A month of parallel Claude Code sessions: eight failures that all looked like success

Automatic_Low_5034 · reddit · 2026-09-10

The author runs several Claude Code sessions side by side on one Mac with custom tooling, and after a month documented eight failure modes that produced no error, no red test, and no log line — each indistinguishable from normal operation, which is why they survived so long:

Watcher blind spot: Claude Code rewrites session state into the same /.claude/sessions/<pid>. file, so a filename-set watcher sees nothing on state changes. Worse, both the filesystem event and the polling fallback funnel into the same comparison — "two independent detectors, one shared blind spot". Fix: include mtime in the snapshot. Lesson: verify a fallback actually diverges from the primary before counting it as redundancy.
A falling number isn't proof of staleness: when reconciling ratelimits (fivehour/sevenday usage and resets) from multiple sessions, the rule "ignore any decrease" discarded every genuine window reset. Use resetsat as the arbiter: a later reset time means a new window where a lower percentage is correct.
Each session reports its own stale snapshot: an idle-for-two-hours session hands you two-hour-old numbers; mixing reads looks like wild limit fluctuation but is really different sessions' stale views.
Hook writes must be atomic and self-cleaning: a killed Stop hook accumulates zero-byte temp files; write to temp + mv, trap on EXIT/INT/TERM/HUP, and sweep leftovers older than an hour (a naive sweep deletes live sessions' files).
Notifications that don't say what finished stop helping at three sessions: six sessions, six identical bells, zero information. Make the Stop hook carry a one-line conclusion ("migration done, tests pass", "waiting on your call about encoding") so you can act without opening anything.
A graceful fallback is a failure-hiding device (text truncated mid-point).

Core theme: the real danger isn't explicit errors but silent failures that look exactly like success.

Original post →

A month of parallel Claude Code sessions: eight failures that all looked like success

More from coding & agent