| name | vault-resync |
| description | Re-pull all of the user's external capture sources into their Obsidian vault inbox and refresh the Command Center dashboards. Use this whenever the user asks to "re-sync", "resync", "pull new stuff", "update my inbox/vault", "refresh my tasks", "what's new in Feishu/WeChat/my tabs", "catch me up", or "pull from the providers again" — even if they only name one source, default to refreshing all of them unless they scope it down. The sources are: Feishu (lark-cli), WeChat self-chat (chatlog), Zen browser tabs (session decode), and active Claude Code sessions (the AI build queue). This skill encodes hard-won auth/format gotchas that are easy to get wrong — always consult it before touching these sources rather than improvising. |
Vault re-sync
This pulls the user's scattered capture sources into one place — the Obsidian vault inbox — and
refreshes the dashboards they read. The whole point of the vault system is that they shouldn't have
to hold "what's new across Feishu / WeChat / ~1500 browser tabs / my Claude projects" in their head.
This skill is the recurring sweep that keeps that promise.
User-specific values (vault path, Feishu app details, the WeChat self-chat hint) live in
local.private.md next to this SKILL.md — it's gitignored and machine-local, kept out of the public
repo on purpose. Read it at the start of a run. Where a value can be discovered at runtime instead
of stored, prefer that (noted per-source below). $VAULT below = the vault root from the sidecar.
The four sources → where they land
| Source | Tool | Output note |
|---|
| Feishu tasks + all-chat messages | lark-cli (read-only scopes) | 01_inbox/feishu/feishu-inbox-<date>.md + feishu-messages-<date>.md |
| WeChat self-saves | chatlog HTTP API | 01_inbox/wechat/wechat-inbox-<date>.md |
| Zen browser tabs | scripts/decode_zen_tabs.py | 01_inbox/tabs/zen-tabs-<date>.md |
| Active Claude sessions | list_sessions MCP | 01_inbox/projects/claude-projects.md (the AI Build Queue) |
<date> is today as YYYY-MM-DD. Always read the previous dated note first so you can frame the
pull as "what's new since last sync" rather than re-dumping everything.
Orchestration pattern (do it in this order)
- Snapshot first.
cd "$VAULT" && git add -A && git commit -m "snapshot before re-sync". This
gives a restore point before you write anything.
- Pre-flight: disk space. See "Disk pre-flight" — a full disk silently breaks everything,
including your own shell. Check before you start.
- Fan out the slow pulls. Feishu and WeChat each involve many API calls and per-conversation
summarization — launch them as parallel background subagents (one per source) using the prompts
in
references/sources.md, so they run while you do the fast local work.
- Do the fast local pulls inline while the agents run: decode Zen tabs (one script call) and
rebuild the AI Build Queue from
list_sessions.
- Fold in the agent results as they land. Read each agent's summary; don't re-read the full note
unless something looks off.
- Update the Command Center. Repoint the capture-source links in
homepage.md to the new dated
notes, and surface any genuinely actionable finding (a deadline, an escalation) onto Tasks Board.md
with an area tag + due date — don't let a real action item rot inside a digest.
- Commit + push.
git add -A && git commit -m "Re-sync all capture sources (<date>): ..." && git push.
If push fails with an HTTP/2 framing error, retry with git -c http.version=HTTP/1.1 push.
Scale to what was asked. "Re-sync everything" → all four. "What's new in Feishu" → just that one,
still following its source rules.
Disk pre-flight
A 100%-full disk on macOS makes the shell itself fail (ENOSPC: no space left on device) — commands
can't even write their output, so you'll be blocked before you start. If a trivial Bash call errors
with ENOSPC, the disk is the problem, not your command.
The user has Burrow for exactly this — its MCP tools run out-of-process so they work even when the
shell can't. Use burrow_doctor (read-only health) and burrow_analyze (size-ranked tree) to see the
situation, and burrow_clean to clear caches. Caches rarely free enough alone; the big wins are usually
VM images, app DB backups (e.g. ~/Library/Application Support/Dayflow/backups + timestamped
Dayflow.backup-* copies), and stale build dirs. Only auto-delete caches. For real user data — VMs,
project folders, app data — show the user the sizes and let them choose. Never nuke a directory you
didn't create without their say-so.
Per-source details
Each source has non-obvious auth/format traps. Read references/sources.md before pulling — it has the
exact commands, the subagent prompts, and the gotchas. The traps that bite hardest:
- Feishu — read-only scopes ONLY. Re-auth (only if needed) with
lark-cli auth login --scope task:task:read,im:message:readonly,im:chat:read,contact:user.base:readonly.
Never use --domain — that requests write scopes, and authorizing write scopes triggers a slow
enterprise admin review that has cost hours. A needs_refresh token auto-refreshes on the first valid
API call; only if that fails do you re-auth (you can drive the logged-in browser to click Authorize).
App/enterprise IDs and the dev-console URL are in local.private.md.
- WeChat — discover the self-chat at runtime. It's the chat flagged
isSelf in the chatlog session
list, NOT filehelper / 文件传输助手 (which is empty). The chatlog server must be running
(chatlog server, binary at ~/go/bin/chatlog) on http://127.0.0.1:5030. Quote URLs in curl — a bare
? gets eaten by zsh globbing and looks like a dead server when it isn't.
- Zen tabs — use the bundled script.
python3 scripts/decode_zen_tabs.py --out "$VAULT/01_inbox/tabs/zen-tabs-<date>.md" --date <date>.
It handles the mozlz4 decode with no dependencies. Don't pip install lz4 (PEP 668 blocks it).
- Claude sessions — dedupe by folder, flag status.
list_sessions returns ~85 with heavy duplication
(recurring "weekly sync" etc.) and many merged-PR sessions that are actually done. Group by cwd,
collapse dupes, flag ✅-done (merged PRs) as archive candidates, surface what's new since last sync.
Quality bars (what makes this "the same high quality")
- Dedupe and categorize, don't dump. 1,500 raw URLs or 250 raw messages is just moving the mess.
Bucket, summarize, count. The note should be skimmable in 30 seconds.
- Surface deadlines as tasks. A real commitment with a clock (a demo date, an escalating reminder)
belongs on the board with a 📅 due date, not buried in prose.
- Frame as a delta. "9 new since 06-23" beats re-listing 259 old items. Read the prior note first.
- Keep Chinese text as-is but make it scannable.
- Don't loop on Feishu auth. If read-only re-auth still fails, stop and tell the user — repeated
login attempts re-trigger the admin-review trap. One clean attempt, then escalate.
After a re-sync, give a tight recap: what each source surfaced, the count of new items, and the one or
two things that actually need attention this week.
