Project Tracker

Project Tracker

Standard: Task table standard.

Tasks

ID	Task	Type	Area	Size	Status	Owner	Next step
B-133	Keep a stolen task in the run queue of the CPU that owns it.	Bug	Scheduler	S	Open	here	try_steal_from_cpu (kernel/sched/scheduler.c:346-381) sets assigned_cpu without moving the task out of its slot, so the stealing CPU's find_task_local never finds it; move the slot with the assignment — or search the queue by assignment — and assert a steal in the scheduler's own self-test.
C-02	Keep network_stack.c state inside its module.	Bug	Concurrency	S	Open	here	The socket map now returns a copy and no exported function returns a pointer into module state; add the test that dereferences a row after the guard is released and must fail, then re-run it under C-01's multi-core acceptance.
B-89	Answer a CertificateRequest with a real client certificate when one is configured.	Improvement	WebTransport	S	Open	here	Send a CertificateVerify after the Certificate message; RFC 8446 section 4.4.2's empty Certificate is already what a client with no certificate sends and is conformant, so the change is only for a configured certificate.
OD-007	Pin the official immutable Qwen 3.8 fixtures.	Investigation	Models	S	Open	here	Record the hashes and source revisions for the official Qwen 3.8 config, tokenizer, SafeTensors and parity corpus before any adapter work (B-143).
OD-010	Define the opt-in approximate-mode contract.	Investigation	AI server	S	Open	here	Name the modes and define their quality reporting, telemetry and acceptance before any approximate mode ships.
B-85	Encode and validate QUIC transport parameters instead of carrying them as opaque bytes.	Improvement	WebTransport	M	Open	here	Encode the client's own parameters and parse and check the server's against RFC 9000's rules; userspace/wt/ today requires only that the extension is present, so no flow-control limit, idle timeout or connection ID the peer set is known.
B-88	Handle a HelloRetryRequest instead of refusing it by name.	Improvement	WebTransport	M	Open	here	Rewrite the transcript with the synthetic message_hash (RFC 8446 section 4.4.1), send a second ClientHello and refuse a second retry.
B-143	Pin the immutable official Qwen config, tokenizer, SafeTensors and parity corpus.	Improvement	Models	M	Open	here	Record the hashes and source revisions for the official Qwen 3.8 package.
B-145	Give the model package its own tokenizer.	Improvement	Models	M	Open	here	Make trusted tokenizer IDs match the package-owned tokenizer (B-144).
B-147	Reach Python-reference parity for scalar embedding, RMSNorm and the first projection.	Improvement	Models	M	Open	here	Add the three scalar operators and diff every output against the Python reference.
B-150	Make 32-step deterministic decode match the trusted continuation.	Improvement	Models	M	Open	here	Decode 32 steps and compare against the trusted continuation within the documented tolerance.
B-91	Qualify the WebTransport module on RISC-V by running it, not just building it.	Investigation	WebTransport	M	Open	here	The module now compiles to RISC-V objects in compile-check's riscv64 userspace leg; close it with a booted guest that completes a handshake (make qemu-quic-handshake-gate, B-84).
B-139	State the Qwen 3.8 support boundary and its completion gate.	Improvement	Models	M	Open	here	Pin an immutable official configuration before tokenizer, tensor, layer, prefill-logit, decode, session, backend and physical parity work (B-143).
B-140	State the Kimi K3 text support boundary and its completion gate.	Improvement	Models	M	Open	here	Prove KDA/MLA/MoE/MXFP4/operator and target-token parity on a real checkpoint (B-155 to B-159).
B-141	State the Kimi K3 multimodal support boundary and its completion gate.	Improvement	Models	M	Open	here	Stand up a separate official vision and multimodal golden acceptance (B-159).
B-84	Drive a QUIC handshake from a socket inside the guest.	Improvement	WebTransport	L	Open	here	A guest opens a datagram socket, sends a real QUIC Initial, receives a ServerHello and completes a handshake; exit gate make qemu-quic-handshake-gate boots a guest that says so on the console.
