Frequent streaming stalls and ECONNRESET during complex tasks

Problem

During complex, long-running tasks, Claude Code frequently experiences streaming stalls and ECONNRESET errors, causing session interruptions and workflow disruption.

Environment

• Claude Code version: 2.1.70
• OS: WSL2 (Linux 5.15.167.4-microsoft-standard-WSL2) on Windows
• Model: claude-opus-4-6 (1M context)

Symptoms

1. Streaming Stalls (most frequent)

From debug logs (~/.claude/debug/), over 120+ streaming stall events observed across sessions from Feb 25 to Mar 9:

[WARN] Streaming stall detected: 367.5s gap between events (stall #1)
[WARN] Streaming stall detected: 282.2s gap between events (stall #1)
[WARN] Streaming stall detected: 272.0s gap between events (stall #1)
[WARN] Streaming stall detected: 252.8s gap between events (stall #1)

• Stall duration ranges from 30s to 367s (up to ~6 minutes)
• Some sessions have 30+ stalls in a single conversation (e.g., session 8037cf20 on Mar 5)
• Stalls are especially frequent during extended thinking with large context windows

2. API Connection Errors

[ERROR] API error (attempt 1/11): undefined Connection error.
[ERROR] API error (attempt 2/11): undefined Connection error.

3. ECONNRESET

Unable to connect to API (ECONNRESET)

This happens mid-conversation during complex tasks, forcing the user to resume or restart.

Impact

• Complex refactoring tasks that require large context are frequently interrupted
• Users cannot complete multi-step tasks without repeated disconnections
• No way to configure timeout or keep-alive behavior to mitigate the issue

Feature Request

1. Configurable API timeout / keep-alive settings — Allow users to configure streaming timeout, connection timeout, and keep-alive intervals via settings.json or environment variables (e.g., CLAUDE_API_TIMEOUT, CLAUDE_STREAM_TIMEOUT)
2. Better stall recovery — When a streaming stall is detected, automatically attempt to reconnect and resume rather than hanging indefinitely
3. Stall threshold configuration — Allow users to set the threshold for when a stall triggers a reconnection attempt

Reproduction

1. Start a complex task with large context (e.g., multi-file refactoring with extended thinking)
2. Observe streaming stalls in debug logs after ~30-60 seconds of inactivity
3. Eventually encounter ECONNRESET or indefinite hang

Workarounds Attempted

• claude --resume after disconnection
• Reducing thinking depth (/think low)
• Splitting tasks into smaller conversations
• None of these fully resolve the issue for complex tasks that inherently require large context