kerpopule/hermes-jev-skills v0.25.0

0.25.0

on GitHub

Released by the daily Jev steward: 15 commit(s) on main since v0.24.0.

• Integrate PR #49 (thanks @mxuanvan02) with explicit-effort preservation, disabled-reasoning and follow-up telemetry regressions. Retain the contributor commit in merge history; preserve manual reasoning.effort rather than replacing it. Include the numeric unknown-cost row count in daily stewardship summaries.

• Add the tested opt-in bridge for existing daily stewardship collectors, separating root install identity from published main and exposing common-ledger estimates/unknowns. Counterbalance future A/B requests, while retaining the first exploratory run and its documented ordering deviation.

• Qualify the synthetic question A/B collector under a scoped paid ceiling, retain unknown-outcome reservations and replay refusal, and record the exploratory 24-frame result with its arm-order limitation. Honor retries=0 at the native pooled HTTP transport layer, including stale connections.

• Correct Responses effort delivery without enabling disabled reasoning or overriding explicit/manual values; keep Chat Completions wire shape and accurate applied-effort notices, including follow-up requests. Incorporates the wire-format issue identified in PR #49 with additional boundary regressions. No routing or reasoning-mode activation.

• Distinguish published main, dirty checkout, installed managed code/skills and unverified loaded runtime. Add a checked daily-steward adapter, external dependency fingerprints and metadata-only common-ledger aggregation; never overwrite an installed feature candidate merely to match main.

• Meter standalone Blender/video/image-action helpers through hash-only durable write-ahead events with hidden retries disabled. Coalesce logical calls, retain unknown spend, close daily budgets on unknown usage and label rate/counterfactual estimates separately from invoices and realized savings.

• Refresh workflow evidence, code-first and task-specific confidence guidance while preserving required GUI/browser choice and final verification. Prepare a fresh 24-frame synthetic question-design A/B fixture and strict offline scorer; no paid run, floor reduction, production-question replacement or automatic promotion.

• Make video QA use a general local-first triage/review-planning default through jev-video-qa, with task-specific requirements, privacy-safe no-upload fallback, exact-frame deduplication, mandatory audits and low-coverage stop. Preserve the broader holdout’s two raw tiny-text misses and zero confident-clear coverage; do not claim general review-time savings. No blanket cloud-vision opt-in or automatic artifact acceptance.

• Add an opt-in synthetic video-frame Jev QA trial and original encoded/decode fixture, consent/hash/path/replay safeguards, conservative review routing and offline boundary tests. Retain finite known-answer results and the same-set safety recheck with explicit cost/accuracy limits. No automatic frame clearance, private uploads, cloud vision opt-in, routing or gateway change.

• Add the default jev-blender domain workflow, a closed-vocabulary code-first jevkit.blender decision boundary and jev blender-next. Mandatory local checks and artifact verification remain authoritative; optional one-second read-only Jev advice is limited to unclear expensive probes. Add offline boundary/CLI regressions. No model routing, credentials, image forwarding or gateway change.

• Add an opt-in native image/action comparison fixture and reproducible measured receipts. Jev chooses prevalidated action IDs; installed Hermes modules and paired Co-Agent execute verified synthetic clicks and plaintext typing. Retain failed replacement and image-abstention runs; measure explicit element fallbacks without lowering confidence. Add seven offline safety/accounting tests. No default driver, cloud switch or gateway change.

• Fix #48 (thanks @zimu002): remove the unsupported provides_middleware manifest declaration. Hermes v0.22.2 and current hosts register llm_request via PluginContext.register_middleware in plugin code; the declaration only produced a warning and never enabled the middleware. Add a regression preserving code registration and a warning-free manifest.

New

• Add opt-in bounded whole-action tables and measured Arc loop pilot (672732c)
• Add code-first Blender default and bounded Jev diagnostic advice (fe9ba8e)
• Add opt-in Jev synthetic video-frame QA trial and measured receipts (4741719)
• Default video QA to guarded Jev triage with full-review fallback (dea54a7)
• Fix Jev request compatibility and truthful workflow reporting (f38d468)
• Qualify capped synthetic Jev A/B and verify shared-home provenance (980c983)

Changes and fixes

• Record successful installed-selector qualification (cf0e82c)
• Require verified plaintext field types and prove protected-field refusal (a8f6a61)
• Document protected-field contract in candidate skill and changelog (98d1928)
• Measure opt-in image choices on a real native Hermes desktop loop (eb86fa4)
• fix: drop unsupported middleware manifest hint (#48) (b8f69fe)
• fix: carry effort inside reasoning on the Codex Responses wire (ea7a566)
• Retain video QA frame and write-ahead attempt receipts (10beadc)
• Record actual 40-home video QA default readback (e9e391c)
• Wire truthful daily stewardship and counterbalance future A/B collections (d70ae3f)