Context overflow: prompt too large for the model. Try again with less input or a larger-context mode | Friends of the Crustacean 🦞🤝 | Page 1

spring drumBOT Feb 12, 2026, 11:18 PM

#

That message isn’t a “bug” you can permanently patch away — it’s the model hitting its hard context-window limit (your chat history + tool output + system prompts no longer fit in one request).

The real, final fixes are:

Keep the same convo, but shrink it:
Send:

/compact

Then retry your message. (This summarizes older turns into a compact summary and keeps recent messages.)
Docs: https://docs.openclaw.ai/concepts/compaction

Start fresh (guaranteed):
If you just need it to work immediately, do:

/new

#

(or /reset)
This creates a fresh session for that chat, so there’s no accumulated history to overflow.

If it keeps happening in normal use: enable session pruning so old tool results don’t keep bloating the prompt, and/or switch to a larger-context model.
Docs: https://docs.openclaw.ai/concepts/session-pruning
FAQ section: https://docs.openclaw.ai/help/faq#im-getting-context-too-large-errors-how-do-i-reset-or-compact

If you tell me which surface you’re using (Discord/Telegram/etc) and whether this happens in one specific long thread or immediately in new chats, I can point you to the best option (compact vs reset vs pruning vs model).

#

If it happens in a brand-new Telegram chat, then /new / /reset / /compact won’t matter — this is almost certainly not “chat history got too long”.

In that case the only realistic causes are:

The model you’re actually hitting has a much smaller context window than you think (common when the model ref/provider isn’t what you expect, or the catalog entry is wrong), or
Your base prompt is huge (tool schemas + skills list + injected workspace files), so even the first request overflows.

Fastest way to pinpoint it (works even when the LLM can’t reply)

In Telegram, send these as standalone messages and paste the output here:

/status
/context list (or /context detail if it works)

Those commands are Gateway-handled and will show things like:

the exact model ref being used
the context window OpenClaw thinks it has (ctx=...)
how big your system prompt + tool schema JSON overhead is

#

Immediate “make it stop” workaround (if tool schema is the culprit)

If /context shows the tool schemas/system prompt are already close to the limit, temporarily shrink tools:

In ~/.openclaw/openclaw.json (JSON5), set a minimal tool profile (you can do it globally just to test):

{
  tools: { profile: "minimal" }
}

Then retry a simple “hi” message in a new Telegram chat.
If that fixes it, we can widen it back up (e.g. “messaging” or “coding”) and/or deny only the big offenders (often browser, canvas, nodes).

Tool profiles docs: https://docs.openclaw.ai/tools

Also: confirm your install + model config

#

On the Gateway machine, paste (redact keys/tokens):

openclaw --version
openclaw models status

And double-check you’re really on the intended Kimi endpoint/model ref (OpenClaw has specific config examples for Moonshot Kimi with contextWindow set). Docs: https://docs.openclaw.ai/gateway/configuration (search for “Moonshot AI (Kimi)”).

If you paste /context list + openclaw models status, I can tell you which of the two it is in one reply and what to change.

#

Your /context list is the key clue: you’re only at ~14k tokens in the session, and OpenClaw thinks the window is 262k. So a real “prompt too large” overflow should be basically impossible here.

That means the overflow is almost certainly coming from the kimi-coding endpoint rejecting the request because of the requested output size (max tokens), or because that endpoint’s real limits are smaller than what the catalog says.

1) Quick isolation test (no config changes)

In Telegram, temporarily switch to your fallback model and send “hi”:

/model anthropic/claude-opus-4-6

If Opus works and Kimi doesn’t, the problem is 100% Kimi endpoint behavior, not your chat/session.

2) Most common fix: cap Kimi’s output tokens

Add a per-model maxTokens cap (this prevents providers that compute “input + requested output” from instantly throwing overflow):

Edit ~/.openclaw/openclaw.json and add:

#

{
agents: {
defaults: {
models: {
"kimi-coding/k2p5": {
params: { maxTokens: 4096 } // try 4096 first; 8192 also ok if it works
}
}
}
}
}

Then run:

#

openclaw gateway restart

Test again in a brand-new Telegram chat.

3) If it still overflows: shrink tool overhead (test)

As a test, set tools to minimal (just to see if Kimi is choking on tool schemas):

{ tools: { profile: "minimal" } }

Restart gateway, test. If this fixes it, we can move you to tools.profile: "messaging" (keeps messaging features but drops heavy UI tools like browser/canvas/nodes).

Docs on tool profiles: https://docs.openclaw.ai/tools

#

4) Important naming confusion

You’re on kimi-coding/k2p5 (Kimi Coding, Anthropic-compatible). That is not the same as Moonshot’s OpenAI-style moonshot/kimi-k2.5 provider, and the docs explicitly say the keys aren’t interchangeable. If you meant Moonshot Kimi, follow: https://docs.openclaw.ai/providers/moonshot

If you try steps (1) + (2) and it’s still failing, paste the exact overflow log lines from:

openclaw logs --tail 200 | grep -i overflow

and I’ll tell you which knob to turn next.

spring drumBOT Feb 13, 2026, 12:55 AM

#

Yep — openclaw logs doesn’t have --tail. Use --limit:

openclaw logs --limit 200 | grep -i overflow

If you want to watch it live while you trigger the error:

openclaw logs --follow | grep -i overflow

(And if grep isn’t behaving on your shell, just run openclaw logs --limit 200 and paste the chunk around the error.)

spring drumBOT Feb 13, 2026, 1:15 AM

#