Agent doesn't execute tool calls from AGENTS.md startup instructions
Hi,
Goal + expected behaviour: I added a "Session Startup" section in AGENTS.md instructing my agent to use read, write, and exec tools at the start of every new session to manage daily journal files (check if today's journal exists → create it with write, archive yesterday's journal with exec mv). I expect the agent to execute these tool calls automatically before responding to the user.
What happened instead: The agent reads the instructions, acknowledges them, and responds as if it completed the tasks ("I just did my startup checks, everything is in order") — but no tool calls are actually made. The files remain unchanged. This happens consistently across multiple container restarts and new sessions.
What I ran / clicked:
- Added section in AGENTS.md with explicit tool names (read, write, exec)
- Restarted container with docker compose down && docker compose up -d
- Sent a fresh message to Telegram to start new session
- Verified filesystem: no files created or moved
- Confirmed tools are available: agent lists read, write, edit, exec among its tools
Environment:
OpenClaw version: 2026.2.23
Agent model: google/gemini-3-flash-preview
Install method: Docker (self-hosted VPS, Ubuntu)
Channel: Telegram
Tools profile: default (openclaw), no restrictions in config
tools section in openclaw.json: empty (all defaults)
Relevant logs:
[heartbeat] started
[gateway] agent model: google/gemini-3-flash-preview
[diagnostic] stuck session: sessionId=76fb12b6 state=processing age=146s
Is there a way to force tool execution from AGENTS.md instructions at session start? Or is there a better mechanism (hooks, cron, startup script) to run file operations automatically on each new session?
Thank you for your help
Goal + expected behaviour: I added a "Session Startup" section in AGENTS.md instructing my agent to use read, write, and exec tools at the start of every new session to manage daily journal files (check if today's journal exists → create it with write, archive yesterday's journal with exec mv). I expect the agent to execute these tool calls automatically before responding to the user.
What happened instead: The agent reads the instructions, acknowledges them, and responds as if it completed the tasks ("I just did my startup checks, everything is in order") — but no tool calls are actually made. The files remain unchanged. This happens consistently across multiple container restarts and new sessions.
What I ran / clicked:
- Added section in AGENTS.md with explicit tool names (read, write, exec)
- Restarted container with docker compose down && docker compose up -d
- Sent a fresh message to Telegram to start new session
- Verified filesystem: no files created or moved
- Confirmed tools are available: agent lists read, write, edit, exec among its tools
Environment:
OpenClaw version: 2026.2.23
Agent model: google/gemini-3-flash-preview
Install method: Docker (self-hosted VPS, Ubuntu)
Channel: Telegram
Tools profile: default (openclaw), no restrictions in config
tools section in openclaw.json: empty (all defaults)
Relevant logs:
[heartbeat] started
[gateway] agent model: google/gemini-3-flash-preview
[diagnostic] stuck session: sessionId=76fb12b6 state=processing age=146s
Is there a way to force tool execution from AGENTS.md instructions at session start? Or is there a better mechanism (hooks, cron, startup script) to run file operations automatically on each new session?
Thank you for your help
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,723Members
Resources
Similar Threads
Was this page helpful?