B-86	Build a QUIC connection runtime.	Improvement	WebTransport	L	Open	here	Add packet number spaces, CRYPTO stream reassembly, ACK and loss recovery, connection IDs, Retry and Version Negotiation; the TLS module is handed messages already reassembled and in order, and nothing does the reassembling today.
B-87	Implement 0-RTT and the session-ticket store it needs.	Improvement	WebTransport	L	Open	here	Add the c e traffic and binder branches of the key schedule, a session-ticket store for the PSK and an early-data encryption level.
B-129	Make the scheduler tick preempt a task on every architecture.	Improvement	Scheduler	L	Open	here	Each architecture fills the scheduler's frame from its own trap frame, calls the tick from its timer interrupt unless that CPU carries the network tick, and writes the frame back; RISC-V is landed and behaviourally tested, and x86-64 and AArch64 still name their missing half (B-132).
B-130	Isolate mediated functions behind the RISC-V IOMMU.	Improvement	IOMMU	L	Open	here	Mediated functions share one identity-mapped table; give their buffers distinct mappings, prove one cannot reach another's memory, and repeat the proof on a second board; virtio-mmio stays unmediateable and is stated so in docs/RISCV-IOMMU.md.
B-132	Make a user process a scheduled task on x86-64 and AArch64 too.	Improvement	Scheduler	L	Open	here	Port RISC-V's working dispatch to the other two ports, each with its own per-task kernel context (a per-task user_resume_rsp/TSS rsp0 record and floating-point area on x86-64, SP_EL1 loaded from the frame on AArch64), and settle the timer_mask_local() question.
E6	Execute distributed work across the cluster data plane.	Improvement	Cluster	L	Open	here	Framing, sealing and peer state already cross a real network in make qemu-cluster-two-node-gate and -three-node-gate; the remaining piece is distributed execution, which waits on D-05.
D-05	Run real local inference.	Improvement	AI server	L	Open	here	Prove real Qwen correctness, typed state, scheduling, cancellation, backpressure and metrics.
D-06	Control an authenticated cluster across nodes.	Improvement	Cluster	L	Open	here	The QEMU-testable tranche is done (join, partition, recovery, ownership and quorum); the remaining piece is execution across nodes, which waits on D-05.
D-07	Place and execute experts across the cluster.	Improvement	Cluster	L	Open	here	Hosted tests cover deterministic ownership, grouped routing, simulated node-loss rerouting and stable reduction; end-to-end distributed activation execution needs D-05 and D-06 first.
D-09	Serve production inference.	Improvement	AI server	L	Open	here	Implement the authenticated API, streaming, cancellation, saturation, loss and long-lived tests.
B-144	Stream SafeTensors, config and tokenizer imports.	Improvement	Models	L	Open	here	Reach bounded RSS and deterministic package output.
B-146	Probe the official architecture and build an ordered, config-derived layer plan.	Improvement	Models	L	Open	here	Make unknown fields fail closed.
B-148	Implement every configured operator through a complete layer.	Improvement	Models	L	Open	here	Cover attention, recurrent and convolution operators, position encoding, FFN, residual and norm/head, and gate on complete-layer and prefill-logit parity.
B-149	Separate prefill and decode plans with real per-layer state.	Improvement	Models	L	Open	here	Prove state and reload continuity.
B-152	Run a real model plan in a native, model-executing macOS process, with optional Metal.	Improvement	Models	L	Open	here	Run a real model plan end to end while CPU fallback remains authoritative.
B-154	Add typed paged state, prefix COW, ragged batching and exact speculation.	Improvement	Models	L	Open	here	Make target-only and speculative deterministic outputs match.
B-155	Build a separate kimi_k3 adapter from the immutable official config.	Improvement	Models	L	Open	here	Make config and tensor roles reject unsupported fields.
