jarvis-knowledge-os-v2

Classify vector-similar pages against each orphan page via Haiku 4.5 and add typed inbound references (link_type="related", relation encoded in context). Goal: chip away at the 1803/1952 orphan pile (v1 wiki flat import) across multiple sweep runs. Default is --dry-run; --apply does DB link inserts + markdown sentinel upsert on candidate pages that exist on disk, then commits ~/brain changes.

Merge WITHIN-SOURCE duplicate entity-variant pages (e.g. people/lucien + people/lucien-chen, companies/tp-link-system-inc + tp-link-systems-inc) into one canonical node: repoint links + facts + display-name aliases, register a slug redirect, soft-delete the alias pages. An LLM disambiguates each cluster (merge / ambiguous / distinct); ambiguous first-name-vs-multiple-full-names cases are quarantined, never merged. Default --dry (simulate + report, zero writes). Cross-source pages are NEVER touched — the two-axis design keeps them legitimately separate.

One-shot sweep that reads every page in the brain, extracts every person/company/concept/project mentioned via Haiku NER, dedupes, tiers by mention count + context, and creates stub pages for entities that lack one. Tier 2 candidates get Tavily web-search augmentation; Tier 3 stays brain-only (confidence: low). This is the primary G1 payoff skill from JARVIS-NEXT-STEPS.md §5.

Daily sweep: runs kos-lint + staleness detection + knowledge-gap analysis, then produces a MEMORY-format digest ready to be appended to OpenClaw's MEMORY.md by the digest-to-memory skill (Sun 22:00 weekly). Also writes a dashboard to ~/brain/agent/dashboards/knowledge-health-<date>.md.

Bridge an upstream gap: extract_atoms never stamps `frontmatter.concepts` on page-derived atoms, but synthesize_concepts consumes ONLY that field — so concept synthesis is permanently skipped ("no atoms with concept refs"). This script batch-labels atoms lacking the field with 1-3 kebab-case shared-vocabulary topic labels (sonnet) and stamps them, so upstream synthesize_concepts can tier + materialize concepts/<label> pages. Idempotent: only touches atoms WITHOUT `concepts`; safe to rerun after each extract_atoms drain. Retires when the upstream fix lands.

One-command bulk-ingest pipeline (D→I→K→W) for a single isolated source. Sequences the 7 fork stages — register → ingest → embed → enrich → graph → (per-entity synth, opt-in) → corpus-synth — over a folder of frontmatter'd .md files, each idempotent/incremental so re-runs only do new work. Turns a systematic corpus (product docs, FAQ, a research library) into a queryable knowledge asset: facts (RAG) + entities + graph + corpus-level viewpoints.

Corpus-level Wisdom synthesis. Where synthesis-sweep deepens ONE entity, corpus-synth synthesizes a whole imported knowledge base into queryable "观点与洞察" pages — the corpus's major themes, cross-document tensions, transferable patterns, and open questions. Top-down: Opus reads a corpus map, proposes the themes, then deeply synthesizes each. Turns a freshly imported systematic KB into knowledge AND wisdom (not just RAG facts).

The "Wisdom layer" sibling of enrich-sweep. Walks each significant entity, gathers EVERY source page that mentions it via the by-mention link graph, feeds the whole cluster into Opus 4.8 in one large-context shot, and writes back a deep, evidence-cited high-level dossier (kind: synthesis). Turns thin entity stubs into synthesized knowledge: role, relationships, decision timeline, open questions, and cross-source insights a single page can't show.

RETIRED 2026-05-17. The Notion → gbrain delta poller produced 0 net pages for at least 24 h (10.4 MB of "0 ingested (empty)" logs); brain has 0 pages with source_of_truth=notion and 0 pages tagged notion-ingest. Worker moved to workers/_archived/notion-poller/. Replacement: mailagent v4 SQLite SSoT push to kos-compat-api/ingest directly (方案 B), tracked in GitHub issue on ChenyqThu/jarvis-knowledge-os-v2. Story in §6.27.

