Ship local capability layer (gstack-inspired) — 4 components
Decision
Built gstack-inspired local capability layer: aj-checkpoint CLI (companion to existing continuous-checkpoint.py hook), bench-model CLI (cross-model latency/cost/quality compare), aj-taste skill+CLI+weekly-decay-cron (5%/wk soft preference signal), aj-doc-release skill+CLI+post-ship-hook (deterministic product-doc drift detector). Risk (a) NOISE-list shell-noise mitigated in-session; risks (b)-(f) deferred to 30/90d outcome review.
Rationale
Comparison vs garrytan/gstack identified 4 high-value gaps in AJ’s enterprise stack: (a) no cross-model cost/quality routing intelligence, (b) no soft preference layer with decay (vault/graphiti are permanent), (c) no targeted product-doc drift detector (Pristine Sweep is setup-tree only), (d) gap between auto-firing checkpoint hook and manual recovery CLI. Build held to zero-redundancy (Law 1 caught existing continuous-checkpoint.py before duplicating), zero-compromise (DRY_RUN cron, async-detached hook, dual-config parity), zero-tech-debt (single 119L topic file vs 4 feedback files = bloat avoided). All deterministic — LLM cost only when bench-model invoked. Smoke-tested all 4 capabilities end-to-end. PRISTINE verdict on Phase 1.5 audit.
Alternatives Rejected
Rejected (1) gstack /critic-codex cross-model agent — deferred until ChatGPT Max plan + Codex CLI subscribed. (2) 4 separate feedback files — chose single topic file to avoid bloat per setup-curator format contract. (3) Continuous-checkpoint hook duplicate — caught existing implementation, built CLI companion only. (4) Auto-edit prose mode for aj-doc-release — deemed too risky, advisory-only output.
Outcome
Pending