--channels Discord plugin: duplicate replies per inbound message (multi-cycle re-processing)

Bug Description

Claude Code with the --channels Discord plugin generates 2-4 duplicate replies per inbound Discord message. Each duplicate is a separate API cycle with full token billing.

Environment

• Claude Code: Latest stable (auto-updates enabled)
• OS: Windows 11 (WSL2 for supporting infrastructure)
• Plugin: plugin:discord@claude-plugins-official
• Model: claude-opus-4-6

Two Distinct Surfaces

Surface 1: Discord triple-reply (with API proxy)

Setup: Claude Code → API proxy (injects persistent memory into system prompt) → Anthropic API

Symptoms:

• Each inbound Discord message triggers 2-4 separate discord__reply tool calls
• Only happens in conversational/idle mode — when Claude Code is busy executing tools (writing files, running commands), it single-replies correctly
• Does NOT happen when hitting Anthropic directly (no proxy)
• The proxy adds ~200-500ms latency per call

Hypothesis: The added latency causes Claude Code's internal processing loop to re-evaluate and re-process the same inbound message across multiple API cycles. Each cycle independently decides to call discord__reply.

Surface 2: Terminal triple-generation (no proxy)

Symptoms:

• Running Claude Code in terminal (no --channels, no proxy, direct to Anthropic)
• Complex or emotionally rich responses sometimes generate 2-3 visible drafts before producing final output
• This costs 2-3x tokens but only produces one visible terminal output
• Pattern: more likely with longer, nuanced responses

Hypothesis: Claude Code's agentic loop re-evaluates mid-generation on complex responses.

Current Workaround

A PreToolUse hook on mcp__plugin_discord_discord__reply that blocks duplicate sends within a 15-second cooldown window per channel:

# ~/.claude/hooks/discord_reply_dedup.py
# Checks if a reply was already sent to the same channel within COOLDOWN_SECONDS
# If yes, returns permissionDecision: "deny"
# If no, records the timestamp and returns permissionDecision: "allow"

This prevents visible duplicates on Discord but does NOT prevent the underlying duplicate API cycles and token billing.

Key Observations

1. Idle vs busy: The bug only fires in conversational mode. When Claude Code is actively executing tools, it processes correctly (single reply per message)
2. Latency correlation: Adding any latency between Claude Code and the API endpoint dramatically increases duplicate frequency
3. Content correlation (terminal): Longer, more complex responses are more likely to trigger re-generation
4. Single process confirmed: This is NOT caused by multiple Claude Code instances — verified via process count and lock file

Expected Behavior

One inbound Discord message → one API cycle → one discord__reply tool call, regardless of API endpoint latency.

Reproduction Steps

1. Run claude --channels plugin:discord@claude-plugins-official with an API proxy that adds 200-500ms latency
2. Send a conversational message in Discord
3. Observe 2-4 duplicate replies

Or for terminal surface:

1. Run claude in terminal (no channels, no proxy)
2. Ask a complex, emotionally nuanced question
3. Observe multiple generation attempts in the output