This repo is a workspace for building and maintaining reusable agent harnesses (Claude Code + Codex), studying strong reference harness repos, and keeping the resulting learnings in one place. It is a meta-repo — there is no first-party application code.
mvp-harness/— the harness plugins under active development:mvp-plugin/— the reusable harness installer./mvp-plugin:adoptcopies the.claude+.codexsetup, rules, hooks, and beads tracking into any repo.codex-adapter/— call OpenAI Codex (gpt-5.x) from Claude Code viacodex exec.code-intel/— graph-first code intelligence plugin (serena + CBM + ast-grep).
harness_learnings/— the synthesized canon, collaboration notes, and reference repo learnings.reference_harnesses/— third-party harness repos tracked as git submodules (read-only references).harness_lifecycle/— curation tooling for the reference harnesses: capability scanner, gap report, and the adoption ledger..claude/+.codex/— the harness installed into this repo itself (rules, skills, agents, commands, hooks, and theproject/overlay of repo facts)..beads/— Beads issue tracker store (see.beads/beads.md).
harness_learnings/coding-harness-best-practices.mdharness_learnings/claude-codex-collaboration.mdharness_learnings/reference-harness-workflow.md.claude/project/brief.md— repo facts for agents
Start in mvp-harness/plugins/mvp-plugin/ and use harness_learnings/ as the design
reference.
Follow harness_learnings/reference-harness-workflow.md.
Run /harness-status (what changed upstream since our pins) then
/harness-scan <repo> (what it ships that we don't), and triage candidates with
the harness-evaluate skill. Deterministic tooling lives in harness_lifecycle/
(scan.py, gap.py).
Reference repos stay under reference_harnesses/ as git submodules — clean
external references, never copied into the local harness. Borrow only the
smallest durable pattern that improves the harness.