Problem

apm install always writes relative to $PWD. There's no flag to redirect deployment to a different root.

This forces scratch-dir verification workflows (see #684) to physically stage every input file in the scratch tree and cd into it, rather than running apm install --root <scratch> against inputs that stay in place.

Concrete workaround in the wild

juspay/kolu runs apm install in a scratch directory as part of just ci — the motivating use case on #684. The workaround currently spans ~20 lines of a justfile recipe (agents/ai.just:49-77):

scratch=$(mktemp -d)
trap 'rm -rf "$scratch"' EXIT
rsync -a \
    --exclude='.git/' \
    --exclude='node_modules/' \
    --exclude='.direnv/' \
    --exclude='.logs/' \
    --exclude='.claude/' \
    --exclude='.codex/' \
    --exclude='.agents/' \
    --exclude='.opencode/' \
    --exclude='CLAUDE.md' \
    "$PWD/" "$scratch/"
mkdir -p "$scratch/.claude" "$scratch/.codex" "$scratch/.opencode" "$scratch/.agents"
[[ ! -e .claude/launch.json ]] || cp .claude/launch.json "$scratch/.claude/launch.json"
[[ ! -e .codex/config.toml ]] || cp .codex/config.toml "$scratch/.codex/config.toml"
(cd "$scratch" && apm install >/dev/null)
(cd "$scratch" && apm compile --target codex,opencode >/dev/null)

Every line between scratch=$(mktemp -d) and (cd "$scratch" && apm install ...) exists only because apm install can't take a target directory. With a --root flag this collapses to roughly:

scratch=$(mktemp -d)
trap 'rm -rf "$scratch"' EXIT
apm install --root "$scratch"
apm compile --root "$scratch" --target codex,opencode

Why --root (not cd-based workaround) is the right primitive

• No input duplication. The current workaround copies the entire worktree into scratch (7 exclude rules, a pair of cp fallbacks for user-owned runtime files, and mkdirs to satisfy target auto-detection). A --root flag keeps sources in place — apm.yml, the lockfile, and local-path packages all resolve from $PWD while writes land under --root.
• Uses beyond verification. Scripted setup scenarios (bootstrap scripts, fixture generation for tests, CI dry-runs) all benefit from installing into a specific directory without chdir gymnastics.
• Matches precedent. pip install --target, npm install --prefix, cargo install --root — redirecting output root is a standard package-manager affordance.

Scope

Two sub-questions the implementation will need to resolve:

1. Which paths does --root redirect? I'd expect: deployed-file writes (.claude/, .codex/, .agents/, .opencode/), apm_modules/, and the lockfile. Sources (apm.yml, local-path package roots) should continue resolving from $PWD unless a separate --manifest flag is added.
2. Interaction with apm compile. compile has the same scratch-dir use case (see the (cd "$scratch" && apm compile ...) line above). Worth applying the same flag to compile and any other command that writes to the working directory.

References

• Context: audit --ci: verify deployed file content, not just existence #684 (content verification in audit --ci — Sergio + Daniel both mentioned --root in their triage replies)
• Reference implementation of the workaround: ci(apm-sync): verify .claude/ without wiping it juspay/kolu#469
• Workaround source: agents/ai.just:49-77
• apm 0.9.2 install entry points that assume Path.cwd() / get_deploy_root(scope): src/apm_cli/commands/install.py:1191,1271,1283,1424

Environment

• apm 0.9.2 (via uvx --from git+https://github.com/juspay/apm apm — pinned to juspay fork for an unrelated fix)