docs: update AGENTS.md with accurate file layout and current conventions

AGENTS.md

@@ -6,14 +6,17 @@
 
 ## Layout
 
-- `index.ts` — Extension entrypoint, footer registration, refresh timers, slash command, auth lookup.
+- `index.ts` — Extension entrypoint, footer registration, refresh timers, debounced rendering, slash command, auth lookup.
 - `config.ts` — User-facing layout/config. Prefer changing this for layout tweaks.
 - `types.ts` — Shared config/module/state types.
 - `render.ts` — Width allocation, ANSI-aware truncation, dim text rendering.
+- `utils.ts` — Shared utility helpers (`isRecord`, `toNumber`).
+- `usage.ts` — Provider-agnostic usage type definitions (report, limit, credential, etc.).
 - `modules/basic.ts` — Directory, context, model, thinking, and cost modules.
 - `modules/usage.ts` — Usage bars/text modules and color gradient rendering.
 - `usage/` — Provider usage fetchers adapted from `oh-my-pi`: https://github.com/can1357/oh-my-pi/tree/main/packages/ai/src/usage
-- `providers/openai-codex/constants.ts`, `usage.ts`, `utils.ts` — Local shims/support for copied usage code.
+- `usage/shared.ts` — Shared helpers for usage fetchers.
+- `providers/openai-codex/constants.ts` — OpenAI Codex base URL constant.
 
 ## Conventions
 
@@ -26,14 +29,15 @@
 { type: "usage", window: "week", style: "text", parts: ["percent", "time"], separator: " | " }
 ```
 
-- Usage fallback behavior should be stable/no-pop-in:
-  - unknown/loading line bars render like 0% usage
-  - unknown/loading percent renders `0%`
-  - unknown/loading time renders `∞`
+- Usage data is preserved during refresh (no clearing before fetch) to prevent pop-in:
+  - Old data stays visible until new data arrives.
+  - When no data exists yet: line bars render like 0% usage, percent renders `0%`, time renders `∞`.
 - Usage bar colors are a continuous green → yellow → red gradient derived from usage percentage, not hard thresholds.
 - Rendering must remain ANSI-aware. If adding colors, update `visibleWidth`/`truncate` behavior as needed.
 - Text modules and separators should remain dim; filled usage bars may use gradient color.
 - Thinking level comes from `pi.getThinkingLevel()` and `thinking_level_select`, not `ctx.thinkingLevel`.
+- `rerender()` is debounced via `queueMicrotask` — multiple rapid calls coalesce into a single render pass.
+- All providers use the same refresh interval (`REFRESH_MS`).
 
 ## Testing