continuous-review

Pattern for running an independent, sandboxed agent that keeps auditing the main agent's generated code over the course of a session — not a one-off review, a recurring one. Use when a session is prod

SKILL.md

GitHub

Continuous review

The correctness gate (scripts/run-checks.sh) catches mechanical issues. /code-review and the swarm review pattern (.claude/skills/swarm/SKILL.md) catch a single diff's logic/security/simplicity issues on demand. Neither is designed to keep watching while a long session keeps producing more code — by the time you remember to ask for a review, ten more features have landed on top of what you meant to check.

This pattern closes that gap with sandbox-auditor (.claude/agents/sandbox-auditor.md): an agent that runs in its own isolated git worktree, so it can freely inspect and test the repository without any risk of colliding with work still in progress on the main branch, and that tracks its own scope between runs via a local git tag rather than relying on conversation memory.

When to use it

  • A session is autonomously implementing a long list of features or fixes and you want an independent check running alongside it, not just at the end.
  • You're delegating a large chunk of work (via /swarm or a big task list) and want assurance that nothing slipped through while you weren't watching every diff yourself.
  • You want drift detection — not just "is this one diff correct" but "has anything violated a documented decision or invariant since I last checked."

Don't use it as a substitute for the correctness gate (still required before any turn is "done") or for /code-review on a specific PR-sized diff you actually want a thorough one-time pass on — this is for the in-between, ongoing case.

How to run it

One-off audit of everything since the last check:

/audit

(or, without the slash command, ask directly: "spawn the sandbox-auditor agent to review what's changed since the last audit")

Recurring, during a long autonomous session:

/loop 15m /audit

Fifteen minutes is a reasonable default for a session that's actively generating code continuously; lengthen it for slower-paced work. The loop skill handles the actual recurring invocation — this skill and the agent are what it calls each time.

What you get back

A findings report scoped explicitly to "everything since commit X" (or "since the beginning," on a first run), covering: whether the gate is currently clean, correctness/security/simplicity issues in the new code, and anything that drifted from a decision recorded in .claude/knowledge-graph/. An empty findings list is a normal, good outcome — it means the audit ran and found nothing, not that it didn't run.

Why a worktree, and why a tag instead of memory

Isolation via isolation: "worktree" on the Agent tool means the auditor can run the full test suite, check out arbitrary commits, and poke around without any chance of stepping on files the main agent is mid-edit on. The git tag (sandbox-audit/last, local-only, never pushed) is what lets a stateless agent — one with no memory of prior runs — pick up exactly where the last run left off. Don't replace the tag with an in-conversation note or a memory entry; the whole point is that this keeps working correctly even across sessions that share no context.