v1.8.0-beta.1

v1.8.0-beta.1 — Token Optimization (BETA)

This is a pre-release for community testing. Install only if you want to help validate that SDD phases still work correctly with optimized prompts.

What changed

~63% reduction in SDD prompt token overhead — from ~20,650 tokens to ~7,700 tokens per session.

SDD phase skills (sdd-explore, sdd-spec, sdd-apply, etc.) previously duplicated the same boilerplate in every file: skill loading, engram retrieval patterns, persistence contracts, return envelopes. This bloated every sub-agent's context with ~1,500 tokens of repeated instructions.

Now all shared patterns live in sdd-phase-common.md and each skill references it. Skills contain only phase-specific content — what artifacts to read, what work to do, what to output.

Changes

• Expanded sdd-phase-common.md with 4 shared sections:
  • A: Skill loading (self-discovery via engram/registry)
  • B: Artifact retrieval (mem_search → mem_get_observation pattern)
  • C: Artifact persistence (mem_save template, mode switching)
  • D: Return envelope format
• Slimmed 8 SDD phase skills — net -340 lines across all files
• Unified orchestrator templates — Claude and Generic are now identical and agent-agnostic

What to test

1. Run a full SDD flow: /sdd-new some-feature through to /sdd-verify
2. Verify sub-agents still read artifacts correctly (engram search → get observation)
3. Verify sub-agents still persist their artifacts (check engram after each phase)
4. Verify sub-agents still load project skills from the registry
5. Test on both OpenCode and Claude Code

How to install

brew tap Gentleman-Programming/homebrew-tap
brew install gentle-ai --HEAD  # or build from source on this branch

Report issues at https://github.com/Gentleman-Programming/gentle-ai/issues

Full changelog: v1.7.24...v1.8.0-beta.1