B-156	Implement K3's KDA, gated MLA, AttnRes, exact top-16 routing, shared experts, SiTU and MXFP4.	Improvement	Models	L	Open	here	Prove scalar operator, router and expert parity.
B-157	Make K3 expert shards independently addressable with async residency and prefetch.	Improvement	Models	L	Open	here	Keep authoritative routing unchanged by prediction.
B-159	Implement the K3 MoonViT-V2 multimodal pipeline.	Improvement	Models	L	Open	here	Cover separate golden image and text cases.
B-166	Support Qwen 3.8.	Improvement	Models	L	Open	here	Roadmap target whose acceptance criteria are B-143 to B-154; work starts at B-143.
B-167	Support Kimi K3 text.	Improvement	Models	L	Open	here	Interface only today; the acceptance criteria are B-155 to B-159, starting at B-155.
B-168	Support Kimi K3 multimodal.	Improvement	Models	L	Open	here	Roadmap only; it needs a separate golden image and text suite (B-159).
V-06	Capture the Virtualization.framework graphical console.	Improvement	Hypervisor	S	Blocked	operator	The display device is claimed and make vz-framebuffer-gate captures it through ScreenCaptureKit; grant Screen Recording permission to the terminal that runs the gate, then collect the evidence.
P-05	Provide physical Apple NEON evidence.	Investigation	Hardware	S	Blocked	operator	Run the NEON canary on physical Apple silicon; QEMU cannot satisfy this physical gate.
B-142	Verify the DeepSeek V4 Flash 0731 source before architecture work.	Improvement	Models	S	Blocked	maintainer	An authoritative, maintainer-approved immutable official source must exist before any architecture work (B-160).
B-160	Verify the DeepSeek V4 Flash 0731 official source.	Improvement	Models	S	Blocked	maintainer	Record the maintainer-approved immutable official source; B-142 and B-169 wait on it.
OD-005	Define SSH fleet limits, identity, audit retention, lockout and recovery.	Investigation	Network	S	Blocked	operator	Required before production SSH exposure; the operator must set the policy.
OD-008	Pin the official Kimi and DeepSeek sources.	Investigation	Models	S	Blocked	upstream	Required before the corresponding adapters; the exact DeepSeek label is unresolved upstream.
V-02	Deliver MSI-X for virtio on PCI.	Improvement	Hypervisor	M	Blocked	operator	Every PCI virtio device receives a distinct vector through QEMU's translation service and three defects are fixed; the remaining evidence is physical ARM PCIe, and Virtualization.framework has no ITS so its queues stay polled.
V-03	Make the guest's IPv6 globally routable.	Improvement	Network	M	Blocked	upstream	Apple issues com.apple.vm.networking only with a provisioning profile and ad-hoc signing cannot provide it; with the entitlement a bridged Fusion guest autoconfigures a routable address (F-03).
V-04	Qualify multiple vCPUs.	Investigation	Hypervisor	M	Blocked	operator	Boots come up 1/1, 4/4 and 8/8 with byte-identical smptest signatures across ten runs, and make vz-gate requires 4/4; the host has eight cores against a 128-256 core target and the platform offers no interrupt-affinity control.
C-01	Hold shared kernel state correct under genuine parallelism.	Bug	Concurrency	M	Blocked	operator	The network stack, service records and CPU-AI runtime now take reentrant guards, the resolver shares the network guard and the pointer-holding counters became atomics; the remaining evidence is a sustained multi-core run on a machine with more cores than this one.
D1	Prove the USB image on every architecture.	Investigation	Hardware	M	Blocked	operator	make release-image-gate boots each shipped image as a disk across five environments and the installed-disk gate boots the partition layout twice, but write-usb.sh has never been run against a real device; write a stick per architecture and boot it.
D2	Network-boot a blank machine and let it install itself.	Investigation	Network	M	Blocked	operator	make qemu-netboot-gate runs DHCP/TFTP, a loader-only medium, an install onto a blank disk and a standalone boot to SSH; the remaining evidence is the same sequence on physical hardware.
