Chameleon

Changed

• Plugin surface aligned with current Claude Code best practice (audited
  against the community best-practice reference at
  shanraisshan/claude-code-best-practice and Anthropic's official plugin
  conventions).
  • Skill frontmatter modernized: the six argument-taking commands
    (deep-work, explain, pr-review, status, journey, teach) now
    declare argument-hint for slash-menu autocomplete, hinting only flags
    that exist today (--shadow for status; journey's --list /
    --dry-run / --max-budget-usd; teach's future --deprecate is
    deliberately not hinted).
  • using-chameleon is now marked user-invocable: false +
    disable-model-invocation: true: it is background knowledge the
    SessionStart hook injects verbatim, so listing it in the / menu and the
    model's skill listing duplicated content that was already in context. The
    injection path reads the raw file and is unaffected. On an older CLI the
    fields degrade to no-ops; nothing stops loading.
  • Stop/SubagentStop hook entries carry an explicit "timeout": 60,
    documenting the contract the stop-backstop's internal 55s SIGKILL budget
    is designed against instead of relying on the harness default.
  • CLAUDE.md trimmed to under 200 lines (the adherence ceiling the
    best-practice guide recommends): the full environment-variable operator
    reference and the effectiveness-eval guide moved verbatim to
    .claude/rules/environment-variables.md and
    .claude/rules/effectiveness-eval.md, lazy-loaded by path (mcp/**,
    hooks/**, bin/**, tests/** and tests/effectiveness/**
    respectively). .gitignore narrowed from .claude/ to .claude/* +
    !.claude/rules/ so the rules ship while local state (settings,
    agents, worktrees, caches) stays ignored.
  • README "Proof, not promises" stats refreshed to measured reality
    (5,311 unit tests, 181 released versions through v2.68.0, 6,500+
    changelog lines).
  • Consistency: exec bit set on hooks/_resolve-python.sh and
    scripts/libcst_dump.py (both are interpreter-invoked, so this is
    hygiene, not a behavior change); .qa_scratch/ and .ruby-lsp/ added
    to .gitignore.

Fixed

• Two stale journey-harness act assertions (surfaced by a live journey
  run): act 01 still expected the pre-v2.13 20-tool MCP registry (now 48,
  pinned with an update note), and act 02 phase 7 still asserted the
  pre-trust-persistence behavior where a force re-bootstrap flips trust to
  stale — trust is one-time and persists across profile changes, so the
  phase now asserts trusted remains and treats stale as the failure.