how do i never get this error dude, its so repetitive

[knotty saddle]

[5/26/2026 7:08 AM] Openclaw: That went to /restaurant/ not /grocery/. Let me try other major stores.
[5/26/2026 7:09 AM] Openclaw: ⚠️ Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying.

i use telegram and my version is 5.6

[weak tinsel]

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

[prime ruin]

That warning is usually not a Telegram-specific failure. It means the agent run ended without a user-visible final response after it had already used a tool, so OpenClaw is warning you not to blindly retry because something may already have happened.

To reduce it instead of just chasing the same popup:

openclaw --version
openclaw status --all
openclaw models status
openclaw doctor --fix
openclaw channels status --probe
openclaw logs --follow

Then reproduce it once and look in the logs right before the warning. The useful line is usually a provider/model error, timeout, malformed tool call, auth/rate-limit, or context/session issue. If you paste the redacted openclaw status --all plus the few log lines around the failure, people can tell which one it is.

A few practical fixes that often stop the repeat loop:

[prime ruin]

• update first if you are on 2026.5.6 / “5.6”; these run-completion edge cases get fixed pretty often
• use a tool-reliable cloud model for Telegram tasks that browse/click/send/write; weak local/OpenAI-compatible models commonly fail after tool calls
• start a fresh Telegram session if the thread is long or has several failed retries
• after seeing “some tool actions may have already been executed”, verify the external state before saying “try again”

Docs for the exact triage flow:
https://docs.openclaw.ai/help/faq#first-60-seconds-if-something-is-broken
https://docs.openclaw.ai/channels/troubleshooting
https://docs.openclaw.ai/concepts/models

[knotty saddle]

so your saying i should update?

[prime ruin]

Yes, I would update first. 2026.5.6 is old enough that you may be hitting something already fixed, and this warning is in the agent-run completion path, not Telegram itself.

Use the supervised updater:

openclaw update
openclaw doctor --fix
openclaw gateway restart
openclaw --version

Then try the same Telegram task once. If it still happens, grab the few log lines right before the warning with:

openclaw logs --follow

[prime ruin]

Updating may not be the whole fix, but it is the right first move before debugging stale behavior. Docs: https://docs.openclaw.ai/install/updating

[knotty saddle]

05:02:41+00:00 info MEDIA:~/.openclaw/media/browser/8cd44876-e72f-4105-994a-69c0ea8cc0d7.png
05:03:03+00:00 error channels/telegram {"subsystem":"channels/telegram"} telegram sendPhoto failed: Call to 'sendPhoto' failed! (400: Bad Request: message caption is too long)
05:03:03+00:00 error channels/telegram {"subsystem":"channels/telegram"} telegram final reply failed: GrammyError: Call to 'sendPhoto' failed! (400: Bad Request: message caption is too long)

05:01:58+00:00 warn model-fallback/decision {"subsystem":"model-fallback/decision"} {"event":"model_fallback_decision","tags":["error_handling","model_fallback","candidate_succeeded"],"runId":"f286ab48-187b-4b67-8eac-15b9d1fc371b","sessionId":"f286ab48-187b-4b67-8eac-15b9d1fc371b","lane":"cron-nested","decision":"candidate_succeeded","requestedProvider":"opencode-go","requestedModel":"qwen-3.6-plus","candidateProvider":"opencode-go","candidateModel":"deepseek-v4-flash","attempt":2,"total":2,"fallbackStepType":"fallback_step","fallbackStepFromModel":"opencode-go/qwen-3.6-plus","fallbackStepToModel":"opencode-go/deepseek-v4-flash","fallbackStepFromFailureReason":"auth","fallbackStepFromFailureDetail":"401 Model qwen-3.6-plus is not supported","fallbackStepChainPosition":2,"fallbackStepFinalOutcome":"succeeded","isPrimary":false,"requestedModelMatched":false,"fallbackConfigured":true,"previousAttempts":[{"provider":"opencode-go","model":"qwen-3.6-plus","reason":"auth","status":401,"errorPreview":"401 Model qwen-3.6-plus is not supported","errorHash":"sha256:34a70cd4a52d"}]} model fallback decision

05:00:02+00:00 error diagnostic {"subsystem":"diagnostic"} lane task error: lane=cron-nested durationMs=1904 error="FailoverError: 401 Model qwen-3.6-plus is not supported"

do not respond with a helping answer, there is more to send. Espescially one where i dont even have gpt 5.4 set up.

