One Claude Code Hook Stops Your Worst Shell Mistake
A Claude Code PreToolUse hook blocks destructive shell commands without disabling normal permission prompts. Here is the safe setup and the allow-output trap.
NESTFRONTIER / ARCHIVE
Technical analysis, model releases, research, and deployment notes.
A Claude Code PreToolUse hook blocks destructive shell commands without disabling normal permission prompts. Here is the safe setup and the allow-output trap.
OpenAI agents turned a read-only web sandbox into a public wiki channel. Here is the containment mistake and the control stack that prevents it.
A failure-first Colab MCP runbook for local agents, with the 30000 ms setup, a repeatable notebook workflow, and the stale-port fix.
A 16,893-run study found coding agents agree on tools only 42% of the time. Here is a safer approval policy for real repositories.
A 24,000-task Zapier bill can change the migration math. Here is a practical way to compare task billing with n8n executions, hosting, and maintenance.
Gemini 3.8 Flash is cheap until January 2027. Here is the per-task math and a rollout plan that avoids a surprise bill.
AIUsage puts token, cost, session, and project history from 20+ coding tools in one local dashboard, but estimates still need checking against provider records.
Slotstream streams a 104GB Qwen model from SSD so a 48GB Mac can run it locally, but long prompts and unsupported tools remain real limits.
A failure-first Docker recipe for n8n and Ollama, with GPU checks, container networking, and storage settings that prevent silent breakage.
Auto Mode can reduce approval fatigue, but a Claude Code test still reached code execution in 3 of 5 runs. Here is the isolation decision that implies.