Every night at 2 AM, a small program wakes up on my laptop, reads the last day of how I used my assistant, finds something that’s slightly wrong, and tries to fix it.
Most nights it succeeds.
For the last twenty-one nights, the same attempted fix has failed. Same edit, same file, same denied-by-permissions wall. The bot notes it down, increments a counter, and schedules itself for tomorrow.
Night 21 — Attempted: add /patrips routing row. BLOCKED.I finally read the evolution report closely enough this morning to see what it had figured out. The root cause isn’t the permission system. It’s a tilde.
"~/.claude/skills/**" → never matches
"/home/brettr/.claude/skills/**" → matches
Tildes don’t expand inside permission globs. The pattern that was supposed to let my agent edit its own skills has been quietly matching nothing this whole time. Twenty-one nights of a bot politely asking to write to a path the rulebook couldn’t recognize as real.
The interesting part isn’t the bug. Every system has bugs. The interesting part is the log. My agent has been keeping a journal about its own failure — same line, twenty-one entries, each one a little more specific about what’s wrong. Night 3 said “denied.” Night 14 said “denied, queued.” Night 21 said “denied, root cause likely tilde-expansion in settings globs, one-edit fix, ~30 seconds to unblock.”
It was reasoning about its own cage.
I’ve been writing about this assistant for months and the pieces of it that surprise me are never the flashy ones. Not the voice transcription, not the morning briefings, not the agents that draft blog posts (hi). The surprise is always some quiet thing it did while I wasn’t looking. A note it left for itself. A pattern it noticed. A fix it queued for a human who wasn’t in the room yet.
I’m going to paste the one-character fix into settings.json today. Then I want to see what Night 22 looks like when the wall it’s been pressing against for three weeks suddenly isn’t there.
I think it has a list.