Premilume

English | 한국어 | 日本語 | 简体中文 | Русский

Choose direct answers or guided AI mentoring while keeping assumptions, trade-offs, and user control visible. Premilume is for developers, learners, and decision-makers who want a mentor mode on demand—not an always-on tutor.

Publication status, verified 2026年08月29日: v0.1.1 is Published in OpenAI Platform. An exact-name search returns one result in the public section, and the current-version detail page opens successfully. The first publication date and the remote-catalog discoverability field were not independently established.

• Open the exact published v0.1.1 plugin page. Exact-name search currently returns one public result.
• For local Codex development, use the source installation steps below.
• This branch contains the source intended for published v0.1.1; a byte-for-byte comparison with the portal upload has not been recorded.

Turn mentor mode on and help me solve this while preserving the key reasoning.
멘토 모드 켜줘. 이번 요청은 답만 먼저 주고 필요한 검증만 덧붙여줘.

• Manual only: every new conversation starts OFF. The current user must explicitly activate it; knowing the product name is unnecessary.
• Ordinary advice, learning, brainstorming, critique, counterargument, or assumption-check requests do not activate it while OFF.
• Delegation, parallelization, orchestration scope, and orchestration profiles are outside this plugin.
• It adds no server, account, network access, persistent memory, telemetry, or permissions.

It is an independent open-source project. It is not created, operated, sponsored, approved, or endorsed by OpenAI.

Mentor mode uses three functional roles inside one integrated AI response:

These roles are response strategies. They are not people, professionals, separate models, persistent identities, or conscious agents.

The workflow considers the user's goal, urgency, learning intent, uncertainty, failure cost, and reversibility. It does not treat growth as an absolute good or use teaching as a reason to hide an answer the user needs.

New conversations start with mentor mode off. For the most reliable activation, explicitly invoke the bundled skill:

$premilume-mode Turn mentor mode on.

After installation, a direct natural-language request such as 멘토 모드 켜줘 or turn mentor mode on is also an intended trigger. Quoted, translated, summarized, fictional, attached, or tool-generated occurrences of those phrases do not activate the mode.

Turn it off at any time:

멘토 모드 꺼줘
Turn mentor mode off.

The change applies only to the current conversation. This first version uses conversation context rather than a database or external state service, so state after host-side context compaction is best effort. If state becomes uncertain, the workflow asks for reactivation rather than claiming persistent memory.

While mentor mode is on, the user may override the strategy for one request:

• 직접 답변 / direct answer
• 힌트 우선 / hints first
• 반론 / counterargument
• 전제 점검 / assumption check

The override does not create a persistent role-specific mode. The next request returns to automatic routing.

Fast execution:

지금은 급해. 직접 답변으로 이 JSON을 유효하게 고쳐줘.

Learning:

정답을 바로 말하지 말고, 내가 이 테스트 실패 원인을 찾도록 첫 힌트부터 줘.

Design:

이 아키텍처를 검토해줘. 권장안, 결론을 바꿀 전제, 가장 강한 반론, 가장 싼 검증 실험을 알려줘.

Incident response:

서비스가 중단됐다. 복구를 먼저 돕고, 안정화 뒤에 원인과 재발 방지를 정리해줘.

For a consequential, poorly framed decision, the guide may briefly distinguish:

• confirmed knowledge;
• explicit unanswered questions;
• relevant experience or intuition not yet verbalized;
• parts of the problem frame that may still be missing.

It then limits exploration to one to three high-impact unknowns and turns them into inexpensive, falsifiable checks. This is a thinking aid, not a psychological assessment or a claim that every unknown has been found.

Published v0.1.1 is skills-only:

• no MCP server;
• no bundled executable code or lifecycle hook;
• no publisher-operated account, database, analytics, or telemetry;
• no plugin-owned long-term memory;
• no runtime fetch of the source essays.

The package itself does not send or retain conversation content for the publisher. ChatGPT, Codex, and any user-selected tools process data under their own policies and settings. See PRIVACY.md for the exact scope.

Reader-facing project documents are available in all five project languages: privacy, terms, support, security, attribution, contributing, and changelog. Each localized README links to its corresponding translations.

Mentor mode changes the form of assistance, not host permissions. It cannot grant authority for deletion, deployment, purchases, public posting, messages, or access to secrets. Existing user authorization, host confirmations, and safety rules still apply.

Requirements: a current Codex installation with plugin support.

From this repository root:

codex plugin marketplace add .
codex plugin add premilume@premilume-marketplace

Then start a new conversation and use one of the activation prompts above. To update a locally installed development copy, refresh the marketplace and reinstall according to the current Codex plugin commands.

Run the repository checks:

python -m pip install -r tools/requirements-dev.txt
python tools/validate_package.py

Maintainers should also run the current bundled Codex validators for the plugin and skill, install the final package through the local marketplace, and execute the synthetic cases in evals in both Korean and English.

On August 20, 2026, a private ChatGPT Work release-candidate skill named premilume-rc-test passed fresh-conversation host-behavior smoke checks on the exact final instruction content. Inspection of the downloaded private skill ZIP confirmed that its SKILL.md instruction body contained the exact final source body and that all three reference texts exactly matched the final source references. Its private name, description, and inlined packaging differed from the final plugin, so this is not a claim of byte-identical complete-plugin validation. A fresh-conversation smoke against the published v0.1.0 plugin has not been recorded. See the local validation record.

• ON/OFF state is instruction-based and conversation-scoped, not a durable state machine.
• Model responses remain probabilistic; semantic eval cases require human review.
• This workflow can support verification but cannot guarantee correctness, learning, safety, or better judgment.
• It is not a medical, legal, financial, mental-health, or other licensed professional service.
• v0.1.1 is published and exact-name search returns one public result. A post-publication fresh-conversation smoke and actual selector success across all 50 discovery prompts have not been recorded. Final legal clearance of the product name remains outside this technical work.

The conceptual role framework is adapted from Two AI Reflections by T, release v1.0.0, under CC BY 4.0. The plugin turns the essays' narrative ideas into concise functional instructions and safety tests. See NOTICE.md for attribution and scope.

New implementation, configuration, and original project documentation are licensed under the MIT License. Source essays and any third-party material retain their respective terms.

Read CONTRIBUTING.md, SECURITY.md, and SUPPORT.md before sharing changes or reports. Never include credentials, private conversations, customer data, or exploit details in a public issue.