claude-code-harness

Plan. Work. Review. Ship.
Turn Claude Code into a disciplined development partner.

English | 日本語

---

v4.2 Update — Claude Code 2.1.99-110 + Opus 4.7

Hokage line carries forward. Full Claude Code 2.1.99-2.1.110 + Opus 4.7 integration. Plugin manifest now matches the official plugins-reference schema.

Anthropic shipped Opus 4.7 with literal-instruction-following semantics, plus Claude Code 2.1.105 added the PreCompact hook and the monitors/monitors.json manifest. v4.2 makes Harness fit those changes precisely:

Area	Before (v4.1)	After (v4.2)
Long-running workers	Could be silently compacted mid-task	PreCompact hook blocks compaction while a worker is active or Plans.md is dirty
Plugin validation	claude plugin validate rejected monitors/agents blocks	Public-spec compliant: monitors/monitors.json + agents/ auto-discovery
Sync regression	harness sync could silently strip declared blocks (4 prior incidents)	Two-layer guard: shell idempotency test + Go struct test for phantom fields
Long sessions	Default 5-minute prompt cache only	bash scripts/enable-1h-cache.sh opt-in to 1-hour TTL (CC v2.1.108)
Reviewer/Advisor effort	medium / high	xhigh (CC v2.1.111, Opus 4.7) — sharper review, with high fallback for other models
Agent prompts	Worked on Opus 4.6 implicit semantics	Re-tuned for Opus 4.7 literal instruction following — explicit thresholds, schemas, command names
Guardrails (R01-R13)	Conformed to CC 2.1.98 contract	Re-conformed to CC 2.1.110 (PermissionRequest updatedInput, PreToolUse additionalContext, Bash bypass closures) — 17 new regression tests

What you'll notice:

• Long-running tasks no longer get cut off mid-flight by automatic compaction
• claude plugin validate runs clean on Harness for the first time since monitors was added
• harness sync stops mysteriously deleting your monitors/agents block
• Reviewer agent gives sharper feedback (xhigh effort) when running on Opus 4.7

Update with the usual flow:

/plugin update claude-code-harness

For detailed Before/After in Japanese, see CHANGELOG.md [4.2.0] entry.

---

v4.0 "Hokage" — What's New

Go-native engine. 25x faster hooks. Zero Node.js dependency.

Every tool call Claude makes passes through Harness hooks. In v3, each pass cost 40-60ms of bash + Node.js overhead — a subtle drag you feel across hundreds of calls per session. v4 replaces the entire stack with a single Go binary:

	v3 (bash + Node.js)	v4 "Hokage" (Go)
PreToolUse	40-60ms	10ms
SessionStart	500-800ms	10-30ms
PostToolUse	20-30ms	10ms
Node.js	Required (18+)	Not needed

What you'll notice:

• The micro-pauses between tool calls disappear — Claude feels more responsive
• No more npm install or Node.js version issues on setup
• Session startup is instant (10-30ms vs almost a second)
• Optional harness-mem integration: sessions remember what you worked on last time

Just update the plugin — no configuration changes needed:

/plugin update claude-code-harness

---

Hokage Core Extraction Status

Claude Code Harness remains a Claude-first product. "Hokage" currently names the v4 Go-native runtime line, not a separate product claim.

Hokage Core extraction underway: the shared workflow contract is being separated from host-specific adapters, while public support claims stay limited to the gates that are already proven. No public Hokage Harness spin-off is claimed until the Claude/Codex/OpenCode readiness gates pass.

Current gate status: Hokage Spin-Off Readiness.

---

Why Harness?