circle-ir-ai 2.37.9 → 2.37.11
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +55 -0
- package/package.json +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -5,11 +5,66 @@ All notable changes to this project will be documented in this file.
|
|
|
5
5
|
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
|
|
6
6
|
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
|
|
7
7
|
|
|
8
|
+
## [2.37.11] - 2026-07-07
|
|
9
|
+
|
|
10
|
+
### Re-forwarded — repin `circle-ir` `3.155.0 → 3.156.0`
|
|
11
|
+
|
|
12
|
+
Zero-code re-forward. Undoes the 2.37.10 rollback after the
|
|
13
|
+
rollback-verification sweep proved the "regression" flagged in
|
|
14
|
+
[cogniumhq/cognium-ai#202](https://github.com/cogniumhq/cognium-ai/issues/202)
|
|
15
|
+
was a false alarm.
|
|
16
|
+
|
|
17
|
+
**Root cause of the false alarm.** The 2.31.8 golden baseline was run
|
|
18
|
+
against a warm verifier cache and produced 200+ zero-token
|
|
19
|
+
`verifyTaintPathsBatch` entries via `logVerifyCacheHit()` in
|
|
20
|
+
`src/llm/verification.ts:1168`. Comparing those cache-hit log lines
|
|
21
|
+
against the cold-cache 2.31.9 sweep on 3.156.0 inflated the apparent
|
|
22
|
+
"real call" delta by ~2x. The rollback-verification sweep (cold cache
|
|
23
|
+
on 2.31.10 / 3.155.0) came in **higher** than the "regression" it was
|
|
24
|
+
meant to fix (5,543 realCalls vs 4,272), confirming the engine's
|
|
25
|
+
static pipeline is invariant across 3.155.0 → 3.156.0 and the delta
|
|
26
|
+
was cache-warmth artifact, not a signature-shape regression.
|
|
27
|
+
|
|
28
|
+
- All three sweeps (2.31.8 warm, 2.31.9 cold, 2.31.10 cold) share the
|
|
29
|
+
same 12,026 prefilter-skipped count → static pipeline unchanged.
|
|
30
|
+
- `verifyTaintPathsBatch` is still present in circle-ir-ai 2.37.11 —
|
|
31
|
+
it was not removed in circle-ir 3.156.0.
|
|
32
|
+
|
|
33
|
+
### Tests
|
|
34
|
+
|
|
35
|
+
- 90 files, 1350 passed / 3 skipped.
|
|
36
|
+
- Typecheck + build clean.
|
|
37
|
+
|
|
38
|
+
## [2.37.10] - 2026-07-06
|
|
39
|
+
|
|
40
|
+
### Rolled back — repin `circle-ir` `3.156.0 → 3.155.0`
|
|
41
|
+
|
|
42
|
+
Zero-code rollback. Restores the last known-good engine after the Tier-2
|
|
43
|
+
sweep on 3.156.0 surfaced a sharp LLM-verifier regression (see
|
|
44
|
+
[cogniumhq/cognium-ai#202](https://github.com/cogniumhq/cognium-ai/issues/202)):
|
|
45
|
+
|
|
46
|
+
- Verifier calls +83.7% on same 12-repo panel with only +0.12% findings.
|
|
47
|
+
- USD cost +89.6% (\$2.47 → \$4.68).
|
|
48
|
+
- Root cause candidate: `verifyTaintPathsBatch` signature removed in
|
|
49
|
+
3.156.0; work folded into heavier `verifyBatch` (0-token → full-token).
|
|
50
|
+
- Two amplifier repos (`openapi-generator` +385%, `hutool` +137%) drove
|
|
51
|
+
~102% of the aggregate call increase.
|
|
52
|
+
|
|
53
|
+
Holding `circle-ir` at `3.155.0` until a fix ships upstream.
|
|
54
|
+
|
|
55
|
+
### Tests
|
|
56
|
+
|
|
57
|
+
- 90 files, 1350 passed / 3 skipped (unchanged).
|
|
58
|
+
- Typecheck + build clean.
|
|
59
|
+
|
|
8
60
|
## [2.37.9] - 2026-07-06
|
|
9
61
|
|
|
10
62
|
Bump `circle-ir` `3.155.0` → `3.156.0`. Zero-code refresh — pulls in
|
|
11
63
|
upstream engine improvements published in `circle-ir@3.156.0`.
|
|
12
64
|
|
|
65
|
+
**Superseded by 2.37.10** — 3.156.0 caused a Tier-2 sweep regression
|
|
66
|
+
(cognium-ai#202); do not use 2.37.9.
|
|
67
|
+
|
|
13
68
|
### What changed
|
|
14
69
|
|
|
15
70
|
- **`package.json`**:
|