You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Independent issue review against main 61303fa8150472dfbda62ec8d77908bc05c73f3b. This is a bounded contract/evidence review, not contract acceptance, implementation authorization, or a campaign PASS. I used the isolated requirement set and inspected code/tests independently; I did not rely on discussion claims or select an upstream source.
The existing snapshot primitives are useful, but do not by themselves establish an external dependency contract.
_source_snapshot.py:12–60 fixes the repository, registry path and library layout to InferRSI. Its rejection of another repository is appropriate for that internal projection; it is not proof that a separately selected external source has been pinned, authorized or packaged.
Bundle.manifest():190–209 verifies included dependency bytes against their declared digests. The resource-change test also rejects an old revision after local content changes. Those checks establish byte identity, not permission to redistribute or default-enable imported material. The documentation explicitly separates permission, status and validation.
Concrete contract gaps to resolve before calling this requirement covered:
Specify source identity confirmation, immutable source revision, complete file/dependency digest inventory, license/notice evidence, allowed-use/distribution decision and explicit unknown states. Keep an unconfirmed source unresolved; do not convert an illustrative package name into a URL. Do not silently reuse the existing internal-source schema as the external schema.
Define reference-only, retained snapshot and locally distilled knowledge as different modes. A live link is not an offline artifact; a gitlink pins a commit but does not supply its bytes or permission evidence; a retained snapshot can support offline reads only with the complete closure and notices. Distilled material needs a distinct identity and lineage back to the pinned source, rather than becoming an ambiguous second authority.
Specify the update boundary: validate a proposed successor before publishing it, preserve the previous usable snapshot when acquisition/validation fails, and make rollback select retained verified bytes rather than re-resolve a mutable upstream name. Keep this separate from the local store's documented transaction guarantees.
Make acceptance fixtures show an allowed pinned source resolving with network unavailable, and rejection of substituted source/pin, changed dependency bytes, missing closure, unknown permission/default-enable requests and failed updates without advancing the active pointer. Include distinct reference/snapshot/distilled identities. Existing reader tests are reusable lower-level evidence, not an end-to-end external importer/update test.
Checks actually run:PYTHONPATH=src python -m pytest -o addopts='' -q -rs tests/test_knowledge_operations.py tests/test_source_snapshot.py in an existing test environment: 58 passed, 2 skipped (Windows-only NTFS junction cases). The initial default Python attempt could not run because pytest was absent. No network-isolated external-source fixture, external update/rollback integration, license authorization verification, full suite or Windows run was performed. Tracked checkout remained unchanged.
Suggested next scope is a contract plus explicit acceptance fixtures/comparison, not selecting an upstream or committing to an implementation. No source import, bulk distillation, automatic promotion, Harness/Benchmark relocation, or completion claim follows from this review.
Activity
amy-why-3459 commentedon Nov 10, 2026
感谢整理,范围和「已有覆盖」写得很清楚。外部技能依赖和 #91 的蒸馏路径分开处理,我赞成。另外,接受 RFC 不等于选定实现,这个前提也合理。在评审进入外链 / submodule 的比较之前,有几点想先讨论一下:
license: UNKNOWN材料对齐:这些历史材料不能因为新机制上线而自动获得分发资格。read/collect读到,冲突时以哪一边为准?建议在契约里规定,任一知识条目只能有一个权威来源,并要求登记时声明和已蒸馏内容的对应关系(比如引用 knowledge: ground CannBot stage evidence in Ascend analysis #114 的 squash commit)。另外想确认一下:本 RFC 的契约是只约束 Knowledge 读取侧(#387 的
read/collect),还是也要约束开发期的同步工具?如果只管读取侧,同步工具的行为(比如怎么生成 digest)是否另开实施切片?zengchuang-hw commentedon Nov 17, 2026
Independent issue review against main
61303fa8150472dfbda62ec8d77908bc05c73f3b. This is a bounded contract/evidence review, not contract acceptance, implementation authorization, or a campaign PASS. I used the isolated requirement set and inspected code/tests independently; I did not rely on discussion claims or select an upstream source.The existing snapshot primitives are useful, but do not by themselves establish an external dependency contract.
_source_snapshot.py:12–60fixes the repository, registry path and library layout to InferRSI. Its rejection of another repository is appropriate for that internal projection; it is not proof that a separately selected external source has been pinned, authorized or packaged.Bundle.manifest():190–209verifies included dependency bytes against their declared digests. The resource-change test also rejects an old revision after local content changes. Those checks establish byte identity, not permission to redistribute or default-enable imported material. The documentation explicitly separates permission, status and validation.Concrete contract gaps to resolve before calling this requirement covered:
Checks actually run:
PYTHONPATH=src python -m pytest -o addopts='' -q -rs tests/test_knowledge_operations.py tests/test_source_snapshot.pyin an existing test environment: 58 passed, 2 skipped (Windows-only NTFS junction cases). The initial default Python attempt could not run because pytest was absent. No network-isolated external-source fixture, external update/rollback integration, license authorization verification, full suite or Windows run was performed. Tracked checkout remained unchanged.Suggested next scope is a contract plus explicit acceptance fixtures/comparison, not selecting an upstream or committing to an implementation. No source import, bulk distillation, automatic promotion, Harness/Benchmark relocation, or completion claim follows from this review.