HyLogger 0.1.4 preview — simpler expert review

A simpler expert review flow for the local Codex preview.

• Keep one persistent evidence book beside the conversation; the plugin and tools tell the agent to reuse it after continuation.
• Review notes autosave as drafts in the current browser tab, including reload. Saved/unavailable feedback is visible. Drafts are not a server backup and closing the tab or changing its origin can lose them.
• Write an optional note, then click Continue, Change the investigation, or Set aside. That action commits the exact choice/note and turns the page; there is no separate record button.
• Contents entries display the choice that led to each successor, without rewriting saved page titles/history.
• Automatic wake-ups are labelled book events. The note is retrieved from the saved decision; it is not copied into the notification. The local host still displays a notification message, so invisible delivery is not claimed.
• Manual copy/paste recovery is collapsed by default.

PR #22 passed full local checks (343 Python / 50 frontend tests, strict typing, lint, docs and installed-wheel checks) and both hosted head checks. A separate real SWMB007s browser acceptance verified exact-note reload, no decision while typing, one action/one successor, live same-tab scientific continuation and unchanged sealed history. Technical test choices were labelled; no expert geological validation is claimed. Host tab reuse is an agent instruction, not a server-enforced guarantee.

The model and its hashes are unchanged. The small installer reuses checksum-pinned model assets from v0.1.2-preview.1 and the pinned upstream base. No retraining or source data is included. The experimental model remains weak.

Install from the account running Codex, without sudo. If macOS blocks uv, click Done, go to System Settings → Privacy & Security → Security → Open Anyway for uv, then rerun Install.command. See the README installation help for full steps. This Apple Silicon preview is unsigned; cloud/web ChatGPT support remains unverified.

HyLogger 0.1.3 preview — verified chat return binding

Fix automatic decision return for new expert investigations. In the previous preview, an agent could omit session_id, create a disconnected book and leave a saved decision waiting indefinitely. Runtime 0.1.3/plugin 0.1.5 rejects that omission before science or book creation, explains how the calling agent can obtain its actual conversation UUID, and requires explicit manual_return=true for manual follow-up. The shared MCP server environment is never inferred as the caller. Start/open results expose the return mode and destination; the skill verifies them before presenting automatic review.

Install or update

Download hylogger-macos-arm64.zip, extract with Finder and run bash Install.command in its folder, without sudo, under the account running Codex. Existing books and the identical model are preserved/reused. Start a new local conversation after update if the current tools have not refreshed. Requires a signed-in Codex desktop on Apple Silicon macOS; the preview is unsigned/not notarized. If macOS blocks uv, see installation security steps.

The installer automatically fetches the unchanged model from checksum-pinned v0.1.2-preview.1 assets and the pinned upstream base. This update includes only the new runtime/skill/UI kit and wheel; no retraining or duplicate model upload is needed. About 3.2 GB of model downloads on first install; existing installations reuse their model. No raw HyLogger data is bundled.

Validation

Source main: ed6d8786e2e47f2f3e031b7e030622f163d4fd3b. Both exact-head PR checks passed: 343 Python tests, 45 frontend tests, strict runtime/training typing, lint/format, docs and installed-wheel validation. Regressions reproduce the missing-session call, validate manual mode, prevent shared-process environment misattribution and verify one exact notification when repairing a saved case without changing its history. The owner's unbound case was repaired in place; its original conversation automatically received the decision and completed a clarification page without calculating the comparison. Actual default installation completed, verified the existing model, registered/enabled plugin 0.1.5 and the new installed MCP interpreter. An installed-only real-model stdio check exposed six expert tools, refused the exact missing-session request, accepted a corrected retry against the owner’s real case, preserved the locked page and continuation records, and retained exactly one dispatch attempt. Start/open reported the correct originating conversation and automatic mode. No model/assets were retrained or replaced. A fresh-agent recovery from the new error is guided and unit-tested; autonomous recovery by a newly started host conversation remains a separate observation.

Limits

Automatic mode requires the calling host to expose a verified conversation ID and support the local queue route; otherwise explicit manual follow-up is required. ChatGPT cloud/web compatibility is not established. The model remains experimental (all held-out predictions clarify, 38.75%, majority baseline); it supplies no geological judgement or execution permission. Science is bounded QA and TSG-interpreted SWIR group comparison, not assays, resource estimates or new mineral identification. No expert geological validation or signed/notarized distribution is claimed.

HyLogger 0.1.2 preview — installed Codex evidence review

An Apple Silicon Codex desktop preview for expert HyLogger evidence review. Install once, open a local chat in a supported data workspace and ask for QA and interval comparison. Scientific calculations produce the evidence; the expert decides in the browser book; the bound conversation is notified and verifies the exact committed decision before continuing.

Install

Requires macOS on Apple Silicon and an installed, signed-in Codex desktop. Download hylogger-macos-arm64.zip, extract it, and run bash Install.command inside the extracted folder (or open Install.command). No Git, Node, administrator access or source checkout is required. The first installation downloads about 3 GB of model assets and managed Python/dependencies. Setup registers the installed tools and skill; start a new local chat if tools have not refreshed. Existing HyLogger books/models and unrelated host settings are preserved. This preview is unsigned and not notarized.

