Browse every SikiT article in reverse chronological order, or use search and the topic shortcuts below.
Runtime & failure · Delegation & permissions · Models & tooling · Open questions
-
Three times a measurement was wrong and the system looked fine
An audit reported 148 empty sections that did not exist. A second check counted 100 items when there were 128. A third deleted records it had proved were…
-
When an external write ends ambiguously, query instead of retrying
Two records one second apart with identical content digests. The first write succeeded and the caller could not tell, so it wrote again and the server helpfully renamed…
-
Give your evidence an expiry date before an agent acts on it
Work prepared on Monday and executed on Thursday runs on Monday’s facts. A guard that re-checks the evidence before the action, and demotes rather than edits when the…
-
What a completion receipt needs before it can catch a false claim
An agent’s summary is not evidence. Four fields turn a log line into a receipt that can be checked without the agent’s cooperation, and a fifth catches the…
-
Your agent timeout kills the agent and leaves its grandchildren
A timeout that kills the agent PID, or the PID plus its direct children, leaves anything one level deeper still running. Four measured variants on the same three-level…
-
Exit code 0 from an agent that did nothing
Three permission modes, one task, identical exit status. Two of the three ended the turn after stating intent and never ran anything, and the calling script had no…
-
My read-only agent wrote to disk, and the sandbox flag was set
A delegated agent run with a read-only sandbox created a file anyway. The sandbox was set correctly; a separate permission flag on the same command line silently outranked…
-
Why disown and nohup did not keep a backgrounded agent alive
A backgrounded agent CLI dies when its terminal goes away. disown does not stop it, and nohup only helps until something re-arms the handler. Measured on macOS with…
-
Why an idempotency guard let an agent create 122 duplicate uploads
A scheduled agent uploaded the same two images 122 times past a reuse guard built to prevent it. The guard compared filenames, and the server renames colliding uploads,…
-
When an AI rule changes, who gets the correction?
A four-way change test and five continuity duties for deciding when an AI correction should reach people affected by an earlier version.
-
When does a local AI answer measure the model?
Use a four-layer test to tell whether a local AI failure belongs to the model, runtime, application, or host.
-
Exit, Ctrl+C, or taskkill: stop the right Windows command
Learn how to distinguish Cmd.exe, a batch script, and a running console process, choose the least destructive stop control, and verify the target ended.