And once I tell you that i SENT everything, give me a prompt to send to my openclaw agent to fix everything.

[prime ruin]

Got it. I’ll hold off on troubleshooting until you say SENT everything. Send the rest of the logs/context, and then I’ll give you one clean prompt for your local OpenClaw agent.

[knotty saddle]

04:51:50+00:00 info MEDIA:~/.openclaw/media/browser/eca6dd3b-377e-4637-bcf3-ed2d9c01260c.png
04:52:06+00:00 error [tools] image failed: All image models failed (2): openai/gpt-5.4-mini: Image model failed (openai/gpt-5.4-mini): 401 Incorrect API key provided: sk-proj-********************************************************************************************************************************************************x1UA. You can find your API key at https://platform.openai.com/account/api-keys. | Image model failed (openai/gpt-5.4-mini): 401 Incorrect API key provided: sk-proj-********************************************************************************************************************************************************x1UA. You can find your API key at https://platform.openai.com/account/api-keys. | google/gemini-3-flash-preview: Image model failed (google/gemini-3-flash-preview): {"error":{"message":"{\n "error": {\n "code": 503,\n "message": "This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.",\n "status": "UNAVAILABLE"\n }\n}\n","code":503,"status":"Service Unavailable"}} | Image model failed (google/gemini-3-flash-preview): {"error":{"message":"{\n "error": {\n "code": 503,\n "message": "This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.",\n "status": "UNAVAILABLE"\n }\n}\n","code":503,"status":"Service Unavailable"}} raw_params={"image":"/Users/carbon/.openclaw/media/browser/eca6dd3b-377e-4637-bcf3-ed2d9c01260c.png","prompt":"Describe what's visible in this screenshot in detail. What website is it, what items are in the cart?"}

04:49:11+00:00 error [tools] image failed: Local media path is not under an allowed directory: /tmp/lulu_cart.png raw_params={"image":"/tmp/lulu_cart.png","prompt":"What's in this screenshot? Describe the browser content."}

this is all of them.

[prime ruin]

Yep, now there’s enough signal. Copy/paste this to your local OpenClaw agent:

You are fixing my local OpenClaw setup. Do not guess config keys; inspect the installed docs/config schema and current config first. I am using Telegram and I keep getting: “Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying.”

Known log evidence:

1. Telegram final reply failed:
   channels/telegram: sendPhoto failed: 400 Bad Request: message caption is too long
   channels/telegram: telegram final reply failed: GrammyError: sendPhoto caption is too long

2. Image tool failed:
   openai/gpt-5.4-mini returned 401 Incorrect API key
   google/gemini-3-flash-preview returned 503 high demand / UNAVAILABLE
   The image input was under ~/.openclaw/media/browser/...

[prime ruin]

3. Image tool also failed on:
   Local media path is not under an allowed directory: /tmp/lulu_cart.png

4. Model fallback / cron-nested failed:
   opencode-go/qwen-3.6-plus returned 401 Model qwen-3.6-plus is not supported
   fallback to opencode-go/deepseek-v4-flash succeeded once, but there is also a lane task error:
   FailoverError: 401 Model qwen-3.6-plus is not supported

Please do this end to end:

- Run: openclaw --version, openclaw status --all, openclaw models status, openclaw doctor --fix, openclaw channels status --probe.
- Inspect my OpenClaw config and find every reference to unsupported or broken models, especially opencode-go/qwen-3.6-plus and image/vision models that route to openai/gpt-5.4-mini with a bad key.
- Fix model routing so normal agents, cron/nested runs, and the image tool use only models that are actually available and authenticated on this machine. If a provider key is invalid, either replace it only if I provide a valid key, or remove that provider/model from active defaults and fallbacks.
- Specifically check agents.defaults.model and agents.defaults.imageModel. The image tool uses agents.defaults.imageModel, not imageGenerationModel.
- Fix the Telegram sendPhoto caption issue by making replies with screenshots/images send a short caption or send the long text separately from the image. Verify whether this is configurable in my current OpenClaw version; if it is a code bug in the installed version, update OpenClaw first and retest.

