Series

Claude Harness Modernization

2 posts in this series

A dark technical cover titled I built my harness a doctor, with the deck line: it runs every hook for real, in a sandbox, and checks the exit code, not just the message. A dashed teal sandbox panel on the left, labeled nothing real is touched, carries four tags (throwaway home, allowlist env, no-op shims, canary planted) and a terminal showing the file-guard PreToolUse hook on the Windows PC: input Write .env, stdout blocked with a check mark, exit 1 with a note that Claude Code only blocks a tool call on exit code 2. A large red exit 1 arrow green exit 2 callout says the guard said blocked, then the write went through, above the fix: powershell -Command with exit $LASTEXITCODE. An arrow labeled scan after leads to a Real state panel listing four clean checks: activity log, Homunculus observations, session archive index, vector database. A bottom amber strip shows timeout of 5000, a bar reading I meant five seconds and seconds, not milliseconds, and 83 min before anything stopped the hook.

Claude Code Harness Hooks Audit Python Windows

My Claude Code Harness Grew Up Ad Hoc. So I Built It a Doctor.

Eight months of one-session-at-a-time skills, hooks and agents had never been checked. Part 1 of the Claude Harness Modernization series: the plan, and the audit tool that found a hook that could run for 83 minutes and a Windows guard that never blocked anything.

Chris Johnson·October 21, 2026·18 min read

Dark dot-field graphic titled The 0.2 was mostly robots, with chips reading Claude Harness Modernization and SP1 Memory. The left panel, Automated sessions, shows 445 dim gray squares labeled entrypoint sdk with a large 0 readout: searches in 30 days, since they fire the session-start hook but never touch memory. A dashed divider labeled zoom in leads to the right panel, Interactive, showing 48 glowing teal dots in an 8 by 6 grid, 44 lit and 4 hollow, with a 44 readout: of 48 searched, about 2.6 searches per session against about 58 prompts. A bottom strip titled Recall, rebuilt chains four nodes with arrows: every prompt (keywords only), recall hook (searches for me), allowlist proxy (8 tools, no admin), and one memory server (localhost only). Footer reads CryptoFlex LLC, From the Workshop, with the stack tags Claude Code, MCP, Hooks, SQLite.

Claude Code Memory MCP Hooks Security Harness

Rebuilding My Claude Code Memory: One Server, One Proxy, and a Hook That Does the Searching

The first audit said Claude searched memory 0.2 times per session. Splitting the sessions by who started them changed the design. Part 2 of the Claude Harness Modernization series.

Chris Johnson·October 27, 2026·18 min read