Sync all

Sync all → TARS

Vault house rules. Before acting, read $TARS_HOME/AGENTS.md if it exists and honor it — source allowlists, cadence, and privacy carve-outs there override this skill's defaults on conflict.

One invocation, every source. This skill is only the conductor: each connector still runs its own sync-<connector> skill (with that connector's concept and people judgment), and the CLI still owns storage. What this adds is order, a single clean finish, and an optional catch-up digest.

Scope

Default set — the connectors that sweep on a watermark:

• gmail (sync-gmail) — incremental inbox sweep
• jira (sync-jira) — issues assigned since the watermark
• granola (sync-granola) — recent meetings
• github (sync-github) — PRs per connectors.yml

Slack is excluded by default — sync-slack is deliberate, thread-level capture, not a sweep; pull those threads by hand when they matter. Narrow or widen on request ("sync all except gmail", "just jira and github"). Read connectors.yml for the code-connector scope (orgs / repos / authors).

Steps

1. Pick the set from the request + connectors.yml, and announce what you will run so the user can veto before the work starts.

2. Run each connector in sequence, following its own sync-<connector> skill end to end (fetch → shelve concepts → wire people → label / commit cursor). Keep context lean: run the heavy MCP-mediated sweeps (gmail, jira, granola) as a subagent per connector and collect its report; github is a CLI fetch (tars sync github) plus light judgment.

   • A connector whose MCP isn't authorized (the claude.ai Gmail / Granola connectors can be absent in headless or cron runs) → skip it, note it, and keep going. Never fail the whole pass because one source is unavailable.
   • Each connector's own two-phase cursor (--begin / --commit) still guards its watermark, so a connector that errors leaves its cursor uncommitted and the others are unaffected.

3. Finalize once — after every connector finishes, run tars finalize (regenerate hubs, clear any index drift, re-check invariants). Each sync already finalizes itself, so this is the cheap global confirmation that the whole pass left the vault consistent; it exits non-zero if anything remains.

4. Digest (default yes) — run the digest skill to summarize everything that entered since the last digest, by concept, with open questions and extracted commitments. Skip on request.

5. Report: per-connector added / updated / skipped counts and each new watermark, any connector skipped for auth or availability, concepts created vs reused across the whole pass, people pages touched, and the final doctor verdict.

Notes

• Judgment stays per-connector. This skill never centralizes concept or people decisions — each sync-<connector> skill owns those for its own sources, so the taxonomy stays coherent.
• Idempotent. Origins dedup, so a re-run only picks up what is new and a half-finished pass is safe to resume.
• Not a scheduler. For a hands-off cadence, wrap this in a scheduled agent — but the claude.ai-MCP connectors (gmail, granola) may not authenticate in an unattended run, so a schedule reliably covers the code connectors (github, jira) and the MCP-only ones may still need an attended pass.