Skip to content

Token Efficiency Overhaul

Choose a tag to compare

View all tags
@stevebrownlee stevebrownlee released this 28 Feb 20:15
· 7 commits to main since this release
v1.3.0
aa0b50b

What's New

Strategy file overhaul — 80–90% cold-start token reduction

The behavioral instruction files agents load at startup have been completely rearchitected:

  • Deduplicated template architecture — 7 IDE-specific strategy files (Roo, Cline, Cursor, Windsurf, Claude Code, Claude Desktop, Generic) are now generated from one _core_strategy.yaml + 7 thin delta*.yaml files merged at install time. Single source of truth, no more drift between tool variants.
  • Just-in-time strategy delivery — Strategy sections are now served on-demand from a built-in BUILTIN_STRATEGY dict in the MCP server via get_custom_data(category='engrams_strategy'). The installed bootstrap stub drops from ~10,000 tokens to ~800. Full section details are fetched only when the agent actually triggers that workflow.
  • Semantic compression — Remaining instruction text rewritten in flat, terse format. Eliminated 8 empty thinking_preamble blocks and over-structured YAML nesting that added tokens with no semantic benefit.

Breaking Changes
None. All 166 tests pass. Existing strategy files continue to work; the new architecture applies to freshly initialized projects via engrams init.

Installation

pip install engrams-mcp==1.3.0
Assets 2
Loading