v0.14.0 - The Harvest

We scanned all 408 forks of this repo and shipped the gold back upstream

Round 2 of the fork sweep (round 1 was at 166 forks, 2026-05-30). Every diverging fork was compared and deep-read; the best ideas landed upstream the same day, each credited to the fork that built it in FORK_INSIGHTS.md - and where a stalled PR existed, its author is a co-author on the merge.

New thinking tool: /obsidian-brainstorm (#153)

The skill's 45th command and its first stateful, multi-turn tool: a Socratic interview - one question per turn, six question categories, convergence-gated - that ends in a design note with named alternatives and exactly one (Recommended). Grounds its questions in your vault's own decisions and contradictions.

Bounded vault recall (#155, opt-in)

A UserPromptSubmit hook that injects a small brief of relevant vault notes into every prompt: max 4 notes / ~900 chars, abstains when confidence is low, fail-closed, and logs every decision for audit. Inert until you arm it (OBSIDIAN_RECALL_ENABLED=1).

Research toolkit upgrades (#147, #148)

• Brave Search and Tavily join /research free mode when keyed; the pool stays key-less without them
• Full-page source reading: /research-deep fetches its top cited pages via Tavily Extract and synthesizes from actual page text, not snippets
• /youtube summarizes via Gemini free tier with automatic Grok fallback
• The usage ledger is fail-soft and now covers Perplexity + Gemini, labeled per command

Search and safety (#151, #152)

• supersedes: [[B]] in an ADR now fades B in search results even if B's own status was never updated; the pure-lexical path finally gets the freshness rerank it was documented to have
• The write-time hook detects secret material (keys, tokens, private-key blocks) in vault notes; the unattended bg-agent routes sensitive content to a staging note and never writes raw secrets anywhere

Operations (#149, #150, #154)

• scripts/update-vault-integration.sh: a guarded updater for copied dist builds - ff-only pull, smoke-test gate, backup + rollback
• The retrieval eval benchmarks any external engine (--mode external) on the same cases and metrics
• Portuguese (pt-BR) triggers for all 45 commands (translation base by @renatofaria-ia)

Also

• codex-cli and opencode builds deprecated in favor of the unified agent-skills tree (thanks @Litash, #146)
• Full-repo hygiene sweep: 8 privacy fixes, ~15 staleness fixes, SEO/AEO surfaces brought current incl. a four-releases-frozen CITATION.cff (#156)

Fork credits this round: the local-first memory fork, the api-ledger fork, the brave-source fork, the web-reader fork, the updater fork, the brainstorm fork, the structured-rag eval fork, the OKF-first fork. Test suite: 191.

Full detail in CHANGELOG.md.