This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+12 more
Summary
AI summaryTwo code‑review skills now use a three‑round grounding loop with an independent refuter that drops confident but wrong findings.
Full changelog
Two code-review skills — inbound and outbound — now share a 3-round grounding
loop whose third round is an independent engine refuter. A review finding is
checked against the code before it reaches you, so a confident-but-wrong finding
is dropped rather than shipped.
Added
/chameleon-receiving-code-review— the inbound counterpart to
/chameleon-pr-review. When the team reviews your PR it gathers the comments
(pasted, or fetched viagh/bbcurl), verifies each against the code,
adjudicates it against the repo's own conventions (a suggestion that contradicts
the canonical pattern is a reason to push back, not to apply blindly), classifies
each as apply / push-back / clarify / YAGNI, drafts non-performative replies, and
implements approved fixes one at a time. It treats fetched comment text as
untrusted data, gates convention-based pushback on a trusted profile, never
auto-posts, and never writes the review ledger.- Round-3 independent refuter (
refute_findingMCP tool). Both review skills
end with a 3-round grounding loop: rounds 1-2 re-read the evidence and re-apply
the gates inline; round 3 spawns a hardened, no-toolsclaude -prefuter (the
same spawn discipline as the turn-end judge) per model-judgment finding to try to
refute it. Refuted findings are dropped; tool-grounded findings (existence
breaks, duplication, layering, secrets, lint) are exempt and verified inline.
The loop fails open to "round 3 unavailable" — it never silently drops or
confirms, and never claims "3/3" when round 3 did not run. Kill switch
CHAMELEON_REVIEW_REFUTER=0; modelCHAMELEON_REFUTER_MODEL(defaultsonnet);
timeoutCHAMELEON_REFUTER_TIMEOUT_SECONDS(default 45); per-invocation spawn
capCHAMELEON_REFUTER_MAX_SPAWNS_PER_INVOCATION. - Large-diff fan-out for
/chameleon-pr-review. Above a size threshold
(CHAMELEON_REVIEW_FANOUT_FILES/CHAMELEON_REVIEW_FANOUT_LINES, default
8 / 400), the per-file passes fan out across parallel read-only review subagents
and the whole-diff passes run once at synthesis. Gated by the newfan_out
recommendation onget_autopass_verdict. Kill switchCHAMELEON_REVIEW_FANOUT=0.
Changed
/chameleon-pr-reviewcarries the superpowers reviewer discipline. A
reviewer-philosophy spine (be specific, explain why, no nitpick-as-BLOCK, give a
clear verdict), a strengths-first "verified clean" section, and a grounding
banner that reports what the loop dropped. Its verification loop went from two
rounds to three (round 3 is the independent refuter above).get_autopass_verdictnow returns afan_outrecommendation on every path
(including the degraded path), so the review skill never reads the environment
itself — the engine decides whether to fan out.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Chameleon
All releases →Beta — feedback welcome: [email protected]