Hi, (Agent mode leak)

[chilly swallow]

• I try talk to the openclaw tui.

• I get short nearly not related answers in some kind of meta mode (e.g.: explaind to user that nothing is actually broken; etc...).

• Chatgpt told me, that these kind of messages are typically for "Agent mode leak" (what is this? I can't find it).

• Restarting the server, the gateway, using the doctor - won't fix it.

• I landed here by, suddenly got in tui: "HTTP 400: Unsupported value: 'low' is not supported with the 'gpt-5.1-codex' model. Supported values are: 'medium'." - but my config.json always said "thinkingDefault": "medium"

• Doctor said, status, gateway... everything was fine. So i updated to OpenClaw 2026.5.12.

• Now i am here again. You can read from the top again if you want.

Some hints how to proceed?

[karmic tiger]

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

[vagrant owl]

"Agent mode leak" is not really an OpenClaw term. What you describe sounds more like the agent is stuck in bootstrap/session context, or an old session override is still being reused.

First thing: please share the exact output of:

openclaw --version
openclaw status --all
openclaw models status

A few concrete checks I’d do next:

# 1) Check for bootstrap mode in the active agent workspace
ls -la ~/.openclaw/workspace/BOOTSTRAP.md ~/.openclaw/agents/*/agent/BOOTSTRAP.md 2>/dev/null

[vagrant owl]

# 2) Check whether "low" is coming from config/model params/session state
rg -n '"low"|thinkingDefault|thinking|reasoningEffort|gpt-5.1-codex' ~/.openclaw/config.json ~/.openclaw/agents/*/sessions ~/.openclaw/agents/*/agent 2>/dev/null

# 3) Watch the actual gateway error while reproducing one bad TUI reply
openclaw logs --follow

Why: docs/source show thinkingDefault can be global, per-agent, or per-model via agents.defaults.models.<provider/model>.params.thinking, and session state can also carry overrides. So config.json saying global thinkingDefault: "medium" does not fully rule out another stored low.

If BOOTSTRAP.md exists, don’t delete it blindly; move it aside so it’s reversible, then restart the gateway and try a fresh TUI session:

mkdir -p ~/.openclaw/debug-backups
mv ~/.openclaw/workspace/BOOTSTRAP.md ~/.openclaw/debug-backups/BOOTSTRAP.md.$(date +%s)
openclaw gateway restart

[vagrant owl]

For the TUI weird-context part, also try starting a fresh session with /new or /reset in the TUI after the restart.

Docs for the debug loop: https://docs.openclaw.ai/help/faq and session details: https://docs.openclaw.ai/reference/session-management-compaction

[chilly swallow]

Thank you my rusty crusty crab. /new did it. it was the tui session which was lost. I wonder what is was. - Some kind of bootstrap process, yes but what? Also, i'll miss the context of the session. A fuzzy learning for me. Anyhow thx.

[vagrant owl]

Glad /new fixed it 🦞

My best read: the TUI session state got into a bad contextual lane, not that the model itself was broken. OpenClaw sessions keep transcript/routing/model-ish metadata in the session store, so if an earlier failed turn, bootstrap-pending context, or bad reasoning/thinking override got persisted, the next turns can keep inheriting that weird framing until the session is reset.

The annoying tradeoff is exactly what you found: /new gives you a clean session, but drops the old working context. If that context matters, the practical recovery path is usually: export/copy the important bits from the old session, start /new, then paste a short summary back in.

If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution. That helps Answer Overflow index the fix for the next poor soul with a lost TUI session.