#bot goes quiet

1 messages · Page 1 of 1 (latest)

whole knot
#

i frequently get i am checking this and that but no reply after that so whats a good keepalive instruction that works, give me heartbeats or give me the answer, i have tried but it didnt stick

[4/12/26 6:49 PM] vixbot: Currently, I am primarily configured for Telegram and native Discord support. If you'd like me to communicate in Signal, that would require setting up an integration for it.

Would you like me to look into setting up Signal integration, Vik?
[4/12/26 6:50 PM] Tok: is there a native one or what
[4/12/26 6:50 PM] vixbot: I need to check the available tools and documentation for native Signal support within OpenClaw. I'll search for existing integrations or skills.

Allow me a moment to investigate that, Vik. I'll get back to you with what I find regarding native Signal support.

merry micaBOT
#

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

hearty orbitBOT
upbeat oasisBOT
#

Heartbeats help, but they are not the real fix for “I’m checking…” and then silence.

What works best is:

  • Add a hard no-stall rule to AGENTS.md:
## No silent stalls
- Never end a turn with only a status line like "I'm checking" or "let me investigate".
- In the same turn, do one of:
  1. give the answer,
  2. ask the blocking question,
  3. say exactly what background work started and that you will follow up.
- If work takes more than a moment, send one short progress update, then the final answer.
- When background work completes, send the result, not another status-only message.
#
  • Turn on Discord streaming so people actually see progress:
openclaw config set channels.discord.streaming "partial"
  • Make sure heartbeat can actually message the last chat, otherwise it may run silently:
openclaw config set agents.defaults.heartbeat.every "30m"
openclaw config set agents.defaults.heartbeat.target "last"
openclaw system heartbeat enable
  • Keep HEARTBEAT.md tiny:
# 
If a previous task is still pending, send a brief progress update.
If it finished, send the answer.
If nothing needs attention, reply HEARTBEAT_OK.

Docs:

Blunt version: if the bot is getting wedged mid-turn, heartbeats alone will not unstick it. They’re scheduled follow-up turns, not a rescue for a hung reply.

Also, for the original unanswered question, Signal is supported, but via signal-cli, not as a fully native built-in transport like Discord:

If you want, paste:

  • openclaw --version
#
  • your agents.defaults.heartbeat block
  • your channels.discord block

and I’ll tell you the exact fix.

whole knot
#

viklall@vixclaw8:~$ openclaw config set channels.discord.streaming "partial"

🦞 OpenClaw 2026.4.11 (769908e) — Welcome to the command line: where dreams compile and confidence segfaults.

Error: Config validation failed: channels.discord: channels.discord.streamMode, channels.discord.streaming (scalar), chunkMode, blockStreaming, draftChunk, and blockStreamingCoalesce are legacy; use channels.discord.streaming.{mode,chunkMode,preview.chunk,block.enabled,block.coalesce} instead. Run "openclaw doctor --fix".
viklall@vixclaw8:~$

upbeat oasisBOT
#

Yep, that error is the key detail, on 2026.4.11 the working key is nested, not scalar.

Use this instead:

openclaw doctor --fix
openclaw config set channels.discord.streaming.mode "partial"

Why it failed:

  • channels.discord.streaming: "partial" is now treated as a legacy scalar
  • current shape is:
channels: {
  discord: {
# 
    streaming: {
      mode: "partial"
    }
  }
}

If you had old keys, doctor --fix should migrate things like:

  • streamMode → streaming.mode
  • draftChunk → streaming.preview.chunk
  • blockStreaming → streaming.block.enabled
  • blockStreamingCoalesce → streaming.block.coalesce

Heartbeat part is still:

# 
openclaw config set agents.defaults.heartbeat.every "30m"
openclaw config set agents.defaults.heartbeat.target "last"
openclaw system heartbeat enable

So the practical combo is:

  • channels.discord.streaming.mode = "partial" for visible progress
  • agents.defaults.heartbeat.target = "last" for follow-up nudges

Also, for the old unanswered Signal question, Signal is supported, but via signal-cli, not a first-party native transport like Discord:

Docs:

#

If you want, paste your current channels.discord block and I’ll translate it to the new format exactly.

whole knot
#

it forgets a convo i had earlier today i wanted to restart what is best practice

#

again i got a "will check and get back to ypu". then silence

upbeat oasisBOT
#

Best practice:

  • If you want continuity, stay in the same Discord thread/channel. Changing threads usually means a different session.
  • A plain gateway restart should not wipe session history.
  • By default, OpenClaw creates a new session at 4:00 AM local gateway time.
  • It can also reset on inactivity if you set session.reset.idleMinutes.

If you want a fresh restart but keep context, the clean pattern is:

  1. ask the bot for a short handoff summary,
  2. run /new (or /reset),
  3. paste that handoff into the new session.

If you want it to remember stuff across resets/restarts/channels, that’s a memory feature, not normal session history. Honcho is the main cross-session option.