D3	Ship ready-to-run images per environment.	Investigation	Release	M	Blocked	operator	make vm-package-gate boots all five VM kits out of their archives and make boot-media-gate boots all three shipped netboot binaries; the remaining evidence is physical media (D1).
E1	Install XAIOS onto a blank disk from a running XAIOS.	Improvement	Storage	M	Blocked	operator	The installer partitions, formats the EFI System Partition and copies loader, kernel, initfs and entropy seed across, and make qemu-installed-disk-gate boots the result; the remaining run is on physical media (D1).
E3	Prove NUMA placement on both nodes.	Investigation	NUMA	M	Blocked	operator	The self-test now asserts placement inside node 1's range and exactly one node per CPU through both lookups, held by make qemu-x86_64-numa-gate across two- and four-node machines; the remaining evidence is physical NUMA hardware.
E4	Qualify multiqueue and RSS networking.	Improvement	Network	M	Blocked	operator	Per-queue state and round-robin polling are in, RSS is negotiated from the device's own config and make qemu-rss-steering-gate measures four pairs spreading [64, 62, 66, 64] against a one-bucket control; the gate needs Linux, root and a multiqueue tap, so it runs on the Intel host only, and virtio-mmio multiqueue is unexercised.
E5	Take xaibootFS and xaiFS to scale.	Improvement	Storage	M	Blocked	operator	xaibootFS v6 records a file as extents, lifting a volume to 1 GiB and 1024 nodes, and the write path deliberately refuses above 256 KiB until it streams; the remaining evidence is scale on physical storage.
E7	Measure storage throughput, caching and power-loss durability.	Investigation	Storage	M	Blocked	operator	Raising buffer_size to a mebibyte and reading each byte once changed the throughput by orders of magnitude, and make qemu-power-loss-gate replays QEMU's blklogwrites journal; the remaining evidence is physical storage, because emulator figures are not throughput evidence.
P-07	Provide the SVE/SVE2 backend's physical evidence.	Improvement	SIMD	M	Blocked	operator	make qemu-aarch64-sve2-gate executes the SVE2 canary and preserves per-task Z/P/FFR across scheduling and interrupts, and the packed SVE GEMV kernel is verified against the scalar reference; the remaining evidence is physical Apple silicon, since nothing here is performance evidence.
P-14	Provide physical Intel/Xeon evidence.	Investigation	Hardware	M	Blocked	operator	Physical firmware, ISA, NUMA, storage, network, thermal and sustained-load gates remain and need the machines chosen by OD-002 and OD-003.
S-11P	Qualify physical production NVMe.	Investigation	Storage	M	Blocked	operator	make qemu-qualification-readiness consolidates the QEMU NVMe and crash-recovery evidence; named physical devices must still pass queue scaling, interrupt affinity, FUA/flush/discard semantics, reset recovery, power-loss durability, sustained-load and performance gates.
S-12	Establish the production xaiFS trust root and signing-key custody.	Investigation	Storage	M	Blocked	operator	Offline trusted-replica payload repair is implemented and QEMU/hosted-tested; production trust-root enrollment, private-key custody, replica authorization and rotation decisions need named operators and deployment credentials.
N-F3P	Qualify SSH and network security on physical hardware.	Investigation	Network	M	Blocked	operator	make qemu-qualification-readiness consolidates the QEMU evidence; physical lossy-link, sustained-load, side-channel analysis and independent SSH/cryptography review remain open.
D-08	Produce benchmark and diagnostic evidence on real machines.	Investigation	AI server	M	Blocked	operator	QEMU benchmark telemetry and a hashed qualification-readiness report exist; physical metadata-rich NUMA, bandwidth, PMU, thermal, storage, network and redacted support-bundle evidence remain.
D-10	Qualify and clean up support.	Investigation	AI server	M	Blocked	operator	Documentation contracts and the consolidated make qemu-qualification-readiness gate exist; physical, model, cluster, thermal, PMU and durability qualifications remain.
