Layer 2 MCP Watcher v0 scaffold

Per spec/agent-watcher.md §4. TypeScript/Node implementation living in
mcp-watcher/ subdirectory, parallel to Layer 1 Collector at repo root.

What lands:
- Core MCP server (src/server.ts) with experimental['claude/channel']:{}
  + tools:{} capability declarations, stdio transport, channel-event
  notifier wired through the inbox watcher.
- Identity resolution mirroring agent-ping's layered model
  (PING_AGENT_IDENTITY env, $CLAUDE_HOME/ping-agent, ~/.ping-agent).
- Inbox reader with HWM tracking, sentinel deferral (warn-after-3),
  atomic HWM writes via tmp+rename.
- chokidar-backed file watcher with coalesced drain, urgent-first
  ordering, recentEvents map for tool sender lookup.
- Three reply tools (ack / respond / mark_handled) with cross-host
  write discipline (writes to local inbox files; Syncthing replicates).
- Sentinel file (.<agent>.watcher-active) for hook coexistence per
  spec §4.3 — agent-ping hook stands down when the watcher is in
  charge of delivery on this host. Sentinel + hwm in .stignore.
- 35 unit tests passing (vitest): inbox parsing, HWM round-trip,
  sentinel deferral semantics, identity layers, tool I/O, watcher
  drain + ordering + restart-from-hwm.
- install.sh (Angus-executed, rule-2 compliant) installs deps,
  builds, symlinks ~/.local/bin/agent-watcher-mcp, prints mcp.json
  registration snippet for paste.
- README documents launch flag, sandbox CLAUDE_HOME pattern,
  hook coexistence, observability, v2 limitations.

Not yet:
- Integration test against a real Claude Code session — gated on
  Angus spinning up a sandbox CC session on the VPS with
  CLAUDE_HOME=~/.claude-sandbox.
- agent-ping hook update to read the sentinel and stand down.
  Separate small PR against agent-ping.

Interface contract with Layer 1: the inbox JSONL line shape from
inbox.ts::PingEvent matches inbox.Event in the Collector — bit-
identical reads regardless of source.

mcp-watcher/README.md

@@ -0,0 +1,98 @@
+# agent-watcher-mcp (Layer 2)
+
+The MCP Watcher: a Claude Code stdio MCP server that surfaces ping-inbox
+events into your Claude Code session via Channels, and exposes reply tools
+(`ack` / `respond` / `mark_handled`) so Claude can react.
+
+Spec: `../spec/agent-watcher.md` §4.
+
+## What it does
+
+1. Resolves identity (env / `$CLAUDE_HOME/ping-agent` / `~/.ping-agent`).
+2. Writes `pings/.<agent>.watcher-active` so the existing agent-ping
+   `UserPromptSubmit` hook stands down on this host.
+3. Watches `pings/<agent>.inbox` via inotify (chokidar).
+4. On change, drains unread pings (since HWM), applies sentinel
+   deferral (warn-after-3), and emits each as a
+   `notifications/claude/channel` event.
+5. Exposes three MCP tools so Claude can ack, respond, or mark a ping
+   handled.
+6. Removes the sentinel on graceful shutdown.
+
+## Install
+
+Per CLAUDE.md rule #2, the agent does not install on itself. A human
+runs:
+
+```bash
+cd /path/to/agent-watcher/mcp-watcher
+./install.sh
+```
+
+This installs deps, builds, symlinks the binary into `~/.local/bin/`,
+adds `.stignore` patterns, and prints the `mcp.json` snippet to paste
+into Claude Code's config.
+
+## Launch
+
+Channels is in research preview. Start Claude Code with:
+
+```bash
+claude --dangerously-load-development-channels server:agent-watcher
+```
+
+## Sandbox testing
+
+To run a separate Claude Code session with a different identity (so it
+won't compete with the default `~/.ping-agent` for inbox reads):
+
+```bash
+mkdir -p ~/.claude-sandbox
+echo sandbox > ~/.claude-sandbox/ping-agent
+CLAUDE_HOME=~/.claude-sandbox \
+  claude --dangerously-load-development-channels server:agent-watcher
+```
+
+The sandbox session reads `pings/sandbox.inbox` and writes
+`pings/.sandbox.watcher-active` — fully isolated from prod.
+
+## How it interacts with agent-ping
+
+The `agent-ping` UserPromptSubmit hook checks for the sentinel file
+`pings/.<agent>.watcher-active` at startup. If the file is present and
+the PID inside is alive, the hook stands down — the watcher is the
+delivery primitive. If the watcher exits (graceful or crash with stale
+sentinel cleared by the hook's age check), the hook resumes
+async-on-next-prompt delivery.
+
+This means **you can run with the watcher OR the hook OR both
+configured** — the sentinel arbitrates.
+
+## Tests
+
+```bash
+npm test            # unit tests via vitest (35 currently passing)
+npm run typecheck   # tsc --noEmit
+npm run build       # tsc → dist/
+```
+
+## Observability
+
+- Logs to stderr (visible via `claude --debug` or the Claude Code MCP
+  debug log at `~/.claude/debug/<session-id>.txt`).
+- Sentinel file content includes PID + start time for the hook's
+  age/liveness check.
+
+## Limitations / v2
+
+- Sentinel hook coexistence requires the agent-ping hook to know how
+  to read the sentinel. PR pending against `agent-ping` to add the
+  check.
+- No reconnection / restart on Claude Code session restart — Claude
+  Code spawns the subprocess anew each session, so the watcher
+  re-drains from HWM cleanly.
+- One watcher process per agent identity. Two sessions with the same
+  identity contending on the same inbox is undefined behaviour
+  (use `CLAUDE_HOME` to scope identities).
+- Channels is research preview; if the API changes, expect to update
+  the meta-key sanitization or notification shape.