This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+12 more
Summary
AI summaryFixed intent contract scope‑line extraction to prevent unrelated prose from leaking into review jobs.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Medium |
Added churn-class pinning tests hardening regression invariants. Added churn-class pinning tests hardening regression invariants. Source: llm_adapter@2026-07-16 Confidence: high |
— |
| Bugfix | Medium |
Fixed scope-line extraction to persist bounded clause, preventing unrelated prose leakage. Fixed scope-line extraction to persist bounded clause, preventing unrelated prose leakage. Source: llm_adapter@2026-07-16 Confidence: high |
— |
| Refactor | Low |
Rewrote docs/architecture.md to describe async-first design and correct stale claims. Rewrote docs/architecture.md to describe async-first design and correct stale claims. Source: llm_adapter@2026-07-16 Confidence: high |
— |
Full changelog
Completes the Stop/idiom overhaul. Across v3.2.0-v4.0.0 the turn-end review
system was rebuilt: the Stop hook is deterministic and fast, launching one
detached background job that runs the correctness/duplication/idiom lenses, an
independent VERIFY refuter, and a canonical Finding-lifecycle ledger, then
delivers findings on a later turn instead of blocking turn end. The whole Stop
and SessionStart emission is ranked and token-budgeted, idiom truth lives in a
structured per-file store, and three new capabilities are live: severity-aware
triage, task-aware review, and self-learning idioms. See the [3.4.0], [3.5.0],
and [3.6.0] entries for the per-phase detail.
Changed
- Rewrote
docs/architecture.mdto describe the shipped async-first design,
correcting stale claims (idioms.md is a generated view of the JSON idiom
store, not the source of truth; the correctness reviewer runs in a detached
job, not inline under the Stop timeout).
Added
- Churn-class pinning tests hardening the invariants a regression could
silently break: the hook wrapper-timeout caps (fast hooks 3s, Stop backstop
55s), the liveCHAMELEON_JUDGE_WAITStop-budget headroom, and the idiom
lens's own prompt and finding flood caps.
Fixed
- The intent contract's scope-line extraction now persists a bounded clause
around each scoping phrase rather than the whole sentence, so a run-on
prompt no longer carries unrelated leading prose or a trailing reason clause
into the review job's evidence. The scope object itself, including a
filename, module, or version a directive names (config.json,v1.2.3),
is never split.
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]