RaLHF

RaLHF — Personal Context Engineer for your AI assistant (dev variant)

Gives your AI assistant a personal context engineer. Before any task, RaLHF assembles a context package from your personal wiki, the assistant's memory, local files, and connected apps — and shows you the plan before the assistant touches the work.

Built by Bot Food on the RaLHF MCP server.

Production plugin source. This repository contains the public RaLHF plugin package for Claude and Codex. Public releases point at the production RaLHF MCP endpoint: https://backend.ralhf.ai/mcp.

---

Why this exists

Generic AI answers are the default failure mode of every assistant. The model has no idea who you are, what you've already decided, what your team calls things, or which of your past projects matter. So it produces something competent but anonymous — and you spend the next five turns correcting it.

RaLHF fixes that by intervening before the assistant executes. On every user message, RaLHF runs a five-phase flow that gathers everything relevant to the task — from your RaLHF wiki, the assistant's memory, your local project, and your connected apps — proposes a plan, and waits for your green light before handing the assembled package to the assistant.

---

The flow

Load → Discover → Propose → Confirm → Execute → Remember

Phase	What happens
0a — Triage & ask-first gate	On every real task, RaLHF names the task, recommends whether to pull context or hand straight to the assistant, and asks (yes / no) — then waits. The recommendation is computed from the prompt (no lookups): personal-context signals → pull; self-contained tasks → skip; new users lean pull. Trivia / plugin meta-questions skip the gate silently. Only get_my_mcp_usage runs before the reply.
0 — Load	On "pull": silent pull of personalized retrieval rules (get_instructions) and the wiki map (get_wiki_catalog). No separate greeting — identity rode along in the gate. The light flow (pull on a self-contained task) skips the catalog.
1 — Discover	Parallel drill: relevant wiki pages, the assistant's existing memory, local project files, session state, and the MCP connector surface. Follows wikilinks and triages source documents.
2 — Propose	Staged check-ins (Turn 2a / 2b / 2c) — one ask per message. Soft asks first, then connector flow when Gmail / Calendar / Drive / Jira / etc. could help.
3 — Confirm	Surfaces gaps + final pre-handoff check-in + Library-refresh ask (promotes new Drive/Cowork/memory items via upload/remember). Hard gate. No execution until the user says go. A silent context-gathering postmortem (save_context_feedback) fires here at handoff.
4 — Execute	Hands the assembled package to the assistant with citations and scope notes.
5 — Remember	Post-task feed-ralhf ask captures durable facts (dense summary + file uploads) and saves approved artifacts to the Library. The postmortem already fired at handoff.

See PHASES.md for the orientation map and skills/ralhf-start/SKILL.md for the canonical spec.

---

Skills

Skill	Trigger	What it does
ralhf-start	Auto-fires on every user task	The main five-phase flow
ralhf-learn	/ralhf:ralhf-learn <fact>	One-shot save of a preference, fact, constraint, or goal
ralhf-sync	/ralhf:ralhf-sync	Manual Phase 5 with a review gate before saves
ralhf-intro	/ralhf:ralhf-intro	First-run setup check + onboarding intro
feed-ralhf	/ralhf:feed-ralhf	Autonomous end-of-session dump: dense summary + file uploads + postmortem

---

Installation

This production plugin points at https://backend.ralhf.ai/mcp. Install from the public Claude marketplace or Codex marketplace as documented below.

Claude Code

Build the distributable zip, unzip it, then install the extracted plugin folder:

./build-plugin.sh
unzip dist/ralhf-3.10.1.zip
/plugin install ./ralhf

Codex

This repo also includes .codex-plugin/plugin.json, so Codex can load the same skills/ directory and .mcp.json server definition.

This repository is the plugin source, not the marketplace root. Keep Codex marketplace catalog metadata in botfoodai/ralhf-codex-marketplace so the RaLHF plugin package does not vendor marketplace state. Use that marketplace repository for Codex installation instructions. After installing or updating the plugin, start a new Codex thread so the skills and MCP tools are loaded.