Problem

just ci::apm-sync depends on the apm recipe, which implements its "clean sync" by deleting and recreating the APM-owned tree on the live .claude/ directory:

apm:
    find .claude -mindepth 1 -maxdepth 1 ! -name launch.json -exec rm -rf {} +
    {{ apm_cmd }} install

Between the find ... -exec rm -rf and apm install steps (and for the duration of the reinstall), files under .claude/hooks/, .claude/settings.json, .claude/commands/, .claude/rules/, and .claude/skills/ do not exist on disk.

If a developer is running Claude Code in the same worktree while just ci is running, and Claude Code's stop hook fires during that window, the hook fails with:

Stop hook error: Failed with non-blocking status code: /bin/sh: line 1:
.claude/hooks/agency/scripts/do-stop-guard.sh: No such file or directory

I hit this today while /do was running just ci in the background — the stop hook fired at a turn boundary during the apm-sync step and couldn't find its own script.

Why this is fragile

apm-sync is supposed to be a verification step ("does the vendored tree match the sources?") — but it mutates the live tree to perform that verification. Any concurrent reader of .claude/ during CI sees a half-built filesystem. That includes:

• Claude Code sessions running in the same worktree (stop hooks, pre-tool hooks, skill loads, .claude/rules/*.md reads)
• Editors/LSPs that watch .claude/
• Other just recipes that expect .claude/ to be stable

Suggested fix

Perform the wipe-and-reinstall in a scratch directory, then compare against the live tree without touching it:

apm-sync: apm-audit
    #!/usr/bin/env bash
    set -euo pipefail
    scratch=$(mktemp -d)
    trap 'rm -rf "$scratch"' EXIT
    cp -a .claude/launch.json "$scratch/" 2>/dev/null || true
    (cd "$scratch" && {{ apm_cmd }} install)
    diff -r .claude "$scratch" || {
        echo "ERROR: .claude/ out of sync with .apm/ — run: just ai::apm"
        exit 1
    }

This keeps the CI check's semantics (fail if the live tree diverges from what apm install would produce) without ever wiping the live tree. The just ai::apm developer recipe can keep its destructive behavior — that's an explicit local-dev action, not a CI check.

Alternatively, narrow the window: stage a fresh tree under .claude.new/, then do an atomic mv swap so no single file is ever missing for more than a few milliseconds. That's simpler but still briefly inconsistent (stop hook fires between the swap steps).

Related

• f5dbe1a — the commit that introduced the wipe-and-reinstall pattern (for orphan detection)
• Architecture: Unify integration dispatch, result types, and hook dedup microsoft/apm#561 — upstream lacks --prune, which is why we wipe locally