Capture Ingest Protocol

When Kevin pastes a URL, X post, GitHub repo, article, tool, or similar source and says "add it", the task is full graph integration, not a one-page note.

This protocol exists because Kevin's primary capture path is chat. He should not need to say "update every related page, route the tool, rebuild search, and log it." That is the default. Source: User, 2026-05-11; AGENTS.md Ingest

Trigger Phrases

Treat these as capture ingest unless Kevin explicitly narrows scope:

  • "add it"
  • "add this to the wiki"
  • "remember this"
  • "capture this"
  • pasted URL/repo/post with an imperative
  • "this is important for X"

If Kevin asks only for a summary or opinion, answer the question first. If the result is durable, propose or perform the wiki update depending on repo access.

Non-Negotiables

  1. Resolve the source. Read the source itself through the best available route.
  2. Brain-first. Search existing pages before creating anything.
  3. Choose the right shape. Use Wiki Brain Operating Model before creating pages.
  4. File correctly. Use Filing Decision Tree and target directory README.
  5. Expect multi-page impact. A real capture usually touches 5-15 pages.
  6. Cite every new fact. Use inline [Source: ...].
  7. Update routing and versions. Tools, CLIs, MCPs, skills, SDKs, frameworks, and workflows must land in the relevant category router, SKILL-RESOLVER, Capability Routing Map, Active Stack (What to Actually Use), trigger evals, or config when behavior changes. Versionable tools also need a current release/package/API snapshot from primary sources.
  8. Keep the graph bidirectional where it matters. Add entity timeline references for meaningful new mentions.
  9. Rebuild machine surfaces. Run build-index, qmd update, and qmd embed; follow Generated Surface Contract when packs, registries, redirects, or stateful ledgers are affected.
  10. Log the operation. Append wiki/log.md.
  11. Codify repeats. If the capture class will recur, create or extend a skill/automation.

Standard Flow

Phase Action Output
Resolve Fetch/read the source with agent-reach, browser, GitHub, curl, or primary docs source facts and provenance
Search qmd search, qmd query, _index.md, aliases existing pages to update
Shape Wiki Brain Operating Model page, section, ledger, generated view, redirect, or no write
Classify Filing Decision Tree primary page location
Integrate update primary and related pages compiled truth + timeline entries
Route update skill/tool/action surfaces resolver rows, active-stack, trigger evals, version snapshots
Verify build-index, qmd, targeted doctors machine surfaces current
Record log entry and absorb log if raw source durable operation history

Page Touch Expectations

A capture often affects:

  • primary entity page
  • related project/tool/concept page
  • comparison or stack page
  • active-stack or resolver
  • philosophy/design/style page when it changes Kevin's stance
  • capability routing, active stack, or Dedalus config mirror when behavior changes
  • entity pages for key people/projects/tools
  • wiki/log.md

One-page capture is acceptable only when the source is narrow and has no routing, entity, or graph implications.

Raw vs Chat Captures

Source class Extra step
Chat paste only no absorb-log entry unless later stored under raw/
raw/ source call markAbsorbed(rawPath, pageSlugs) after compile
External page cite URL and access date
X/social post cite post URL, author, and date where available; use React Tweet / ::tweet when the original post is source-critical
GitHub repo cite repo URL and commit/date when exact code facts matter; check releases/tags before calling the tool current
Source-critical external link use Wiki Artifact Embeds / ::artifact when a rendered card is more useful than a bare link

Activation Gate

A captured skill/tool/loop is not done until it is usable:

  1. Routed in SKILL-RESOLVER, the relevant category router (tools/README, design/README, skills/README, etc.), or Capability Routing Map.
  2. Recommended in Active Stack (What to Actually Use) or explicitly marked reference-only.
  3. Covered by config/trigger-evals.json if a trigger phrase should route to it.
  4. Reflected in Capability Routing Map and any relevant harness projection if compact routing changes.
  5. Version/release checked from primary sources when the capture changes a tool, SDK, CLI, framework, platform, or agent skill.
  6. Validated by npm run routing-doctor and npm run trigger-eval when routing changed.

The wiki actively recommends. It does not merely store.

Anti-Patterns

  • Creating a lone page from a rich source.
  • Summarizing the source without synthesizing it into existing pages.
  • Missing citations because the source "obviously" says it.
  • Adding a tool page without a resolver route.
  • Updating compiled truth by appending stale layers.
  • Marking raw material absorbed before it is actually integrated.
  • Promoting a tool from X/social without checking the package registry, release tags, docs, changelog, or API version.

Timeline