v2.4.2 — MemoryVault Discovery + Mining (paired with toolkit v0.10.0)
MemoryVault Discovery + Mining — paired doc-only release with agent-toolkit v0.10.0. Second roadmap item under the MemoryVault parent design closes (ROADMAP item #7b); the parent design is now fully shipped (both #7a + #7b are Completed).
What ships (toolkit side — five new /memory sub-commands)
/memory index-skills— auto-indexer for installed SKILL.md files/memory reflect corpus— historical-transcript-backlog mining with dry-run-default/memory discover-skills— cadence-checked internet skill-discovery scan (4-source operator-confirmed whitelist)/memory adapt-skills— two-pass adapt-don't-import (Python rubric + LLM sub-agent judgment with GitHub metadata + trustworthiness signals)/memory watchlist— promote/dismiss/defer review surface
Adapt-don't-import is architecturally enforced
The adapt-evaluator sub-agent's tool allowlist physically scopes Write to _skill-watchlist/<source-slug>/<pattern-slug>.md only. Agents cannot auto-fork into agent-toolkit/skills/ — only the operator's manual authoring step graduates a pattern to a real skill.
Why this matters for harness users
The harness itself is unchanged. Operators who installed the memory skill via the toolkit gain five new /memory sub-commands on next install. The personal-skills indexer auto-runs from bash agent-toolkit/install.sh ~/their-project; the cadence-checked skill-discovery scan auto-fires from the existing memory-reflect-idle hook; the adapt-don't-import workflow + watchlist review are operator-invoked when ready.
After this release
ROADMAP item #7b moves to Completed; both #7a + #7b are now shipped (full MemoryVault parent design closes). Next per the locked execution order: #13 diataxis-author skill, with smaller items #19-#22 queued in parallel (Ideas.md format redesign / transfer-context × AgentMemory integration / harness self-audit skill / cross-surface AgentMemory protocol).