or gateway changes.

• Harden frozen-screen collector recovery: require agreement between write-ahead attempts, local completed calls, inherited call receipts and terminal sent-row outcomes. Refuse ambiguous evidence before creating output or making requests. Six offline corruption regressions; no replay, paid calls, detector or threshold changes.

• Complete the authorized frozen #25 public regression replay with pinned TypeSafe jev-1.13.0, capped single-POST collection and verified never-sent-row continuation. Actual aggregate withholding: 322/1,051 attack-labeled rows and 1/1,050 clean-labeled rows, retaining 11 fail-opens and two local-only rows. Publish sanitized provenance/accounting and an honest scorecard; no reserve scoring, threshold tuning or detector/default changes. Conservative list-price spend upper bound $0.061657764 under a $5 approval. #25 remains open.

• Add an offline, hash-only screening preflight for #25 that verifies frozen source/selection identities and captures the real screening payload construction without inference or credential lookup. Reverify pinned public receipts and publish a sanitized preflight summary. Reserve rows are not screened, no detection scores are fabricated, and a live collector/budget remains separately gated.

New

• Add tested bounded design-choice reference example (23a5409)
• Reconcile opt-in local-goal pilot with public source (7b7da6c)
• Add verified offline screening preflight without inference (58434c8)
• Add reviewed bounded collector for authorized frozen screening replay (38395b5)

Changes and fixes

• Fix provider-neutral Hermes manifest and installation regression (bf348be)
• Resume only unattempted rows while carrying conservative prior spend (27b7825)
• Publish actual capped screening replay and advisory-only safety findings (3b2e46b)
• Fail closed on inconsistent collector recovery evidence (88dba15)
• Document first-principles PR review and combined offline qualification (c714060)