Summary

After updating to recent versions (worked on ~v2026.2.17, broken on v2026.2.19+), every message — including a single "hi?" on a fresh /reset session — triggers:

Context overflow: prompt too large for the model. Try /reset (or /new) to start a fresh session, or use a larger-context model.

The model is openrouter/moonshotai/kimi-k2.5 (262K context). The system prompt is ~26K chars (~6K tokens). This cannot be a real context overflow.

Environment

• OpenClaw version: coollabsio/openclaw:latest (Docker, pulled ~2026-02-21)
• Model: openrouter/moonshotai/kimi-k2.5 (262,144 token context)
• Channel: Discord
• Hosting: Docker on Hetzner VPS via Coolify
• Previously working: Yes, ~v2026.2.17 (last confirmed working)

Evidence that this is NOT a real context overflow

From the sessions.json system prompt report on this exact session:

{
  "systemPrompt": {
    "chars": 26223,
    "projectContextChars": 15040,
    "nonProjectContextChars": 11183
  },
  "inputTokens": 20286,
  "contextTokens": 262144
}

• 26K chars system prompt against a 262K token context window
• All 8 workspace files total ~16K chars — none truncated
• Tool schemas: ~18K chars
• Skills metadata: ~2K chars
• After /reset, conversation history is empty

20K input tokens vs 262K context = 7.7% utilization. This cannot overflow.

Reproduction

1. Set model to openrouter/moonshotai/kimi-k2.5
2. Run /reset or /new in Discord
3. Send any message (e.g., "hi?")
4. Immediate "Context overflow: prompt too large for the model" error

Likely root cause

This matches two known issues:

• FailoverError on rate limit/billing is misreported as 'Context overflow' #10368: isContextOverflowError() / isLikelyContextOverflowError() misclassifies rate-limit, billing, or provider errors as "context overflow" due to overly broad regex matching on response text
• [Bug]: Context overflow error #5771: Same error detection scans response text rather than API error codes, causing false positives

The actual underlying error is likely a provider-side issue from Chutes (the only OpenRouter provider for Kimi K2.5, running int4 quantized Kimi-K2.5-TEE) being surfaced as a context overflow.

Version correlation

• v2026.2.19 introduced ACP prompt payload changes ("bound ACP prompt text payloads to 2 MiB") and provider routing fixes
• v2026.2.21 modified overflow compaction retry budgeting
• The regression appeared somewhere between v2026.2.17 (working) and v2026.2.19+ (broken)

Expected behavior

• The actual error from the provider should be shown (rate limit, billing, provider unavailable, etc.)
• isContextOverflowError() should not match non-overflow errors
• If the error genuinely is from the provider, a fallback or clearer message would help

Request

Could you check what error OpenRouter/Chutes is actually returning? The "context overflow" label appears to be masking the real issue. The fix in PR #12988 (for #10368) may not cover this specific provider error pattern.