agents

a planning and review loop for claude code: brainstorm, spec, plan, execute, review.

• // spec and plan live in the repo, so the agent and the human read the same document.
• // each plan task runs in a fresh subagent, then two read-only reviewers check it against the spec and for quality.
• // tdd is the default, with a list of testing anti-patterns the agent is told to avoid.

a second brain for claude code: markdown notes, indexed by qmd, searched before every plan.

• // notes are plain markdown in an obsidian-friendly tree. no database, no lock-in.
• // search runs through a bundled qmd mcp server: bm25, semantic, or hybrid, with the cli as fallback.
• // hooks keep the index fresh and nudge the agent to search or record at the right moments.

skill observability: watch skills run, review each run from the live transcript, fold the reviews back into the skill.

• // a mod, not a shell script: function hooks that run inside claude code with no background process.
• // reviews reuse the session's warm prompt cache, so each one costs about its own output.
• // changing a skill is always human-gated. the plugin proposes diffs, you approve them.

guardrails for one repo: hooks that encode its recorded failures, a router to its own procedures.

• // every deny maps to a protection rule in the repo's agents.md or to a dated incident in its session notes.
• // inert outside the repo, so it stays installed at user scope and costs no context.
• // procedures stay in the repo and are shared with other harnesses; the plugin only routes to them.