Vault Change Feed

Give your AI agents a changelog of your vault. Every edit you make is recorded as a machine-readable event feed with per-reader cursors — so any AI can catch up on what changed since its last visit, instead of blindly rescanning thousands of notes.

中文文档 · Obsidian Community Listing

Support / 支持

If this plugin saves you time, you can buy me a coffee — it keeps the development going. 如果这个插件帮你省了时间，可以请我喝杯咖啡 ☕

• International:
• 中国大陆: 爱发电（微信 / 支付宝直达 · WeChat / Alipay）

---

Features

• Live change tracking — create / modify / delete / rename events with line-level diff stats (+added / −removed)
• Offline backfill — startup reconciliation catches edits made while Obsidian was closed (phone, iCloud sync, CLI tools); content-hash rename detection included
• Incremental AI reads — each reader (AI agent) keeps its own cursor and pulls only what's new; per-file merge on read collapses edit bursts into one cumulative line
• Self-describing to agents — a reading-protocol block is auto-installed into AGENTS.md / CLAUDE.md / GEMINI.md on first run, so coding agents discover the feed with zero setup
• Rotation + stale signal — the log is capped (90 days / 50k entries by default); readers are told explicitly when they must do a full rescan
• Fully local — no network, no telemetry, works on desktop and mobile

The problem

Your AI assistant has no idea what you edited between sessions. Scanning the whole vault every time is expensive; not scanning means it works from stale knowledge. This plugin continuously records which file changed, how, and by how much — the AI pulls that incrementally, on demand.

Installation

Community market (recommended): Settings → Community plugins → Browse → search Vault Change Feed → Install → Enable. The AI protocol block installs itself on first run.

BRAT: add kains2866/vault-change-feed as a beta plugin.

Manual: copy main.js and manifest.json from the latest release into <vault>/.obsidian/plugins/vault-change-feed/, then enable the plugin.

How it works

• Live: listens to Obsidian's create / modify / delete / rename events and computes line-level diff stats
• On startup: reconciles against the last baseline snapshot to backfill changes made while Obsidian was closed. A delete+create pair with identical content hash is reported as a rename (binary files re-downloaded by iCloud get a fresh mtime and may degrade to delete+create)
• All data stays local, under <configDir>/plugins/vault-change-feed/:
  • changelog.jsonl — the event stream, one JSON event per line
  • cursors.json — per-reader read cursors
  • baseline.gz — content baseline snapshot (for diffs and reconciliation)

Note: the plugin has built-in standby protection — concurrent instances coordinate through a heartbeat writer lock (writer.lock); only the lock holder records, other instances stand by (read-only API still works) and automatically take over once the lock goes stale (90 s). Still, prefer running it in only one Obsidian instance per vault at a time.

Event format

{"seq": 1284, "ts": 1785000000000, "op": "modify", "path": "ML/overfitting.md", "stat": {"added": 12, "removed": 3}, "source": "live"}

• op: create / modify / delete / rename (carries oldPath) / resync (baseline rebuilt — full rescan advised)
• stat: {added, removed} line counts; null means "changed, magnitude unknown — open the file" (binaries, oversized files)
• source: live / reconcile (startup backfill) / system

Letting AI agents discover the feed

AI agents don't know the feed exists out of the box. On first enable, the plugin automatically installs a reading-protocol block (wrapped in <!-- vault-change-feed:start/end --> markers) into the vault-root AGENTS.md, CLAUDE.md, and GEMINI.md — the conventional discovery points for coding agents (AGENTS.md is the cross-tool standard; CLAUDE.md for Claude Code; GEMINI.md for Gemini CLI, which doesn't read AGENTS.md by default). Zero clicks needed.

• Opt-out: disable Auto-install AI protocol on first run to fall back to a one-time notice; the Install AI protocol for agents / Remove AI protocol from agent files commands remain available
• Idempotent: re-runs update only the marked block; your own content outside the markers is preserved verbatim
• Refreshes with the plugin: after updates, installed blocks are refreshed automatically — but auto-sync only touches files that already have a block, it never creates new ones (can be disabled)
• Clean removal: the remove command strips the block from all three files, deleting a file only if nothing else remains
• Per-file toggles: Sync AGENTS.md / Sync CLAUDE.md / Sync GEMINI.md
• AIs without filesystem access (plain web chats) use the Copy unread changes for AI command instead

Optional: SessionStart hook (enforced, not "please read")

The protocol block relies on the AI choosing to read it. extras/vault-feed-hook.mjs goes further: hooked into an agent's SessionStart event, it injects merged unread changes into the context automatically and advances the cursor. The script walks up from the session cwd to find a tracked vault, and stays completely silent outside a vault or when nothing is unread.

