Chameleon

Deep QA of the chameleon-pr-review and chameleon-receiving-code-review skills
across every supported language and framework, driving each engine surface the
skills call with real tools against real repos. The headline is the round-3
refuter, which was silently inert; several coverage gaps and skill-prose
ambiguities were fixed alongside.

Fixed

• The round-3 refuter never produced a verdict — round 3 was dead on both review
  skills. Three independent defects stacked: (1) run_one scanned the raw
  claude -p --output-format stream-json stdout with _extract_json_array, which
  locked onto the system-init tools array instead of the model's answer (which
  lives inside an assistant result/text block), so every finding came back
  unverified: "unparseable refuter output". It now harvests the assistant text
  blocks first (the same two-step the turn-end judge uses, factored into
  _stream_json_texts) and accepts either the prompted array or a bare object.
  (2) refute_finding pre-gated to unavailable on any --bare auth failure —
  true on every current CLI — even though the spawn falls back to a plain
  claude -p (the exact fallback the judge takes every turn); it now bails only
  when the CLI is genuinely absent (refuter_cli_absent). (3) The plain-fallback
  spawn starts a fresh full session and can transiently return nothing, so
  run_one now retries once on a non-timeout failure. Together these make round 3
  produce real confirmed/refuted verdicts for the first time.
• A deleted file's broken importers were invisible to the cross-file pass.
  get_crossfile_context skipped a module it could not read, conflating a DELETED
  module (which exports nothing, so every importer still referencing it is a
  genuine break — the strongest existence break there is) with a merely-unreadable
  one. A deleted module is now read as a closed empty export set; the per-site
  _live_importer_break re-check keeps only importers that still reference it and
  still resolve there, so a stale index never fabricates a break. The pr-review
  skill (Step 1/1a) now captures file status with --name-status -M so a deletion
  is handled as a sanctioned skip (not a normal source review) and a rename's old
  path enters the Step 2.9c diff-scope set.
• A Python (or Go/Rust/PHP) dependency-manifest change read as reviewed-clean.
  scan_dependency_changes parses only npm and Bundler, and a change to
  requirements*.txt / pyproject.toml / Pipfile / setup.py (Python is a
  first-class language) produced an empty result indistinguishable from a no-op
  diff. It now returns those in a new uncovered_manifests field, and the
  pr-review skill (Step 2.5) hand-reviews the added lines with the same
  severity split the npm path uses: an ACK for the coverage-gap disclosure and a
  routine name-only dependency add, a FIX for a visible red flag the reviewer can
  read directly (a non-registry/git source, an --index-url redirection, an
  install hook). A clean Python dependency add stays APPROVE, a Python manifest
  carrying an off-PyPI source is NEEDS CHANGES — symmetric with the identical
  npm content, instead of a silent clean.
• load_calls_index did not follow a linked worktree to the main profile.
  While get_pattern_context/lint_file resolve a worktree's profile via
  resolve_profile_root, the calls index read the raw worktree root, so a review
  run from a worktree (the pr-review skill's own recommended way to inspect another
  revision) silently degraded every blast-radius / contract-break / caller fact to
  unknown. It now applies the same resolution (identity off a worktree).
• Migration-safety parse ambiguities (pr-review Step 2.7a) could flip a verdict
  tier between two faithful reviewers. change_column_null was wrongly listed as
  irreversible (Rails inverts it) and double-routed to both the 2.7a BLOCK and the
  2.7b table-size FIX; it now lives only in 2.7b. The reversible do |dir| carve-out
  no longer clears the BLOCK for a one-directional block (dir.up with no
  dir.down).

Changed

• receiving-code-review gained handling the skill was silent on: GitHub/Bitbucket
  suggestion blocks (verify and adjudicate, never paste verbatim), same-anchor
  contradictory comments (surface the conflict, route to NEEDS CLARIFICATION), a
  distinct "this file isn't in your PR" outcome, an explicit inline-outdated
  re-resolution mechanism, per-file archetype/canonical resolution on multi-file
  reviews, a wider round-3 refuter-exemption set (canonical-grounded and
  deterministic file-witnessed verdicts verify inline), and an explicit post-apply
  re-lint in Step 8 so a fix that introduces a convention violation is caught
  rather than trusted to a non-blocking hook.