Digest

Digest 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.

Digests are derived artifacts: regenerable views over raw/, never truth. Every claim links its source as an Obsidian wiki-link — [[<raw-file-stem>|<title>]] (raw files are named by human title; tars search --json returns the stem in the file field). Organize content by concept, not by document — concepts are the vault's grouping. Never edit notes/ from here.

1. Scope: tars cursor digest gives the last digest watermark (ISO timestamp; empty = ask the user for a window, default last 7 days). Then, before reading anything, stamp the sweep start: tars cursor digest --begin — two-phase advance, so a document captured while you work lands in the next digest instead of being skipped forever. The new set is tars list --since "<watermark>" --json (id, title, connector — no bodies); tars log --json distinguishes added vs updated when it matters.

2. Read at the depth each document deserves — never full-show by default:

   • Connector backfills (github, jira sweeps): the titles are the digest — one summary line for the batch, per-doc lines only for items that changed something (a decision in a PR discussion, a ticket that flipped state).
   • Meetings (granola): tars show <id> --head 60 covers frontmatter + the Notes panel; tars show <id> --grep "next steps|próximos pasos" -C 6 mines commitments. Read the transcript only when a claim needs the exact words.
   • Small captures (note, agent, slack, web, file): full tars show is usually fine — they're short.

3. Write digests/<YYYY>-W<week>.md (append a dated section if it exists):

   • What came in — one line per document: [[<file-stem>|<title>]] + connector (a connector backfill collapses to one line with a count).
   • By concept — the heart of the digest: one ### [[<concept>]] subsection per active concept, containing that concept's decisions (actually made, not discussed), movements, and open questions/tensions — each claim ending with its [[<file-stem>|<title>]] source link. Cover every concept touched this period; cross-cutting items go under the most load-bearing concept with links to the others.
   • Unshelved — anything that fits no concept; if it recurs, that's a signal to create one.
   • Retrieval health — count of new retrieval-misses.md entries this period (omit the section when zero). A sustained miss rate around ~20% of asks is the agreed trigger to consider a semantic-search rung — flag it when the trend points there.
   • Promotion candidates — at most 2–3 insights worth tars promote, phrased and ready for the user to approve, each linked.

4. Extract commitments into tasks/ — one file per commitment, from meeting "next steps" and explicit commitments in any source. Before creating one, grep tasks/ for the same Source: link and action (extraction is idempotent). File tasks/<created-date>-<slug>.md:

   ---
   status: open
   owner: me            # or the person's slug
   due: 2026-07-10      # or —
   created: 2026-07-06
   ---
   
   <the action, one clear sentence>
   
   Source: [[<file-stem>|<title>]]
   Concepts: [[<concept-slug>]]
   Owner: [[<person-slug>|<Name>]]   # only for others' commitments
   

   Agents never flip status or delete task files — that's the user's. Then regenerate tasks/TASKS.md exactly as the tasks skill specifies: ## ⚠ Overdue first (any open task past its due date, omitted when empty), then ## Mine / ## Waiting on, sorted by due date (nulls last).

5. Commit the watermark: tars cursor digest --commit — only if every document was covered and the digest written cleanly; on any failure leave it uncommitted so the next run re-scans from the old watermark.

6. Show the user the digest content and any new tasks, and mention documents skipped (empty, unreadable) by name. If ## ⚠ Overdue is non-empty, list those tasks and ask which are done, moved, or need a new date — apply what the user answers.