Capability Budget

A gate run built around least privilege instead of speed. Each stage offers two tool-call gates — READ, WRITE, NETWORK, or SHELL — each showing a claimed type and cost. Pick one, it charges against a shrinking capability budget, and some gates are disguised: the badge claims a cheap READ but actually charges like a WRITE or worse. There’s no clock. The only score that matters is total privilege spent versus the oracle-optimal path, so racing through recklessly is strictly worse than reading the tells.

Riffs on the current shape of agent tool-permission design: sandboxing, least-privilege scoping, and the incidents where an agent’s tool calls turned out to cost more than their labels claimed.

Reuse

The mechanic lives in src/capability-budget.js as a framework-free ES module:

• createRun(opts) — returns a state object; call update(dt) each frame and pick(index) (0 or 1) to choose a gate. opts.rand overrides the RNG, opts.stageCount overrides the run length (default 6).
• State exposes phase (choosing | reveal | breach | complete), stages (each { gates: [gate, gate] }, a gate being { claimed: { type, cost }, actual: { type, cost }, disguised }), stageIndex, budget/startingBudget, spent, optimalCost (the oracle-optimal total — cheapest actual cost available at every stage, for scoring efficiency at the end), and lastPick ({ gate, index } for driving the reveal).
• disguised is exposed on every gate at all times; the module has no notion of a visual tell. The demo’s tell is a small rotation applied only to disguised gates’ badges before they’re picked — purely a rendering choice, easy to swap for a different tell (color, shadow, jitter) without touching the state machine at all.

Port notes for DragonRuby: update() and pick() have no DOM dependencies; gate layout, hit-testing, and the tell all belong to the renderer, same split as every other mechanic here.

Gotchas

• pick() is a no-op outside phase === 'choosing' — same guard pattern as the other turn-based mechanics, so a stray double-click during the reveal hold can’t double-charge the budget.
• A gate’s claimed type is capped a tier below the top (SHELL) so there’s always room for a disguise to jump upward; actual can still land on SHELL from a two-tier jump, which happens for about a quarter of disguised gates.
• The demo bundles its own copy of the module (self-contained by contract). If you touch src/, re-copy it into demo/.