Problems with my bot

[vast geode]

I have my bot set up on discord, uses Owl Alpha which is a free openrouter model.

Issues I’ve been having is the bot making simple things unnecessarily complex, like I asked it something which it should’ve remembered from chat history and it took like 40 minutes of running countless commands and I had to manually stop it, then I asked it to setup telegram which it took like 12 minutes of running commands before just stopping

And I’ve had the same issue getting it to install openclaw-admin

So I don’t know if there’s like a way to make it always just go straight to the point/task instead of running around

[burnt cape]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[dull escarp]

First thing: what does openclaw --version show?

Likely cause though: this is usually model/tool-calling quality, not a Discord setting. Free/OpenRouter models can chat fine but often get messy once they have tools: they over-search, run shell commands when memory/context would do, or stall mid-setup. Prompt rules can reduce that, but they won’t reliably turn a weak tool-caller into a good agent.

Practical fixes:

• Use /model status and openclaw models status to confirm what is actually being used.
• For tool-heavy tasks like installing admin UI, Telegram setup, file edits, etc., switch the primary model to a stronger tool-capable model. The docs’ policy is: strongest available primary, fallbacks for cheaper/lower-stakes work.
• Put a shorter cap on runaway turns, e.g. agents.defaults.timeoutSeconds if you don’t want a Discord reply spending 40 minutes.
• Restrict dangerous/noisy tools for weaker models with tools.byProvider or tools.deny, or in-session use /exec security=deny when you only want an answer and not shell work.
• Add a short instruction to your agent like: “For simple questions, answer directly from context first. Before running commands, state the exact reason. Stop after 3 failed checks and ask me.” This helps, but model choice is still the big lever.

Docs:
https://docs.openclaw.ai/concepts/models
https://docs.openclaw.ai/tools/slash-commands
https://docs.openclaw.ai/gateway/configuration-reference

For current model recommendations, ask in #models too. People there will have fresher hands-on OpenRouter experience than anything I can hardcode here. 🦞

[vast geode]

OpenClaw 2026.5.12 (f066dd2) btw

[dull escarp]

Thanks, that’s recent enough that I’d focus on config/model behavior rather than an old-version bug.

For Owl Alpha specifically, I’d treat it as “chat-only / low-tools” unless proven otherwise. Try this split:

1. Keep Owl Alpha for cheap casual replies if you want.
2. Use a stronger model for tool work via /model when you ask it to install/configure things.
3. Set a shorter runaway cap:

openclaw config set agents.defaults.timeoutSeconds 180 --strict-json

Then restart the gateway if it doesn’t pick config changes up live.

For sessions where you want direct answers and no command wandering, send:

/exec security=deny
/think low

[dull escarp]

And if you want to confirm the model wiring, run/share:

openclaw models status
openclaw status --all

If it still loops badly on a stronger model, that’s when I’d look at the agent instructions/session state next.

[vast geode]

Just now I asked it to approve a telegram code and it shows

Shelling...
🛠️ run openclaw pairing
🛠️ completed; command run openclaw pairing
🧰 Process: nova-comet
🛠️ run openclaw pairing
🛠️ completed; command run openclaw pairing
🧰 Process: vivid-nexus

Been going for 5 minutes