Graph ingest

Graph Ingest

Capture an external source into the second brain through the graph. This is the bridge between graphify's add and the gAIOS wiki loop: fetch → admit → graph. One command, three guarded hops, ending with the new knowledge cross-linked in wiki/ and visible in the graph.

When to run

• A URL / paper / tweet / thread / page is worth keeping, not just reading once.
• The user says "save this", "add this to my knowledge", or pastes a link and wants it remembered.
• You spot a source referenced 2-3 times across sessions → suggest ingesting it so it stops being re-fetched.

The output (always this shape)

## Ingested: <source title>
**Source** — <url> (fetched <date>)
**Captured** — raw/<file> (git-ignored inbox)
**Admitted** — wiki/<entry>.md   ⏸ draft — awaiting your OK to commit
   - de-identified: <what was stripped/referenced-out, or "nothing flagged">
   - lint: <wiki_lint.py result — pass / N warnings>
   - cross-links: <related wiki entries linked>
**Graphed** — graphify-out/ refreshed; new node(s): <node names>
   - notable edges (EXTRACTED / INFERRED): <1-2 connections worth seeing>
**Next** — review the draft entry → approve commit → open graph.html

Process

1. Fetch into the capture inbox. Run graphify add <url> [--author --contributor] — it pulls the source into raw/ (the git-ignored capture inbox), then re-extracts. Use --author/--contributor only when you know who wrote it. Verify: the new file landed in raw/ and nowhere else. If the fetch failed, report the error; do not hand a half-fetched file forward.
2. Admit it to the wiki. Hand off to /wiki (references/sops/wiki-translate.md) to ADMIT the capture: read any existing entry first then merge, apply the admission policy (de-identify — strip/reference-out any secret, PHI/PII, named-deal specific, or live financial figure), cross-link related entries, update wiki/_index.md, then run python tools/wiki_lint.py. Gate: lint must pass (no broken-link/IBAN/secret errors) before the entry is eligible to commit. Leave the entry as a draft; do not commit without the human's OK.
3. Refresh the graph. Run /graph (default scope = code + committed wiki/ only) so the admitted entry joins the graph. Verify: graphify-out/ regenerated and the new entry appears as a node; name 1-2 notable edges from GRAPH_REPORT.md, labelled EXTRACTED vs INFERRED so claims stay honest.
4. Move the capture. Per the wiki SOP, the processed file moves raw/ → raw/_archive/ (never deleted).
5. Hand back. Surface the output block, point at the draft entry and graph.html, and offer to commit once approved.

Autonomy

L2 — drafts the wiki entry; admission and commit are human-gated. Fetch and graph-refresh run automatically (reversible, local); the wiki entry stays a draft until the human approves the commit. Never commit de-identified content on the user's behalf without that OK.

Guardrails (from CLAUDE.md) — CRITICAL

• Fetched external content lands ONLY in git-ignored raw/. It reaches committed wiki/ ONLY after de-identification + a passing tools/wiki_lint.py. (Guardrails #1, #2, #7)
• Never graph raw/ directly. raw/ can hold raw PHI/PII/financials before de-identification; the graph's default scope is code + committed wiki/ only — raw/, .env, .tmp/ are hard-excluded.
• No secrets / PHI / confidential figures into wiki/ — reference the source, never transcribe sensitive specifics. The wiki is committed; treat it as public.
• Cite, don't invent. Keep the source URL on the entry; rely on graphify's EXTRACTED / INFERRED / AMBIGUOUS labels rather than asserting unverified links.
• graphify-out/ is a git-ignored derived artifact — don't commit it.