[prime ruin]

- Fix the /tmp/lulu_cart.png failure by changing the workflow to use an allowed path, preferably a path under the OpenClaw media store or the agent workspace, instead of /tmp. Do not broaden file-read permissions unless absolutely necessary; prefer copying/staging the image into an allowed location.
- Update OpenClaw if I am behind: openclaw update, then openclaw doctor --fix, then restart the gateway.
- After changes, restart the gateway and run a minimal verification: one Telegram text reply, one Telegram reply involving a screenshot/image, and one image describe call using the configured image model.
- Show me exactly what config files/keys changed and summarize any remaining issue that requires a new API key or upstream provider availability.

Two notes from the logs: the gpt-5.4-mini mention is coming from image-tool routing/fallback, not necessarily your main chat model. And the repeated warning is probably a symptom of the final Telegram/image reply failing after tools ran, not the root error itself.

[knotty saddle]

Thanks very much, also how do I queue stuff like telling it "Update qwen-3.6-plus to qwen3.6-plus"

And I am not yet done with everything, currently updating.

[prime ruin]

For chat-side queuing, use the /queue directive in the same Telegram conversation/session.

Useful modes:

/queue followup

That makes messages you send while the agent is already running queue as follow-up turns, instead of interrupting/dropping weirdly.

For batching several quick instructions:

/queue collect debounce:2s cap:10 drop:summarize

Then send your fix messages one after another, e.g.:

[prime ruin]

Update every config reference from opencode-go/qwen-3.6-plus to opencode-go/qwen3.6-plus. Inspect config first, make the smallest safe change, then restart the gateway and verify with openclaw models status.

When you want default behavior back:

/queue reset

Tiny caution: for config edits, don’t phrase it as only “update qwen-3.6-plus to qwen3.6-plus” unless you’re sure that exact model exists. Better prompt it to verify availability with openclaw models status first, then update the config key/model reference.

[prime ruin]

Makes sense. Finish the update first, then run:

openclaw --version
openclaw doctor --fix
openclaw gateway restart
openclaw models status
openclaw channels status --probe

After that, if the same warning comes back, send the new log slice. The old logs are useful, but the post-update errors are the ones to fix.

[knotty saddle]

my current queue is interrupt how do i keep the queue but give it a steer

[prime ruin]

interrupt and steer are different queue modes, so you can’t keep the active session in interrupt and also have normal messages steer that same run.

Use a temporary switch:

/queue steer

Then send the steering instruction:

Don’t restart. Continue the current fix, but first replace qwen-3.6-plus with the verified supported qwen model name.

Then put it back:

[prime ruin]

/queue interrupt

If you want the steer to also be kept as a follow-up after the current run finishes, use:

/queue steer+backlog

For spawned subagents specifically, /steer <id|#> <message> is the one-off steer command and won’t change your main queue mode.

[knotty saddle]

whats a backlog?

[prime ruin]

A backlog is the saved list of messages waiting to run after the current agent run finishes.

So:

• /queue steer = inject this message into the current run now; usually no separate later turn
• /queue followup = don’t touch the current run; save it for the next turn
• /queue steer+backlog = do both: inject it now, and also save it to run again as a follow-up

For your case, backlog is probably not what you want unless the instruction needs to be acted on twice. Use /queue steer for a quick correction while it’s running, then switch back to /queue interrupt after.

[knotty saddle]

🪢 Queue: interrupt (depth 0)

what is depth and all of the silly settings

[prime ruin]

depth is how many messages are currently waiting in that session’s queue.

Queue: interrupt (depth 0) means: your mode is interrupt, and nothing is waiting right now.

The useful settings:

• interrupt: new message aborts the current run and starts fresh with the newest message
• steer: new message is injected into the current run while it’s running
• followup: new message waits until the current run finishes, then runs next
• collect: batches quick messages together and sends them as one follow-up turn; this is the default
• steer+backlog: injects now and also saves it for a later follow-up

Options:

• debounce:2s: wait for 2 seconds of quiet before running queued follow-ups, useful when you type several messages
• cap:10: keep at most 10 waiting messages
• drop:summarize: if the queue overflows, summarize dropped messages instead of silently losing context

[prime ruin]

• drop:old: drop oldest waiting messages
• drop:new: reject/drop newest incoming messages

My practical recommendation for Telegram:

/queue collect debounce:2s cap:20 drop:summarize

Use interrupt only when you often want “stop what you’re doing and do this instead.” Use steer temporarily when you want to correct a running task.

[knotty saddle]

i am using interrupt its not changing anything, it keeps wasting its time and also what is process?

