v3.12.1

Refactored

• Refactor(command): tighten search-existing-implementations spec

Revisions after user feedback on the v3.12.0 draft:

Inputs — all four are now MANDATORY:

1. Quoted feature description (first $ARGUMENTS token)
2. Source file(s) (positional, 1+)
3. --diff (now mandatory, was optional)
4. --in (now mandatory, was optional; defaulted to cwd)
   Supports multiple paths via repeated flag or comma-separated
   list. Each entry can be a directory (walked) or a single file

LLM output — drastically simplified:

• One line per finding: NO or YES symbol=<name> lines=<a-b>
• Max 5 YES lines per file if multiple matches
• Special: NO (self-reference) when the LLM recognises the PR
  file itself
• No STATUS categories (EXISTS/SIMILAR/HELPER dropped)
• No RATIONALE field, no REUSE_PATH field
• Ensemble mode trusted for false-positive filtering —
  disagreements between the 3 models are the reviewer's signal

Forwarded options (same as every other LLM Externalizer command):
--free → pass through to code_task as free: true
--output-dir → pass through as output_dir
--exclude-dirs → applied during target filtering
--redact-regex → pass through as redact_regex

Architecture (unchanged):

• instructions_files_paths carries sources + diff (server reads
  them once, orchestrator never loads file contents)
• input_files_paths is the filtered codebase list (Glob + dedupe
  • exclude source files + exclude non-code dirs)
• Auto-batching by the server keeps request count low inside
  max_payload_kb
• answer_mode: 0 → one .md report per input file, each report
  has one section per ensemble model

Verified: claude plugin validate . ✓, CPV remote validation ✓
(CRITICAL=0 MAJOR=0 MINOR=0).