B-163	Enforce SRAT/SLIT/HMAT placement policy and local/remote byte telemetry.	Improvement	NUMA	M	Blocked	operator	The two-node x86_64 QEMU gate validates SRAT/SLIT/HMAT parsing, usable-memory intersection, deterministic preferred-node policy, node-local allocation and local/remote byte accounting; physical locality and performance qualification remain.
B-164	Make NUMA/machine expert ownership stable and failure-aware.	Improvement	Cluster	M	Blocked	operator	Hosted tests validate deterministic owner selection, grouping, simulated owner failure and stable reduction; real NUMA/machine transport, remote activation execution and multi-QEMU exactness remain.
B-169	Support DeepSeek V4 Flash 0731.	Improvement	Models	M	Blocked	maintainer	Roadmap only; it waits on an official immutable source being verified (B-160 and B-161).
OD-001	Select the first physical Apple/ARM target and its firmware/storage/NIC boundary.	Investigation	Hardware	M	Blocked	operator	The operator must choose the machine before physical ARM support can be claimed (D4).
OD-002	Select the representative AVX2 Intel desktop and hybrid-core/device baseline.	Investigation	Hardware	M	Blocked	operator	The operator must choose the machine before Intel desktop support can be claimed (P-14).
OD-003	Select the Xeon generation, sockets/NUMA, memory, NIC and NVMe.	Investigation	Hardware	M	Blocked	operator	The operator must choose the machine before Xeon support can be claimed (P-14).
OD-004	Provision the production update and xaiFS trust roots.	Investigation	Storage	M	Blocked	operator	Rotation, revocation, offline recovery and interrupted-activation rollback are implemented; private operator keys and defined custody and authorization procedures are required before untrusted deployment.
OD-006	Define the supported NVMe/FUA/flush/discard/repair/power-loss contract.	Investigation	Storage	M	Blocked	operator	Required before physical persistent deployment (S-11P).
OD-009	Select the expert-parallel interconnect and failure/ownership model.	Investigation	Cluster	M	Blocked	operator	Required before cluster inference (D-06 and D-07).
D4	Qualify physical Apple/ARM, Intel desktop and Xeon machines.	Investigation	Hardware	L	Blocked	operator	Named hardware must pass firmware, device, durability, security and ISA-static gates; the machines are chosen by OD-001, OD-002 and OD-003 and no physical result is being waited on to close another row.
B-151	Add physical AVX2 and tiled prefill/verification kernels.	Improvement	Models	L	Blocked	operator	Produce physical differential and performance artifacts on the chosen Intel machine.
B-153	Add AVX-512/VNNI/AMX, SVE/SVE2, persistent worker gangs and NUMA autotuning.	Improvement	Models	L	Blocked	operator	Produce capability canaries, a scalar differential and physical evidence on the chosen machines.
B-158	Run a real K3 text checkpoint.	Improvement	Models	L	Blocked	operator	Pass the tokenizer, operator, router, target-token and production-width physical gates.
B-161	Build the DeepSeek adapter and parity suite.	Improvement	Models	L	Blocked	maintainer	Depends on the verified source (B-160).
B-162	Support multi-terabyte sparse allocations and large pages.	Improvement	Models	L	Blocked	operator	Hosted packages represent sparse offsets above 100 GiB and both QEMU targets cover 2 MiB mappings, with x86_64 covering a 1 GiB leaf plus targeted SMP TLB invalidation; physical capacity and performance qualification remain.
B-165	Dispatch real inference to AI Cells and secondary CPUs.	Improvement	Models	L	Blocked	operator	Real model work must execute on leased workers.
F-06	Cover Fusion releases beyond 26H1.	Investigation	Hypervisor	M	Parked	operator	Parked: revive when a Fusion release other than 26H1, an x86_64 guest or physical Apple hardware has to be carried as a compatibility claim; nothing external blocks it.