Dokkabei97/dokkabei-harness

# dokkabei-harness
Claude Code plugin marketplace. It started as a collection of sub agents, but now
it bundles agents, commands, skills, and hooks together into 12 plugins under plugins/,
installed and distributed via .claude-plugin/marketplace.json.
## Installation
/plugin marketplace add
/plugin install @dokkabei-harness
Example: /plugin install base@dokkabei-harness, /plugin install search@dokkabei-harness.
## ★ Featured: observe — the self-measuring harness
Most plugins do work. observe watches the harness itself — which skills and agents actually
fire, why, and how their turns end (committed, cancelled, corrected) — and turns that into concrete improvements.
> Claude Code's native /usage shows cost and tokens. observe shows the layer above it — which
> skills and agents fire, why, how their turns end, and which go stale or uncorrected — the
> harness-health view that cost metrics can't give you.
Inspired by Nous Research's Hermes Agent: where Hermes grows by autonomously creating skills
from experience, observe takes the evaluation-first path — it measures and proposes. It ports the
deterministic half of Hermes's curator (usage lifecycle: stale / archive candidates; correction
signals) and surfaces dead/stale assets, missed activations, and uncorrected skill runs — then drafts
the fix. You approve; it never edits a file silently.
Use it in three steps:
sh
# 1. Turn on tracing (opt-in — records prompt text locally, so it's off by default)
export OBSERVETRACE=1
# 2. Work normally — hooks log prompt → skill → agent → result → turn outcome → session
# into /.tracer/ (self-gitignored). Codex runs the same hooks; Antigravity and
# past Codex sessions come in via observe-backfill.js; git hooks are opt-in
# 3. Ask for the report: deterministic aggregation → LLM interpretation → proposals
/observe-report # unused & stale/archive candidates · missed activations
# correction signals · description-tuning drafts
# then asks you to accept / reject / snooze each recommendation
### Local observability — no SaaS
There is no hosted service; everything runs on your machine. observe's traces live in the repo's
git-ignored .tracer/ folder (prompt text only in raw/, after secret redaction), and nothing is uploaded or sold. OTel (Claude Code's built-in telemetry) is not supported — observe
neither reads nor exports it.
## Structure
text
├── .claude-plugin/marketplace.json # Marketplace registration surface (12 plugins)
├── plugins/ / # Each plugin
│ ├── .claude-plugin/plugin.json # Manifest (name/description/version)
│ ├── agents/ # Sub agents
│ ├── commands/ # Slash commands (legacy format)
│ ├── skills/ /SKILL.md # Skills (recommended format)
│ ├── hooks/hooks.json # Hook registration surface
│ └── bin/hooks/.js # Hook scripts (Node)
├── claude/ # Distribution for Claude Code root settings (CLAUDE.md·settings.json·output-styles)
├── archive/ # Retired plugins; archive/components/ / = retired individual components
└── .github/ # Issue / PR templates (no CI workflow)
## Plugin list
### Foundation·harness building
base — 15 common guard hooks (security warning warn-security, format/compile checks, tmux enforcement, etc.) + LSP. The foundation layer of the loop harness.