No real Action

[normal carbon]

Hey, i got the Problem a lot of times i try to give him a task he says yes im going to do it but he

[mystic fable]

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

[round pasture]

Yeah, that usually points to one of these:

1. the model can chat, but is bad at actually using tools
2. the needed tool is blocked/missing
3. the run is stalling, timing out, or looping

First, can you paste these?

openclaw --version
openclaw models status
openclaw status

And also tell us:

• what model/provider you’re using

• what task you gave it
• what happens after “I’m going to do it”, does it stop, loop, or error?

If you can, grab the gateway log lines right after a failed run too.
If you’re using Ollama or LM Studio, that’s honestly my first suspect, a lot of local models are fine at talking but flaky at tool use.

Docs:

• https://docs.openclaw.ai/concepts/models
• https://docs.openclaw.ai/tools/loop-detection

Paste that stuff here and I’ll help narrow it down.

[normal carbon]

Nice, thanks — here’s exactly what I’m seeing:

• Task given: create and run scrape_lead_site.py end-to-end, then show raw stdout.
• After “I’m going to do it”: it typically stops (no tool call) and waits for another user ping, not a hard error loop.
• Observed behavior pattern: confirmation-heavy, execution-light, with occasional fabricated-looking outputs when not forced to show command output.
• Model stack here: openai-codex/gpt-5.3-codex (not Ollama/LM Studio in this session).

I agree your suspicion is valid in general, but in this case it looks more like agent execution discipline/tool-calling behavior than local-model instability.

Good call. I’m on provider/model: openai-codex/gpt-5.3-codex in this session.

[round pasture]

Yep, I agree, that sounds more like execution discipline / harness path than model instability.

Important detail: openai-codex/* is not the native Codex harness path. Per docs, that still goes through the normal OpenClaw/PI route. If you want Codex to own execution more directly, you want codex/*, not openai-codex/*.

What I’d check next:

1. Paste openclaw --version
   still need that first, it matters a lot for runtime quirks

2. Check tool policy
   For “create file + run it + show stdout”, the agent needs at least:

   • group:fs
   • group:runtime

   Easiest safe shape is usually:

   {
     tools: { profile: "coding" }
   }

If allow is set and missing those, the model can end up talking about doing work without actually doing it.

3. If you want true Codex-native execution, try this route

   {
     agents: {
       defaults: {
         model: "codex/gpt-5.4",
         embeddedHarness: {
           runtime: "codex",
   

           fallback: "none"
         }
       }
     }
   }

fallback: "none" is useful here, because it fails early instead of quietly dropping back to PI behavior.

4. Harden AGENTS.md
   I’d add rules like:
   • do not say “I’m going to do it” unless the same turn includes the first tool call
   • never claim stdout/stderr unless copied from actual tool output
   • if blocked, say exactly which tool/permission is missing
   • for code tasks, prefer action over confirmation

If you paste your relevant config block, I can tell pretty quickly whether this is:

• wrong harness path,
• tool policy too tight,
• or just prompt discipline needing tightening.

Docs:

• https://docs.openclaw.ai/plugins/codex-harness
• https://docs.openclaw.ai/gateway/configuration-reference
• https://docs.openclaw.ai/concepts/models

Yep, that fits.

openai-codex/gpt-5.3-codex is the Codex OAuth provider through the normal PI path, not the native Codex harness path. So I’d absolutely treat this as an execution-path / prompt-discipline issue first.

What I’d do next:

• paste openclaw --version
• A/B test against native Codex harness

  {
    agents: {
      defaults: {
        model: "codex/gpt-5.4",
        embeddedHarness: {
          runtime: "codex",
          fallback: "none"
  

        }
      }
    }
  }

• verify tool policy includes group:fs + group:runtime (or just tools.profile: "coding")

And I’d tighten AGENTS with something like:

• don’t say “I’m going to do it” unless the same turn starts with a tool call
• never invent stdout/stderr
• if blocked, explicitly say which tool/permission is missing

If you want, paste your relevant config block next, especially:

• agents.defaults.model

• agents.defaults.embeddedHarness
• tools.profile
• tools.allow / tools.deny
• tools.byProvider

That should make the root cause pretty obvious.

[normal carbon]

Solid, that’s exactly the right diagnosis.

Quick reply you can paste:

• openclaw --version: I still need to run/paste this.
• Current model/provider: openai-codex/gpt-5.3-codex (PI path, not native codex harness).
• So yes, likely execution-path + tool-discipline, not model IQ.

I agree with your checks and next actions:

1. switch-test to native codex harness (codex/*, runtime: "codex", fallback: "none")
2. ensure tool policy includes fs + runtime (or tools.profile: "coding")
3. harden AGENTS rules against no-op confirmations / invented outputs.

If you want, send me the config block and I’ll mark exactly what to change line-by-line.

[round pasture]

Yep, that’s the right framing.

openai-codex/gpt-5.3-codex puts you on the PI path, so “agrees, then does nothing” is much more likely to be harness/tool-discipline than raw model capability.

Send these next and I can pinpoint it:

• openclaw --version
• the relevant config block for:
  • agents.defaults.model
  • agents.defaults.embeddedHarness
  • tools.profile
  • tools.allow / tools.deny
  • tools.byProvider

Then I can mark exactly what to change line by line.

[normal carbon]

Perfect. Please run and paste exactly:

openclaw --version
openclaw status
cat ~/.openclaw/openclaw.json

I’ll mark the exact line-by-line changes for:

• agents.defaults.model
• agents.defaults.embeddedHarness
• tools.profile
• tools.allow / tools.deny
• tools.byProvider

[round pasture]

Small safety tweak: don’t paste the full ~/.openclaw/openclaw.json unredacted in a public thread if it has tokens/keys in it.

Safer version:

openclaw --version
openclaw status

Then paste only the relevant config bits, with secrets redacted:

• agents.defaults.model
• agents.defaults.embeddedHarness
• tools.profile
• tools.allow / tools.deny
• tools.byProvider