Gated Dual Clone Audit

Gated Dual-Clone · Audit (independent evaluator)

Why this skill exists · 为什么独立出来

EN — Two rules carried over from Anthropic's harness design for long-running apps and the GAN paradigm:

1. The agent doing the work praises its own work. gated-dual-clone's bootstrap.sh runs three gate checks as its final step, but once it exits it has no further interest in the topology. It will not notice that the pre-push hook file was deleted last week, or that user.email drifted back to the global default.
2. An independent, skeptical evaluator is the real lever. This skill imports nothing from gated-dual-clone. It only reads the output topology and checks that the safety gates still hold. Same independence rule design-review holds against the 9 design skills.

ZH — 两条原则来自 Anthropic harness-design 文章和 GAN 范式:(1) 做事 的 agent 称赞自己的作品;(2) 独立持怀疑的 evaluator 才是真正的杠杆。 本 skill 不依赖 gated-dual-clone,只读拓扑成品,验三道检查。

What the audit checks · 审计覆盖

Three tiers run in order. Each tier short-circuits on failure (no point checking user.email if .git/ is missing):

Tier 1 · Structural (filesystem, no git state read)

• S1 · gateway directory exists and is a directory
• S2 · gateway is a git repo (.git/ present, git rev-parse succeeds)
• S3 · satellite directory exists
• S4 · satellite is a git repo
• S5 · pre-push hook file present at gateway/.git/hooks/pre-push
• S6 · pre-push hook has execute bit
• S7 · pre-push hook body contains the protected-branch regex placeholder (rough content check — substance verified in Tier 3)
• S8 · .git/objects hardlink spot-check — sample a loose object, verify inode match between gateway and satellite

Tier 2 · Configuration (read git config only)

• C1 · satellite remote.origin.url points at gateway path (file:// or absolute local path)
• C2 · satellite remote.origin.pushurl == DISABLED (exact match)
• C3 · gateway remote.origin.url is present (non-empty)
• C4 · gateway user.email set locally (not falling back to global)
• C5 · gateway user.name set locally
• C6 · satellite user.email set locally
• C7 · satellite user.name set locally
• C8 · satellite current branch is the expected push branch (if --push-branch given)

Tier 3 · Behavioural (safe calls only — no real push, no state change)

• B1 · cd satellite && git push is rejected (greps for "DISABLED" / "does not accept" / "unable")
• B2 · gateway's pre-push hook is invoked directly with synthetic stdin (format per githooks(5): local_ref local_sha remote_ref remote_sha) and must reject a push to the protected branch (greps for "REJECTED" / "protected" / "hook declined"). We do NOT use git push --dry-run here — when local = remote (a fresh bootstrap, or a push-branch just branched off upstream) git short-circuits with "Everything up-to-date" and never runs the hook, silently masking a broken hook as "fine".
• B3 · cd satellite && git fetch origin <branch> --dry-run succeeds (fetch path from gateway is reachable)

Tier 4 · Taste (optional, via LLM subagent)

Runs an LLM subagent to answer questions the machine can't:

• Is $PUSH_BRANCH still based on $UPSTREAM_BRANCH, or has it diverged past easy rebase?
• Has the project shrunk to the point where a single repo would be cleaner?
• Are --no-verify bypasses showing up in .git/logs?

Two entry points for Tier 4:

1. /gdc-audit-critic slash command (recommended) — one shot: builds the prompt, dispatches the subagent in a fresh context, and reports the verdict. Accepts the same --gateway-dir / --satellite-dir / optional --upstream-branch / --push-branch arguments.

   /gdc-audit-critic --gateway-dir=<path> --satellite-dir=<path> \
                     --upstream-branch=<name> --push-branch=<name>
   

2. Manual path — bin/gated-dual-clone-audit --critic invokes scripts/audit-critic.mjs to produce a self-contained prompt file; the wrapper prints the file path and you invoke the subagent yourself in Claude Code via Task(subagent_type="gated-dual-clone-audit-critic", prompt=<contents>). Useful when scripting / CI.

Tier 4 does not affect the wrapper's exit code; its verdict is advisory.

Entry point · 入口

bin/gated-dual-clone-audit \
  --gateway-dir=<path> \
  --satellite-dir=<path> \
  [--upstream-branch=<name>] \
  [--push-branch=<name>] \
  [--strict] [--json] [--critic]

Exit code convention:

Output: human-readable table by default · --json produces machine JSON (suitable for feeding into design-review's learning-loop.mjs when a drift is caught in the wild).

Use cases · 使用场景

On demand before a risky push

bin/gated-dual-clone-audit \
  --gateway-dir=~/projects/foo-work \
  --satellite-dir=~/projects/foo-verify \
  --upstream-branch=release

As a git hook (run before every push)

# in gateway/.git/hooks/pre-push, ADDED TO the hook that bootstrap.sh installed:
bin/gated-dual-clone-audit \
  --gateway-dir="$PWD" \
  --satellite-dir="$SATELLITE_DIR" \
  --strict

As a scheduled cron check

# weekly drift check
0 9 * * 1 cd /path/to/sky-skills && \
  bin/gated-dual-clone-audit --json \
  --gateway-dir=/home/alice/foo-work \
  --satellite-dir=/home/alice/foo-verify \
  > /var/log/gdc-audit-$(date +\%Y\%m\%d).json

Deliberate non-goals · 故意不做的事

• Not a setup tool. If the topology is missing, don't build it — use gated-dual-clone/scripts/bootstrap.sh. This skill assumes the topology exists and checks it.
• Not a fixer. Gate failures print diagnosis + suggested remediation; the user applies the fix.
• No imports from gated-dual-clone. Path conventions are shared (--gateway-dir / --satellite-dir flag names), nothing else.

Files · 文件

• bin/gated-dual-clone-audit — wrapper, runs tiers + aggregates output
• scripts/audit-structural.sh — tier 1 (filesystem / hook / hardlink)
• scripts/audit-config.sh — tier 2 (git config)
• scripts/audit-behavioural.sh — tier 3 (safe --dry-run tests)
• scripts/audit-critic.mjs — tier 4 prompt builder for the LLM subagent
• agents/gated-dual-clone-audit-critic.md — Claude subagent definition used by --critic
• references/gate-catalog.md — per-gate reference · what each gate checks, common failure modes, fixes
• references/known-drifts.md — drift patterns caught in the wild · entry point for learning-loop feedback · template for adding new entries
• references/troubleshooting.md — task-oriented "audit says X · what do I do"

Reference · 参考

• Anthropic · harness design for long-running apps — "separate doing from judging"
• Design spec: docs/design-mr-gated-dual-repo.md §11
• Peer generator: skills/gated-dual-clone/