[prdai]

Summary

• Stop splicing memory into event.system at index 1; append it after the stable prefix instead
• Update plugin tests to find the memory chunk by content rather than a fixed index

Fixes #46

Context

#46 reported two cache-busting behaviors:

1. Per-turn toISOString() in <memory_metadata> — already removed on main / 0.3.0 (thanks)
2. Early splice at system index 1 — still present; any memory edit invalidates tools + core instructions for prompt cache

This PR addresses (2). Appending keeps tools/core instructions cacheable; memory (and journal notes) only touch the trailing system chunk.

Test plan

• bun test src/plugin.test.ts --timeout 120000
• bun test src/prompt.test.ts
• Manual: long OpenCode session should show growing cacheRead rather than a flat ~22k ceiling after memory is injected

Made with Cursor

[prdai]

Closing — early placement of memory blocks is expected for salience; the per-turn ISO timestamp bust is already fixed on main / 0.3.0.