v1.23.0

gentle-ai v1.23.0

This release makes Claude Code a first-class SDD environment in gentle-ai.

Highlights

• Claude now ships native SDD sub-agents

  • installs ~/.claude/agents/sdd-{explore,propose,spec,design,tasks,apply,verify,archive}.md
  • each phase is now discoverable directly in Claude's native agent system
  • the SDD contract lives in the agent layer instead of depending on orchestrator-only prompt prose

• Per-phase Claude model assignment is now enforced natively

  • design / propose run on opus
  • spec / tasks / apply / verify run on sonnet
  • archive runs on haiku
  • invalid or missing aliases now fall back safely to validated defaults

• Better phase boundaries and backend-aware tool scopes

  • planning phases now have the minimum Edit / Write access needed for openspec
  • Claude SDD agents expose the exact Engram MCP tools they actually use (mem_search, mem_get_observation, mem_save, and mem_update where needed)
  • recursive delegation remains blocked because phase agents do not carry Task

• Claude install, sync, uninstall, golden tests, and E2E are now aligned

  • native sub-agent capability was formalized on the shared adapter contract
  • Claude goldens now snapshot generated native agent files
  • E2E coverage now validates real Claude SDD agent installation and robust tool-scope expectations

• OpenCode SDD permissions are stricter and safer

  • wildcard task permissions were replaced with explicit allowlists
  • profile overlays now discard stale wildcard task permissions during sync

Reliability improvements

• fixed Claude native sub-agent model stamping so generated agent frontmatter matches balanced defaults
• hardened Claude E2E assertions so they validate stable tool presence instead of brittle exact ordering
• kept uninstall and verification paths aware of native sub-agent assets through the shared adapter contract

Community contributors in this release

• Gerardo Franco (@GerardoFC8) for the Claude native SDD sub-agent work in PR #338
• Luis Diego Cubero Chavarría (@LCubero) for tightening SDD task permissions in PR #287

Thanks

This release is a really nice example of the project getting sharper through collaboration: new runtime capability, stronger guardrails, and better tests all landed together. Thanks to everyone pushing on the real workflow details — that's what makes the tool feel solid.