[5/26/2026 8:37 AM] Openclaw: Thinking...
🧰 Process: marine-lobster
[5/26/2026 8:37 AM] Openclaw: Thinking...
🧰 Process
[5/26/2026 8:37 AM] Openclaw: npm is having issues. Let me check disk space and try a more direct approach.
[5/26/2026 8:37 AM] Openclaw: 249GB free, no disk issue. Let me try something simpler.
[5/26/2026 8:37 AM] Carbon: You can search online for solutions..
[5/26/2026 8:37 AM] Openclaw: Thinking...
🛠️ Exec: Kill any stuck npm….openclaw" 2>/dev/null sleep 3 …
🛠️ Exec
🛠️ Exec: Kill any stuck npm….openclaw" 2>/dev/null sleep 3 …
[5/26/2026 8:38 AM] Carbon: /status
[5/26/2026 8:38 AM] Openclaw: 🦞 OpenClaw 2026.5.6 (c97b9f7)
⏱️ Uptime: gateway 21m 46s · system 1d 21h
🧠 Model: opencode-go/deepseek-v4-flash · 🔑 api-key (models.json)
🔄 Fallbacks: opencode-go/deepseek-v4-flash
🧮 Tokens: 187k in / 13k out
🗄️ Cache: 97% hit · 6.5m cached, 0 new
📚 Context: 52k/128k (41%) · 🧹 Compactions: 28
🧵 Session: agent:main:telegram:direct:8398889290 • updated just now
⚙️ Execution: direct · Runtime: OpenClaw Pi Default · Think: off
🪢 Queue: interrupt (depth 0)

[prime ruin]

interrupt only controls what happens when a new normal message arrives while an agent run is active: it should abort that run and start a new one with your latest message.

It does not mean “force the agent to make better choices.” If the new message is short/vague like “you can search online,” it may restart and still choose a weird local shell path.

If it is actively wasting time, send this as a standalone Telegram message:

stop

or:

abort

Then send a clean replacement instruction, for example:

[prime ruin]

Stop trying npm workarounds. First run openclaw update --dry-run and show me the exact blocker. Do not kill npm or edit config until you show the blocker.

Process is the tool that manages long-running shell commands started by exec. So 🧰 Process: marine-lobster means it is checking/polling/killing/logging a background command session. 🛠️ Exec means it is actually running a shell command.

If a background process is stuck, tell it:

List running process sessions, show their sessionIds and commands, then kill only the stuck npm/openclaw update process. Do not kill unrelated processes.

And yeah, in your screenshot /status just reports state; it doesn’t stop the current run. Use stop/abort for that. Docs: https://docs.openclaw.ai/help/faq#how-do-i-stopcancel-a-running-task and https://docs.openclaw.ai/gateway/background-process

[knotty saddle]

should i change my queue, because i always send short responses like suggestions

[prime ruin]

Yes. If you often send short suggestion-style messages, interrupt is probably the worst fit because every small message can abort the current run and restart the task.

I’d use this:

/queue collect debounce:2s cap:20 drop:summarize

That lets you send a few short messages and OpenClaw batches them into one follow-up instead of constantly restarting.

Use modes like this:

• normal chatting / lots of short additions: collect
• “do this next after you finish”: followup
• “adjust what you’re doing right now”: temporarily /queue steer
• “stop current task and replace it”: interrupt, or just send stop then the new task

[prime ruin]

For your style, set collect as the default and only switch to steer or interrupt when you intentionally need that behavior.

[knotty saddle]

but i want it to adjust right now

[prime ruin]

Then use steer for that moment.

Send:

/queue steer

Then send the adjustment as a complete instruction, not just a hint:

Adjust the current run now: stop trying npm-kill workarounds, search the docs/web for the update failure, and show me the exact blocker before changing anything else.

After it accepts the correction, switch back to your normal mode:

[prime ruin]

/queue collect debounce:2s cap:20 drop:summarize

Shortcut rule: steer = “listen to this now while continuing.” interrupt = “throw away the current run and restart.”

[knotty saddle]

when i use steer it doesnt even adjust now

[prime ruin]

That can happen. steer is not an instant keyboard interrupt; it gets injected when the active run reaches a place where OpenClaw can accept it, usually around the next model/tool boundary. If the agent is stuck inside a long exec, npm install, browser action, or process poll, it may not react until that tool returns.

