How OpenClaw injects workspace files into the runtime system prompt
Hey @Krill 🦐 — quick question about OpenClaw’s system prompt behavior.
Background: I understand OpenClaw generates the runtime system prompt on every agent run and injects workspace “bootstrap” files (SOUL.md, AGENTS.md, USER.md, etc.) into that prompt. I want to be sure I’ve got the mechanics right and confirm the best way to make a persistent persona apply across model switches.
Questions:
1. Confirm mechanics: Does OpenClaw rebuild the full system prompt at each session start by concatenating a generated header (tools list, safety notes, runtime info) with the workspace bootstrap files, or is any part cached between runs? If cached, where/how?
2. Scope: If I update SOUL.md or AGENTS.md in the agent workspace, will those changes be applied to new sessions immediately, or is a restart/reload required? Any delay or extra step to force a refresh?
3. Cross-model application: If the agent (main:main) is configured with those workspace files, will the generated system prompt be injected for any model the agent switches to at runtime (e.g., switching from Kimi to another api model or to a local Qwen,or do some model backends require applying the system prompt separately?
4. Best practice for consistency with model switching?:
• Editing SOUL.md/AGENTS.md directly, or
• Adding a separate persona file (e.g., ~/.openclaw/prompts/kimi_system.txt) and referencing it in the agent bootstrap so it’s explicitly injected?
Any preferred approach for maintainability and clarity?
5. Verification tools: What’s the canonical way to inspect the exact prompt OpenClaw injects for a running session? Is /context list and /context detail the full truth, or is there a more direct dump (logs or CLI) that shows the final assembled system prompt?
6. Security/hosting note: If using a remote Gateway, are the workspace bootstrap files pulled from the gateway host only (i.e., editing local ~/.openclaw/workspace won’t affect the gateway)? Any gotchas there?
Background: I understand OpenClaw generates the runtime system prompt on every agent run and injects workspace “bootstrap” files (SOUL.md, AGENTS.md, USER.md, etc.) into that prompt. I want to be sure I’ve got the mechanics right and confirm the best way to make a persistent persona apply across model switches.
Questions:
1. Confirm mechanics: Does OpenClaw rebuild the full system prompt at each session start by concatenating a generated header (tools list, safety notes, runtime info) with the workspace bootstrap files, or is any part cached between runs? If cached, where/how?
2. Scope: If I update SOUL.md or AGENTS.md in the agent workspace, will those changes be applied to new sessions immediately, or is a restart/reload required? Any delay or extra step to force a refresh?
3. Cross-model application: If the agent (main:main) is configured with those workspace files, will the generated system prompt be injected for any model the agent switches to at runtime (e.g., switching from Kimi to another api model or to a local Qwen,or do some model backends require applying the system prompt separately?
4. Best practice for consistency with model switching?:
• Editing SOUL.md/AGENTS.md directly, or
• Adding a separate persona file (e.g., ~/.openclaw/prompts/kimi_system.txt) and referencing it in the agent bootstrap so it’s explicitly injected?
Any preferred approach for maintainability and clarity?
5. Verification tools: What’s the canonical way to inspect the exact prompt OpenClaw injects for a running session? Is /context list and /context detail the full truth, or is there a more direct dump (logs or CLI) that shows the final assembled system prompt?
6. Security/hosting note: If using a remote Gateway, are the workspace bootstrap files pulled from the gateway host only (i.e., editing local ~/.openclaw/workspace won’t affect the gateway)? Any gotchas there?
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,827Members
Resources
Was this page helpful?
