Releases: gurinderu/craft
Releases · gurinderu/craft
Release list
v0.16.0
0.16.0 (2026-08-04)
Features
- review: write run records with a script, and checkpoint them as the run goes (9ff972c)
Bug Fixes
- review: a pre-existing advisory reports, it does not block (8c57e2a)
- review: let the logger be told where craft lives (de05551)
- review: make preflight probe with commands, and on the real PATH (95dbcb6)
Performance Improvements
- decide a High on the pair that can decide it, fold same-spot duplicates (99d50cd)
- review: deadline every agent, verify in one wave, keep preflight cheap (0373987)
- review: make the Rust gate probe its environment before it compiles (ae8a6ec)
- review: resolve the environment in one preflight step, and honour CI for security tools too (0e9d9c5)
v0.15.0
0.15.0 (2026-08-02)
Features
- build profile review rules (95a7ca7)
- catch build-profile divergence in the Rust review rubric (fd16514)
- exclusion catalog, mirror walk, and measured severity magnitude (13f07ed)
- fix-completeness checks, control/attack proof, reachability route (6294e44)
- require findings to pin their off-site premise (whereChecked) (8b3c45c)
- stamp the engine version on run records; bound and CI-defer the gate (c783149)
Bug Fixes
- distinguish a dead lens from a quiet one; cap low-value rule floods (ce8ba77)
- narrow the under-call hatch, find CI by SHA, record per-round lens yield (cd68a65)
- NOISE section ranked precise lenses as noisy; lost records were silent (4c76c9d)
Performance Improvements
- route verification by verdict power; stop the logger retyping records (a4029ce)
v0.14.0
v0.13.1
v0.13.0
0.13.0 (2026-07-19)
Features
- review: adjudicate schema requires invariant+attack; add ATTACK_SCHEMA + lint guard (cb6a394)
- review: fix-completeness adjudication — invariant+attack method, red-team pass for resolved Crit/High priors (34c5253)
Bug Fixes
- review: canonicalize prior-ledger severity for the verdict path, preserve identifiers in prompt fields, persist sources for strict re-review escalation (9b14d67)
- review: guard model-authored fields in the round-1 verifyPrompt (873f8dd)
- review: harden adjudicate track — demote contradictory resolved+attack, audit red-team death, reject empty-attack defeat, bound attack text (bc5dd1b)
- review: sanitize all adjudicate-track text, audit every degraded verdict, harden why bookkeeping (0733b07)
- review: sanitize model-authored fields at all three adjudicate prompt sites; cover still-open path; reconcile baseWhy comment (ea44ad7)
- review: shq/flatten all model-authored refs into shell + prompts (round 3b) (af29dc5)
- review: single-quote shell args in the carry-check git diff, validate ledger head (f207a14)
- review: stop baseWhy residue accreting when attack text echoes a marker (6c11987)
- review: strip new why-suffixes in baseWhy, flatten file/severity into prompts, audit carry-death, case-insensitive severity gate (363e76a)
v0.12.0
0.12.0 (2026-07-15)
Features
- review: adjudicate track for prior-round findings (f7b50e1)
- review: auto-detect re-review round with ancestor guard (b0b1d9d)
- review: delta-scoped lenses and re-review report shape (a558d8f)
- review: disposition mapping and re-review verdict (675ae5e)
- review: fuzzy cross-round finding matching (83b8958)
- review: line-tolerant finding fingerprint helper (187dff8)
- review: persist full finding ledger with branch/head/round (2a2f5c1)
- review: prior-round selection and index branch/head/round (7e99f86)