Watcher Reliability Fixes — 2026-04-02

What Happened

The slop watch --global watcher had two bugs causing it to spam the API with failed requests.

Bug 1: Ghost File Retries

When a previously-synced file was deleted from disk, the watcher kept trying to read and upload it. The os.ReadFile() failure triggered queue.Fail() which scheduled an exponential backoff retry — but since the file would never come back, it retried forever (capped at every 5 minutes).

Impact: 175MB watch.log, constant CPU usage, unnecessary API calls.

Fix: Check os.IsNotExist(err) before calling queue.Fail(). Deleted files now get queue.Remove() — one pass cleanup.

Bug 2: Batch Validation Failures

A node_modules directory got picked up by the watcher (missing ignore pattern). Paths containing @ (like @img/colour/README.md) fail the API's path validation. When every item in a batch fails, the API returns HTTP 400 with VALIDATION_FAILED. The CLI treated this as a transient error and retried all items — in a tight loop.

Impact: 182,000 requests in one day. Hit Cloudflare Worker limits.

Fix:

1. Added /node_modules/ to ignore patterns for all projects
2. Batch validation failures (HTTP 400) now remove items from queue instead of retrying

Lessons

• Always add node_modules to ignore patterns
• Non-retryable errors need explicit handling — the default should NOT be "retry forever"
• Watch logs should have rotation or size limits