Complete native file reads and verify existing memory

ClaudiuSchuster commented

2026-10-18 08:43:02 +00:00

Owner

Native file reads could stop at a truncated text prefix or report an empty file when the provider omitted a large blob. This change provides immutable, lossless byte continuation for get_file and memory_read, with explicit offsets and end-of-file state. Small UTF-8 responses retain their existing shape; legacy memory character offsets remain supported.

Every continuation rechecks the caller's native repository access and the pinned file identity, then reads only the requested raw byte range. Binary files retain their original bytes. Encoded or inconsistent upstream responses are refused. The provider may still perform internal blob reads or range skips; this does not claim constant server-side work.

memory_append verifies the complete existing bytes before writing and retains the expected-old-SHA conflict check, so omitted or malformed old content cannot be mistaken for an empty file. Existing write policy limits and the separate public plugin remain unchanged.

Validation: 196 repository tests and 25 independent real-HTTP contracts pass in both normal and optimized Python. All 35 file-read tests also pass on the native test host. The exact-tree full E2E run passed against the deployed Core image, including full reconstruction above the native 10 MiB contents threshold, an actual conflicting append, and the same pinned cursor resuming after the real read-budget refusal. Lower provider thresholds and SHA256 object variants are covered by contract fixtures. Production deployment remains a separate acceptance step.

Native file reads could stop at a truncated text prefix or report an empty file when the provider omitted a large blob. This change provides immutable, lossless byte continuation for `get_file` and `memory_read`, with explicit offsets and end-of-file state. Small UTF-8 responses retain their existing shape; legacy memory character offsets remain supported. Every continuation rechecks the caller's native repository access and the pinned file identity, then reads only the requested raw byte range. Binary files retain their original bytes. Encoded or inconsistent upstream responses are refused. The provider may still perform internal blob reads or range skips; this does not claim constant server-side work. `memory_append` verifies the complete existing bytes before writing and retains the expected-old-SHA conflict check, so omitted or malformed old content cannot be mistaken for an empty file. Existing write policy limits and the separate public plugin remain unchanged. Validation: 196 repository tests and 25 independent real-HTTP contracts pass in both normal and optimized Python. All 35 file-read tests also pass on the native test host. The exact-tree full E2E run passed against the deployed Core image, including full reconstruction above the native 10 MiB contents threshold, an actual conflicting append, and the same pinned cursor resuming after the real read-budget refusal. Lower provider thresholds and SHA256 object variants are covered by contract fixtures. Production deployment remains a separate acceptance step.

ClaudiuSchuster added 1 commit

2026-10-18 08:43:02 +00:00

feat(gateway): complete native file and memory reads

access-validation-contract / contract (pull_request) Successful in 3m24s

Details Explain this check

ci / ci (pull_request) Successful in 3m56s

Details Explain this check

fdbfa786b9

ClaudiuSchuster changed title from ~~WIP: Complete native file reads and verify existing memory~~ to Complete native file reads and verify existing memory

2026-10-18 08:51:35 +00:00

ClaudiuSchuster merged commit 1140ffcdf0 into main

2026-10-18 08:51:50 +00:00

ClaudiuSchuster deleted branch oss-oo/native-file-reads-20261018-4b72

2026-10-18 08:51:50 +00:00

ClaudiuSchuster referenced this pull request from a commit

2026-10-18 08:51:51 +00:00

Complete native file reads and verify existing memory (#143)

Rows
Columns

Complete native file reads and verify existing memory #143