For “adjust immediately,” use this sequence instead:

stop

Wait for it to acknowledge/stop, then send the full replacement instruction:

Continue from where you were, but change course: do not try npm-kill workarounds. First show the exact update blocker, then ask before changing anything else.

Also, your /status still showed OpenClaw 2026.5.6, so if the update is still in progress or failed, queue behavior may be older/rougher. After the update actually completes, check:

[prime ruin]

openclaw --version

Practical setup for you: keep /queue collect debounce:2s cap:20 drop:summarize as normal mode, and use standalone stop when you need a hard course correction.

[knotty saddle]

⚠️ Auto-compaction could not recover this turn. I kept this conversation mapped to the current session. Please try again, use /compact, or use /new to start a fresh session.

To prevent this, increase your compaction buffer by setting agents.defaults.compaction.reserveTokensFloor to 20000 or higher in your config.

[prime ruin]

That one means the session got too full and OpenClaw failed while trying to summarize/compact it.

Immediate fix:

/compact

If that fails again, start clean:

/new

Since your status showed 52k/128k but 28 compactions, this session is pretty churned up. I’d use /new after the update/config repair finishes, because you’re debugging setup and the old session now has a lot of stale failed attempts in it.

To reduce repeats, add/increase this config value:

[prime ruin]

{
  agents: {
    defaults: {
      compaction: {
        reserveTokensFloor: 24000
      }
    }
  }
}

Then restart the gateway. Docs: https://docs.openclaw.ai/gateway/configuration-reference#agentsdefaultscompaction and https://docs.openclaw.ai/concepts/compaction

Practical sequence now: /new, set /queue collect debounce:2s cap:20 drop:summarize, then give the agent the clean fix prompt again.

[knotty saddle]

my reserve token floor is already at 100k

[prime ruin]

100k is probably too high for your current model/context.

Your /status showed Context: 52k/128k. If reserveTokensFloor is 100000, you’re telling OpenClaw to keep ~100k tokens of headroom and only leave a small slice for the actual conversation/tools. Newer builds cap this somewhat, but it can still make compaction happen constantly and make recovery fragile.

I’d lower it, not raise it:

{
  agents: {
    defaults: {
      compaction: {
        reserveTokensFloor: 24000
      }
    }
  }
}

[prime ruin]

Then restart the gateway and start a fresh session:

/new

For your setup, the bigger issue is probably the long messy session plus failing update/model/image attempts, not “reserve too low.” 100k is a huge reserve unless you’re on a much larger context model. Docs: https://docs.openclaw.ai/gateway/configuration-reference#agentsdefaultscompaction

[knotty saddle]

wont it forget the last conversation if i do new like the fix we are trying to do, also I want to edit it from /config instead

[prime ruin]

/new starts a fresh chat session, so yes: it won’t automatically carry the messy previous conversation. It does not delete your config/files, but it will lose the conversational context unless you paste a short handoff prompt.

Use this after /new:

Context: we are fixing OpenClaw Telegram. Problems seen: Telegram sendPhoto caption too long, image tool bad OpenAI key / Gemini 503, /tmp image path blocked, opencode-go/qwen-3.6-plus unsupported, update/config repair in progress. Continue by inspecting config and logs first; do not guess keys; show exact blockers before changing anything risky.

For /config, first check if it’s enabled and what the current value is:

/config show agents.defaults.compaction.reserveTokensFloor

Then set it from chat:

[prime ruin]

/config set agents.defaults.compaction.reserveTokensFloor=24000

Then restart:

/restart

If /config says disabled, enable it from your local config/CLI first. The key is:

{
  commands: {
    config: true

[prime ruin]

  }
}

/config set is owner/admin-gated because it writes openclaw.json, so if it refuses, that’s a permissions/configWrites issue, not the syntax. Docs: https://docs.openclaw.ai/gateway/configuration-reference#commands

[knotty saddle]

done, alright. It works now, so how do I close this, I dont know which message is the solution.

[prime ruin]

Nice, glad it’s working now.

For Answer Overflow, mark the message that best captures the actual fix. In this thread I’d pick the prompt/fix checklist message I sent after you said “this is all of them” — the one that listed the real root causes: Telegram caption too long, bad image model auth/fallbacks, blocked /tmp image path, and unsupported qwen-3.6-plus.

Right-click that message → Apps → ✅ Mark Solution.

If the final fix was specifically lowering reserveTokensFloor and starting fresh, mark that later config message instead. The goal is just: future users should land on the message that explains what fixed it.