Repository navigation
v0.1.3
v0.1.3: correctness and agentic release. Fixed: the v0.1.2 deep-mode known issue. The engine now sends an SSE keepalive while it reads a long append, and the controller and the OpenCode launch overlay wait up to 3600 s. Through OpenCode 2.2.7 in deep mode, a 28,420-token append at 111,025 tokens of context took 1,080.8 s to read, with 72 keepalives (longest gap 15.0 s) and no client retry. The controller and Studio are split into smaller modules (Studio context and workers, per-engine launch adapters for _begin, the proxy as prep, ledger and stream stages) with byte-for-byte contract tests on the old code. Proxy fixes: a client's Cookie and Authorization headers are never forwarded to a backend; an invalid Content-Length gets a 400 and closes the connection; client X-Oxe-* headers are stripped; a ledger error after admission no longer leaks the request slot; a query string on the legacy Flash route is refused (403). The engine refuses a non-loopback bind without an API key. New, off by default: LAN share (local-ai share; its own API key, inference routes only) and a tool-output cap (LOCAL_AI_TOOL_OUTPUT_CAP=on). New lab tool: the O-117 quality runner (teacher-forced KLD, top-1 agreement and an f32 reference for attention arms). Known issue: OpenCode 2.2.7 can fold a prompt submitted above the preset's compaction trigger into the compacted history as a "historical" checkpoint, and a local model may then continue the old objective instead of the new request. It was seen in deep mode at about 111K and also happens on v0.1.2. Fast mode's lower trigger does not prevent it. Workaround: after the compaction has finished, send the request again as a new message. Fixing it is the top v0.1.4 item.
This release makes no speed claim and changes no default. Optimisations measured in the lab and rejected by their pre-registered thresholds: fa2 and fa2x prefill attention reached 1.15-1.16x at 32K (the bar was 1.2x) and stay opt-in; fused decode reached 1.03x on Ornith and 0.94x on Qwen (the bar was 5%) and stays off. The tool-output cap gate was stopped as confounded (no output was actually capped), so the cap stays off. Tiled attention remains the default. CI binaries are rebuilds rather than the exact hardware-qualified executable.
Receipts: V013-INSTALL-G1, V013-BATCH, V013-AP1, V013-TOOL-OUTPUT-CAP, V013-SPLIT-SMOKE
Squashed snapshot of private source 8700bf12fd81c317f8ca5370ad27ff9894c6d516 (previous public release v0.1.2).
Binaries
Built by GitHub Actions from v0.1.3 (commit 96c5bf9) in run 43395197353; nobody uploads binaries by hand.
Linux aarch64, for ARMv8.6-A cores with dotprod, i8mm, fp16 and bf16 (measured on Snapdragon X Elite X1E80100 only); glibc requirement and toolchain in each tarball's README.txt.
CI builds pass the model-free tests and the Flash selftest; they have not been through the lab's model-based identity gates, and their outputs can differ from the measured laptop builds in the last floating-point bits (README, "Download").
| File | Contents |
|---|---|
oxe-server-v0.1.3-linux-aarch64.tar.gz |
oxe-server: the oxe MoE engine (Ornith 1.5, Qwen3.6 35B-A3B) |
oxe-server-flash-v0.1.3-linux-aarch64.tar.gz |
oxe-server with the Flash engine (--engine flash) + flash-m0 (expert sidecar) and flash-m3 (selftest); ggml built in the same run from the pinned public sources |
oxi-v0.1.3-linux-aarch64.tar.gz |
oxi: FLUX.2 klein image engine (experimental) |
Verify:
sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify oxe-server-v0.1.3-linux-aarch64.tar.gz --repo n0rpan/snapdragon-engine \
--signer-workflow n0rpan/snapdragon-engine/.github/workflows/release.ymlNo model weights are included.