Kimi Code (~/.kimi-code/config.toml):

[[hooks]]
event = "SessionStart"
command = "node /path/to/extras/vault-feed-hook.mjs --reader=kimi-code --format=kimi"
timeout = 10

Claude Code (hooks.SessionStart in ~/.claude/settings.json):

{ "type": "command", "command": "node /path/to/extras/vault-feed-hook.mjs --reader=claude-code --format=claude", "timeout": 10 }

--reader is the agent's stable cursor name. Use an absolute node path.

Protocol for AI agents (this README is the interface doc)

This vault is tracked by the vault-change-feed Obsidian plugin. Before editing notes, catch up on what the user changed since your last visit:

1. Read .obsidian/plugins/vault-change-feed/cursors.json and find your reader name (use your agent id, e.g. "kimi-cli"; absent means cursor 0).
2. Read .obsidian/plugins/vault-change-feed/changelog.jsonl (one JSON event per line) and take events with seq greater than your cursor.
   • If your cursor is > 0 and the smallest seq in the file is greater than cursor + 1, the log was rotated and you missed events — stop and do a full vault rescan instead.
   • If you see an event with op: "resync", the plugin rebuilt its baseline — a full rescan is advised.
3. Event shape: {"seq", "ts", "op": "create"|"modify"|"delete"|"rename"|"resync", "path", "oldPath"?, "stat": {"added", "removed"} | null, "source"}. stat: null means "changed, magnitude unknown — open the file if you care".
4. After processing, write the largest seq you saw back to cursors.json under your reader name. Write atomically: write cursors.json.tmp, then rename it to cursors.json.

Inside Obsidian, other plugins/scripts can use the JS API instead of files:

const api = app.plugins.plugins['vault-change-feed'].api;
const { events, stale, latestSeq } = await api.getChanges('my-plugin');
// ...handle events...
await api.markRead('my-plugin', latestSeq);

The JS API's getChanges merges unread events per file by default (api.getChanges(name, { merge: false }) returns the raw stream; groups that can't be merged losslessly are passed through, see below). External agents reading changelog.jsonl directly see the raw event stream and may implement the same merging:

• Sort by seq first (the log may be out of order), then group by path (resync is never merged); merged events take the group's max seq/ts and the last event's source
• Created and deleted within the window → group dropped; last event is delete and the group contains a rename → delete on the first rename's oldPath (no oldPath field, stat from the delete itself); last event is delete → delete (stat from the last delete)
• Deletes and renames interleaved beyond the case above → not merged, group emitted as-is (any merge would lose some path's fate)
• Deleted then re-created → modify (stat null); starts with create → create; contains a rename → rename (keeps the first rename's oldPath); otherwise → modify
• For the last three, stat is the per-line sum if all entries are non-null, else null; output sorted by merged seq

Commands

• Copy unread changes for AI — copies a compact summary of unread changes (reader manual) to the clipboard, ready to paste into any AI chat
• Install AI protocol for agents / Remove AI protocol from agent files — manage the discovery blocks in AGENTS.md / CLAUDE.md / GEMINI.md

Settings

Setting	Default	Description
Tracked text extensions	md, markdown, txt, canvas, json, csv	These extensions get diff stats
Exclude globs	empty	Extra exclusion rules; the vault config folder is always excluded
Large file threshold	1024 KB	Larger files get stat: null
Baseline content budget	102400 KB (100 MB)	Text kept in memory for diffing; beyond the budget, files store hash only (stat falls back to null)
Retention days / max entries	90 / 50000	Log rotation, whichever limit hits first
Baseline flush interval	300 s	Baseline persistence period
Auto-install AI protocol on first run	on	Install the protocol block on first enable
Sync AGENTS.md / CLAUDE.md / GEMINI.md	on	Per-file install targets
Auto-sync protocol block	on	Refresh installed blocks after plugin updates

Privacy

Fully local: no network calls, no uploads, no data collection. Everything lives in your own vault.

Development

npm install
npm run build   # typecheck + bundle main.js
npm test        # vitest

Works on desktop and mobile (all file operations go through the Obsidian vault API).

License

MIT © tiyukains