v1.8.0 — Leaner, Smarter, Tested
This release was battle-tested with full SDD flows on both Claude Code and OpenCode before going stable. Every phase ran, every artifact persisted, every sub-agent loaded the right skills.
The Numbers
| Metric | Before (v1.7.x) | After (v1.8.0) |
|---|---|---|
| SDD prompt tokens per session | ~20,650 | ~7,700 |
| Lines in SDD skills | 1,825 | 1,421 |
| Orchestrator template variants | 2 (97% duplicated) | 1 unified |
| Sub-agent skill discovery | Orchestrator-dependent | Self-sufficient |
63% fewer tokens means more of your context window goes to actual work, not instructions.
What Changed
Token Optimization — All shared patterns (skill loading, engram retrieval, artifact persistence, return envelopes) extracted to sdd-phase-common.md. Phase skills now contain only what's unique to each phase. Net: -340 lines across 11 files.
Self-Sufficient Sub-Agents — Phase sub-agents (sdd-explore, sdd-spec, sdd-apply, etc.) no longer wait for the orchestrator to provide skill paths. They discover the skill registry themselves via engram search or .atl/skill-registry.md fallback, then load project-relevant skills automatically. A sdd-apply on React loads React 19 patterns; on Go it loads Go testing conventions.
Per-Agent Model Routing (OpenCode) — resolveAgentModel() reads each agent's model from config and passes it through session.prompt(). Configure Opus for spec/design, Sonnet for everything else — each phase actually uses its assigned model now.
Upgrade Progress — gentle-ai upgrade shows real-time status instead of a blank screen while brew/binary upgrades run.
Unified Orchestrators — Claude Code and Generic orchestrator templates are now identical and agent-agnostic. One source of truth.
Validated With Real SDD Flows
Claude Code — Feature: add-dashboard-view on a Bun + React 19 + Tailwind 4 project
| Phase | Status | Engram | Time |
|---|---|---|---|
| explore | PASS | #1333 | ~79s |
| propose | PASS | #1334 | ~39s |
| spec | PASS | #1335 | ~50s |
| design | PASS | #1336 | ~62s |
| tasks | PASS | #1337 | ~48s |
| apply | PASS | #1338 | ~129s |
| verify | PASS | #1339 | ~89s |
- Spec + Design ran in parallel without conflicts
- Apply loaded React 19, Tailwind 4, TypeScript skills from the registry
- Verify flagged missing tests as CRITICAL (correct behavior)
- Zero regressions vs v1.7.24
OpenCode — Feature: add-notification-system on the same project
| Phase | Status | Engram |
|---|---|---|
| explore | PASS | #1341 |
| propose | PASS | #1343 |
| spec | PASS | #1344 |
| design | PASS | #1346 |
| tasks | PASS | #1347 |
| apply | PASS | #1348 |
| verify | PASS | #1349 |
- 10/10 verify checks, zero criticals, zero warnings
- Apply used
React 19 use() hook,<Context value={}>syntax,cn()from Tailwind 4 - Zero regressions vs v1.7.24
Companion Releases
- engram v1.10.5 — Eager tool loading + UserPromptSubmit hook for Claude Code
- engram v1.10.8 — FTS5 now indexes
topic_keysosdd/{change}/proposalsearches work directly - agent-teams-lite v4.1.1 — Per-phase model assignments in example configs
Upgrade
brew upgrade gentle-ai
brew upgrade engramFull changelog: v1.7.24...v1.8.0