harness-configs

Portable AI coding-harness configs — shared skills, commands, rules, agents, and settings — provisioned from one source of truth across multiple Claude Code and pi profiles.

A personal dotfiles-style repo, published so others can borrow the architecture. Paths are hardcoded to one machine — adapt before adopting (see Adopt it).

Ideas worth stealing

• One source of truth, many harnesses. A single skills/ tree feeds Claude Code and pi; edit once, every harness sees it immediately.
• Two Claude profiles via CLAUDE_CONFIG_DIR. pclaude / wclaude wrappers keep personal and work accounts isolated while sharing the same skills and rules.
• Per-resource source tracking. Each vendored skill is pinned in skills/vendored.json (repo, subpath, commit) and materialized on install — so an upstream update is one command and a bare clone stays reproducible.
• Generated catalog, enforced by a doctor. skills/README.md is generated from that metadata; make skills-doctor fails when anything drifts.
• Guardrails as code, not hope. Extensions for pi confirm before destructive actions, block dirty-repo commits, and protect paths; a PreToolUse safety hook does the same for Claude Code (deny/ask on dangerous shell, secrets, and VCS internals); plugins.txt declares the plugin set and make plugins-check reports drift per profile.
• Skills are scanned before they land. make skills-scan runs NVIDIA SkillSpector over every skill, skills-fetch/skills-update refuse a skill that fails, and accepted findings carry a reason in security/skillspector/.
• Measured, not assumed. make context-budget reports the always-loaded context per session (capped by the doctor), and a SubagentStop/Stop telemetry hook plus make usage-report show spend by agent type × model and cache-hit ratio. CI runs the doctor, catalog, script tests, and the security scan on every push.

Layout

harness-configs/
├── Makefile        # primary interface — every <resource>-<action> target
├── CLAUDE.md       # guide for agents working IN this repo (full Makefile reference)
├── claude/         # Claude Code config, symlinked into both profiles
│   ├── CLAUDE.md   #   shared global user instructions
│   ├── commands/   #   slash commands
│   ├── agents/     #   subagents (+ .source.json sidecars)
│   ├── rules/      #   model routing, git hygiene, delegation
│   ├── scripts/    #   hooks (routing, safety guard, usage telemetry) + statusline
│   └── plugins.txt #   desired-state plugin list
├── skills/         # shared skills for Claude + pi — the source of truth
│   └── README.md   #   generated catalog — start here
├── suites/         # curated skill-suite landing pages
├── security/       # SkillSpector baselines (accepted findings, with reasons)
├── pi/             # pi agent config, stowed into ~/.pi
└── scripts/        # repo tooling (not symlinked into profiles)

make install symlinks skills/ and claude/* into both Claude profiles and ~/.pi, and stows pi/ — safe to re-run. Because everything is symlinked, one edit here applies to every profile and both harnesses at once.

What's inside

Skills — browse the catalog (grouped by category, generated from each skill's metadata). Most are authored here; the rest are vendored from pi-skills, cursor-team-kit, anthropics/skills, and the skills.sh ecosystem — each linked to its upstream source at a pinned commit.

Suites — curated skill sets with their own landing pages:

• Go CI Health — Keep a Go repo’s CI green, fast, and secure
• Agent-Maintained Hugo Site — Write, illustrate, verify, optimize, and measure a Hugo blog with an agent
• OSS Maintainer Copilot — Backlog triage, CodeQL remediation, and security advisories for repos you maintain
• PR Shepherding — Get a pull request from pushed to merged: clean diff, green CI, resolved reviews
• Second Brain — An Obsidian knowledge base your agent maintains for you

Agents

Agent	Purpose
plan-reviewer	Pre-execution plan review against the actual codebase; APPROVE/REVISE with evidence.
bulk-mechanic	Haiku executor for mechanical, judgment-free batches.
pr-shepherd	Drives the push → CI → bot-review loop to green.
skill-auditor	Audits a skill directory against this repo's conventions.

Rules

Rule	Governs
model-routing.md	Cheapest reliable model tier per task, and effort calibration.
git-hygiene.md	Staging, commit, and push discipline.
delegation.md	When to hand work to the agents above instead of doing it inline.

Plus the shared CLAUDE.md (secrets hygiene, semantic-line-break markdown per sembr.org), the /history command, helper scripts, and pi guardrail extensions vendored from pi-mono.

Adopt it

Prerequisites: git, GNU stow, jq (plus gh / python3 / npx for the skills that use them).

git clone https://github.com/sanketsudake/harness-configs.git
cd harness-configs
# Before installing:
#   1. Edit CLAUDE_CONFIG_DIRS in the Makefile (your profiles)
#   2. Copy the pclaude/wclaude snippets from scripts/claude-multi-account.sh into your shell profile
#   3. Make claude/CLAUDE.md yours — it's opinionated
make install      # symlink claude/ + skills/ into profiles, stow pi/ into ~/.pi
make skills-list  # see each skill's source and status

make uninstall reverses it.

Working with it

Everyday targets — CLAUDE.md carries the full <resource>-<action> reference, and every skills-* target has an agents-* twin.

Target	Does
install / uninstall	Link/stow everything into the profiles, or reverse it.
skills-find / skills-add	Discover skills on skills.sh and vendor them.
skills-catalog / skills-doctor	Regenerate the catalog / validate every skill and its freshness.
skills-update[-all]	Re-fetch vendored skills whose upstream moved.
plugins-check / plugins-sync	Report plugin drift per profile / emit the /plugin install lines.

Two gotchas: vendored skills/ and pi/extensions/ are overwritten on re-sync — diverge intentionally and note it durably; and use SUBPATH=, never PATH=, on fetch targets (the latter clobbers the shell PATH). Plugin installation stays manual per profile — Claude Code has no headless /plugin install.

License

Apache-2.0. Vendored skills and extensions remain under their upstream licenses; see each resource's source metadata.