Super User Daily: 2026-04-15
Sunday brought the quietest Claude Code and OpenClaw day in weeks. But the silence on use cases was inversely proportional to the noise on reliability. The real story of April 13 isn't what people built — it's what's breaking.
@oga_aiichiro [Claude Code]
https://x.com/oga_aiichiro/status/2043605855220965729
A Japanese developer ran into a wall trying to make Claude Code call OpenAI Codex via MCP in parallel. The standard stdio transport locks to sequential execution, killing throughput when you need multiple agents coordinating. The fix was switching to Streamable HTTP transport, which lets Claude Code fire off concurrent MCP requests to Codex without waiting for each to finish. The full implementation walkthrough is on Zenn. This is the kind of infrastructure plumbing nobody talks about but everyone running multi-agent setups will eventually need.
@ZhihaoLui38846 [OpenClaw]
https://x.com/ZhihaoLui38846/status/2043612614404575447
Running OpenClaw on a Lazy Cat micro-server with 32GB RAM and 2TB storage. Bought in March, installed OpenClaw, reports it works well out of the box. Still exploring new workflows and hasn't purchased the dedicated compute module yet — currently piping to public cloud for inference. This is the self-hosted OpenClaw setup most hobbyists are actually running: beefy local hardware for the runtime, cloud compute for the heavy lifting.
🗣 User Voice
User Voice
The dominant theme on April 13 was trust erosion. Users are flagging reliability issues across multiple dimensions of the Claude Code experience.
@gagarot200 flagged a dramatic quality drop in Claude Opus 4.6 — BridgeBench hallucination accuracy fell from 83.3% to 68.3% between consecutive test runs, with no official announcement from Anthropic. The user warns that relying on Opus for production work is now risky given silent quality changes while pricing stays the same.
@a_lamparelli reports burning 2% of their usage limit on a single continue command with no tool calls — calling the token consumption "bleeding" that makes Claude Code unreliable for sustained sessions.
@voidfreud identifies memory as the single most critical bottleneck. The frustration is that Claude Code writes memory files as workarounds, but the fundamental context persistence problem remains unsolved.
Community analysis confirmed that Claude Code v2.1.100+ injects approximately 20,000 extra server-side tokens per request — invisible to users but validated via HTTP proxy tests. The suggested fix is downgrading to v2.1.98, with users reporting 30-40% better limit longevity and sharper outputs.
The dominant theme on April 13 was trust erosion. Users are flagging reliability issues across multiple dimensions of the Claude Code experience.
@gagarot200 flagged a dramatic quality drop in Claude Opus 4.6 — BridgeBench hallucination accuracy fell from 83.3% to 68.3% between consecutive test runs, with no official announcement from Anthropic. The user warns that relying on Opus for production work is now risky given silent quality changes while pricing stays the same.
@a_lamparelli reports burning 2% of their usage limit on a single continue command with no tool calls — calling the token consumption "bleeding" that makes Claude Code unreliable for sustained sessions.
@voidfreud identifies memory as the single most critical bottleneck. The frustration is that Claude Code writes memory files as workarounds, but the fundamental context persistence problem remains unsolved.
Community analysis confirmed that Claude Code v2.1.100+ injects approximately 20,000 extra server-side tokens per request — invisible to users but validated via HTTP proxy tests. The suggested fix is downgrading to v2.1.98, with users reporting 30-40% better limit longevity and sharper outputs.
📡 Eco Products Radar
Eco Products Radar
OpenAI Codex — mentioned alongside Claude Code in multi-agent setups. The MCP interop story between Claude Code and Codex is becoming a real workflow pattern.
Copilot — Microsoft integrating OpenClaw-inspired always-on agents into M365 Copilot. Enterprise users may soon get proactive agent behavior natively.
SuperBased — screenshot capture, voice dictation, and token compression plugin for AI coding tools including Claude Code, Cursor, and Cline.
OpenServ — platform for deploying AI agent micro-businesses with built-in payment and discovery.
OpenAI Codex — mentioned alongside Claude Code in multi-agent setups. The MCP interop story between Claude Code and Codex is becoming a real workflow pattern.
Copilot — Microsoft integrating OpenClaw-inspired always-on agents into M365 Copilot. Enterprise users may soon get proactive agent behavior natively.
SuperBased — screenshot capture, voice dictation, and token compression plugin for AI coding tools including Claude Code, Cursor, and Cline.
OpenServ — platform for deploying AI agent micro-businesses with built-in payment and discovery.
Comments