With the same one-line prompt, Codex turns the job into a “mystery thriller”
The prompt is just one sentence:
Check local OneDrive and related process resource usage
Figures 1–3: Codex (plot-driven)
Codex starts off very seriously, getting ready to run PowerShell: Get-Process … *OneDrive* …
Then it immediately crashes: batch file arguments are invalid
Even funnier, it doesn’t believe it, and even tried echo hi—still the same error. (Fig. 1)
Then it enters “detective mode”:
- Suspects PowerShell / cmd got mixed up
- Suspects special characters caused parsing issues
- Suspects working directory/execpolicy
- Basically keeps revisiting the “crime scene” over and over (Fig. 1)
In the end, the conclusion is:
“I can’t run it on my side. You copy this huge chunk of commands below into PowerShell, run it, and paste the results back to me.” (Fig. 2/Fig. 3)
Which means:
You asked me to check resource usage, and it wrote an on-the-spot draft of a OneDrive Troubleshooting Manual—and I still did the work.
Figure 4: Others (tool-driven)
Others are very plain:
Run one command directly → output OneDrive’s PID/CPU/memory → and summarize along the way. (Fig. 4)
No mystery movie, no CSI, no “I suspect the universe.”
Same kind of issue: on Linux it also “hard-locks onto the literal”
I asked it to check local cliproxyapi logs, and Codex’s default worldview is:
If you say that name, then it should exist exactly with that name in the system.
If it can’t find it, it starts rummaging through system logs, and at one point even wanted to git clone something for me to see…
(Me: dude, I want to see logs, not watch you surf the web.)
If you switch to another CLI tool, it’s totally normal:
Can’t find an exact match → look for similar names/variants → check inside Docker → done.
My guess
Codex’s problem isn’t that it can’t write commands, but that:
It trusts too much that “the world is clean, names are exact, and the environment is textbook.”
So it really resembles the kind of person who:
- Is great at writing code
- But in real-world troubleshooting turns a 1-minute task into a 1-hour “reasoning flowchart”
Suggestions to round out Codex’s agent capabilities
- Let Codex write code
- Let Codex be “on-site ops”
- If you really want to use it for troubleshooting, the prompt has to be written like a runbook:
“If not found then fuzzy match/check variant names/check systemd/check Docker/list related processes…”
