b1abd87c3d54de1c229887d6521062ee88387e60
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
| 43147b796a |
feat(rust): site-owned permissions — the site is the author, the game is the cache
R2, and the first phase where this module WRITES to a game. Groups and grants are authored on the website and pushed into each server's own permission store, so every plugin that already calls `UserHasPermission` honours them with no adapter, and a wipe stops being a data-loss event. **Seven org-lead decisions (D28-D34).** A grant is keyed to the website USER and resolved to every Steam id they have linked at push time (D28); every authored row carries a scope — a server or `*` (D29); groups are mirrored as real groups rather than flattened (D30); a holder the site did not author is REPORTED, never undone, with adopt and revoke offered (D31); one verb, with the plugin diffing locally (D32); a permission no server has registered is reported unresolved and never self-registered (D33); authoring is people and groups by hand, with rules deferred (D34). **Three sets, and every interesting question is a difference between two.** `desired − pushed` is what to apply; `pushed − desired` is what to RETIRE, because the site put it there and has since withdrawn it; `present − desired` is drift. The middle one is why `rust_perm_pushed` exists: a name in the store that is not in the desired set is either something the site retired or something a human granted, and those two have opposite correct answers. **What lands is not what was sent.** A grant naming a permission the server has not registered did not land — `GrantUserPermission` no-ops silently — and a member the store has never seen could not be placed. Neither is recorded as pushed, so the site never believes it gave a privilege it did not. The loop asks a cheap question every thirty seconds — does the digest of the desired set still equal what this server last confirmed — and syncs on a change, a restart, a wipe, a drift hook, a failed attempt past its backoff, or the fifteen-minute audit that finds drift on a server nobody has touched. **This module's first admin page**, because a permission model is the first thing here that has to be composed rather than configured. What is on it is decided by what an operator can get wrong: four states are invisible from the game and from a list of grants, and each is a sentence rather than a number. Walked end to end against a real core at the pinned ref, the real sidecar, and a stand-in speaking protocol 4 — including a restart that emptied the store and was fully re-pushed. Four defects the browser found that 133 green tests did not. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PMH6bw1jXMgbyF3ZWGEzSM |
|||
| f211969ee1 |
feat: ingest protocol 2, and keep the record a wipe cannot erase
The module half of the read path. Seven tables, an ingest cursor, four public routes, and one file whose only job is deciding who may see what. **The record and the window are different things.** `rust_player_wipe_stats` and `rust_gather_totals` are permanent and per-wipe, so all-time is those rows SUMmed rather than a second set of counters that can disagree with them — that is R12's "per-wipe detail plus all-time rollups" in one table instead of two. `rust_events` is a bounded 30-day window of raw frames for the killfeed, and `rust_presence` is a board: replaced wholesale, never appended. **The feed is a cursor, not a socket, and the header says why.** Core runs Node 20, where a global WebSocket is still behind a flag, so a socket means taking `ws` — against a release that asserts it has no runtime dependencies (D5). The deciding argument is the other one though: a socket needs a cursor anyway, for whatever it missed while the module was restarting, and the catch-up path is the one that has to be right. A cursor alone is one mechanism exercised every five seconds rather than two where the second only runs after an outage. **The cursor advances after the batch, never before.** A crash between the two re-reads events already counted, which inflates a total; the other order loses them silently and for ever. One is visible and bounded, the other is invisible and permanent, so the code fails in the visible direction. A server with no cursor starts at the sidecar's current END rather than at zero — replaying a fortnight of deaths into stats for wipes the site never saw is not a catch-up. **`catalogue.js` is a security boundary, default-deny.** Protocol 2 carries IP addresses (login attempts, approvals, bans), one player's report about another, and the grid reference of somebody's base. They are stored, because an operator chasing ban evasion needs them; they are not served below the admin tier. The allowlist lives here rather than as a field on the wire, because a boundary declared by the sender is one a compromised or merely out-of-date game host can widen — the same reason core's own shard fan-out filters on the serving side. A kind this build has never heard of is not public, and a test holds the list against PROTOCOL.md §8.4 so that adding a kind to the protocol without classifying it fails a build. `PROTOCOL_VERSION` goes to 2 here in the same change as the emitters, though this module consumes none of the new frames yet: the sidecar refuses a mismatched client with a 409, so a module left on 1 would stop being able to read the board it has been reading all along. A constant that lags the deployment is an outage with a version number on it. 95 server tests, 20 client tests, every guard green, and `routes.manifest.json` regenerated against a real core at the pinned ref: 10 routes, all documented, none of core's moved. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016wDDVXWMDz82WqE1i969r4 |
|||
| b33d21d71b |
feat(ci): packaging, release and the frozen manifest
Phase 2 of docs/modules/rust/PLAN.md. Phase 1 built five guards and ran them by hand; this repo had no workflows at all, so nothing gated the branch that gets released and there was no way to release it. Three pieces: - **release.yml** — the derived-version engine link, installer and Module-uo already run (conventional-commit subjects since the newest tag; module.json's version survives as a floor; workflow_dispatch as the backdoor), assembling the bundle from an include list and publishing the tarball, the install manifest carrying its sha256, and SHA256SUMS. The tag is the number that ships and CI stamps it into the bundle's own module.json. - **pr-checks.yml** — server tests, check:imports, check:bundle, check:swagger, the client build, client tests and check:externals, plus frozen-manifest. - **frozen-manifest** — clones core at the sha pinned in ci/core-ref.json, generates its route table without this module and with it, and takes the difference. It ran locally against that exact ref: six routes, all documented, no core route moved. That is the first proof by a running core that /rust collides with nothing — phase 1 could only check it by reading, because core mounts /status and /version at a tier root where the loader's own collision probe cannot see them. The bundle carries no node_modules, because the shipped half declares no runtime dependencies (org lead, phase 2). checkBundle.js holds both halves of that: the include list still covers everything server/index.js reaches, and no dependency has appeared without the release learning to pack it. Verified by breaking it — dropping "model" from the list names the exact edit and exits 1. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016wDDVXWMDz82WqE1i969r4 |
