Harness

Automated fix loops. Each harness runs a make target, spawns sub-agents to fix failures, commits on success, reverts on exhaustion.

Pattern: .claude/harness/<name>/run + fix.prompt.md

Escalation: sonnet:think → opus:think → opus:ultrathink

History: Each harness maintains history.md for cross-attempt learning. Truncated per-file, accumulates across escalation. Agents append summaries after each attempt so higher-level models can avoid repeating failed approaches.

Quality Checks

The check-* namespace is reserved for quality checks. Only these scripts use the check- prefix. Each has a corresponding fix-* script that spawns sub-agents to fix failures.

Core Quality Checks

The 6 core quality checks are the default exit gate for all work. Run these in order when work is complete.

Full Quality Suite

The full quality suite is the 6 core quality checks plus these 5 additional checks. Run only when explicitly requested.

Other Harnesses

Not part of the quality suite. Do not use the check-* prefix.

• prune — dead code detection (.claude/scripts/prune)
• notify — push notifications via ntfy.sh
• pluribus — multi-agent orchestration
• reset-repo — reset jj working copy to fresh state

CLI

• .claude/scripts/check-<name> — symlinks to quality check harness run scripts
• .claude/scripts/fix-<name> — symlinks to fix harness run scripts
• .claude/scripts/<name> — symlinks for non-quality harnesses

Development Inner Loop

After changing a file, run the relevant check with --file=PATH on that file:

• check-compile --file=PATH after every edit
• check-unit --file=PATH when a test file exists
• Other checks as relevant to the change

Stay in this single-file loop. It takes seconds. Do not run project-wide checks during active development.

Running check-* Scripts

All check scripts are on PATH via .claude/scripts/.

• Single file: check-compile --file=PATH — scopes the check to one file. Use this during development for fast feedback.
• Project-wide: check-compile (no args) — checks everything. Use this as the exit gate when work is complete.
• Timeout: Use 60 minute timeout (timeout: 3600000)
• Foreground: Always run in foreground (never use run_in_background)
• Blocking: No output until completion — do not tail or monitor, just wait
• Output format: Structured JSON: {"ok": true/false, "items": [...]}