hrmtz/claude-harness

Three defense-in-depth hooks: credential value scrub (PostToolUse), admission keyword reminder (UserPromptSubmit), bash command guard (PreToolUse). Built after repeated credential leak incidents — structural fix when behavioral rules in CLAUDE.md keep failing.

Peer-pane Claude Code worker orchestration via tmux + jsonl mailbox. Spawns long-running coordinated workers from natural-language triggers. Bundles formation skill + njslyr7 CLI + UserPromptSubmit hook that auto-suggests the skill when user prompts match worker-spawn intent. Session-scoped identity prevents mailbox drift across resume/compact.

Three-perspective preflight review skill (MELCHIOR/BALTHASAR/CASPAR personas, parallel Task spawn) for high-stakes changes — front-loads post-hoc optimization insights before walltime / cost / destructive state is committed. Triggers on walltime >= 2h, >= 100M row DML, non-reversible cutover, new pipeline layer, >= $10 spend, or long-poll scripts.

Operational safety rails for long-running operations. Pre-flight algorithm fitness CLI (working set vs RAM ceiling, with concrete alternatives), in-flight heartbeat helper + cron watcher (stale + ETA overrun detection), Discord + gh issue auto-emit. Human-in-loop only — no auto-kill. Built after a 23h sunk-cost incident where philosophy-level early-bleeding-detection lived only in CLAUDE.md memory, not in structural rails.