Capture

Capture into TARS

Vault house rules. Before acting, read $TARS_HOME/AGENTS.md if it exists and honor it — per-vault rules there (source allowlists, tone, privacy, output layout) override this skill's defaults on conflict.

Everything lands through tars add; the CLI owns storage, dedup, and provenance. Route by whose words it is — that decides the connector, and who authored it is load-bearing (never store your synthesis as the user's words, or their words as yours):

What it is	Command
External source — a URL, a file	tars add "<url>" / tars add "<path>"
External text pasted from elsewhere (an excerpt, someone else's message)	pipe to tars add - --title "<short title>", with --origin "<source-url-or-ref>" when it has one so provenance survives
The user's own words — a thought, note-to-self, decision, running note	pipe verbatim to tars add - --title "..." (lands under the note connector); a living note they keep adding to → pin --origin "note:<stable-slug>"
A synthesis you authored — a work-log or decision record, kept at the user's request	tars add - --connector agent --origin "agent:<stable-slug>" --title "..."
Files dropped in inbox/	tars sweep, then shelve what landed like any capture

Route rules:

• External sources: zero interpretation — never summarize or edit content on the way in. If extraction fails (paywall, JS-only page), fetch it yourself (WebFetch / agent-browser) and pipe the text through tars add - with --origin "<url>". Optional labels the user gives you: --tag <tag>.
• User's words: verbatim. Light shaping only (split a dictated run-on into paragraphs, keep an obvious list a list) — store what they said, don't reword or interpret.
• Agent notes: a faithful record of work actually done — what was done, decisions and their why, open follow-ups; don't pad or speculate. This is the one sanctioned agent-authored path, and it's user-initiated — never a silent summary of a doc you just synced (that would be write-time distillation). The stable slot refreshes in place across sessions (agent:proj-1234-worklog); omit --origin only for a frozen one-shot record.
• Stable slots vs content-addressing: with no --origin, pasted text content-addresses (note:<hash>/agent:<hash>) so re-adding identical text dedupes. A note the user updates in place wants a named slot — a slug they'd say out loud and reuse verbatim (note:onboarding-questions).
• Piped text needs --title (nothing to extract one from). For multi-line text, write a scratch file and pipe it — heredocs mangle special characters.

Shelve to concepts (mandatory, no ceremony)

Every capture gets 1–4 concepts — pass --concept <slug> at add, or tars tag <doc_id> --concept <slug> ... after. ls wiki/concepts/ first and reuse aggressively; mint a new concept only when nothing fits, as a general speakable slug (generic-service, not translation-context-admin-api). Then tars hubs once — it regenerates every hub's ## Sources (and skeletons for new concepts). Hub polish — titles, descriptions, relevance clauses — is the gardener's job, not capture's.

Report

One line: the add result (added/updated/unchanged), the origin slot if named, and the concepts attached — flag newly minted ones so the taxonomy growing stays visible.