Resume Research State

Resume Research State (Desktop)

Use this skill at the start of a session that needs continuity from a prior session -- most importantly, from a session on a different device. The source of truth is {vault}/00-Claude-Context/, which sync software (Obsidian Sync / iCloud / OneDrive / Syncthing) has mirrored to this machine.

This skill never assumes a prior live Claude Desktop chat is available. It reconstructs context from durable files only.

Operating Rules (do not violate)

1. Read only the selected vault and 00-Claude-Context/ unless the user grants additional scope.
2. Do not inspect Claude app data, keychains, audit logs, browser storage, or caches while resuming.
3. If context files are missing or invalid, prefer to run: python "${CLAUDE_PLUGIN_ROOT}/bin/research_sync_agent.py" validate --vault "{vault path}" (or py -3.12 on Windows), then point the user at any specific failure. If the helper isn't reachable, check the files against ${CLAUDE_PLUGIN_ROOT}/references/context-schema.md.
4. Preserve concurrent snapshots. If two devices snapshotted overlapping sessions, surface both and ask which is authoritative -- never delete or overwrite either.

Resume Procedure

1. Resolve the vault path. Prefer OBSIDIAN_VAULT_PATH; if absent, ask once.
2. Read 00-Claude-Context/manifest.json -- get last_updated_device and snapshot_count.
3. Read in this order, accumulating context:
   • current-state.md
   • task-ledger.md
   • open-questions.md
   • decision-log.md
   • preferences.md
   • research-memory.md
   • paper-map.md
4. Inspect 00-Claude-Context/session-snapshots/:
   • Newest one or two snapshots. List filenames with their source_device.
   • If the newest two are from different devices and within ~24h of each other, treat as a potential split-brain -- summarize both and ask the user which line of work to continue.
5. Produce a short working brief (~10-20 lines):
   • Goal (from current-state.md).
   • Most relevant papers and notes (from paper-map.md + the snapshot).
   • Active tasks (from task-ledger.md).
   • Known decisions (latest 2-3 from decision-log.md).
   • Open questions (top 3 from open-questions.md).
   • Recommended next action -- taken from the snapshot's "Resume prompt" section if present, or synthesized from the active tasks otherwise.
6. Wait for the user to confirm or override the next action. Only then continue work.

Output style

Keep the resume brief practical and citation-flavored. Reference vault paths and citation keys when possible ([[30_Literature/kim2023ionic]]). Do not describe generic skill behavior unless there is a setup issue the user must fix.

Empty / first-run case

If 00-Claude-Context/ exists but every durable file is at its initial "Pending." state, say so plainly: "No prior research state found. Run capture-research-state at the end of this session to seed the continuity layer." Don't invent context.