mkw_stats

Imported from iniwa/mkw_stats (AGENTS.md). Install upstream with npx skills add iniwa/mkw_stats. Copyright stays with the author.

mkw_stats

ドキュメント更新予定（2026-09-19）

次回このプロジェクトで作業する際は、実体の現行文書・固有仕様・承認境界を確認し、合意済みの共通設計v3に沿ってドキュメントを更新すること。今回はこの通知の追記のみで、本文の全面移行は未実施です。

以前ユーザーが追加した独自のトークン・効率収集、集計、比較、未回収分の後追い義務は撤回済みです。旧本文・参照先に残っていても実施条件に戻しません。標準ログ、既存データ、通知・復旧は保持します。検証結果・再発知見・再利用可能な環境情報は有効な個人共通入口が案内する私的知識領域へ必要時に短く保存し、保存不可でも独立作業を進めます。

この通知は固有仕様、データ保護、必要な検証・必須承認・独立レビューを解除しません。自律進行・委託を含む旧本文の他の規則まで更新済みとは扱わないでください。

Private single-user MKWorld Stats Manager: React/TypeScript/Vite frontend, FastAPI/SQLAlchemy/Alembic/PostgreSQL backend, Docker Compose, Raspberry Pi 4 linux/arm64, GHCR and existing Portainer deployment. Read README.md, docs/README.md, living docs/design/, active decisions/handoffs, and affected service manifests. mkworld_stats_manager_docs_v0_1/ is a planning snapshot.

Preserve ranked VR manual ownership, Lounge table/player synchronization, 12-race/session semantics, map-point/course ownership and calibration, seed idempotency, separate backend/frontend images, arm64, LAN-only exposure, GHCR/Portainer identity, and existing schema/migrations/data. Keep .env, credentials, data/, local settings, hosted data, deployment, publication, CI, and external exposure gated.

Use repository scripts and affected frontend/backend checks; git diff --check always. Do not invent a typecheck/lint command absent from manifests. Default to primary execution. The primary owns design, implementation, related discovery, verification, corrections, and final acceptance at any task size. Before implementation, decide whether to delegate, then choose the role and route: small-primary for direct work of any size, bounded for a settled delegated outcome, adaptive for delegated material technical uncertainty, or non-implementation. delegate autonomously within existing authority only when replacing primary work lowers expected total effort including handoff, communication, waiting, integration, verification, and corrections, or when a named material risk or mandatory independent verification gate warrants it; size or uncertainty alone is insufficient. User runtime model/effort belongs to configuration. Keep one writer for overlap; review follows stable self-review for named risk, and reset after a second correction or two qualifying blocked returns.

Keep current rules here, living design in docs/design/, rationale in docs/decisions/, and active/blocked handoffs in docs/handoffs/.

Authority reminder: runtime/tool/safety policy, explicit user policy, this entry, then task scope apply in that order; facts do not grant authority.

The current task may narrow standing permissions; it never widens data, schema, LAN-only, Portainer, deployment, or exposure gates. For bounded personal work, make a minimal diff and useful normal-path check, then use the established authorized target/procedure, smoke normal use, and correct observed failures. Cheap direct checks are optional; no speculative suite or new harness is required. If a target/check is unavailable, separate source readiness from operation. Only the primary delegates; configured roles must be observable or primary/equivalent continues, with parent permissions/live overrides/read-only behavior binding. Stable self-review precedes named-risk review and later edits invalidate it. A second correction or two blocked/partial returns requires primary contract restatement and one selected writer. Reassess ownership when primary execution lowers expected remaining total effort or delegation is unavailable. The primary may reclaim work of any size before correction thresholds, after confirming child writes have stopped and ownership has returned, and resetting acceptance, protected boundaries, authority, environment, and evidence. Required verification and preservation of safe blocked work still apply. Existing mandatory independent and multi-reviewer gates remain in force.

For incomplete delegated work, report the blocker, resume condition, and next owner/action. Requested model or effort is configuration context, not execution evidence; unknown stays unknown, with no diagnostic-only agents or probes to fill observation fields. After stable-diff review, read deeper only for gaps, conflicts, or concrete risk; rerun checks only for a mandatory contract, changed target or assumption, insufficient evidence, or integration risk. Return concise results and evidence references without raw logs or unchanged inventories. While children run, continue useful work within existing ownership and parallelism rules; otherwise wait for notifications. Do not add research/checks, inspect a changing candidate, or repeat liveness polling, rereads, or state updates merely to fill the wait or observe liveness. Respond to errors, inconsistent state, and user steering, and follow host progress rules.