Place your unpacked supported TSG package in the folder connected to the conversation, including nested package folders. Ask: “Review SWMB007s: check data quality and compare 5–10 m with 0–5 m.” No source HyLogger data is included. HYLOGGER_SKIP_HOST=1 bash Install.command provides standalone setup.

Verified

Source: 280125a0020c61e44389f0130ea73e0f78abbae8 on main, integrating PRs #15, #16 and #17 with actual ancestry. Full checks: 336 Python tests, 45 frontend tests, strict runtime/training typing, lint/format, documentation and installed-wheel checks. Actual default installation loaded the real local checkpoint, registered plugin 0.1.4 and its installed MCP, and preserved the original owner book byte-for-byte. Real SWMB007s data was prepared in a fresh nested workspace; the browser committed an explicitly labelled test decision; installed MCP read the same exact note/IDs. Notification queueing to the same active conversation was observed. The queued notification subsequently resumed the originating Codex conversation automatically after the active turn ended. The receiving agent verified the exact page/decision/continuation IDs and note through the installed case interface, continued that step, and observed the book update live to a ready successor. There was one dispatch attempt and one worker attempt; the sealed page stayed identical and exact replay returned already_completed with no state change. This was a labelled automated acceptance test, not expert geological validation. The receiving chat used the installed agent command fallback; installed MCP discovery/read was independently verified, while first-class MCP selection in a fresh chat remains unverified. Explicit follow-up remains available.

Limits

The model was fine-tuned on organiser-derived measurements using teaching labels. Held-out predictions all collapse to clarify (38.75%, majority baseline); advice is experimental and never execution permission or geological judgement. Scientific output covers QA and TSG-interpreted SWIR interval comparison, not assays or resource estimates. No expert geological validation, ChatGPT/Work compatibility or physical fresh-machine certification is claimed. Source data, books and reference measurements remain private/local. Qwen and third-party notices accompany the installer.

Model parts are checksum-verified fine-tuned assets. Identical base weights are fetched from their pinned upstream revision; no training is required on the expert's computer.

Published-download acceptance

The public archive was downloaded anonymously and its checksum verified, extracted with macOS's native archive tool, and installed into another clean XDG profile on the existing M1 Max Mac. Installed-only resources, actual checkpoint preflight/inference, real SWMB007s QA (5/5 valid spots in each interval), and the ready book passed. The verified model bundle and dependency cache were reused; this is not a claim of another fully uncached 3 GB download or a new physical machine. Default Codex registration/old-marketplace upgrade was separately tested against the owner's installed host. Installed footprint observed: about 3.0 GiB model plus 1.3 GiB runtime, with managed Python and caches additional. No decision in this standalone-profile check was fabricated.

macOS security prompt

The bundled uv helper may be blocked by Gatekeeper in this unsigned preview. If you trust the downloaded release, click Done, then open System Settings → Privacy & Security, scroll down to Security and authorise only uv with Open Anyway. Rerun bash Install.command. If the entry is missing, run ./uv --version from the extracted installer folder, dismiss the warning and check Security again. Do not use sudo or disable Gatekeeper globally. Install under the same macOS account that will run Codex, and wait for Installation finished. See the maintained installation guide and Apple guidance. The owner confirmed this per-app recovery; no signing/notarization is claimed.

HyLogger 0.1.1 — standalone preview for Apple Silicon

HyLogger Data Bridge — public standalone preview for Apple Silicon Macs.

Download hylogger-macos-arm64.zip, extract it, and run bash Install.command inside the extracted folder. The installer downloads about 3 GB of checksum-verified model files and installs managed Python, scientific libraries and the browser book. No source checkout, Git, Node or administrator access is required. Internet and disk space for downloads, runtime and model are required.

The installer prints the installed hylogger command path. Use:

hylogger review --workspace /path/to/folder
hylogger review --workspace /path/to/folder --package SWMB007s_chips --target 5 10 --reference 0 5

The first command lists supported packages. The second opens a real-data evidence book. Enter the expert decision in the browser; then continue explicitly:

hylogger review --workspace /path/to/folder --case CASE_ID

Books are separated by workspace. No source HyLogger data, saved books or reference measurement records are included. The bundled Qwen3.5-4B model was fine-tuned on organiser-derived measurements with teaching-rule labels; predictions collapse to clarify (38.75%, majority baseline). Advice is experimental and non-authoritative.

Limits: unsigned/not notarized; macOS Apple Silicon only; connected-folder chat, automatic return and ChatGPT compatibility are unfinished. Host registration is opt-in and is not a supported end-to-end route in this preview. No PyPI package is published. This is a pinned preview from the package branch, not a merge of the pending chat adapter.

Validation: full local check passed: 307 Python tests, 42 frontend tests, Ruff, strict mypy, documentation links and installed-wheel validation. Public-download acceptance passed in a fresh folder/profile on the existing M1 Max Mac: managed Python, 90 compatible runtime packages, actual model preflight and real SWMB007s inference, rendered book, labelled UI test decision with exact note preservation, explicit interval-comparison continuation, duplicate replay and server restart with identical persisted state. Imports came from the installed runtime; no checkout imports were used. This is not a fresh physical-machine or expert geological validation. The 3 GB base download took about 16 minutes on the test connection. Pre-action model advice is conservatively labelled stale after the final page revision. Third-party model, uv and frontend notices are included. SHA256SUMS covers downloadable assets; the installer verifies its contents and model parts.