v2.3.12 — External Codebase Packing + chom v2 Rigor

One new skill and one refactor. meow:pack fills a real gap — exporting external repositories as single AI-friendly files for handoff to other LLMs or reviewers. meow:chom adopts selective rigor from ck:xia: 4 user-explicit modes, speed flags, sequential-thinking escalation, and honest boundary rules that correct a previously-fabricated claim.

What Changed

NEW: meow:pack

Pack an external repository into a single AI-friendly file (markdown/xml/json/plain).

• /meow:pack <source> — auto-named output at .claude/packs/{YYYYMMDD-HHMM}-{slug}.{ext}
• Sources: owner/repo, full GitHub URL, or local path (must be outside current git root)
• Formats: markdown (default), xml, json, plain
• --compress — Tree-sitter signature extraction (classes/functions/interfaces, bodies omitted). Ideal for "what's the public API of library X" queries where full files would exceed context.
• Secret scanning on by default (Secretlint, origin sourced from repomix docs). Disable with --no-security-check (explicit flag + emitted warning).
• Self-pack guard — refuses to pack the current git root unless --self is passed. Prevents accidental self-reingest that would burn context.
• No global install required — npx --yes repomix@^1.11 per invocation, caret-pinned.

Use cases:

• Pasting a third-party library into an external LLM (ChatGPT, Gemini, claude.ai web)
• Security audit of vendor/library before adoption
• Research snapshots for offline reading or review handoff

NOT for packing the current project to re-read in the same session — meow:scout is the correct tool for inbound analysis because Explore subagents read files in isolated contexts and return distilled summaries. Pack dumps raw content into the caller's context.

REFACTORED: meow:chom v2

chom preserves its strengths (7-question challenge framework, HARD GATE, 5-type input routing, security anchor) and adopts selective rigor from ck:xia:

• 4 user-explicit modes: --compare / --copy / --improve / --port. chom does NOT auto-derive adaptation depth — user picks explicitly to bias Phase 3 focus.
  • --compare: analysis only (phases 1–3)
  • --copy: transplant-minimal focus
  • --improve: anti-pattern detection focus
  • --port: idiomatic translation focus
  • no flag: full 6-phase workflow, emits Replication Spec without mode declaration
• Speed flags (default off):
  • --lean: skip Phase 1 researcher-agent background gathering (effective for freeform inputs; no-op for git/local/web/image)
  • --auto: auto-approve non-HARD-GATE steps
• HARD GATE non-bypassable: no flag, including --lean or --auto, skips Phase 4 human approval
• Intent detection table — keyword hints (e.g. "port from", "like how X does it") map to suggested mode flags
• Sequential-thinking escalation — for features with ≥3 architectural layers or stateful workflows, chom emits handoff text directing the user to /meow:sequential-thinking, then return to chom. Never auto-invokes.
• Boundary Rules (new section): chom emits handoff text only; does NOT invoke /meow:plan-creator, /meow:brainstorming, /meow:cook, or /meow:sequential-thinking mid-flow. This is chom's design choice for phase ownership, not a MeowKit platform rule.
• Enriched handoff output: Replication Spec now carries challenge-reds summary + risk score. plan-creator decides adaptation depth.
• Migration note: --analyze (v1 default) now aliases the no-flag default with a deprecation notice. Removed in v1.2.

Flow: pack vs chom vs scout

Why This Matters

Two quiet fixes correct false statements in the previous release:

1. "Skills can't call skills" was not a MeowKit rule. Red-team verification showed this claim (previously in chom's Boundary Rules) is unsourced — lessons-build-skill.md explicitly says skills can reference each other. chom's handoff-text-only pattern is a design choice for phase ownership, not a platform restriction. Framing corrected.

2. "Packing burns 40–70% of context" was fabricated. pack SKILL.md now carries honest context-isolation framing: scout wins for inbound analysis because Explore subagents use isolated contexts and return distilled summaries, not because "lazy reading" is inherently better.

Files Changed

File	Change
.claude/skills/meow:pack/	NEW skill (SKILL.md, references/options.md, references/gotchas.md, scripts/self-pack-guard.sh)
.claude/skills/meow:chom/SKILL.md	Refactored (+74/-22 lines) — 4 modes, speed flags, boundary rules, honest framing
.gitignore / .gitignore.meowkit	Added .claude/packs/

Challenge framework unchanged — references/challenge-framework.md untouched.

Breaking Changes

None. All changes are additive. --analyze alias preserved with deprecation notice for one release.

Migration

No action required for existing meow:chom workflows — --analyze continues to work (emits deprecation notice). Re-invoke with --copy, --improve, or --port to bias analysis toward a specific adaptation depth.

Related

• meow:pack reference
• meow:chom reference
• meow:scout reference — correct tool for inbound analysis
• Full changelog