Give two-session Qwen serving a bounded, opt-in decode grace for large cold prompts. Returning to normal interleaving at the deadline bounds the wait when the running reply takes longer than expected.
Bead: nix-overlay-qfi
Update the upstream flake pin and correct version reporting docs.
Evaluation reports 1.8.26. Build and test verification remains pending
on a remote builder or CI; do not build on this workstation.
Strata 0.2.14 leaves llama.cpp's MMQ prompt kernels out by default, so
STRATA_MMQ_KQ does nothing and UD-Q5_K_XL's Q6_K layer runs on Strata's
int8 kernel. On ms-s1-max, 0.2.15 with the gfx1151 prompt set and
STRATA_MMQ_KQ=1 gives first-token logits bit-identical to 0.2.13 over
15 cold prompts at the same speed (~1180 tok/s at 16K); without MMQ an
agent prompt reached KL 0.199 against 0.2.13. The engine grows from 39
to 54 MB.
Bead: nix-overlay-481
Strata 0.2.14 runs the prompt on its own kernels and vendors llama.cpp
(third_party/strata-llama), so the separate llama.cpp fetch for the
vision encoder goes. The build keeps upstream's halo flags: MMQ stays
out (STRATA_MMQ defaults to OFF), 47 of 48 UD-Q5_K_XL layers run on the
fused kernels, and the ROCm 10.1 hipBLASLt table still ships.
UD-Q5_K_XL on ms-s1-max, 16K cold prompts: 715 -> ~810 tok/s on the
defaults, ~1180 tok/s with the gfx1151 prompt set (unchanged).
Bead: nix-overlay-ydg
Upstream's gfx1151-hipblaslt-100401.txt was calibrated on TheRock 7.14.1.
therock-rocm 10.1.0rc3 reports the same hipBLASLt version, so the engine
accepted the table, but its solution ids do not exist in that build and
every prompt GEMM fell back to the heuristic. The table now comes from
tools/hip/tune_hipblaslt run on gfx1151 against therock-rocm 10.1 and is
installed in bin/tuning/, where the engine finds it by itself.
UD-Q5_K_XL, 17.5K cold prompt on ms-s1-max: 418 -> 700-725 tok/s.
Bead: nix-overlay-2yv
Maintain the gfx1151 SDK locally using the official AMD archive. Use the TheRock patchelf pin and adapt LLVM auxiliary tools for pure Nix builds.
Verify SDK installation, HIP and rocBLAS GPU operations, and rebuild its consumers. Document the internal rc3 stamp and upstream migration findings.
nix-overlay-v09 keeps the completed migration evidence and the outstanding
ms-s1-max build gate; nix-overlay-y5r is tombstoned after the duplicate
codex-oauth-proxy work was superseded.
1.1.1 omits the model from normalizeModel's allowlist, so requests for gpt-6.1-terra are rewritten to gpt-5.5. ChatGPT accounts are not entitled to gpt-5.5, so the request fails with model_not_found.
Add the model to the allowlist, allowed reasoning efforts, and default effort. Covered by the transform tests.
Serena rewrote its project config: new language-server IDs in the
comment list plus included_apis, excluded_apis and agent_interface keys.
All new keys keep their empty defaults, so no setting changes.
Remove content that no longer matches the workflow: package templates
already covered by the "Adding a New Package" steps, per-package notes,
the duplicated beads command reference, and the mandatory git push steps.
Drop the sequentialthinking MCP requirement and add a rule that AGENTS.md
changes need an explicit user request.
Packages dvcrn/codex-oauth-proxy, which exposes a ChatGPT Codex account through OpenAI-compatible APIs, as a buildGoModule derivation. Builds only cmd/codex-oauth-proxy because the worker package is js/wasm code, and passes -buildvcs=false because the Nix source has no .git. Upstream Go tests run offline, so doCheck is enabled.
Add magg, an MCP aggregator / meta-MCP server, to the overlay. It is a
Python package on the hatchling backend, built with
python3Packages.buildPythonApplication from the v1.3.0 source tarball,
so no upstream release artifacts are fetched. Ships both console
scripts declared upstream: magg (the aggregator CLI/server) and mbro
(the MCP browser). License is AGPL-3.0-or-later.
The install check runs the upstream pytest suite with
-m "not integration", because the integration-marked tests need a live
server. A forced nix build .#magg --rebuild passes with 329 passed,
3 skipped, 3 deselected, 0 failed, and both binaries smoke-test clean.
Also adds the README and AGENTS.md package rows, the upstream and build
research notes, and the verification report in docs/magg_build_report.md.
Track upstream main at afcc1dc971ae4acd59b157f031d4bdbc9d52a7d0
(2026-11-07), 22 commits past the de3e99c8 pin. Upstream still ships no
tags or releases and Cargo.toml still reads version 1.0.0, so the dated
0-unstable-* string stays. cargoHash moves on its own: the new
crates/zim-search workspace member grew Cargo.lock from 399 to 621
entries, and cargo vendor hashes the whole workspace.
