folder-audit

folder-audit

What it does

The folder-audit pipeline inspects a target directory, builds a tree of folders up to a configurable depth, and asks an LLM auditor to classify every immediate child as fit, misplaced, duplicate, naming_mismatch, etc.

It emits two artifacts in the plan directory:

• audit.json — machine-readable audit with summary and per-folder entries
• audit.md — human-readable nested annotated tree

Running it

The pipeline is built around the native Arnold AgentStep. The audit stage injects a worker callable that dispatches to Codex by default. A local profile is provided at profiles/standard.toml.

Default mode (Codex worker)

python -m arnold run folder-audit \
  --inputs target_dir=/path/to/repo,max_depth=2,chunk_size=5,max_workers=3 \
  --plan-dir /tmp/folder-audit-plan

The default profile resolves the audit stage to codex. If you want to use a different model, override the profile:

python -m arnold run folder-audit \
  --inputs target_dir=/path/to/repo,max_depth=2,chunk_size=5,max_workers=3 \
  --profile @folder-audit:standard \
  --plan-dir /tmp/folder-audit-plan

Using pre-generated subagent results

For faster iteration or when you want to supply audits from another agent harness, pass an agent_results JSON file:

python -m arnold run folder-audit \
  --inputs target_dir=/path/to/repo,agent_results=/tmp/agent_results.json \
  --plan-dir /tmp/folder-audit-plan

The agent_results file must be a JSON object with a top-level folders array. The pipeline will compute a summary and reconcile the entries with the real filesystem tree when rendering audit.md.

Inputs

Key	Default	Description
target_dir	required	Directory to audit
max_depth	8	How many levels deep to walk
chunk_size	5	Folders per Codex call
max_workers	3	Parallel Codex calls per level
agent_results	none	Path to external audit JSON