Bug type

Behaviour bug

Summary

When an upstream provider terminates the connection (TCP reset / stream abort), OpenClaw receives rawError=terminated. This error string is not matched by any entry in ERROR_PATTERNS, so classifyFailoverReason("terminated") returns null and the model fallback chain is never invoked.

Steps to reproduce

1. Configure a provider behind a proxy that can terminate connections (tested with NVIDIA inference API at inference-api.nvidia.com/v1)
2. Configure model fallback chain:
   • primary: claude-opus-4-6
   • fallbacks: [gemini-3.1-pro-preview, gpt-5.4, claude-sonnet-4-5]
3. Send a message when the upstream provider is under load
4. Observe gateway logs

Expected behaviour

When the primary model returns error=terminated, the gateway should classify this as a transient timeout-class error and attempt the next model in the fallback chain — the same way "connection error", "socket hang up", and "fetch failed" are handled today.

Actual behaviour

[agent/embedded] embedded run agent end: runId=5cac9994-... isError=true
  model=azure/anthropic/claude-opus-4-6 provider=nvidia-proxy
  error=terminated rawError=terminated

The same runId retries 4 times on the same model (Opus). The fallback chain (Gemini → GPT → Sonnet) is never attempted. After exhausting retries, the user sees the error.

Root cause

In src/agents/pi-embedded-helpers/failover-matches.ts (compiled to reply-payloads-dedupe-*.js in v2026.3.28):

// ERROR_PATTERNS.timeout includes:
"timeout", "timed out", "connection error", "network error",
"fetch failed", "socket hang up", /\beconn(?:refused|reset|aborted)\b/i, ...

// But does NOT include:
"terminated"   // ← missing

When classifyFailoverReason("terminated") is called:

1. isRateLimitErrorMessage("terminated") → false
2. isOverloadedErrorMessage("terminated") → false
3. isServerErrorMessage("terminated") → false (v2026.3.28 added "connection reset" here, but not "terminated")
4. isTimeoutErrorMessage("terminated") → false ← the gap
5. Returns null → no fallback triggered

Suggested fix

Add "terminated" to ERROR_PATTERNS.timeout:

  timeout: [
+   "terminated",
    "timeout",
    "timed out",
    "service unavailable",

This is consistent with the existing pattern — "terminated" is functionally equivalent to "connection error" and "socket hang up" (all represent unexpected connection closure by the upstream).

Version

2026.3.28 (also reproducible on 2026.3.23-2)

OS

Ubuntu 24.04.3 LTS / Linux 6.17.0

Model

azure/anthropic/claude-opus-4-6 (via NVIDIA inference API)

Provider / routing chain

openclaw gateway → nvidia inference-api.nvidia.com/v1 (openai-completions) → anthropic claude

Impact

This affects any deployment where the LLM provider sits behind a proxy that can terminate connections — including NVIDIA NIM, Cloudflare AI Gateway, Azure API Management, and corporate inference endpoints.

The terminated error is the most common transient failure in proxy environments, yet it's the one error that doesn't trigger fallback. Users configured a fallback chain specifically for resilience, but it never activates for the most frequent failure mode.

Workaround

Manually patch reply-payloads-dedupe-*.js to add "terminated" to ERROR_PATTERNS.timeout. Must reapply after every upgrade.

Related issues

• #29429 — network_error stop reason not triggering failover (closed, similar root cause)
• #48213 — error=terminated with openai-codex (open, documents symptom but not fallback gap)
• #45834 — generic provider errors not triggering fallback (open)
• #18868 → PR #19077 — added ENETUNREACH/ECONNREFUSED to failover (merged, precedent for this fix)
• #28051 — terminated leaking into user-visible replies (open)