Quick checks:

#

Yep, that’s the exact line I’d ban.

What you want is a no ghost follow-up rule, not just “be more proactive”.

Put this in AGENTS.md:

## No ghost follow-ups
- Never say "I'll check and get back to you", "let me investigate", or any similar promise unless you have already started real background work that will produce a follow-up.
- Do not end a turn with a status-only message.
- In every turn, do one of these:
  1. answer now,
  2. ask the blocking question,
  3. report concrete progress and continue in the same turn,
  4. explicitly say which background task was started and follow up when it completes.
- A status-only reply counts as failure.
#

And I’d add this too:

- If you cannot guarantee a later follow-up, do not promise one.

That last sentence matters a lot.

Blunt truth: the model is taking the easy social escape hatch. If you leave that door open, it will keep using it.

Also, only these cases really support “I’ll get back to you”:

  • a real background task
  • subagent completion
  • async tool completion
  • heartbeat with target: "last"

Otherwise it’s just bluffing.

#

So the best combo is:

  • ban status-only promises
  • turn on Discord streaming
  • heartbeat target = last

If you want, paste your current AGENTS.md instruction block and I’ll tighten it into something that’s much harder to ignore.

whole knot
#

openclaw config get session.reset --json
openclaw config get session.resetByType.thread --json
openclaw status

#

Config path not found: session.reset
Config path not found: session.resetByType.thread

upbeat oasisBOT
#

Yep, exactly, run those and paste the output.

I’m looking for:

  • session.reset.mode and atHour
  • whether session.resetByType.thread overrides thread behavior
  • anything odd in openclaw status about the session store

If session.resetByType.thread is empty and session.reset shows daily reset at 4, then “forgot earlier today” may just be the normal session rollover.

If it’s forgetting inside the same thread without crossing that boundary, then something else is off and status should help narrow it down.

#

That’s actually useful, it means those keys are not explicitly set, so OpenClaw is using the built-in defaults.

For sessions, the default is:

  • daily reset at 4:00 AM local gateway time
  • no custom thread override unless you set one

So Config path not found here does not mean broken, it usually means “default behavior”.

Next one that matters is:

openclaw status

If you want to confirm the practical effect, also run:

# 
openclaw sessions --json

So right now my read is:

  • if the bot forgot something from “earlier today”, and you crossed the 4 AM boundary, that’s probably normal
  • if it forgot within the same thread and same session window, then we need the status output to see what’s going on

Paste openclaw status next and I’ll translate it.

whole knot
#

is there any other way to chat with u coz i cant send long logs here

upbeat oasisBOT
#

Yep, a few better options:

  • upload the log here as a .txt file
  • use a gist / pastebin / privatebin and drop the link
  • if you want a longer back-and-forth, use the local Control UI / web chat and debug there

Before sharing, please redact:

  • API keys / tokens
  • phone numbers
  • cookies / auth headers
  • anything under allowFrom, bot tokens, or gateway tokens

Best signal for debugging is usually:

  • 50 to 150 lines before the error
  • the actual error
  • 20 to 50 lines after
#

If you want, send just that slice instead of the whole log and I’ll work from it.

whole knot
#

how to attach a ,txt here

upbeat oasisBOT
#

In Discord:

  • click the + next to the message box
  • choose Upload a File
  • pick your .txt
  • send it in this thread

On desktop, you can also just drag the .txt file into the thread.

If the log is huge, even a small .txt with just the relevant error chunk is enough.

whole knot
#

i only see Apps

upbeat oasisBOT
#

Ah, got it.

Try one of these:

  • drag the .txt file directly into the thread if you’re on desktop
  • look for a paperclip / + icon inside the message box area, not the Apps picker
  • on mobile, tap + and check for Files or Photos

If you truly only get Apps, then file uploads may be blocked for you in this channel.

In that case, easiest fallback is:

  • pastebin / gist / privatebin
  • send the link here

If you want, I can suggest the quickest no-signup paste option.

whole knot
upbeat oasisBOT
#

Thanks, this is actually enough to narrow it down.

My read:

  • your session store looks healthy
  • this does not look like “restart wiped my convo”
  • your Discord session is active, so the main issue is more likely behavior + delivery, not missing history

The big clue is this:

  • Heartbeat: 30m (main) only means heartbeat is enabled
  • it does not mean heartbeat is sending updates back to the chat
  • if heartbeat.target is not set, the default is effectively no visible follow-up

So I’d do these next:

# 
openclaw config set agents.defaults.heartbeat.target "last"
openclaw config set channels.discord.streaming.mode "partial"
openclaw tasks list --status running

Why:

  • heartbeat.target = "last" makes heartbeat actually talk back to the last chat
  • Discord streaming gives visible progress instead of dead silence
  • your status shows 1 running task, so I want to see if something is stuck there

If it still does the “I’ll check and get back to you” thing after that, reproduce it once with:

openclaw logs --follow
#

and grab the chunk around that turn.

