cpv-batch-full-scan-and-fix

/cpv-batch-full-scan-and-fix — Maximum-coverage same-turn sweep

For fleet operators who want the deepest possible parallel sweep across every plugin in a marketplace, this command bundles validate + security + caching audit + caching optimize + FP verification + fix into ONE per-plugin agent turn. Each cpv-plugin-fixer-agent subagent reads each source file ONCE and triggers every applicable in-process checker, classifies findings via the v2.100.x context classifier, verifies uncertain findings via llm-externalizer with file-range syntax (≤ 200 LOC per call), applies confirmed-real fixes inline, then runs a clean-room re-check.

The four separate batch skills run separately would read every source file 4× (once per scan type) plus the fix pass — this command does it in one pass.

Per the iron-rule: same-turn = optimisation. Every finding still walks the full classifier chain; FPs are verified by external LLM before being silenced. Real findings are still fixed.

Same input grammar as the rest of the batch family.

You are the orchestrator

You — the model running THIS turn — drive the batch. You do NOT scan or fix anything yourself.

Step 0 — Resolve arguments

If no target was given, ask plain-text:

What should I full-scan-and-fix? Provide an absolute path, a GitHub URL, a marketplace, or a list file like @/tmp/plugins.txt.

Step 1 — Build the batch plan

# Separate the positional target specs from the --max-parallel flag in a
# single pass. We must NOT collapse to "$1": the shared batch input grammar
# lists a whitespace-separated LIST (`./a ./b ./c`) as a valid shape, and the
# orchestrator's `plan` subcommand declares its inputs as `nargs="+"`, so
# every positional must be forwarded — using only "$1" would silently
# scan-and-fix the first plugin and drop the rest, and would mis-bind a
# leading `--max-parallel` token as the input spec.
#
# Both --max-parallel N (two tokens) and --max-parallel=N (one token) are
# handled, because argparse accepts the `=` form and a user may type it.
# Without this loop the advertised `--max-parallel N` knob is silently
# ignored (MAX_PARALLEL stays at the default 8). Default 8; the orchestrator
# re-caps at 16, so a larger N is safe to pass.
BATCH_SPECS=()
MAX_PARALLEL=8
while [ "$#" -gt 0 ]; do
  case "$1" in
    --max-parallel)
      MAX_PARALLEL="$2"
      shift 2
      ;;
    --max-parallel=*)
      MAX_PARALLEL="${1#--max-parallel=}"
      shift
      ;;
    *)
      BATCH_SPECS+=("$1")
      shift
      ;;
  esac
done

if [ "${#BATCH_SPECS[@]}" -eq 0 ]; then
  echo "ERROR: no target spec given to /cpv-batch-full-scan-and-fix" >&2
  exit 1
fi

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/cpv_batch_orchestrator.py" plan \
  "${BATCH_SPECS[@]}" \
  --agent cpv-plugin-fixer-agent \
  --mode batch_same_turn_full \
  --max-parallel "$MAX_PARALLEL"