Skip to content

fix: complete release artifact comparison and select native clients - #173

Merged
HengYangDS merged 12 commits into
devfrom
proposal/20261015-responses041-closeout
Oct 16, 2026
Merged

HengYangDS merged 12 commits into
devfrom
proposal/20261015-responses041-closeout

Conversation

@HengYangDS

Copy link
Copy Markdown
Owner

Windows native client acceptance installed Hermes and uv even though the supported Windows clients are Claude Code and Codex. Select clients once in the native acceptance owner and use that selection in execution and both CI projections. Release matrix comparison now compares every declared file, including the original checksums.txt bytes.

Adopt the already accepted Claude Desktop Chat, Cowork and Code journeys for the original installed AIGW 0.4.1 subject. Keep final source task 9.2 open for exact-source acceptance; retain the failed c355cef proof.

Validation: focused matrix/client regressions and CUE projection checks pass. The stale Windows tool-closure assertion is corrected and passes. Strict OpenSpec and the native formatter pass. Existing signed-source repository proof 68b9c6df for 865eaea passes all required gates; independent review reports no hold.

This source carrier declares 0.4.2. Original 0.4.1 release artifact and manual/performance acceptance remain bound to 03e210c; they do not qualify new 0.4.2 binaries.

GitHub dev push run 38758007867 for 3a2fffd failed Native macOS acceptance
once: TestEntrypointVerificationKeepsMemoryBounded/ensure-existing allocated
272,953 bytes per operation against its 256 KiB bound, and passed on a re-run.
Executable reads borrow a pooled 256 KiB buffer; when a collection empties
the pool, a command allocates a fresh one, which alone fills the bound.

Reads now use 64 KiB pooled buffers. The memory test empties the pool before
every measured operation, the worst case the hosted runner hit: with 256 KiB
buffers it fails at 275,166 bytes per operation, with 64 KiB buffers it passes
at 78,591, and verification throughput is unchanged.

Change: single-pass-reader-verification, task 1.5.

Generated with Claude Code
GitHub run 38777446724 measured all eight native workloads of the published
v0.4.2 against v0.4.1 on macOS, Linux and Windows. Its first attempt failed
only on Windows: the baseline Hyperfine run hit its deadline, and the candidate
env/projection p95 reached 677 and 407 ms. The re-run attempt measured 99.8
and 98.1 ms, below the baseline's 106.9 and 107.5 ms, and kept every candidate
p95 within budget on all three platforms. On Linux, setup, sync and projection
became 21 to 23 percent faster than 0.4.1.

Change: single-pass-reader-verification, task 2.1.

Generated with Claude Code
The fix landed at release head 3a2fffd and shipped as v0.4.2 on GitHub and
GitLab. After the operator's Homebrew upgrade on 2026-10-16, the installed
`aigw sync` kept Hermes' own Claude Opus 5.5 choice, and every projected
credential command returned its Token before and after the upgrade.

Change: hermes-selection-adoption, tasks 2.1 and 2.2.

Generated with Claude Code
@HengYangDS
HengYangDS merged commit 17d9bdf into dev Oct 16, 2026
7 checks passed
@HengYangDS
HengYangDS deleted the proposal/20261015-responses041-closeout branch October 16, 2026 06:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant