capture

Capture — Universal Source Intake

Purpose

Ingest one source into raw/<type>/<slug>.<ext> (immutable, SHA-256 hashed) and create wiki/sources/<slug>.md (source-summary with key_claims[]).

Used by wiki-scribe (primary, passive intake) and wiki-researcher (via research workflow).

Trigger phrases (agent perspective)

• "capture this URL"
• "save this PDF"
• "intake source"
• "file clipboard"
• "hash and store this"
• "wiki-scribe should ingest X"

Inputs

One of:

• URL (any https://)
• YouTube link
• GitHub link (issue / PR / repo)
• PDF (path or URL)
• Raw text (clipboard, session log, transcript)

Outputs

Two files per capture:

1. raw/<type>/<slug>.<ext> — immutable, SHA-256 in frontmatter, full content
2. wiki/sources/<slug>.md — source-summary with key_claims: [{quote, anchor, confidence}], quality, captured_by

Workflow

1. Detect source type from URL or path:

   • *.pdf → capture-pdf.sh
   • youtube.com/* or youtu.be/* → capture-youtube.sh
   • github.com/*/issues/* or */pull/* → capture-github.sh
   • github.com/*/pulls → capture-prs.sh
   • text without URL → capture-text.sh
   • other URL → capture-url.sh

2. Invoke the appropriate script with --vault $VAULT_PATH and --source $INPUT. Scripts handle SHA-256 and frontmatter.

3. Extract key_claims for source-summary:

   • If source <500 words: preserve full content as single claim
   • If 500–5000 words: top-5 verbatim quotes with anchors (paragraph index)
   • If >5000 words: top-10 quotes with anchors

4. Set quality field in source-summary:

   • high if extraction succeeded with >500 chars
   • medium if 200-500 chars (paywall preview, partial)
   • low if <200 chars or extraction errors

5. Set captured_by to invoking agent name (wiki-scribe or wiki-researcher).

Web→MD stack (for capture-url.sh)

The script tries three extractors in order:

1. defuddle (Node CLI, MIT, local) — primary
2. trafilatura (Python CLI, Apache-2.0, local) — fallback if defuddle fails OR content <500 chars
3. r.jina.ai (cloud) — opt-in via env WIKI_ALLOW_CLOUD=1

If all three fail, exit 3 and tell user to provide content via capture-text.sh (manual clipboard).

Quality gate

• Every output has SHA-256 in frontmatter
• wiki/sources/<slug>.md is in tier 4
• key_claims is never empty for sources >500 words (if empty, log warning)
• captured_by is set

Scripts used

• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-url.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-youtube.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-github.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-prs.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-pdf.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-text.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/capture-git-log.sh

Anti-patterns

• DO NOT interpret content during capture (preserve verbatim)
• DO NOT synthesize across sources (that's research's job)
• DO NOT edit existing wiki pages (only create raw/ + wiki/sources/)