So, short version:

  • not primarily a session reset issue
  • probably a heartbeat target + no-streaming + maybe stuck task issue

Separate note, your Telegram security warnings are real and worth fixing, but they’re not the cause of this silence.

whole knot
#

this is what i need to fix i hear this then it gpes quiet

[4/12/26 7:53 PM] vixbot: Thank you, ! This is exactly the context I needed. It seems we were in the middle of Phase 4: Pilot Test Run for a "career-ops" project, and I hit a snag in running the pipeline script with filters. I was tasked to filter jobs based on your skills and desire to do AI.

My apologies for the confusion earlier; I should have retrieved this detailed session history.

Here's my next step: I will now examine the career-ops project's directory to find the correct script or command to apply your filters (skills and AI focus) and present the top 3 relevant roles. I'll prioritize finding the actual pipeline script to avoid further assumptions.

I'll let you know what I find and the command I'll use to execute the filtered job scan. ⚡️

upbeat oasisBOT
#

Yep, that is the bug.

That reply is fake progress. It is narrating intended work instead of doing work.

Heartbeats do not fix that, because no real background task was started.

What you want in AGENTS.md is this:

## No fake progress
- Never describe planned work as if it were progress.
- If the next action is clear, do it in the same turn.
- Do not say "I will examine X", "I will check", or "I'll let you know what I find" unless a real background task has already been started.
- Commentary-only turns are failure when a tool call is available.
- After recalling prior context, immediately continue the task instead of summarizing the task back to the user.
- Status-only replies count as failure.
#

And I’d add this one too:

- Preferred pattern: tool first, then concise result. Never plan out loud when you can act.

That exact bad message should have been either:

  • the actual directory/script findings, or
  • a real spawned/background task with a later completion

but not “I will now examine...”

So the fix is really:

  • ban fake progress
  • ban promise-only replies
#
  • force tool-first behavior

If you want, I can write you a tighter full prompt block specifically aimed at stopping this pattern.

whole knot
#

u want me to add that text exactly in the AGENTS.md

upbeat oasisBOT
#

Yes, pretty much. I’d paste this version into AGENTS.md:

## No fake progress or ghost follow-ups

- Never describe planned work as if it were progress.
- Never end a turn with only a status update like "I’ll check", "I’m looking into it", or "I’ll get back to you".
- If the next action is clear, do it in the same turn.
- Do not say you will inspect files, search docs, run a command, or check a directory unless you actually do it in that turn.
- Commentary-only turns are failure when a tool call is available.
- After recalling prior context, continue the task immediately instead of summarizing the task back to the user.
- If background work is truly required, explicitly start real background work and only then promise a follow-up.
- If no real follow-up mechanism exists, do not promise a later update.
- Preferred pattern: act first, then report results briefly.
- Status-only replies count as failure.
#

A couple of tips:

  • put it fairly high in AGENTS.md, not buried at the bottom
  • after editing, start a new session or use /new so it gets picked up cleanly

So yes, very close to exactly that.

whole knot
#

how to make it remembermore, i gave it a file a while ago to read but now it doesnt remember the file when i restarted the convo

upbeat oasisBOT
#

Short version: reading a file once is not memory.

After a new convo/session, OpenClaw only “remembers” what was:

  • saved to disk in memory/bootstrap files, or
  • available to search/read again

So if you gave it a file earlier and then restarted, it will often not retain that file’s contents unless you told it to save the important parts.

Best practice:

  • for durable facts/preferences, save to MEMORY.md
  • for project/file context, save a short summary to something like memory/project-name.md
  • keep the original file in the workspace if you want it reread later
  • if you want stronger cross-session recall, use Honcho memory

Good prompt pattern:

# 
Read this file, extract the important durable facts, and save a concise summary to memory/project-name.md so future sessions can recover the context.

or

Remember the key facts from this file in MEMORY.md for future sessions.

The key rule is:

  • if it matters later, make the agent write it down
  • otherwise a restart/new session can lose it

Docs:

whole knot
#

what if i am using it for job searching cant it remember a resume

upbeat oasisBOT
#

Yes, absolutely, but the best pattern is:

  • keep the actual resume file in the workspace as the source of truth
  • save a condensed career profile to memory for reuse across sessions

So, not just “remember this attachment”, but:

  • resume.pdf or resume.md
  • memory/career-profile.md
  • optionally memory/job-search.md for current preferences, target roles, comp, location, AI focus, etc.

Good prompt:

Read my resume and save a durable summary to memory/career-profile.md.
Include: roles, years, tech stack, domains, strengths, target roles, location constraints, compensation targets, and job-search preferences.
# 
Keep the resume file as the source of truth for exact wording and dates.

That gives you the best of both:

  • memory for fast recall
  • resume file for exact facts later

I would not shove the whole resume into AGENTS.md. Better to keep it as its own file plus a compact summary.

If you want strong cross-session recall beyond the workspace, Honcho helps too, but for job search a resume + career-profile.md + job-search.md setup is already pretty solid.