Daily cron that scans IMAGE_SOURCE_DIR, imports new images (jpg/png/heic/webp/gif) into gbrain via importImageFile, embeds them through Voyage multimodal-3 (1024d), and writes them as pages with cross-modal search capability (gbrain search --image).

Provider-neutrality adapter. GBrain's src/core/embedding.ts hardcodes OpenAI (`new OpenAI()`) for 1536-dim vector generation. Lucien's stack uses Gemini embeddings via Google Generative Language API through NANO_BANANA_API_KEY. This shim is a tiny local HTTP server that speaks OpenAI /v1/embeddings on the client side but calls Gemini batchEmbedContents on the backend, preserving the 1536 dimension via outputDimensionality. Keeps upstream src/ untouched.

Auto-score a page's confidence (high/medium/low) based on evidence levels, backlinks, recency, and citation density. Also assigns a compile grade (A/B/C/F) for the latest ingest event touching this page. Driven primarily by dikw-compile and kos-patrol; exposed as a standalone command too.

KOS-flavored post-ingest compilation pass. After any ingest skill creates a source page, walk the DIKW (Data → Information → Knowledge → Wisdom) compilation: identify impacted concept/synthesis/decision pages, enforce strong-link type annotations (supplements/contrasts/implements/extends), 2-5 link budget per page, and bidirectional citations. Runs after idea-ingest / media-ingest / meeting-ingestion.

Gatekeeper that blocks low-evidence claims from landing on durable pages. Ports KOS v1 evidence levels E0-E4 to GBrain. Different page types have different minimum evidence thresholds. Fails fast on violations; advises how to upgrade the claim.

Normalize relative-path `*.md` references inside KOS page frontmatter (`related:`, `source_refs:`, etc.) into canonical slug form. v1-wiki legacy left ~220 refs in `../X/Y.md` shape across the brain; gbrain's link-extraction DIR_PATTERN doesn't accept `sources/` (plural) so frontmatter refs to source pages all silently fail to resolve. v2 adds: (1) EXTERNAL_POINTER_KEYS allowlist for fields that legitimately point at brain-external paths (e.g. `raw_path:`); (2) bare-slug fuzzy resolve via basename index; (3) optional deletion of brain-external refs (`../../X.md`) and dead links (`X/Y.md` with missing target).

Six-check lint pass ported from KOS v1 lint-light.sh + staleness-check.sh. Extends GBrain's `maintain` with checks specific to KOS quality rules (frontmatter required fields, duplicate KOS ids, weak-link detection with type annotations, evidence coverage gaps). Runs standalone or chained by kos-patrol.

One-shot slug normalization for KOS-v2 — fixes 7 root-level stray pages by adding the missing `kind/` prefix (`ai-jarvis` → `concepts/ai-jarvis`, 6 URL-slug sources → `sources/<slug>`). Preserves `frontmatter.id`, updates 1 intra-brain compiled_truth reference. Part of Step 1.5 of the filesystem-canonical migration track.

Migration adapter for the existing OpenClaw 飞书 skill. The skill lives at ~/.openclaw/workspace/skills/knowledge-os/ and was designed for KOS v1 `kos` CLI. This file documents the minimum changes needed when pointing that skill at GBrain.

Queue schema for entity-mention signals that need later enrichment. Producer: OpenClaw Feishu signal-detector fork (external). Consumer: skills/kos-jarvis/enrich-sweep/run.ts. This file pins the on-disk format so both sides stay honest without shared code.

KOS-Jarvis adapter that bridges kos-compat-api `/ingest` to UltimateSearchSkill's Tavily Extract → FireCrawl Scrape three-tier fetcher (with native fetch fallback). Replaces v2 KOS's bare `fetch(url)` which had no deadline and caused Cloudflare 524 timeouts on slow/reflective targets like X/Twitter.

KOS → OpenClaw MEMORY reflux. Port of KOS v1 `kos digest` writeback flow. Takes the latest kos-patrol digest and appends `[knowledge-os]` entries to ~/.openclaw/workspace/MEMORY.md 近期层 so daily Jarvis decisions are informed by what KOS compiled that week. Runs weekly Sun 22:00 via cron-scheduler, or on demand.