The real packaging work is the TMPDIR override appended to preCheck.
Upstream's new harness::fs::path_tests scratch-root helper puts its
scratch dirs under <cwd>/target/ and first asserts that root lies
outside std::env::temp_dir(). The build sandbox exports TMPDIR=/build,
which is an ancestor of the unpacked source at /build/source, so that
premise check tripped and all 8 path-security tests failed with
"scoped root /build/source/target/marmel-path-tests-N sits inside the
temp allow-list entry /build". Pointing TMPDIR at
/tmp/marmel-check-tmp, outside $NIX_BUILD_TOP, restores upstream's
premise, so all 8 tests keep running for real instead of being filtered
out of cargoTestFlags. Build-time only; the installed binary is
untouched.
Two derivation comments carried stale facts and are corrected:
- meta.license claimed the repository shipped no LICENSE file and that
Cargo.toml had no license field. Upstream c98179d023 (Na1w/marmel#40)
added a root LICENSE file and declared license = "MIT" in Cargo.toml,
so the comment now records that MIT is asserted in the README,
Cargo.toml and LICENSE alike; lib.licenses.mit was already the right
value.
- the cargoTestFlags comment quoted "the lib suite passes at 1055
tests". Roughly 500 test fns landed since (TUI, fs-path, PTY and
dispatch-integrity suites), so the figure is dropped rather than
refreshed: nothing checks it and it rots on every bump.
Also adds the two research notes behind the bump under docs/: the
de3e99c8..afcc1dc9 upstream delta audit and the update-tooling findings.
Verified: clean nix build .#marmel including the serialised test suite
(--test-threads=1) — 2146 passed, 0 failed.
Editors and MCP clients need the language server and the MCP server, and
upstream's release job builds all three crates in one cargo invocation.
Build -p agnix-cli -p agnix-lsp -p agnix-mcp in the same derivation so the
workspace compiles once. Closes nix-overlay-ocf.
Add the agnix agent-config linter to the overlay. Built with
rustPlatform.buildRustPackage from the v0.60.3 source tarball
(cargoBuildFlags -p agnix-cli), so no upstream pre-built release
binaries are fetched. Closes nix-overlay-wfv.
Packages the storytold/pdfcraft Cargo workspace at tag v0.18.0: the pdfcraft
desktop app and pdfcraft-cli, with the desktop entry, AppStream metadata,
icons, licences and the two CC-BY-SA OCR models the app reads from
share/pdfcraft/models. No upstream binaries are downloaded.
Two deviations from upstream's build instructions:
- The three [patch.crates-io] entries upstream's own Cargo.lock records as
[[patch.unused]] (egui_demo_lib, egui_inspection, egui_system_font_provider)
have no package in nixpkgs' partial vendor tree to resolve against, so cargo
aborts. postPatch drops them; they are unused, and cargo drops the matching
lock entries itself, so the build runs without --locked.
- nixos-unstable's rustc 1.98.1 and 1.99.0 reject every AVX-VNNI intrinsic
("intrinsic signature mismatch for llvm.x86.avx512.vpdpbusd/vpwssd"), and
rten-gemm, pic-scale and zlib-rs all use one. The package therefore builds
against a new nixpkgs-rust195 input (nixos-26.05, rustc 1.95.0), the same
pattern as nixpkgs-torch211. Follow-up: nix-overlay-g4w.
Also adds a Home Manager module (programs.pdfcraft) to
homeManagerModules.default, and documents both in the README.
Upstream 1.8.4/1.8.5 are iGPU decode and returning-session speedups with no
packaging change: setup.py still reads EXL3_ROCM_DEV_INCLUDE/EXL3_HIP_DEFINES,
the pyproject runtime deps are unchanged, and tools/{glm,mimo,qwen}/serve.py
are still the serve entry points. Only the pin and source hash move.
Verified on gfx1151: nix build .#kyojin, importing exllamav3_ext against the
pinned ROCm torch 2.11, kyojin-serve-qwen --help, and nix build
.#kyojin.tests.jit.
The agent instructions told agents to run two skills that are no longer
part of the workflow:
- remove the "Communication: Caveman Mode" section, so replies are no
longer required to be compressed through the caveman skill
- replace the `conventional-commit` skill instruction with a plain
requirement to follow Conventional Commits concisely
Add guregodevo/memdoor, a Go coding/automation agent, built from source:
- the binary embeds the React UI through web/embed.go (`//go:embed
all:dist`), and upstream gitignores web/dist, so the bundle is built
with buildNpmPackage and copied into web/dist before the Go build
- CGO stays off: the SQLite driver is modernc.org/sqlite, so FTS5 needs
no C flags and the binary is static
- upstream's whole Go suite runs (buildGoModule would trim checks to
subPackages, where cmd/cli has no tests); it needs git on PATH and
separate scratch dirs for HOME and the Go temp root, because one test
writes ~/.memdoor and ProjectRoot walks up looking for that marker
Track the package in beads as nix-overlay-yxj.