Is knowledge-base-enrich safe to install?

SaferSkills independently audited knowledge-base-enrich (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.

CategoryWeightCategory scoreContribution

Securityprompt, exec, net, exfil, eval

35%

100

35.0 pts

Supply chainhash, typosquat, maintainer, lockfile

20%

100

20.0 pts

Maintenancestaleness, pinning, CI

15%

100

15.0 pts

TransparencySKILL.md, perms, README

15%

100

15.0 pts

Communityinstalls, verify, response

15%

100

15.0 pts

Findings & checks · 0 flagged

Securityscore 100 · 0 findings

✓—All security checks passedNo findings in this category for the latest scan.pass

Supply chainscore 100 · 0 findings

✓—All supply chain checks passedNo findings in this category for the latest scan.pass

Maintenancescore 100 · 0 findings

✓—All maintenance checks passedNo findings in this category for the latest scan.pass

Transparencyscore 100 · 0 findings

✓—All transparency checks passedNo findings in this category for the latest scan.pass

Communityscore 100 · 0 findings

✓—All community checks passedNo findings in this category for the latest scan.pass

The primary manifest — the file an agent reads to learn what this artifact does.

skills/knowledge-base-enrich/SKILL.md · 1 file

skills/knowledge-base-enrich/SKILL.md4.0 KB · Markdown

name: knowledge-base-enrich

description: >-

Work a knowledge-forge knowledge base in two modes: (A) ENRICH — automatic

Ollama/local-model summaries (enrich-worker) or manual override; (B) DISCOVER —

search the catalogue and install recommended tools/skills. Do NOT auto-write

summaries in the agent; the Ollama worker handles the queue.

Knowledge Base — Enrich & Discover (tool-agnostic)

Mode A — Enrich (automatic): queued entries are enriched by **`knowledge-forge

enrich-worker** (Ollama) or Docker enrich-worker`. Cursor/Claude hooks do not prompt agents to write summaries.

Mode B — Discover: search the catalogue for best practices and install tools

the videos recommend.

CLI: knowledge-forge <cmd> · MCP: kf_ingest, kf_search, kf_queue, …

Capture: knowledge-forge ingest <url> (MCP: kf_ingest) — Telegram bridge in integrations/telegram/.

Browse: knowledge-forge serve → http://localhost:8077 · Docker: docker compose up -d.

Automatic enrichment (default — Ollama only)

After update/ingest, entries needing 300-1000 word summaries land in the enrich queue. The worker drains it via Ollama (default gemma3:4b), gated by the same validator.

knowledge-forge enrich-worker            # loop
knowledge-forge enrich-worker --once     # single pass
docker compose logs -f enrich-worker     # Docker

Provider is pluggable via env (no code change): KB_LLM_PROVIDER = ollama (default) | openai (any OpenAI-compatible host) | anthropic. Set KB_LLM_API_KEY/KB_LLM_BASE_URL for hosted models, KB_ENRICH_MAX_WORDS for longer summaries. Full table: integrations/docker/README.md.

Dual pipeline (KB_PIPELINE=auto|llm|rules):

Hosted credentials (auto/llm): ingest always transcribes, gathers GitHub

sources in tiers (caption URLs → transcript URLs incl. spoken URL normalization → spoken owner slash repo → GitHub API search → optional GLM llm-extract), then one-pass hosted LLM enrich at ingest. Transcripts cached under transcripts/by-url/<hash>/. KB_GITHUB_LLM_EXTRACT=false disables LLM tier. Optional GITHUB_TOKEN / KB_GITHUB_TOKEN for API rate limits. Hosted empty-response fallbacks: full prompt → compact transcript prompt (KB_LLM_COMPACT_PROMPT_FALLBACK) → iterative chunk map-reduce (KB_LLM_ITERATIVE_FALLBACK, KB_LLM_ITERATIVE_CHUNK_CHARS). Compact/iterative also run after validation rejections and transient API errors. HTTP retries: KB_LLM_HTTP_RETRIES, KB_LLM_HTTP_RETRY_DELAY. Apply payload carries sourceContext for consistent re-validation. Telegram/knowledge-forge ingest picks this when credentials are in .env.

No hosted credentials (rules): rule-based enrich + Ollama enrich-worker queue.

Do not process the queue in the agent unless the user explicitly requests manual override.

Hooks rebuild the queue after update silently — no agent follow-up to enrich.

Anti-hallucination rules (mandatory)

Only use `sourceContext` from queue entries.
Web search only for names/URLs already in sourceContext.
Source URLs must trace to the post or sourceContext.
No engagement bait in summaries or learnings.
`validate` before `apply` always.

Mode A — Manual override (user asks only)

knowledge-forge queue
knowledge-forge validate --file payload.json
knowledge-forge apply --file payload.json

Write 300-1000 words from sourceContext only. MCP: kf_validate then kf_apply.

Mode B: Discovery

Frame the use case.
Search: knowledge-forge search "<use case>" or kf_search(...).
If needsScrape: true, run knowledge-forge update --force --transcribe, then

knowledge-forge run, wait for enrich-worker, re-search.

Answer from matched entries; cite source URLs.
Install tools (after confirm) using grounded installCommands only.

Validation gate

Word count 300-1000, per-paragraph grounding, no bait, URLs from source only.

knowledge-base-enrich— agent skill

Is knowledge-base-enrich safe to install?

Knowledge Base — Enrich & Discover (tool-agnostic)

Automatic enrichment (default — Ollama only)

Anti-hallucination rules (mandatory)

Mode A — Manual override (user asks only)

Mode B: Discovery

Validation gate

Audit the pieces. Scan the whole. Decide.