138 Tests, 7 Repos, Zero Failures: Strengthening the Hawk Ecosystem

The Hawk ecosystem is seven Go projects that work together: hawk (agent), eyrie (LLM runtime), tok (compression), yaad (memory), sight (code review), inspect (site auditor), and trace (session capture).

Today we ran a deep health check across all of them, then merged improvements to every single one. Hawk remains in development — no public install.

What we merged

Prompt caching (eyrie) Anthropic's cache_control mechanism saves 90% on repeated system prompts. We added full support — cache breakpoints on system prompts, tool definitions, and conversation prefixes. Also parsing OpenAI's cached_tokens from responses.

MCP servers for sight and inspect Any MCP-compatible agent (Claude Code, Cursor, etc.) can now use our code reviewer and site auditor. Just connect via MCP protocol — no hawk required.

Yaad MCP resources and prompts Yaad's MCP server now exposes 3 resources and 2 prompts. The documented API is now fully implemented.

Cascade routing Automatic model selection based on task type. Planning goes to Opus, coding to Sonnet, summaries to Haiku. No manual model switching needed.

Integration tests Five end-to-end tests that exercise the full pipeline: hawk calls eyrie (with mock), compresses via tok, stores in yaad, reviews with sight, and scans with inspect.

The numbers

Before: 129 test packages. After: 138 test packages. Failures: 0. Race conditions: 0.

All seven repos now have CI triggering on both main and dev branches, all green.

Building in public

We ship every day. Every commit is public. Every test must pass before merge.