Repository navigation
Research and Freshness
Every version number and API claim in minewright comes from a dated research pass, and each skill re-checks its claims on a schedule. The skills are "evergreen" units: each one carries the files that keep it current, and a freshness check at the start of every use.
| File | Holds |
|---|---|
SKILL.md |
the working document, under 200 lines |
RESEARCH.md |
Current understanding, Open questions, a four-track Search plan, and a Findings log of dated R- entries with sources |
CHANGELOG.md |
every change to the skill, as C- entries, each with its reason |
LEARNINGS.md |
procedural lessons as L- entries, each with a trigger and a hypothesis |
TESTS.md, evals/evals.json
|
the eval suite and the record of its runs (T- entries) |
evergreen.json |
tier, interval, last_checked, next_due, the volatile claims, and the history |
MAINTENANCE.md |
where to find the evergreen protocol, and the condensed rules when it is not installed |
Each skill reads its evergreen.json at Step 0 of every use. If today is on or after next_due, or a contradiction is set, the agent says so in one line, does your task with the current content, and then refreshes the skill in the same session. Below the due date it spends nothing more.
State on 2026-10-27, the day all five were written:
| Skill | Tier | Interval | Next due | Volatile claims it watches |
|---|---|---|---|---|
minewright-project |
fast | 3 days | 2026-10-30 |
@minecraft/server 2.11.0 and server-ui 2.3.0 stable; min_engine_version [1, 26, 50]; core-build-tasks 5.7.1; eslint-plugin-minecraft-linting 2.0.13; the Windows GDK path; MCPE-227874; consoles only through a Realm |
minewright-hud |
fast | 3 days | 2026-10-30 |
hud_screen.json still JSON UI; the Ore UI migration; server-ui 2.3.0; no item-render-by-id control; custom glyphs unofficial |
minewright-workflow |
moderate | 14 days | 2026-11-10 | the stable module versions; GameTest has no stable release; custom add-ons earn no achievements; experiments are permanent; Creator Tools' MCP server and skills |
minewright-checks |
moderate | 14 days | 2026-11-10 | Creator Tools 0.20.0 and its validate flags; Creator Tools needs Node 22+; the Blockception schemas release |
minewright-server |
moderate | 30 days | 2026-11-26 | BDS 1.26.60.6 and its download URL; the module versions it accepts; the default permissions.json; no /gametest runall; the loopback SIDs; the console line formats |
The two fast skills track what moves with every game release: module versions, engine versions, install paths and the JSON UI to Ore UI move. A refresh moves the interval: a finding that proves a core claim wrong shortens it a lot, a minor one leaves it, and a quiet pass lengthens it by half. Each tier has a floor and a ceiling (fast: 3 to 21 days; moderate: 14 to 90).
A refresh reads RESEARCH.md, then runs four to eight searches scoped to the time since last_checked, at least one on each track:
| Track | What it searches |
|---|---|
| Subject | The thing itself: npm dist-tags for @minecraft/server, Mojang/bedrock-samples, the Microsoft Learn changelog with the stable view, mojira.dev for MCPE-227874, the BDS download-links API |
| Tooling | What else helps an agent do the job: Creator Tools releases and its own skills, path:SKILL.md searches on GitHub, skills.sh, the MCP registry |
| Practice | How others build Bedrock add-ons with agents, from dated sources only |
| Testing | How the job is verified: @minecraft/server-gametest's dist-tags (does it ever go stable?), others' scriptevent self-tests, and a re-run of bds_run.py on the fixture after each BDS update |
Each material finding becomes an R- entry with its track, sources and a magnitude. A change to the skill becomes a C- entry that cites it. Each RESEARCH.md names its best sources (npm registry JSON, Microsoft Learn, Mojang's GitHub repositories, mojira.dev) and its noisy ones (video tutorials pinned to 2023 APIs, sites that rehost .mcaddon files, hosting companies' blog posts).
| Skill |
R- findings |
L- lessons |
|---|---|---|
minewright-workflow |
7 | 11 |
minewright-project |
7 | 8 |
minewright-hud |
7 | 5 |
minewright-checks |
6 | 5 |
minewright-server |
6 | 7 |
Building the example add-on taught the skills 24 lessons, logged in their LEARNINGS.md and CHANGELOG.md on 2026-10-27. Examples: minecraft:pushable is rejected from entity format 1.26.10; vanilla item_texture.json keys differ from item ids; a hidden JSON UI control stops its children's bindings.
Each RESEARCH.md keeps the questions it could not settle. Among them:
- When does the MCPE-227874 import fix reach a stable release?
- Does a base Realm (not Realms Plus) accept an uploaded world with add-ons?
- Does
bds_run.pybehave the same on the Linux build of BDS? Only Windows was run. - Is the resource pack applied on BDS at all? Its pack stack never lists it.
- Does
kill()bypass a Totem of Undying on 1.26.60? - Which HUD screens has Ore UI taken over in each release?
- Do the example's JSON UI tricks (the
title_framereplace, the factory-child animation, the blank texture) work in the game?
Each evergreen.json sets "protocol": "plugin", so the skills follow whichever evergreen plugin is installed (evergreen-protocol) rather than carrying their own copy. They work without it: MAINTENANCE.md holds the condensed rules, including the interval arithmetic.
Every skill carries six eval cases: two trigger prompts, two decoys that must not fire it (one of them a sibling's job), an action case proven by a tool call or a file, and an outcome case. On 2026-10-27 no eval run had been recorded yet: each evergreen.json shows "last_run": null.
This wiki describes minewright 0.1.0 (no release tag yet) and was last updated on 2026-10-27. The plugin is MIT licensed. Report problems in the issues.
Using it
The skills
How it stays current
Contributing
Elsewhere