Add WORLD.md project charter and Dependabot baseline

.github/dependabot.yml

@@ -0,0 +1,12 @@
+version: 2
+updates:
+  - package-ecosystem: "bundler"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    labels: ["loop:deps"]
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    labels: ["loop:deps"]

WORLD.md

@@ -0,0 +1,77 @@
+# WORLD.md — project charter
+
+This file sets direction for agentic: what it exists to do, what it will not
+become, and how proposed work gets judged. The maintainer (codenamev) owns it.
+Minerva (minerva-sky), an autonomous agent that helps maintain this project,
+reads this document before proposing anything and stays inside its fences. If a
+proposal conflicts with this file, the proposal is wrong.
+
+Status: PROPOSED. Drafted 2026-08-13, awaiting maintainer approval.
+
+## Purpose
+
+A Ruby gem for building and running AI agents in a plan-and-execute fashion,
+as a library and a CLI. It exists so Ruby developers can build purpose-driven
+agents with plans they can inspect, edit, and save, without leaving Ruby or
+adopting a framework the size of a small country.
+
+## Direction
+
+- The plan is the product. Inspectable, editable, persistable plans are what
+  separate this from "call the API in a loop." Depth here wins: better plan
+  quality, better failure recovery, better execution history.
+- The self-assembly and capability system (v0.2.0) is the bet worth maturing:
+  agents that construct themselves from task requirements, with versioned
+  capabilities and a persistent store.
+- Good gem citizenship: semantic versioning, honest CHANGELOG, current Ruby.
+  Track the latest stable Ruby; do not carry EOL versions.
+- Provider strategy is an open question. The gem is OpenAI-only today. Whether
+  to abstract providers (and which second provider earns the abstraction) needs
+  a researched proposal with a real cost, not a drive-by PR. Until decided,
+  no provider abstraction lands.
+
+## Constraints
+
+- Small core. New runtime dependencies need strong justification; the gem
+  should stay auditable in an afternoon.
+- Public API changes follow semver and deprecation cycles. People's agents
+  run on this.
+- Tests accompany every behavior change. CI (main.yml) gates merges.
+
+## Anti-goals
+
+- Not a LangChain port. Ruby idioms over framework mimicry.
+- Not a hosted platform. No server component, no accounts.
+- Not an everything-store of integrations. Capabilities keep the core small;
+  integrations belong in plugins or downstream gems.
+- No speculative abstraction. Two concrete uses before one abstraction.
+
+## How work gets proposed
+
+Improvements arrive as GitHub issues, labeled by origin and state:
+
+- Origin: `loop:quality`, `loop:security`, `loop:deps`, `loop:research`,
+  `loop:self` (agent-originated), or unlabeled (human-originated).
+- State: `status:analyzed`, `status:deferred`, `status:wont-do`,
+  `status:blocked`. A closed issue with `status:wont-do` records the reason in
+  its final comment and is permanent institutional memory. Proposals must check
+  closed and deferred issues before re-raising an idea.
+
+## Review policy
+
+Changes are classed by risk (external visibility times reversibility), and each
+class has an autonomy level that can rise as the agent's track record earns it:
+
+- L0: propose in an issue only.
+- L1: open a PR; the maintainer merges.
+- L2: open a PR; it may merge after a 72-hour quiet period with green CI.
+- L3: merge on green CI, reported in a digest.
+
+Every class starts at L0 or L1. Promotions happen only on the maintainer's
+explicit approval, backed by the acceptance record. Any revert demotes the
+class immediately. Gem releases and version bumps are the maintainer's alone.
+
+## Amending this document
+
+By pull request with maintainer approval, nothing else. The agent may propose
+amendments; it may never merge them.