Bug Report

Version: 2026.3.2
Channel: Discord
Model: anthropic/claude-opus-4-6

What happens

When an Anthropic API call errors out (Internal server error / overloaded) after the agent has already delivered a response to Discord, OpenClaw retries the run. The retry produces a second response that also gets delivered — resulting in the user seeing two different replies to the same message.

Evidence

• 12 Anthropic API errors in a single session (mix of Internal server error and The AI service is temporarily overloaded)
• User consistently sees two replies with slightly different wording for the same input
• Only one active session exists for the channel (confirmed via sessions_list)
• Single gateway process running
• No duplicate agent configs

Config context

• compaction.mode: "safeguard" with reserveTokensFloor: 40000 — safeguard re-runs may also trigger duplicate delivery
• contextPruning.mode: "cache-ttl" with ttl: "5m"

Expected behavior

If a response has already been delivered to Discord before the error occurs, the retry should detect this and skip re-delivery (or deduplicate).

Workaround

None found — no retry config exposed to users.

Happy to provide full logs if helpful.